From f245cc28d4eb900efad0bc740145f58d713c6e4f Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sat, 16 Nov 2024 10:32:50 +0200
Subject: [PATCH 001/102] scripts : fix missing key in compare-llama-bench.py
 (#10332)

---
 scripts/compare-llama-bench.py | 7 ++-----
 1 file changed, 2 insertions(+), 5 deletions(-)

diff --git a/scripts/compare-llama-bench.py b/scripts/compare-llama-bench.py
index 6125a080af137..5069ae6382799 100755
--- a/scripts/compare-llama-bench.py
+++ b/scripts/compare-llama-bench.py
@@ -303,14 +303,11 @@ def get_rows(properties):
 
     show = []
     # Show CPU and/or GPU by default even if the hardware for all results is the same:
-    if "gpu_blas" not in properties_different and "n_gpu_layers" not in properties_different:
-        gpu_blas = bool(rows_full[0][KEY_PROPERTIES.index("gpu_blas")])
+    if "n_gpu_layers" not in properties_different:
         ngl = int(rows_full[0][KEY_PROPERTIES.index("n_gpu_layers")])
 
-        if not gpu_blas or ngl != 99 and "cpu_info" not in properties_different:
+        if ngl != 99 and "cpu_info" not in properties_different:
             show.append("cpu_info")
-        if gpu_blas and "gpu_info" not in properties_different:
-            show.append("gpu_info")
 
     show += properties_different
 

From bcdb7a23862b61aa307fc462fadfe1e2e653d010 Mon Sep 17 00:00:00 2001
From: MaggotHATE <clay1326@gmail.com>
Date: Sat, 16 Nov 2024 18:26:54 +0500
Subject: [PATCH 002/102] server: (web UI) Add samplers sequence customization
 (#10255)

* Samplers sequence: simplified and input field.

* Removed unused function

* Modify and use `settings-modal-short-input`

* rename "name" --> "label"

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
---
 examples/server/public/index.html | 19 ++++++++++++++++---
 examples/server/server.cpp        | 20 ++++++++++++++------
 2 files changed, 30 insertions(+), 9 deletions(-)

diff --git a/examples/server/public/index.html b/examples/server/public/index.html
index 65a915d59bfd5..6216c08410a28 100644
--- a/examples/server/public/index.html
+++ b/examples/server/public/index.html
@@ -212,6 +212,9 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
           <details class="collapse collapse-arrow bg-base-200 mb-2 overflow-visible">
             <summary class="collapse-title font-bold">Other sampler settings</summary>
             <div class="collapse-content">
+              <!-- Samplers queue -->
+              <settings-modal-short-input label="Samplers queue" :config-key="'samplers'" :config-default="configDefault" :config-info="configInfo" v-model="config.samplers"></settings-modal-short-input>
+              <!-- Samplers -->
               <template v-for="configKey in ['dynatemp_range', 'dynatemp_exponent', 'typical_p', 'xtc_probability', 'xtc_threshold']">
                 <settings-modal-short-input :config-key="configKey" :config-default="configDefault" :config-info="configInfo" v-model="config[configKey]" />
               </template>
@@ -231,6 +234,7 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
             <summary class="collapse-title font-bold">Advanced config</summary>
             <div class="collapse-content">
               <label class="form-control mb-2">
+                <!-- Custom parameters input -->
                 <div class="label inline">Custom JSON config (For more info, refer to <a class="underline" href="https://github.com/ggerganov/llama.cpp/blob/master/examples/server/README.md" target="_blank" rel="noopener noreferrer">server documentation</a>)</div>
                 <textarea class="textarea textarea-bordered h-24" placeholder="Example: { &quot;mirostat&quot;: 1, &quot;min_p&quot;: 0.1 }" v-model="config.custom"></textarea>
               </label>
@@ -253,7 +257,7 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
     <label class="input input-bordered join-item grow flex items-center gap-2 mb-2">
       <!-- Show help message on hovering on the input label -->
       <div class="dropdown dropdown-hover">
-        <div tabindex="0" role="button" class="font-bold">{{ configKey }}</div>
+        <div tabindex="0" role="button" class="font-bold">{{ label || configKey }}</div>
         <div class="dropdown-content menu bg-base-100 rounded-box z-10 w-64 p-2 shadow mt-4">
           {{ configInfo[configKey] || '(no help message available)' }}
         </div>
@@ -282,6 +286,7 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
       apiKey: '',
       systemMessage: 'You are a helpful assistant.',
       // make sure these default values are in sync with `common.h`
+      samplers: 'dkypmxt',
       temperature: 0.8,
       dynatemp_range: 0.0,
       dynatemp_exponent: 1.0,
@@ -305,6 +310,7 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
     const CONFIG_INFO = {
       apiKey: 'Set the API Key if you are using --api-key option for the server.',
       systemMessage: 'The starting message that defines how model should behave.',
+      samplers: 'The order at which samplers are applied, in simplified way. Default is "dkypmxt": dry->top_k->typ_p->top_p->min_p->xtc->temperature',
       temperature: 'Controls the randomness of the generated text by affecting the probability distribution of the output tokens. Higher = more random, lower = more focused.',
       dynatemp_range: 'Addon for the temperature sampler. The added value to the range of dynamic temperature, which adjusts probabilities by entropy of tokens.',
       dynatemp_exponent: 'Addon for the temperature sampler. Smoothes out the probability redistribution based on the most probable token.',
@@ -352,10 +358,16 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
       { props: ["source"] }
     );
 
-    // inout field to be used by settings modal
+    // input field to be used by settings modal
     const SettingsModalShortInput = defineComponent({
       template: document.getElementById('settings-modal-short-input').innerHTML,
-      props: ['configKey', 'configDefault', 'configInfo', 'modelValue'],
+      props: {
+        label: { type: String, required: false },
+        configKey: String,
+        configDefault: Object,
+        configInfo: Object,
+        modelValue: [Object, String, Number],
+      },
     });
 
     // coversations is stored in localStorage
@@ -546,6 +558,7 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
               ],
               stream: true,
               cache_prompt: true,
+              samplers: this.config.samplers,
               temperature: this.config.temperature,
               dynatemp_range: this.config.dynatemp_range,
               dynatemp_exponent: this.config.dynatemp_exponent,
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index 00f9031dc7a4a..b8e003be9730e 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -927,14 +927,22 @@ struct server_context {
 
         {
             const auto & samplers = data.find("samplers");
-            if (samplers != data.end() && samplers->is_array()) {
-                std::vector<std::string> sampler_names;
-                for (const auto & name : *samplers) {
-                    if (name.is_string()) {
-                        sampler_names.emplace_back(name);
+            if (samplers != data.end()) {
+                if (samplers->is_array()) {
+                    std::vector<std::string> sampler_names;
+                    for (const auto & name : *samplers) {
+                        if (name.is_string()) {
+                            sampler_names.emplace_back(name);
+                        }
+                    }
+                    slot.sparams.samplers = common_sampler_types_from_names(sampler_names, false);
+                } else if (samplers->is_string()){
+                    std::string sampler_string;
+                    for (const auto & name : *samplers) {
+                        sampler_string += name;
                     }
+                    slot.sparams.samplers = common_sampler_types_from_chars(sampler_string);
                 }
-                slot.sparams.samplers = common_sampler_types_from_names(sampler_names, false);
             } else {
                 slot.sparams.samplers = default_sparams.samplers;
             }

From 8ee0d09ae6928d0501765cfc4e430b9236730caf Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sat, 16 Nov 2024 17:58:32 +0200
Subject: [PATCH 003/102] make : auto-determine dependencies (#0)

---
 .gitignore |   1 +
 Makefile   | 401 ++++++++++++++++-------------------------------------
 2 files changed, 120 insertions(+), 282 deletions(-)

diff --git a/.gitignore b/.gitignore
index 1092d097a7542..0b56bcc7ab682 100644
--- a/.gitignore
+++ b/.gitignore
@@ -3,6 +3,7 @@
 *.a
 *.bat
 *.bin
+*.d
 *.dll
 *.dot
 *.etag
diff --git a/Makefile b/Makefile
index 804c1bbec4091..539370e0639f4 100644
--- a/Makefile
+++ b/Makefile
@@ -527,11 +527,11 @@ ifndef GGML_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
 	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
 	ifeq ($(UNAME_S),Darwin)
-		MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS -DGGML_BLAS_USE_ACCELERATE
-		MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
-		MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
-		MK_LDFLAGS  += -framework Accelerate
-		OBJ_GGML    += ggml/src/ggml-blas/ggml-blas.o
+		MK_CPPFLAGS  += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS -DGGML_BLAS_USE_ACCELERATE
+		MK_CPPFLAGS  += -DACCELERATE_NEW_LAPACK
+		MK_CPPFLAGS  += -DACCELERATE_LAPACK_ILP64
+		MK_LDFLAGS   += -framework Accelerate
+		OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
 	endif
 endif # GGML_NO_ACCELERATE
 
@@ -542,44 +542,44 @@ ifndef GGML_NO_OPENMP
 endif # GGML_NO_OPENMP
 
 ifdef GGML_OPENBLAS
-	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
-	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)
-	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
-	OBJ_GGML    += ggml/src/ggml-blas/ggml-blas.o
+	MK_CPPFLAGS  += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
+	MK_CFLAGS    += $(shell pkg-config --cflags-only-other openblas)
+	MK_LDFLAGS   += $(shell pkg-config --libs openblas)
+	OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
 endif # GGML_OPENBLAS
 
 ifdef GGML_OPENBLAS64
-	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
-	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas64)
-	MK_LDFLAGS  += $(shell pkg-config --libs openblas64)
-	OBJ_GGML    += ggml/src/ggml-blas/ggml-blas.o
+	MK_CPPFLAGS  += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
+	MK_CFLAGS    += $(shell pkg-config --cflags-only-other openblas64)
+	MK_LDFLAGS   += $(shell pkg-config --libs openblas64)
+	OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
 endif # GGML_OPENBLAS64
 
 ifdef GGML_BLIS
-	MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_BLIS -I/usr/local/include/blis -I/usr/include/blis
-	MK_LDFLAGS  += -lblis -L/usr/local/lib
-	OBJ_GGML    += ggml/src/ggml-blas/ggml-blas.o
+	MK_CPPFLAGS  += -DGGML_USE_BLAS -DGGML_BLAS_USE_BLIS -I/usr/local/include/blis -I/usr/include/blis
+	MK_LDFLAGS   += -lblis -L/usr/local/lib
+	OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
 endif # GGML_BLIS
 
 ifdef GGML_NVPL
-	MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_NVPL -DNVPL_ILP64 -I/usr/local/include/nvpl_blas -I/usr/include/nvpl_blas
-	MK_LDFLAGS  += -L/usr/local/lib -lnvpl_blas_core -lnvpl_blas_ilp64_gomp
-	OBJ_GGML    += ggml/src/ggml-blas/ggml-blas.o
+	MK_CPPFLAGS  += -DGGML_USE_BLAS -DGGML_BLAS_USE_NVPL -DNVPL_ILP64 -I/usr/local/include/nvpl_blas -I/usr/include/nvpl_blas
+	MK_LDFLAGS   += -L/usr/local/lib -lnvpl_blas_core -lnvpl_blas_ilp64_gomp
+	OBJ_GGML_EXT += ggml/src/ggml-blas/ggml-blas.o
 endif # GGML_NVPL
 
 ifndef GGML_NO_LLAMAFILE
-	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
-	OBJ_GGML    += ggml/src/ggml-cpu/llamafile/sgemm.o
+	MK_CPPFLAGS  += -DGGML_USE_LLAMAFILE
+	OBJ_GGML_EXT += ggml/src/ggml-cpu/llamafile/sgemm.o
 endif
 
 ifndef GGML_NO_AMX
 	MK_CPPFLAGS += -DGGML_USE_AMX
-	OBJ_GGML    += ggml/src/ggml-amx/ggml-amx.o ggml/src/ggml-amx/mmq.o
+	OBJ_GGML_EXT += ggml/src/ggml-amx/ggml-amx.o ggml/src/ggml-amx/mmq.o
 endif
 
 ifdef GGML_RPC
-	MK_CPPFLAGS += -DGGML_USE_RPC
-	OBJ_GGML    += ggml/src/ggml-rpc.o
+	MK_CPPFLAGS  += -DGGML_USE_RPC
+	OBJ_GGML_EXT += ggml/src/ggml-rpc.o
 endif # GGML_RPC
 
 OBJ_CUDA_TMPL      = $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-wmma*.cu))
@@ -604,9 +604,9 @@ ifdef GGML_CUDA
 	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
 	MK_NVCCFLAGS += -use_fast_math
 
-	OBJ_GGML += ggml/src/ggml-cuda/ggml-cuda.o
-	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML += $(OBJ_CUDA_TMPL)
+	OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
+	OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
+	OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
 
 ifdef LLAMA_FATAL_WARNINGS
 	MK_NVCCFLAGS += -Werror all-warnings
@@ -723,9 +723,9 @@ ggml/src/ggml-cuda/ggml-cuda.o: \
 endif # GGML_CUDA
 
 ifdef GGML_VULKAN
-	MK_CPPFLAGS += -DGGML_USE_VULKAN
-	MK_LDFLAGS  += $(shell pkg-config --libs vulkan)
-	OBJ_GGML    += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o
+	MK_CPPFLAGS  += -DGGML_USE_VULKAN
+	MK_LDFLAGS   += $(shell pkg-config --libs vulkan)
+	OBJ_GGML_EXT += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o
 
 ifdef GGML_VULKAN_CHECK_RESULTS
 	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
@@ -821,9 +821,9 @@ ifdef GGML_CUDA_NO_PEER_COPY
 	HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY
 endif # GGML_CUDA_NO_PEER_COPY
 
-	OBJ_GGML += ggml/src/ggml-cuda/ggml-cuda.o
-	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML += $(OBJ_CUDA_TMPL)
+	OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
+	OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
+	OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
 
 ggml/src/ggml-cuda/ggml-cuda.o: \
 	ggml/src/ggml-cuda/ggml-cuda.cu \
@@ -922,9 +922,9 @@ ifdef GGML_CUDA_FA_ALL_QUANTS
 	MUSAFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
 endif # GGML_CUDA_FA_ALL_QUANTS
 
-	OBJ_GGML += ggml/src/ggml-cuda/ggml-cuda.o
-	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML += $(OBJ_CUDA_TMPL)
+	OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o
+	OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
+	OBJ_GGML_EXT += $(OBJ_CUDA_TMPL)
 
 ggml/src/ggml-cuda/ggml-cuda.o: \
 	ggml/src/ggml-cuda/ggml-cuda.cu \
@@ -945,9 +945,9 @@ ggml/src/ggml-cuda/%.o: \
 endif # GGML_MUSA
 
 ifdef GGML_METAL
-	MK_CPPFLAGS += -DGGML_USE_METAL
-	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
-	OBJ_GGML	+= ggml/src/ggml-metal/ggml-metal.o
+	MK_CPPFLAGS  += -DGGML_USE_METAL
+	MK_LDFLAGS   += -framework Foundation -framework Metal -framework MetalKit
+	OBJ_GGML_EXT += ggml/src/ggml-metal/ggml-metal.o
 
 ifdef GGML_METAL_USE_BF16
 	MK_CPPFLAGS += -DGGML_METAL_USE_BF16
@@ -956,8 +956,8 @@ ifdef GGML_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
 ifdef GGML_METAL_EMBED_LIBRARY
-	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
-	OBJ_GGML    += ggml/src/ggml-metal-embed.o
+	MK_CPPFLAGS  += -DGGML_METAL_EMBED_LIBRARY
+	OBJ_GGML_EXT += ggml/src/ggml-metal-embed.o
 endif
 endif # GGML_METAL
 
@@ -987,36 +987,41 @@ ggml/src/ggml-metal-embed.o: \
 endif
 endif # GGML_METAL
 
-OBJ_GGML += \
-	ggml/src/ggml.o \
-	ggml/src/ggml-aarch64.o \
-	ggml/src/ggml-alloc.o \
-	ggml/src/ggml-backend.o \
-	ggml/src/ggml-backend-reg.o \
-	ggml/src/ggml-quants.o \
-	ggml/src/ggml-threading.o \
-	ggml/src/ggml-cpu/ggml-cpu.o \
-	ggml/src/ggml-cpu/ggml-cpu-cpp.o \
-	ggml/src/ggml-cpu/ggml-cpu-aarch64.o \
-	ggml/src/ggml-cpu/ggml-cpu-quants.o
+DIR_GGML = ggml
+DIR_LLAMA = src
+DIR_COMMON = common
+
+OBJ_GGML = \
+	$(DIR_GGML)/src/ggml.o \
+	$(DIR_GGML)/src/ggml-aarch64.o \
+	$(DIR_GGML)/src/ggml-alloc.o \
+	$(DIR_GGML)/src/ggml-backend.o \
+	$(DIR_GGML)/src/ggml-backend-reg.o \
+	$(DIR_GGML)/src/ggml-quants.o \
+	$(DIR_GGML)/src/ggml-threading.o \
+	$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \
+	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o \
+	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \
+	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \
+	$(OBJ_GGML_EXT)
 
 OBJ_LLAMA = \
-	src/llama.o \
-	src/llama-vocab.o \
-	src/llama-grammar.o \
-	src/llama-sampling.o \
-	src/unicode.o \
-	src/unicode-data.o
+	$(DIR_LLAMA)/llama.o \
+	$(DIR_LLAMA)/llama-vocab.o \
+	$(DIR_LLAMA)/llama-grammar.o \
+	$(DIR_LLAMA)/llama-sampling.o \
+	$(DIR_LLAMA)/unicode.o \
+	$(DIR_LLAMA)/unicode-data.o
 
 OBJ_COMMON = \
-	common/common.o \
-	common/arg.o \
-	common/log.o \
-	common/console.o \
-	common/ngram-cache.o \
-	common/sampling.o \
-	common/build-info.o \
-	common/json-schema-to-grammar.o
+	$(DIR_COMMON)/common.o \
+	$(DIR_COMMON)/arg.o \
+	$(DIR_COMMON)/log.o \
+	$(DIR_COMMON)/console.o \
+	$(DIR_COMMON)/ngram-cache.o \
+	$(DIR_COMMON)/sampling.o \
+	$(DIR_COMMON)/build-info.o \
+	$(DIR_COMMON)/json-schema-to-grammar.o
 
 OBJ_ALL = $(OBJ_GGML) $(OBJ_LLAMA) $(OBJ_COMMON)
 
@@ -1117,246 +1122,78 @@ endif
 # Build libraries
 #
 
-# ggml
+# Libraries
+LIB_GGML   = libggml.so
+LIB_GGML_S = libggml.a
 
-ggml/src/ggml.o: \
-	ggml/src/ggml.c \
-	ggml/include/ggml.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
-
-ggml/src/ggml-threading.o: \
-	ggml/src/ggml-threading.cpp \
-	ggml/include/ggml.h
-	$(CXX) $(XXCFLAGS)   -c $< -o $@
+LIB_LLAMA   = libllama.so
+LIB_LLAMA_S = libllama.a
 
-ggml/src/ggml-cpu/ggml-cpu.o: \
-	ggml/src/ggml-cpu/ggml-cpu.c \
-	ggml/include/ggml.h \
-	ggml/src/ggml-common.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
-
-ggml/src/ggml-cpu/ggml-cpu-cpp.o: \
-	ggml/src/ggml-cpu/ggml-cpu.cpp \
-	ggml/include/ggml.h \
-	ggml/src/ggml-common.h
-	$(CXX) $(CXXFLAGS)   -c $< -o $@
+LIB_COMMON   = libcommon.so
+LIB_COMMON_S = libcommon.a
 
-ggml/src/ggml-alloc.o: \
-	ggml/src/ggml-alloc.c \
-	ggml/include/ggml.h \
-	ggml/include/ggml-alloc.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
+# Targets
+BUILD_TARGETS += $(LIB_GGML) $(LIB_GGML_S) $(LIB_LLAMA) $(LIB_LLAMA_S) $(LIB_COMMON) $(LIB_COMMON_S)
 
-ggml/src/ggml-backend.o: \
-	ggml/src/ggml-backend.cpp \
-	ggml/src/ggml-backend-impl.h \
-	ggml/include/ggml.h \
-	ggml/include/ggml-backend.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+# Dependency files
+DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d)
 
-ggml/src/ggml-quants.o: \
-	ggml/src/ggml-quants.c \
-	ggml/include/ggml.h \
-	ggml/src/ggml-quants.h \
-	ggml/src/ggml-common.h
-	$(CC) $(CFLAGS)    -c $< -o $@
+# Default target
+all: $(BUILD_TARGETS)
 
-ggml/src/ggml-aarch64.o: \
-	ggml/src/ggml-aarch64.c \
+# Note: need this exception because `ggml-cpu.c` and `ggml-cpu.cpp` both produce the same obj/dep files
+#       g++ -M -I ./ggml/include/ -I ./ggml/src ggml/src/ggml-cpu/ggml-cpu.cpp | grep ggml
+$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o: \
+	ggml/src/ggml-cpu/ggml-cpu.cpp \
+	ggml/include/ggml-backend.h \
 	ggml/include/ggml.h \
-	ggml/src/ggml-aarch64.h \
-	ggml/src/ggml-common.h
-	$(CC) $(CFLAGS)    -c $< -o $@
+	ggml/include/ggml-alloc.h \
+	ggml/src/ggml-backend-impl.h \
+	ggml/include/ggml-cpu.h \
+	ggml/src/ggml-impl.h
+	$(CXX) $(CXXFLAGS)   -c $< -o $@
 
-ggml/src/ggml-blas/ggml-blas.o: \
-	ggml/src/ggml-blas/ggml-blas.cpp \
-	ggml/include/ggml-blas.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+# Rules for building object files
+$(DIR_GGML)/%.o: $(DIR_GGML)/%.c
+	$(CC) $(CFLAGS) -MMD -c $< -o $@
 
-ifndef GGML_NO_LLAMAFILE
-ggml/src/ggml-cpu/llamafile/sgemm.o: \
-	ggml/src/ggml-cpu/llamafile/sgemm.cpp \
-	ggml/src/ggml-cpu/llamafile/sgemm.h \
-	ggml/include/ggml.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@ -I ggml/src -I ggml/src/ggml-cpu
-endif # GGML_NO_LLAMAFILE
+$(DIR_GGML)/%.o: $(DIR_GGML)/%.cpp
+	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
 
-ifndef GGML_NO_AMX
-ggml/src/ggml-amx/ggml-amx.o: \
-	ggml/src/ggml-amx/ggml-amx.cpp \
-	ggml/include/ggml-amx.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+$(DIR_LLAMA)/%.o: $(DIR_LLAMA)/%.cpp
+	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
 
-ggml/src/ggml-amx/mmq.o: \
-	ggml/src/ggml-amx/mmq.cpp \
-	ggml/src/ggml-amx/mmq.h \
-	ggml/include/ggml.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif
-
-ifdef GGML_RPC
-ggml/src/ggml-rpc.o: \
-	ggml/src/ggml-rpc.cpp \
-	ggml/include/ggml-rpc.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif # GGML_RPC
+$(DIR_COMMON)/%.o: $(DIR_COMMON)/%.cpp
+	$(CXX) $(CXXFLAGS) -MMD -c $< -o $@
 
-$(LIB_GGML): \
-	$(OBJ_GGML)
+# Rules for building libraries
+$(LIB_GGML): $(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
-$(LIB_GGML_S): \
-	$(OBJ_GGML)
+$(LIB_GGML_S): $(OBJ_GGML)
 	ar rcs $(LIB_GGML_S) $^
 
-# llama
-
-src/unicode.o: \
-	src/unicode.cpp \
-	src/unicode.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-src/unicode-data.o: \
-	src/unicode-data.cpp \
-	src/unicode-data.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-src/llama.o: \
-	src/llama.cpp \
-	src/llama-impl.h \
-	src/llama-vocab.h \
-	src/llama-grammar.h \
-	src/llama-sampling.h \
-	src/unicode.h \
-	include/llama.h \
-	ggml/include/ggml-cuda.h \
-	ggml/include/ggml-metal.h \
-	ggml/include/ggml.h \
-	ggml/include/ggml-alloc.h \
-	ggml/include/ggml-backend.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-src/llama-vocab.o: \
-	src/llama-vocab.cpp \
-	src/llama-vocab.h \
-	src/llama-impl.h \
-	include/llama.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-src/llama-grammar.o: \
-	src/llama-grammar.cpp \
-	src/llama-grammar.h \
-	src/llama-impl.h \
-	src/llama-vocab.h \
-	src/llama-sampling.h \
-	include/llama.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-src/llama-sampling.o: \
-	src/llama-sampling.cpp \
-	src/llama-sampling.h \
-	src/llama-impl.h \
-	include/llama.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-$(LIB_LLAMA): \
-	$(OBJ_LLAMA) \
-	$(LIB_GGML)
+$(LIB_LLAMA): $(OBJ_LLAMA) $(LIB_GGML)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
-$(LIB_LLAMA_S): \
-	$(OBJ_LLAMA)
+$(LIB_LLAMA_S): $(OBJ_LLAMA)
 	ar rcs $(LIB_LLAMA_S) $^
 
-# common
-
-common/common.o: \
-	common/common.cpp \
-	common/common.h \
-	common/console.h \
-	common/sampling.h \
-	common/json.hpp \
-	common/json-schema-to-grammar.h \
-	include/llama.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-common/arg.o: \
-	common/arg.cpp \
-	common/arg.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-common/log.o: \
-	common/log.cpp \
-	common/log.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-common/sampling.o: \
-	common/sampling.cpp \
-	common/sampling.h \
-	include/llama.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-common/console.o: \
-	common/console.cpp \
-	common/console.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-common/json-schema-to-grammar.o: \
-	common/json-schema-to-grammar.cpp \
-	common/json-schema-to-grammar.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-common/ngram-cache.o: \
-	common/ngram-cache.cpp \
-	common/ngram-cache.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-$(LIB_COMMON): \
-	$(OBJ_COMMON) \
-	$(LIB_LLAMA) \
-	$(LIB_GGML)
+$(LIB_COMMON): $(OBJ_COMMON) $(LIB_LLAMA) $(LIB_GGML)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
-$(LIB_COMMON_S): \
-	$(OBJ_COMMON)
+$(LIB_COMMON_S): $(OBJ_COMMON)
 	ar rcs $(LIB_COMMON_S) $^
 
+# Include dependency files
+-include $(DEP_FILES)
+
+# Clean rule
 clean:
-	rm -vrf *.dot $(BUILD_TARGETS) $(TEST_TARGETS)
-	rm -rvf src/*.o
-	rm -rvf tests/*.o
-	rm -rvf examples/*.o
-	rm -rvf common/*.o
-	rm -rvf *.a
-	rm -rvf *.dll
-	rm -rvf *.so
-	rm -rvf *.dot
-	rm -rvf ggml/*.a
-	rm -rvf ggml/*.dll
-	rm -rvf ggml/*.so
-	rm -rvf ggml/src/*.o
-	rm -rvf common/build-info.cpp
-	rm -rvf ggml/src/ggml-cpu/*.o
-	rm -rvf ggml/src/ggml-cpu/llamafile/*.o
-	rm -vrf ggml/src/ggml-amx/*.o
-	rm -vrf ggml/src/ggml-blas/*.o
-	rm -vrf ggml/src/ggml-cann/*.o
-	rm -vrf ggml/src/ggml-cpu/*.o
-	rm -vrf ggml/src/ggml-cuda/*.o
-	rm -vrf ggml/src/ggml-cuda/template-instances/*.o
-	rm -vrf ggml/src/ggml-hip/*.o
-	rm -vrf ggml/src/ggml-kompute/*.o
-	rm -vrf ggml/src/ggml-metal/*.o
-	rm -vrf ggml/src/ggml-metal/ggml-metal-embed.metal
-	rm -vrf ggml/src/ggml-rpc/*.o
-	rm -vrf ggml/src/ggml-sycl/*.o
-	rm -vrf ggml/src/ggml-vulkan/*.o
-	rm -vrf ggml/src/ggml-musa/*.o
-	rm -rvf $(BUILD_TARGETS)
-	rm -rvf $(TEST_TARGETS)
-	rm -f vulkan-shaders-gen ggml/src/ggml-vulkan-shaders.hpp ggml/src/ggml-vulkan-shaders.cpp
-	rm -rvf $(LEGACY_TARGETS_CLEAN)
-	find examples pocs -type f -name "*.o" -delete
+	rm -vrf $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -rvf *.a *.dll *.so *.dot
+	find ggml src common tests examples pocs -type f -name "*.o" -delete
+	find ggml src common tests examples pocs -type f -name "*.d" -delete
 
 #
 # Examples

From db4cfd5dbc31c90f0d5c413a2e182d068b8ee308 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sat, 16 Nov 2024 17:58:56 +0200
Subject: [PATCH 004/102] llamafile : fix include path (#0)

ggml-ci
---
 ggml/src/ggml-cpu/llamafile/sgemm.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
index da4146ec4f688..b2ce2e6649479 100644
--- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp
+++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
@@ -50,7 +50,8 @@
 
 #include "sgemm.h"
 #include "ggml-impl.h"
-#include "ggml-cpu-impl.h"
+// hack until moved into the CPU backend
+#include "../ggml-cpu-impl.h"
 #include "ggml-quants.h"
 
 #ifdef _MSC_VER

From 4e54be0ec6cb5cd6ed56e52e927b80b2796ec844 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Sat, 16 Nov 2024 23:00:41 +0100
Subject: [PATCH 005/102] llama/ex: remove --logdir argument (#10339)

---
 common/arg.cpp                     |  11 --
 common/common.cpp                  | 210 -----------------------------
 common/common.h                    |  13 --
 examples/infill/infill.cpp         |  46 -------
 examples/main/main.cpp             |  45 -------
 examples/perplexity/perplexity.cpp |  51 -------
 examples/server/README.md          |   1 -
 include/llama.h                    |   2 -
 scripts/run-with-preset.py         | 146 --------------------
 src/llama.cpp                      |  22 ---
 10 files changed, 547 deletions(-)
 delete mode 100755 scripts/run-with-preset.py

diff --git a/common/arg.cpp b/common/arg.cpp
index 7c5c5e5cd5b88..4115b2f7511d3 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -1939,17 +1939,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.simple_io = true;
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL}));
-    add_opt(common_arg(
-        {"-ld", "--logdir"}, "LOGDIR",
-        "path under which to save YAML logs (no logging if unset)",
-        [](common_params & params, const std::string & value) {
-            params.logdir = value;
-
-            if (params.logdir.back() != DIRECTORY_SEPARATOR) {
-                params.logdir += DIRECTORY_SEPARATOR;
-            }
-        }
-    ));
     add_opt(common_arg(
         {"--positive-file"}, "FNAME",
         string_format("positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str()),
diff --git a/common/common.cpp b/common/common.cpp
index ebd16b600eed2..93037462127d5 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -1890,213 +1890,3 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
     return result;
 }
 
-//
-// YAML utils
-//
-
-void yaml_dump_vector_float(FILE * stream, const char * prop_name, const std::vector<float> & data) {
-    if (data.empty()) {
-        fprintf(stream, "%s:\n", prop_name);
-        return;
-    }
-
-    fprintf(stream, "%s: [", prop_name);
-    for (size_t i = 0; i < data.size() - 1; ++i) {
-        fprintf(stream, "%e, ", data[i]);
-    }
-    fprintf(stream, "%e]\n", data.back());
-}
-
-void yaml_dump_vector_int(FILE * stream, const char * prop_name, const std::vector<int> & data) {
-    if (data.empty()) {
-        fprintf(stream, "%s:\n", prop_name);
-        return;
-    }
-
-    fprintf(stream, "%s: [", prop_name);
-    for (size_t i = 0; i < data.size() - 1; ++i) {
-        fprintf(stream, "%d, ", data[i]);
-    }
-    fprintf(stream, "%d]\n", data.back());
-}
-
-void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data) {
-    std::string data_str(data == NULL ? "" : data);
-
-    if (data_str.empty()) {
-        fprintf(stream, "%s:\n", prop_name);
-        return;
-    }
-
-    size_t pos_start = 0;
-    size_t pos_found = 0;
-
-    if (std::isspace(data_str[0]) || std::isspace(data_str.back())) {
-        data_str = std::regex_replace(data_str, std::regex("\n"), "\\n");
-        data_str = std::regex_replace(data_str, std::regex("\""), "\\\"");
-        data_str = std::regex_replace(data_str, std::regex(R"(\\[^n"])"), R"(\$&)");
-        data_str = "\"" + data_str + "\"";
-        fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
-        return;
-    }
-
-    if (data_str.find('\n') == std::string::npos) {
-        fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
-        return;
-    }
-
-    fprintf(stream, "%s: |\n", prop_name);
-    while ((pos_found = data_str.find('\n', pos_start)) != std::string::npos) {
-        fprintf(stream, "  %s\n", data_str.substr(pos_start, pos_found-pos_start).c_str());
-        pos_start = pos_found + 1;
-    }
-}
-
-void yaml_dump_non_result_info(FILE * stream, const common_params & params, const llama_context * lctx,
-                               const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
-    ggml_cpu_init(); // some ARM features are detected at runtime
-
-    const auto & sparams = params.sparams;
-
-    fprintf(stream, "build_commit: %s\n",        LLAMA_COMMIT);
-    fprintf(stream, "build_number: %d\n",        LLAMA_BUILD_NUMBER);
-    fprintf(stream, "cpu_has_arm_fma: %s\n",     ggml_cpu_has_arm_fma()     ? "true" : "false");
-    fprintf(stream, "cpu_has_avx: %s\n",         ggml_cpu_has_avx()         ? "true" : "false");
-    fprintf(stream, "cpu_has_avx_vnni: %s\n",    ggml_cpu_has_avx_vnni()    ? "true" : "false");
-    fprintf(stream, "cpu_has_avx2: %s\n",        ggml_cpu_has_avx2()        ? "true" : "false");
-    fprintf(stream, "cpu_has_avx512: %s\n",      ggml_cpu_has_avx512()      ? "true" : "false");
-    fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
-    fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
-    fprintf(stream, "cpu_has_fma: %s\n",         ggml_cpu_has_fma()         ? "true" : "false");
-    fprintf(stream, "cpu_has_neon: %s\n",        ggml_cpu_has_neon()        ? "true" : "false");
-    fprintf(stream, "cpu_has_sve: %s\n",         ggml_cpu_has_sve()         ? "true" : "false");
-    fprintf(stream, "cpu_has_f16c: %s\n",        ggml_cpu_has_f16c()        ? "true" : "false");
-    fprintf(stream, "cpu_has_fp16_va: %s\n",     ggml_cpu_has_fp16_va()     ? "true" : "false");
-    fprintf(stream, "cpu_has_riscv_v: %s\n",     ggml_cpu_has_riscv_v()     ? "true" : "false");
-    fprintf(stream, "cpu_has_wasm_simd: %s\n",   ggml_cpu_has_wasm_simd()   ? "true" : "false");
-    fprintf(stream, "cpu_has_sse3: %s\n",        ggml_cpu_has_sse3()        ? "true" : "false");
-    fprintf(stream, "cpu_has_vsx: %s\n",         ggml_cpu_has_vsx()         ? "true" : "false");
-    fprintf(stream, "cpu_has_matmul_int8: %s\n", ggml_cpu_has_matmul_int8() ? "true" : "false");
-
-#ifdef NDEBUG
-    fprintf(stream, "debug: false\n");
-#else
-    fprintf(stream, "debug: true\n");
-#endif // NDEBUG
-
-    fprintf(stream, "model_desc: %s\n", model_desc);
-    fprintf(stream, "n_vocab: %d  # output size of the final layer, 32001 for some models\n", llama_n_vocab(llama_get_model(lctx)));
-
-#ifdef __OPTIMIZE__
-    fprintf(stream, "optimize: true\n");
-#else
-    fprintf(stream, "optimize: false\n");
-#endif // __OPTIMIZE__
-
-    fprintf(stream, "time: %s\n", timestamp.c_str());
-
-    fprintf(stream, "\n");
-    fprintf(stream, "###############\n");
-    fprintf(stream, "# User Inputs #\n");
-    fprintf(stream, "###############\n");
-    fprintf(stream, "\n");
-
-    fprintf(stream, "alias: %s # default: unknown\n", params.model_alias.c_str());
-    fprintf(stream, "batch_size: %d # default: 512\n", params.n_batch);
-    fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
-    fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
-    fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
-    fprintf(stream, "dry_allowed_length: %d # default: 2\n", sparams.dry_allowed_length);
-    fprintf(stream, "dry_base: %.2f # default: 1.75\n", sparams.dry_base);
-    fprintf(stream, "dry_multiplier: %.1f # default: 0.0\n", sparams.dry_multiplier);
-    fprintf(stream, "dry_penalty_last_n: %d # default: -1 (0 = disable, -1 = context size)\n", sparams.dry_penalty_last_n);
-    fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
-    fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
-    fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.penalty_freq);
-    yaml_dump_string_multiline(stream, "grammar", sparams.grammar.c_str());
-    fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
-    fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
-    fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
-    fprintf(stream, "ignore_eos: %s # default: false\n", sparams.ignore_eos ? "true" : "false");
-
-    yaml_dump_string_multiline(stream, "in_prefix", params.input_prefix.c_str());
-    fprintf(stream, "in_prefix_bos: %s # default: false\n", params.input_prefix_bos ? "true" : "false");
-    yaml_dump_string_multiline(stream, "in_suffix", params.input_prefix.c_str());
-    fprintf(stream, "interactive: %s # default: false\n", params.interactive ? "true" : "false");
-    fprintf(stream, "interactive_first: %s # default: false\n", params.interactive_first ? "true" : "false");
-    fprintf(stream, "keep: %d # default: 0\n", params.n_keep);
-    fprintf(stream, "logdir: %s # default: unset (no logging)\n", params.logdir.c_str());
-
-    fprintf(stream, "logit_bias:\n");
-    for (const auto & logit_bias : sparams.logit_bias) {
-        fprintf(stream, "  %d: %f", logit_bias.token, logit_bias.bias);
-    }
-
-    fprintf(stream, "lora:\n");
-    for (auto & la : params.lora_adapters) {
-        if (la.scale == 1.0f) {
-            fprintf(stream, "  - %s\n", la.path.c_str());
-        }
-    }
-    fprintf(stream, "lora_scaled:\n");
-    for (auto & la : params.lora_adapters) {
-        if (la.scale != 1.0f) {
-            fprintf(stream, "  - %s: %f\n", la.path.c_str(), la.scale);
-        }
-    }
-    fprintf(stream, "lora_init_without_apply: %s # default: false\n", params.lora_init_without_apply ? "true" : "false");
-    fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
-    fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
-    fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
-    fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
-    fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
-    fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false");
-    fprintf(stream, "model: %s # default: %s\n", params.model.c_str(), DEFAULT_MODEL_PATH);
-    fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str());
-    fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
-    fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers);
-    fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
-    fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", sparams.n_probs);
-    fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
-    fprintf(stream, "penalize_nl: %s # default: false\n", sparams.penalize_nl ? "true" : "false");
-    fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
-    fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
-    fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);
-    yaml_dump_string_multiline(stream, "prompt", params.prompt.c_str());
-    fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str());
-    fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
-    fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
-    yaml_dump_vector_int(stream, "prompt_tokens", prompt_tokens);
-    fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.penalty_repeat);
-
-    fprintf(stream, "reverse_prompt:\n");
-    for (std::string ap : params.antiprompt) {
-        size_t pos = 0;
-        while ((pos = ap.find('\n', pos)) != std::string::npos) {
-            ap.replace(pos, 1, "\\n");
-            pos += 1;
-        }
-
-        fprintf(stream, "  - %s\n", ap.c_str());
-    }
-
-    fprintf(stream, "rope_freq_base: %f # default: 10000.0\n", params.rope_freq_base);
-    fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
-    fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
-    fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
-    fprintf(stream, "flash_attn: %s # default: false\n", params.flash_attn ? "true" : "false");
-    fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
-
-    const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + llama_max_devices());
-    yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector);
-
-    fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency());
-    fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
-    fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
-    fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
-    fprintf(stream, "xtc_probability: %f # default: 0.0\n", sparams.xtc_probability);
-    fprintf(stream, "xtc_threshold: %f # default: 0.1\n", sparams.xtc_threshold);
-    fprintf(stream, "typ_p: %f # default: 1.0\n", sparams.typ_p);
-    fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
-    fprintf(stream, "display_prompt: %s # default: true\n", params.display_prompt ? "true" : "false");
-}
diff --git a/common/common.h b/common/common.h
index 6289feaeba752..7977cc7a99a78 100644
--- a/common/common.h
+++ b/common/common.h
@@ -209,7 +209,6 @@ struct common_params {
     std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state             // NOLINT
     std::string input_prefix         = ""; // string to prefix user inputs with                             // NOLINT
     std::string input_suffix         = ""; // string to suffix user inputs with                             // NOLINT
-    std::string logdir               = ""; // directory in which to save YAML log files                     // NOLINT
     std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding           // NOLINT
     std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
     std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
@@ -584,15 +583,3 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 static const char * const LLM_KV_SPLIT_NO            = "split.no";
 static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
 static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
-
-//
-// YAML utils
-//
-
-void yaml_dump_vector_float    (FILE * stream, const char * prop_name, const std::vector<float> & data);
-void yaml_dump_vector_int      (FILE * stream, const char * prop_name, const std::vector<int> & data);
-void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
-
-void yaml_dump_non_result_info(
-    FILE * stream, const common_params & params, const llama_context * lctx,
-    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp
index f18362c91c7bf..15b358dc4e854 100644
--- a/examples/infill/infill.cpp
+++ b/examples/infill/infill.cpp
@@ -43,50 +43,6 @@ static std::vector<llama_token> * g_output_tokens;
 
 static bool is_interacting = false;
 
-static void write_logfile(
-    const llama_context * ctx, const common_params & params, const llama_model * model,
-    const std::vector<llama_token> & input_tokens, const std::string & output,
-    const std::vector<llama_token> & output_tokens
-) {
-    if (params.logdir.empty()) {
-        return;
-    }
-
-    const std::string timestamp = string_get_sortable_timestamp();
-
-    const bool success = fs_create_directory_with_parents(params.logdir);
-    if (!success) {
-        LOG_ERR("%s: warning: failed to create logdir %s, cannot write logfile\n",
-                __func__, params.logdir.c_str());
-        return;
-    }
-
-    const std::string logfile_path = params.logdir + timestamp + ".yml";
-    FILE * logfile = fopen(logfile_path.c_str(), "w");
-
-    if (logfile == NULL) {
-        LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
-        return;
-    }
-
-    fprintf(logfile, "binary: infill\n");
-    char model_desc[128];
-    llama_model_desc(model, model_desc, sizeof(model_desc));
-    yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
-
-    fprintf(logfile, "\n");
-    fprintf(logfile, "######################\n");
-    fprintf(logfile, "# Generation Results #\n");
-    fprintf(logfile, "######################\n");
-    fprintf(logfile, "\n");
-
-    yaml_dump_string_multiline(logfile, "output", output.c_str());
-    yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
-
-    llama_perf_dump_yaml(logfile, ctx);
-    fclose(logfile);
-}
-
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
 static void sigint_handler(int signo) {
     if (signo == SIGINT) {
@@ -96,7 +52,6 @@ static void sigint_handler(int signo) {
             console::cleanup();
             LOG("\n");
             common_perf_print(*g_ctx, *g_smpl);
-            write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
 
             // make sure all logs are flushed
             LOG("Interrupted by user\n");
@@ -625,7 +580,6 @@ int main(int argc, char ** argv) {
 
     LOG("\n");
     common_perf_print(ctx, smpl);
-    write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
 
     llama_free(ctx);
     llama_free_model(model);
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index 374ed47ad6311..7c4ce4be2abae 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -62,49 +62,6 @@ static bool file_is_empty(const std::string & path) {
     return f.tellg() == 0;
 }
 
-static void write_logfile(
-    const llama_context * ctx, const common_params & params, const llama_model * model,
-    const std::vector<llama_token> & input_tokens, const std::string & output,
-    const std::vector<llama_token> & output_tokens
-) {
-    if (params.logdir.empty()) {
-        return;
-    }
-
-    const std::string timestamp = string_get_sortable_timestamp();
-
-    const bool success = fs_create_directory_with_parents(params.logdir);
-    if (!success) {
-        LOG_ERR("%s: failed to create logdir %s, cannot write logfile\n", __func__, params.logdir.c_str());
-        return;
-    }
-
-    const std::string logfile_path = params.logdir + timestamp + ".yml";
-    FILE * logfile = fopen(logfile_path.c_str(), "w");
-
-    if (logfile == NULL) {
-        LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
-        return;
-    }
-
-    fprintf(logfile, "binary: main\n");
-    char model_desc[128];
-    llama_model_desc(model, model_desc, sizeof(model_desc));
-    yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
-
-    fprintf(logfile, "\n");
-    fprintf(logfile, "######################\n");
-    fprintf(logfile, "# Generation Results #\n");
-    fprintf(logfile, "######################\n");
-    fprintf(logfile, "\n");
-
-    yaml_dump_string_multiline(logfile, "output", output.c_str());
-    yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
-
-    llama_perf_dump_yaml(logfile, ctx);
-    fclose(logfile);
-}
-
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
 static void sigint_handler(int signo) {
     if (signo == SIGINT) {
@@ -115,7 +72,6 @@ static void sigint_handler(int signo) {
             console::cleanup();
             LOG("\n");
             common_perf_print(*g_ctx, *g_smpl);
-            write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
 
             // make sure all logs are flushed
             LOG("Interrupted by user\n");
@@ -926,7 +882,6 @@ int main(int argc, char ** argv) {
 
     LOG("\n\n");
     common_perf_print(ctx, smpl);
-    write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
 
     common_sampler_free(smpl);
 
diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp
index e803ff143f7d1..64a84607c22d8 100644
--- a/examples/perplexity/perplexity.cpp
+++ b/examples/perplexity/perplexity.cpp
@@ -34,55 +34,6 @@ struct results_log_softmax {
     float  prob;
 };
 
-static void write_logfile(
-    const llama_context * ctx, const common_params & params, const llama_model * model,
-    const struct results_perplexity & results
-) {
-    if (params.logdir.empty()) {
-        return;
-    }
-
-    if (params.hellaswag) {
-        LOG_WRN("%s: logging results is not implemented for HellaSwag. No files will be written.\n", __func__);
-        return;
-    }
-
-    const std::string timestamp = string_get_sortable_timestamp();
-
-    const bool success = fs_create_directory_with_parents(params.logdir);
-    if (!success) {
-        LOG_WRN("%s: failed to create logdir %s, cannot write logfile\n",
-                __func__, params.logdir.c_str());
-        return;
-    }
-
-    const std::string logfile_path = params.logdir + timestamp + ".yml";
-    FILE * logfile = fopen(logfile_path.c_str(), "w");
-
-    if (logfile == NULL) {
-        LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
-        return;
-    }
-
-    fprintf(logfile, "binary: main\n");
-    char model_desc[128];
-    llama_model_desc(model, model_desc, sizeof(model_desc));
-    yaml_dump_non_result_info(logfile, params, ctx, timestamp, results.tokens, model_desc);
-
-    fprintf(logfile, "\n");
-    fprintf(logfile, "######################\n");
-    fprintf(logfile, "# Perplexity Results #\n");
-    fprintf(logfile, "######################\n");
-    fprintf(logfile, "\n");
-
-    yaml_dump_vector_float(logfile, "logits", results.logits);
-    fprintf(logfile, "ppl_value: %f\n", results.ppl_value);
-    yaml_dump_vector_float(logfile, "probs", results.probs);
-
-    llama_perf_dump_yaml(logfile, ctx);
-    fclose(logfile);
-}
-
 static std::vector<float> softmax(const std::vector<float>& logits) {
     std::vector<float> probs(logits.size());
     float max_logit = logits[0];
@@ -2072,8 +2023,6 @@ int main(int argc, char ** argv) {
     LOG("\n");
     llama_perf_context_print(ctx);
 
-    write_logfile(ctx, params, model, results);
-
     llama_free(ctx);
     llama_free_model(model);
 
diff --git a/examples/server/README.md b/examples/server/README.md
index 6f72c6bb83cbf..0936e0b7ba13a 100644
--- a/examples/server/README.md
+++ b/examples/server/README.md
@@ -85,7 +85,6 @@ The project is under active development, and we are [looking for feedback and co
 | `-hfr, --hf-repo REPO` | Hugging Face model repository (default: unused)<br/>(env: LLAMA_ARG_HF_REPO) |
 | `-hff, --hf-file FILE` | Hugging Face model file (default: unused)<br/>(env: LLAMA_ARG_HF_FILE) |
 | `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)<br/>(env: HF_TOKEN) |
-| `-ld, --logdir LOGDIR` | path under which to save YAML logs (no logging if unset) |
 | `--log-disable` | Log disable |
 | `--log-file FNAME` | Log to file |
 | `--log-colors` | Enable colored logging<br/>(env: LLAMA_LOG_COLORS) |
diff --git a/include/llama.h b/include/llama.h
index 5e742642eec8d..bc268e7996baa 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -1244,8 +1244,6 @@ extern "C" {
     LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
     LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);
 
-    LLAMA_API void llama_perf_dump_yaml(FILE * stream, const struct llama_context * ctx);
-
 #ifdef __cplusplus
 }
 #endif
diff --git a/scripts/run-with-preset.py b/scripts/run-with-preset.py
deleted file mode 100755
index 8f0bf8ca8aa55..0000000000000
--- a/scripts/run-with-preset.py
+++ /dev/null
@@ -1,146 +0,0 @@
-#!/usr/bin/env python3
-
-import logging
-import argparse
-import os
-import subprocess
-import sys
-
-import yaml
-
-logger = logging.getLogger("run-with-preset")
-
-CLI_ARGS_LLAMA_CLI_PERPLEXITY = [
-    "batch-size", "cfg-negative-prompt", "cfg-scale", "chunks", "color", "ctx-size", "escape",
-    "export", "file", "frequency-penalty", "grammar", "grammar-file", "hellaswag",
-    "hellaswag-tasks", "ignore-eos", "in-prefix", "in-prefix-bos", "in-suffix",
-    "interactive", "interactive-first", "keep", "logdir", "logit-bias", "lora", "lora-base",
-    "low-vram", "main-gpu", "mirostat", "mirostat-ent", "mirostat-lr", "mlock",
-    "model", "multiline-input", "n-gpu-layers", "n-predict", "no-mmap", "no-mul-mat-q",
-    "np-penalize-nl", "numa", "ppl-output-type", "ppl-stride", "presence-penalty", "prompt",
-    "prompt-cache", "prompt-cache-all", "prompt-cache-ro", "repeat-last-n",
-    "repeat-penalty", "reverse-prompt", "rope-freq-base", "rope-freq-scale", "rope-scale", "seed",
-    "simple-io", "tensor-split", "threads", "temp", "top-k", "top-p", "typical",
-    "verbose-prompt"
-]
-
-CLI_ARGS_LLAMA_BENCH = [
-    "batch-size", "low-vram", "model", "mul-mat-q", "n-gen", "n-gpu-layers",
-    "n-prompt", "output", "repetitions", "tensor-split", "threads", "verbose"
-]
-
-CLI_ARGS_LLAMA_SERVER = [
-    "alias", "batch-size", "ctx-size", "embedding", "host", "lora", "lora-base",
-    "low-vram", "main-gpu", "mlock", "model", "n-gpu-layers", "n-probs", "no-mmap", "no-mul-mat-q",
-    "numa", "path", "port", "rope-freq-base", "timeout", "rope-freq-scale", "tensor-split",
-    "threads", "verbose"
-]
-
-description = """Run llama.cpp binaries with presets from YAML file(s).
-To specify which binary should be run, specify the "binary" property (llama-cli, llama-perplexity, llama-bench, and llama-server are supported).
-To get a preset file template, run a llama.cpp binary with the "--logdir" CLI argument.
-
-Formatting considerations:
-- The YAML property names are the same as the CLI argument names of the corresponding binary.
-- Properties must use the long name of their corresponding llama.cpp CLI arguments.
-- Like the llama.cpp binaries the property names do not differentiate between hyphens and underscores.
-- Flags must be defined as "<PROPERTY_NAME>: true" to be effective.
-- To define the logit_bias property, the expected format is "<TOKEN_ID>: <BIAS>" in the "logit_bias" namespace.
-- To define multiple "reverse_prompt" properties simultaneously the expected format is a list of strings.
-- To define a tensor split, pass a list of floats.
-"""
-usage = "run-with-preset.py [-h] [yaml_files ...] [--<ARG_NAME> <ARG_VALUE> ...]"
-epilog = ("  --<ARG_NAME> specify additional CLI ars to be passed to the binary (override all preset files). "
-          "Unknown args will be ignored.")
-
-parser = argparse.ArgumentParser(
-    description=description, usage=usage, epilog=epilog, formatter_class=argparse.RawTextHelpFormatter)
-parser.add_argument("-bin", "--binary", help="The binary to run.")
-parser.add_argument("yaml_files", nargs="*",
-                    help="Arbitrary number of YAML files from which to read preset values. "
-                    "If two files specify the same values the later one will be used.")
-parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
-
-known_args, unknown_args = parser.parse_known_args()
-
-if not known_args.yaml_files and not unknown_args:
-    parser.print_help()
-    sys.exit(0)
-
-logging.basicConfig(level=logging.DEBUG if known_args.verbose else logging.INFO)
-
-props = dict()
-
-for yaml_file in known_args.yaml_files:
-    with open(yaml_file, "r") as f:
-        props.update(yaml.load(f, yaml.SafeLoader))
-
-props = {prop.replace("_", "-"): val for prop, val in props.items()}
-
-binary = props.pop("binary", "llama-cli")
-if known_args.binary:
-    binary = known_args.binary
-
-if os.path.exists(f"./{binary}"):
-    binary = f"./{binary}"
-
-if binary.lower().endswith("llama-cli") or binary.lower().endswith("llama-perplexity"):
-    cli_args = CLI_ARGS_LLAMA_CLI_PERPLEXITY
-elif binary.lower().endswith("llama-bench"):
-    cli_args = CLI_ARGS_LLAMA_BENCH
-elif binary.lower().endswith("llama-server"):
-    cli_args = CLI_ARGS_LLAMA_SERVER
-else:
-    logger.error(f"Unknown binary: {binary}")
-    sys.exit(1)
-
-command_list = [binary]
-
-for cli_arg in cli_args:
-    value = props.pop(cli_arg, None)
-
-    if not value or value == -1:
-        continue
-
-    if cli_arg == "logit-bias":
-        for token, bias in value.items():
-            command_list.append("--logit-bias")
-            command_list.append(f"{token}{bias:+}")
-        continue
-
-    if cli_arg == "reverse-prompt" and not isinstance(value, str):
-        for rp in value:
-            command_list.append("--reverse-prompt")
-            command_list.append(str(rp))
-        continue
-
-    command_list.append(f"--{cli_arg}")
-
-    if cli_arg == "tensor-split":
-        command_list.append(",".join([str(v) for v in value]))
-        continue
-
-    value = str(value)
-
-    if value != "True":
-        command_list.append(str(value))
-
-num_unused = len(props)
-if num_unused > 10:
-    logger.info(f"The preset file contained a total of {num_unused} unused properties.")
-elif num_unused > 0:
-    logger.info("The preset file contained the following unused properties:")
-    for prop, value in props.items():
-        logger.info(f"  {prop}: {value}")
-
-command_list += unknown_args
-
-sp = subprocess.Popen(command_list)
-
-while sp.returncode is None:
-    try:
-        sp.wait()
-    except KeyboardInterrupt:
-        pass
-
-sys.exit(sp.returncode)
diff --git a/src/llama.cpp b/src/llama.cpp
index dc5dfba0c2e1b..1703104fb3680 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -22075,28 +22075,6 @@ void llama_perf_context_reset(struct llama_context * ctx) {
     ctx->t_p_eval_us = ctx->n_p_eval = 0;
 }
 
-void llama_perf_dump_yaml(FILE * stream, const llama_context * ctx) {
-    fprintf(stream, "\n");
-    fprintf(stream, "###########\n");
-    fprintf(stream, "# Timings #\n");
-    fprintf(stream, "###########\n");
-    fprintf(stream, "\n");
-
-    fprintf(stream, "mst_eval: %.2f  # ms / token during generation\n",
-            1.0e-3 * ctx->t_eval_us / ctx->n_eval);
-    fprintf(stream, "mst_p_eval: %.2f  # ms / token during prompt processing\n",
-            1.0e-3 * ctx->t_p_eval_us / ctx->n_p_eval);
-    fprintf(stream, "n_eval: %d  # number of tokens generated (excluding the first one)\n", ctx->n_eval);
-    fprintf(stream, "n_p_eval: %d  # number of tokens processed in batches at the beginning\n", ctx->n_p_eval);
-    fprintf(stream, "t_eval_us: %" PRId64 "  # total microseconds spent generating tokens\n", ctx->t_eval_us);
-    fprintf(stream, "t_load_us: %" PRId64 "  # total microseconds spent loading the model\n", ctx->t_load_us);
-    fprintf(stream, "t_p_eval_us: %" PRId64 "  # total microseconds spent prompt processing\n", ctx->t_p_eval_us);
-    fprintf(stream, "ts_eval: %.2f  # tokens / second during generation\n",
-            1.0e6 * ctx->n_eval / ctx->t_eval_us);
-    fprintf(stream, "ts_p_eval: %.2f  # tokens / second during prompt processing\n",
-            1.0e6 * ctx->n_p_eval / ctx->t_p_eval_us);
-}
-
 // For internal test use
 const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
     struct llama_context * ctx

From 0fff7fd79818980763a601660f25b01a0cf4b87a Mon Sep 17 00:00:00 2001
From: FirstTimeEZ <179362031+FirstTimeEZ@users.noreply.github.com>
Date: Sun, 17 Nov 2024 12:29:18 +1300
Subject: [PATCH 006/102] docs : vulkan build instructions to use git bash
 mingw64 (#10303)

---
 docs/build.md | 27 +++++++++++++++++++++++++--
 1 file changed, 25 insertions(+), 2 deletions(-)

diff --git a/docs/build.md b/docs/build.md
index 52de2b4e2c224..811bbb409c250 100644
--- a/docs/build.md
+++ b/docs/build.md
@@ -282,9 +282,9 @@ The following compilation options are also available to tweak performance (yes,
 
 #### w64devkit
 
-Download and extract [w64devkit](https://github.com/skeeto/w64devkit/releases).
+Download and extract [`w64devkit`](https://github.com/skeeto/w64devkit/releases).
 
-Download and install the [Vulkan SDK](https://vulkan.lunarg.com/sdk/home#windows). When selecting components, only the Vulkan SDK Core is required.
+Download and install the [`Vulkan SDK`](https://vulkan.lunarg.com/sdk/home#windows) with the default settings.
 
 Launch `w64devkit.exe` and run the following commands to copy Vulkan dependencies:
 ```sh
@@ -302,6 +302,29 @@ EOF
 ```
 Switch into the `llama.cpp` directory and run `make GGML_VULKAN=1`.
 
+#### Git Bash MINGW64
+
+Download and install [`Git-SCM`](https://git-scm.com/downloads/win) with the default settings
+
+Download and install [`Visual Studio Community Edition`](https://visualstudio.microsoft.com/) and make sure you select `C++`
+
+Download and install [`CMake`](https://cmake.org/download/) with the default settings
+
+Download and install the [`Vulkan SDK`](https://vulkan.lunarg.com/sdk/home#windows) with the default settings.
+
+Go into your `llama.cpp` directory and right click, select `Open Git Bash Here` and then run the following commands
+
+```
+cmake -B build -DGGML_VULKAN=ON
+cmake --build build --config Release
+```
+
+Now you can load the model in conversation mode using `Vulkan`
+
+```
+build/bin/release/llama-cli -m "[PATH TO MODEL]" -ngl 100 -c 16384 -t 10 -n -2 -cnv
+```
+
 #### MSYS2
 Install [MSYS2](https://www.msys2.org/) and then run the following commands in a UCRT terminal to install dependencies.
   ```sh

From 5c9a8b22b10132db620529435e3cfa49304b65cc Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sat, 16 Nov 2024 22:16:04 +0200
Subject: [PATCH 007/102] scripts : update sync

---
 scripts/sync-ggml-am.sh | 54 +++++++++++++++++++++++------------------
 scripts/sync-ggml.sh    | 26 +++++++++++---------
 2 files changed, 45 insertions(+), 35 deletions(-)

diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh
index 74d6c6c8b47ce..d0815cf89e2f5 100755
--- a/scripts/sync-ggml-am.sh
+++ b/scripts/sync-ggml-am.sh
@@ -73,17 +73,20 @@ while read c; do
         src/ggml*.h \
         src/ggml*.c \
         src/ggml*.cpp \
-        src/ggml*.m \
-        src/ggml*.metal \
-        src/ggml*.cu \
         src/ggml-amx/* \
+        src/ggml-blas/* \
         src/ggml-cann/* \
+        src/ggml-cpu/* \
         src/ggml-cuda/* \
+        src/ggml-hip/* \
+        src/ggml-kompute/* \
+        src/ggml-metal/* \
+        src/ggml-musa/* \
+        src/ggml-rpc/* \
         src/ggml-sycl/* \
-        src/vulkan-shaders/* \
+        src/ggml-vulkan/* \
         include/ggml*.h \
         tests/test-opt.cpp \
-        tests/test-grad0.cpp \
         tests/test-quantize-fns.cpp \
         tests/test-quantize-perf.cpp \
         tests/test-backend-ops.cpp \
@@ -121,21 +124,22 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml*.c          -> ggml/src/ggml*.c
     # src/ggml*.cpp        -> ggml/src/ggml*.cpp
     # src/ggml*.h          -> ggml/src/ggml*.h
-    # src/ggml*.cu         -> ggml/src/ggml*.cu
-    # src/ggml*.m          -> ggml/src/ggml*.m
-    # src/ggml-amx/*       -> ggml/src/ggml-amx/
-    # src/ggml-cann/*      -> ggml/src/ggml-cann/
-    # src/ggml-cuda/*      -> ggml/src/ggml-cuda/
-    # src/ggml-sycl/*      -> ggml/src/ggml-sycl/
-    # src/vulkan-shaders/* -> ggml/src/vulkan-shaders/
+    # src/ggml-amx/*       -> ggml/src/ggml-amx/*
+    # src/ggml-blas/*      -> ggml/src/ggml-blas/*
+    # src/ggml-cann/*      -> ggml/src/ggml-cann/*
+    # src/ggml-cpu/*       -> ggml/src/ggml-cpu/*
+    # src/ggml-cuda/*      -> ggml/src/ggml-cuda/*
+    # src/ggml-hip/*       -> ggml/src/ggml-hip/*
+    # src/ggml-kompute/*   -> ggml/src/ggml-kompute/*
+    # src/ggml-metal/*     -> ggml/src/ggml-metal/*
+    # src/ggml-musa/*      -> ggml/src/ggml-musa/*
+    # src/ggml-rpc/*       -> ggml/src/ggml-rpc/*
+    # src/ggml-sycl/*      -> ggml/src/ggml-sycl/*
+    # src/ggml-vulkan/*    -> ggml/src/ggml-vulkan/*
     #
     # include/ggml*.h -> ggml/include/ggml*.h
     #
-    # tests/test-opt.cpp           -> tests/test-opt.cpp
-    # tests/test-grad0.cpp         -> tests/test-grad0.cpp
-    # tests/test-quantize-fns.cpp  -> tests/test-quantize-fns.cpp
-    # tests/test-quantize-perf.cpp -> tests/test-quantize-perf.cpp
-    # tests/test-backend-ops.cpp   -> tests/test-backend-ops.cpp
+    # tests/test*.cpp -> tests/
     #
     # LICENSE                -> LICENSE
     # scripts/gen-authors.sh -> scripts/gen-authors.sh
@@ -147,18 +151,20 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.c/\1ggml\/src\/ggml\2.c/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.cpp/\1ggml\/src\/ggml\2.cpp/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.h/\1ggml\/src\/ggml\2.h/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.cu/\1ggml\/src\/ggml\2.cu/g' \
-        -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.m/\1ggml\/src\/ggml\2.m/g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-amx\//\1ggml\/src\/ggml-amx\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-blas\//\1ggml\/src\/ggml-blas\//g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-cpu\//\1ggml\/src\/ggml-cpu\//g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-cuda\//\1ggml\/src\/ggml-cuda\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-hip\//\1ggml\/src\/ggml-hip\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-kompute\//\1ggml\/src\/ggml-kompute\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-metal\//\1ggml\/src\/ggml-metal\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-musa\//\1ggml\/src\/ggml-musa\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-rpc\//\1ggml\/src\/ggml-rpc\//g' \
         -e 's/([[:space:]]|[ab]\/)src\/ggml-sycl\//\1ggml\/src\/ggml-sycl\//g' \
-        -e 's/([[:space:]]|[ab]\/)src\/vulkan-shaders\//\1ggml\/src\/vulkan-shaders\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-vulkan\//\1ggml\/src\/ggml-vulkan\//g' \
         -e 's/([[:space:]]|[ab]\/)include\/ggml(.*)\.h/\1ggml\/include\/ggml\2.h/g' \
-        -e 's/([[:space:]]|[ab]\/)examples\/common\.h/\1examples\/common.h/g' \
-        -e 's/([[:space:]]|[ab]\/)examples\/common\.cpp/\1examples\/common.cpp/g' \
-        -e 's/([[:space:]]|[ab]\/)examples\/common-ggml\.h/\1examples\/common-ggml.h/g' \
-        -e 's/([[:space:]]|[ab]\/)examples\/common-ggml\.cpp/\1examples\/common-ggml.cpp/g' \
+        -e 's/([[:space:]]|[ab]\/)tests\/(.*)\.cpp/\1tests\/\2.cpp/g' \
         -e 's/([[:space:]]|[ab]\/)LICENSE/\1LICENSE/g' \
         -e 's/([[:space:]]|[ab]\/)scripts\/gen-authors\.sh/\1scripts\/gen-authors.sh/g' \
         > ggml-src.patch.tmp
diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh
index f29554c82aced..000270afbfd1e 100755
--- a/scripts/sync-ggml.sh
+++ b/scripts/sync-ggml.sh
@@ -4,21 +4,25 @@ cp -rpv ../ggml/CMakeLists.txt       ./ggml/CMakeLists.txt
 cp -rpv ../ggml/src/CMakeLists.txt   ./ggml/src/CMakeLists.txt
 cp -rpv ../ggml/cmake/FindSIMD.cmake ./ggml/cmake/FindSIMD.cmake
 
-cp -rpv ../ggml/src/ggml*.c          ./ggml/src/
-cp -rpv ../ggml/src/ggml*.cpp        ./ggml/src/
-cp -rpv ../ggml/src/ggml*.h          ./ggml/src/
-cp -rpv ../ggml/src/ggml*.cu         ./ggml/src/
-cp -rpv ../ggml/src/ggml*.m          ./ggml/src/
-cp -rpv ../ggml/src/ggml-amx/*       ./ggml/src/ggml-amx/
-cp -rpv ../ggml/src/ggml-cann/*      ./ggml/src/ggml-cann/
-cp -rpv ../ggml/src/ggml-cuda/*      ./ggml/src/ggml-cuda/
-cp -rpv ../ggml/src/ggml-sycl/*      ./ggml/src/ggml-sycl/
-cp -rpv ../ggml/src/vulkan-shaders/* ./ggml/src/vulkan-shaders/
+cp -rpv ../ggml/src/ggml*.c        ./ggml/src/
+cp -rpv ../ggml/src/ggml*.cpp      ./ggml/src/
+cp -rpv ../ggml/src/ggml*.h        ./ggml/src/
+cp -rpv ../ggml/src/ggml-amx/*     ./ggml/src/ggml-amx/
+cp -rpv ../ggml/src/ggml-blas/*    ./ggml/src/ggml-blas/
+cp -rpv ../ggml/src/ggml-cann/*    ./ggml/src/ggml-cann/
+cp -rpv ../ggml/src/ggml-cpu/*     ./ggml/src/ggml-cpu/
+cp -rpv ../ggml/src/ggml-cuda/*    ./ggml/src/ggml-cuda/
+cp -rpv ../ggml/src/ggml-hip/*     ./ggml/src/ggml-hip/
+cp -rpv ../ggml/src/ggml-kompute/* ./ggml/src/ggml-kompute/
+cp -rpv ../ggml/src/ggml-metal/*   ./ggml/src/ggml-metal/
+cp -rpv ../ggml/src/ggml-musa/*    ./ggml/src/ggml-musa/
+cp -rpv ../ggml/src/ggml-rpc/*     ./ggml/src/ggml-rpc/
+cp -rpv ../ggml/src/ggml-sycl/*    ./ggml/src/ggml-sycl/
+cp -rpv ../ggml/src/ggml-vulkan/*  ./ggml/src/ggml-vulkan/
 
 cp -rpv ../ggml/include/ggml*.h ./ggml/include/
 
 cp -rpv ../ggml/tests/test-opt.cpp           ./tests/test-opt.cpp
-cp -rpv ../ggml/tests/test-grad0.cpp         ./tests/test-grad0.cpp
 cp -rpv ../ggml/tests/test-quantize-fns.cpp  ./tests/test-quantize-fns.cpp
 cp -rpv ../ggml/tests/test-quantize-perf.cpp ./tests/test-quantize-perf.cpp
 cp -rpv ../ggml/tests/test-backend-ops.cpp   ./tests/test-backend-ops.cpp

From 8a43e940ab0daaff198809bf9277289994ec62f5 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Sat, 16 Nov 2024 22:17:59 +0200
Subject: [PATCH 008/102] ggml: new optimization interface (ggml/988)

---
 ggml/CMakeLists.txt                  |    1 +
 ggml/include/ggml-backend.h          |   26 +-
 ggml/include/ggml-opt.h              |  216 ++++
 ggml/include/ggml.h                  |  218 +---
 ggml/src/CMakeLists.txt              |    2 +
 ggml/src/ggml-alloc.c                |   14 +-
 ggml/src/ggml-backend.cpp            |    2 +-
 ggml/src/ggml-cpu/ggml-cpu.c         |   41 +-
 ggml/src/ggml-cuda/opt-step-adamw.cu |   68 +-
 ggml/src/ggml-impl.h                 |   19 +-
 ggml/src/ggml-metal/ggml-metal.m     |    8 +-
 ggml/src/ggml-opt.cpp                |  867 +++++++++++++
 ggml/src/ggml.c                      | 1748 +++++++++-----------------
 tests/test-backend-ops.cpp           |   79 +-
 tests/test-opt.cpp                   |  995 ++++++++++++---
 15 files changed, 2667 insertions(+), 1637 deletions(-)
 create mode 100644 ggml/include/ggml-opt.h
 create mode 100644 ggml/src/ggml-opt.cpp

diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 4fb78e59fa72c..fd94998264547 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -228,6 +228,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-cann.h
     include/ggml-cuda.h
     include/ggml-kompute.h
+    include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
     include/ggml-sycl.h
diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
index 0a65dbfcaf5e8..cef164764bb1a 100644
--- a/ggml/include/ggml-backend.h
+++ b/ggml/include/ggml-backend.h
@@ -86,7 +86,7 @@ extern "C" {
     GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
-    // "offset" refers to the offset of the tensor data for setting/getting data
+    // "offset" refers to the offset in tensor->data for setting/getting data
     GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_memset(   struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
@@ -242,14 +242,20 @@ extern "C" {
         ggml_backend_sched_reserve(sched, reserve_graph);
 
         // compute
-        graph = build_graph(sched);
-        ggml_backend_sched_graph_compute(sched, graph);
+        graph = build_graph(sched); // the graph and its tensors are single-use in terms of allocation, multi-use in terms of computation
+        for (int i = 0; i < 10; ++i) {
+            ggml_backend_sched_graph_compute(sched, graph); // on the first iteration the graph is allocated automatically
+        }
 
         // if there are graph inputs:
-        ggml_backend_sched_reset(sched);
-        ggml_backend_sched_alloc_graph(sched, graph);
-        ggml_backend_tensor_set(input_tensor, ...);
-        ggml_backend_sched_graph_compute(sched, graph);
+        graph = build_graph(sched); // get a new graph that is not allocated (the metadata for the old graph is freed once ggml_free is called)
+        ggml_backend_sched_reset(sched); // clear the allocation of the previous graph
+        ggml_backend_sched_alloc_graph(sched, graph); // explicitly allocate the new graph but do not execute it
+        ggml_backend_tensor_set(input_tensor, ...); // copy data to the newly allocated graph tensors
+        ggml_backend_sched_graph_compute(sched, graph); // execute the graph
+
+        // as an alternative to the above it is also possible to assign the inputs to a dedicated context and
+        // allocate them statically via ggml_backend_alloc_ctx_tensors
     }
     */
 
@@ -264,7 +270,7 @@ extern "C" {
     //
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
-    // Initialize a backend scheduler
+    // Initialize a backend scheduler, backends with low index are given priority over backends with high index
     GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
@@ -289,7 +295,9 @@ extern "C" {
     GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
     GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
 
-    // Reset all assignments and allocators - must be called before changing the node backends
+    // Reset all assignments and allocators - must be called before changing the node backends or allocating a new graph.
+    // This in effect deallocates all tensors that were previously allocated and leaves them with dangling pointers.
+    // The correct way to use this API is to discard the deallocated tensors and create new ones.
     GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
 
     // Set a callback to be called for each resulting node during graph compute
diff --git a/ggml/include/ggml-opt.h b/ggml/include/ggml-opt.h
new file mode 100644
index 0000000000000..eb5eab9de6781
--- /dev/null
+++ b/ggml/include/ggml-opt.h
@@ -0,0 +1,216 @@
+// This file contains functionality for training models using GGML.
+// It is not strictly needed vs. just vanilla GGML but it provides a more high-level interface for common needs such as datasets.
+// At the bottom of this file especially there are relatively high-level functions that are suitable use or adaptation in user code.
+//
+// Module maintainer: Johannes Gäßler (@JohannesGaessler, johannesg@5d6.de)
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include <stdint.h>
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    struct ggml_opt_dataset;
+    struct ggml_opt_context;
+    struct ggml_opt_result;
+
+    typedef struct ggml_opt_dataset * ggml_opt_dataset_t;
+    typedef struct ggml_opt_context * ggml_opt_context_t;
+    typedef struct ggml_opt_result  * ggml_opt_result_t;
+
+    // ====== Loss ======
+
+    // built-in loss types, i.e. the built-in quantities minimized by the optimizer
+    // custom loss types can be defined via mean or sum which simply reduce the outputs for all datapoints to a single value
+    enum ggml_opt_loss_type {
+        GGML_OPT_LOSS_TYPE_MEAN,
+        GGML_OPT_LOSS_TYPE_SUM,
+        GGML_OPT_LOSS_TYPE_CROSS_ENTROPY,
+        GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR,
+    };
+
+    // ====== Dataset ======
+
+    GGML_API ggml_opt_dataset_t ggml_opt_dataset_init(
+            int64_t ne_datapoint, // number of elements per datapoint
+            int64_t ne_label,     // number of elements per label
+            int64_t ndata,        // total number of datapoints/labels
+            int64_t ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
+    GGML_API void ggml_opt_dataset_free(ggml_opt_dataset_t dataset);
+
+    // get underlying tensors that store the data
+    GGML_API struct ggml_tensor * ggml_opt_dataset_data  (ggml_opt_dataset_t dataset); // shape = [ne_datapoint, ndata]
+    GGML_API struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset); // shape = [nd_label,     ndata]
+
+    // shuffle idata first datapoints from dataset with RNG from opt_ctx, shuffle all datapoints if idata is negative
+    GGML_API void ggml_opt_dataset_shuffle(ggml_opt_context_t opt_ctx, ggml_opt_dataset_t dataset, int64_t idata);
+
+    // get batch at position ibatch from dataset and copy the data to data_batch and labels_batch
+    GGML_API void ggml_opt_dataset_get_batch(
+            ggml_opt_dataset_t   dataset,
+            struct ggml_tensor * data_batch,   // shape = [ne_datapoint, ndata_batch]
+            struct ggml_tensor * labels_batch, // shape = [ne_label,     ndata_batch]
+            int64_t              ibatch);
+
+    // ====== Model / Context ======
+
+    enum ggml_opt_build_type {
+        GGML_OPT_BUILD_TYPE_FORWARD,
+        GGML_OPT_BUILD_TYPE_GRAD,
+        GGML_OPT_BUILD_TYPE_OPT,
+    };
+
+    // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
+    struct ggml_opt_optimizer_params {
+        // AdamW optimizer parameters
+        struct {
+            float alpha; // learning rate
+            float beta1;
+            float beta2;
+            float eps;   // epsilon for numerical stability
+            float wd;    // weight decay for AdamW, use 0.0f to disable
+        } adamw;
+    };
+
+    // callback to calculate optimizer parameters prior to a backward pass
+    // userdata can be used to pass arbitrary data
+    typedef struct ggml_opt_optimizer_params (*ggml_opt_get_optimizer_params)(void * userdata);
+
+    // returns the default optimizer params (constant)
+    // userdata is not used
+    GGML_API struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata);
+
+    // parameters for initializing a new optimization context
+    struct ggml_opt_params {
+        ggml_backend_sched_t backend_sched; // defines which backends are used to construct the compute graphs
+
+        struct ggml_context * ctx_compute; // created in user code, holds non-static tensors
+
+        // the forward graph is defined by inputs and outputs
+        // those tensors and all tensors inbetween are not intended to be reusable between multiple optimization contexts
+        struct ggml_tensor * inputs;
+        struct ggml_tensor * outputs;
+
+        enum ggml_opt_loss_type  loss_type;
+        enum ggml_opt_build_type build_type;
+
+        int32_t opt_period; // after how many gradient accumulation steps an optimizer step should be done
+
+        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
+        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+    };
+
+    // get parameters for an optimization context with defaults set where possible
+    // parameters for which no sensible defaults exist are supplied as arguments to this function
+    GGML_API ggml_opt_params ggml_opt_default_params(
+            ggml_backend_sched_t      backend_sched,
+            struct ggml_context     * ctx_compute,
+            struct ggml_tensor      * inputs,
+            struct ggml_tensor      * outputs,
+            enum ggml_opt_loss_type   loss_type);
+
+    GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
+    GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
+
+    // set gradients to zero, initilize loss, and optionally reset the optimizer
+    GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
+
+    // get underlying tensors that store data
+    GGML_API struct ggml_tensor * ggml_opt_inputs(  ggml_opt_context_t opt_ctx); // forward graph input tensor
+    GGML_API struct ggml_tensor * ggml_opt_outputs( ggml_opt_context_t opt_ctx); // forward graph output tensor
+    GGML_API struct ggml_tensor * ggml_opt_labels(  ggml_opt_context_t opt_ctx); // labels to compare outputs against
+    GGML_API struct ggml_tensor * ggml_opt_loss(    ggml_opt_context_t opt_ctx); // scalar tensor that contains the loss
+    GGML_API struct ggml_tensor * ggml_opt_pred(    ggml_opt_context_t opt_ctx); // predictions made by outputs
+    GGML_API struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx); // number of matching predictions between outputs and labels
+
+    GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
+
+    // ====== Optimization Result ======
+
+    GGML_API ggml_opt_result_t ggml_opt_result_init();
+    GGML_API void ggml_opt_result_free(ggml_opt_result_t result);
+    GGML_API void ggml_opt_result_reset(ggml_opt_result_t result);
+
+    // get data from result, uncertainties are optional and can be ignored by passing NULL
+    GGML_API void ggml_opt_result_ndata(   ggml_opt_result_t result, int64_t * ndata);                  // writes 1 value, number of datapoints
+    GGML_API void ggml_opt_result_loss(    ggml_opt_result_t result, double  * loss,     double * unc); // writes 1 value
+    GGML_API void ggml_opt_result_pred(    ggml_opt_result_t result, int32_t * pred);                   // writes ndata values
+    GGML_API void ggml_opt_result_accuracy(ggml_opt_result_t result, double  * accuracy, double * unc); // writes 1 value
+
+    // ====== Computation ======
+
+    // do forward pass, increment result if not NULL
+    GGML_API void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+
+    // do forward pass, increment result if not NULL, do backward pass
+    GGML_API void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+
+    // ############################################################################
+    // ## The high-level functions start here. They do not depend on any private ##
+    // ## functions or structs and can be copied to and adapted for user code.   ##
+    // ############################################################################
+
+    // ====== Intended Usage ======
+    //
+    // 1. Select the appropriate loss for your problem.
+    // 2. Create a dataset and set the data for the "data" tensor. Also set the "labels" tensor if your loss needs them.
+    //    Setting the shard size to 1 will be fine, it's the granularity with which data is shuffled/loaded (bigger values are faster).
+    // 3. Create a GGML graph for your model with no_alloc == true. Use two separate contexts for the tensors.
+    //    The first context should contain the model parameters and inputs and be allocated statically in user code.
+    //    The second context should contain all other tensors and will be (re)allocated automatically.
+    //    Due to this automated allocation the data of the second context is not defined when accessed in user code.
+    //    Note that the second dimension of the inputs/outputs are interpreted as the number of datapoints in those tensors.
+    // 4. Call ggml_opt_fit. If you need more control you can use ggml_opt_epoch instead.
+
+    // signature for a callback while evaluating opt_ctx on dataset, called after an evaluation
+    typedef void (*ggml_opt_epoch_callback)(
+            bool               train,       // true after training evaluation, false after validation evaluation
+            ggml_opt_context_t opt_ctx,
+            ggml_opt_dataset_t dataset,
+            ggml_opt_result_t  result,      // result associated with the dataset subsection
+            int64_t            ibatch,      // number of batches that have been evaluated so far
+            int64_t            ibatch_max,  // total number of batches in this dataset subsection
+            int64_t            t_start_us); // time at which the evaluation on the dataset subsection was started
+
+    // do training on front of dataset, do evaluation only on back of dataset
+    GGML_API void ggml_opt_epoch(
+            ggml_opt_context_t      opt_ctx,
+            ggml_opt_dataset_t      dataset,
+            ggml_opt_result_t       result_train,   // result to increment during training, ignored if NULL
+            ggml_opt_result_t       result_eval,    // result to increment during evaluation, ignored if NULL
+            int64_t                 idata_split,    // data index at which to split training and evaluation
+            ggml_opt_epoch_callback callback_train,
+            ggml_opt_epoch_callback callback_eval);
+
+    // callback that prints a progress bar on stderr
+    GGML_API void ggml_opt_epoch_callback_progress_bar(
+            bool               train,
+            ggml_opt_context_t opt_ctx,
+            ggml_opt_dataset_t dataset,
+            ggml_opt_result_t  result,
+            int64_t            ibatch,
+            int64_t            ibatch_max,
+            int64_t            t_start_us);
+
+    // fit model defined by inputs and outputs to dataset
+    GGML_API void ggml_opt_fit(
+            ggml_backend_sched_t            backend_sched,  // backend scheduler for constructing the compute graphs
+            ggml_context                  * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
+            ggml_tensor                   * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
+            ggml_tensor                   * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
+            ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
+            enum ggml_opt_loss_type         loss_type,      // loss to minimize
+            ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)
+            int64_t                         nepoch,         // how many times the dataset should be iterated over
+            int64_t                         nbatch_logical, // datapoints optimizer step, must be a multiple of ndata_batch in inputs/outputs
+            float                           val_split,      // fraction of the dataset to use for validation, must be in [0.0f, 1.0f)
+            bool                            silent);        // whether or not info prints to stderr should be suppressed
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 3b3f6798a36a7..69e6a24344b97 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -602,7 +602,6 @@ extern "C" {
 
         int32_t flags;
 
-        struct ggml_tensor * grad;
         struct ggml_tensor * src[GGML_MAX_SRC];
 
         // source tensor and offset for views
@@ -615,7 +614,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        // char padding[4];
+        char padding[8];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -1985,28 +1984,20 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * grad,
-            float                 alpha,
-            float                 beta1,
-            float                 beta2,
-            float                 eps,
-            float                 wd); // weight decay
+            struct ggml_tensor  * m,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * adamw_params); // parameters such a the learning rate
 
     //
     // automatic differentiation
     //
 
-    GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
-    GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate);
-
-    GGML_API void ggml_build_opt_adamw(
-            struct ggml_context * ctx,
-            struct ggml_cgraph  * gf,
-            struct ggml_cgraph  * gb,
-            float                 alpha,
-            float                 beta1,
-            float                 beta2,
-            float                 eps,
-            float                 wd); // weight decay
+    GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
+    GGML_API void ggml_build_backward_expand(
+        struct ggml_context * ctx_static,  // context for static gradients (loss + gradient accumulation)
+        struct ggml_context * ctx_compute, // context for gradient computation
+        struct ggml_cgraph  * cgraph,
+        bool                  accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
 
     // graph allocation in a context
     GGML_API struct ggml_cgraph * ggml_new_graph       (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
@@ -2026,7 +2017,9 @@ extern "C" {
     GGML_API size_t ggml_graph_overhead(void);
     GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
 
-    GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
+    GGML_API struct ggml_tensor * ggml_graph_get_tensor  (const struct ggml_cgraph * cgraph, const char * name);
+    GGML_API struct ggml_tensor * ggml_graph_get_grad    (const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
+    GGML_API struct ggml_tensor * ggml_graph_get_grad_acc(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
 
     GGML_API void                 ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
     GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
@@ -2037,198 +2030,15 @@ extern "C" {
     // dump the graph into a file using the dot format
     GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
 
-    // build gradient checkpointing backward graph gb for gf using provided checkpoints
-    // gb_tmp will contain original backward graph with rewritten backward process nodes,
-    // but without the second forward pass nodes.
-    GGML_API void ggml_build_backward_gradient_checkpointing(
-            struct ggml_context   * ctx,
-            struct ggml_cgraph    * gf,
-            struct ggml_cgraph    * gb,
-            struct ggml_cgraph    * gb_tmp,
-            struct ggml_tensor  * * checkpoints,
-            int                     n_checkpoints);
-    //
-    // optimization
-    //
-
-    // optimization methods
-    enum ggml_opt_type {
-        GGML_OPT_TYPE_ADAM,
-        GGML_OPT_TYPE_LBFGS,
-    };
-
-    // linesearch methods
-    enum ggml_linesearch {
-        GGML_LINESEARCH_DEFAULT = 1,
-
-        GGML_LINESEARCH_BACKTRACKING_ARMIJO       = 0,
-        GGML_LINESEARCH_BACKTRACKING_WOLFE        = 1,
-        GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2,
-    };
-
-    // optimization return values
-    enum ggml_opt_result {
-        GGML_OPT_RESULT_OK = 0,
-        GGML_OPT_RESULT_DID_NOT_CONVERGE,
-        GGML_OPT_RESULT_NO_CONTEXT,
-        GGML_OPT_RESULT_INVALID_WOLFE,
-        GGML_OPT_RESULT_FAIL,
-        GGML_OPT_RESULT_CANCEL,
-
-        GGML_LINESEARCH_FAIL = -128,
-        GGML_LINESEARCH_MINIMUM_STEP,
-        GGML_LINESEARCH_MAXIMUM_STEP,
-        GGML_LINESEARCH_MAXIMUM_ITERATIONS,
-        GGML_LINESEARCH_INVALID_PARAMETERS,
-    };
-
-    typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
+    // TODO these functions were sandwiched in the old optimization interface, is there a better place for them?
     typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
     GGML_API void ggml_log_set(ggml_log_callback log_callback, void * user_data);
 
-    // optimization parameters
-    //
-    //   see ggml.c (ggml_opt_default_params) for default values
-    //
-    struct ggml_opt_params {
-        enum ggml_opt_type type;
-
-        size_t graph_size;
-
-        int n_threads;
-
-        // delta-based convergence test
-        //
-        //   if past == 0 - disabled
-        //   if past > 0:
-        //     stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|)
-        //
-        int past;
-        float delta;
-
-        // maximum number of iterations without improvement
-        //
-        //   if 0 - disabled
-        //   if > 0:
-        //     assume convergence if no cost improvement in this number of iterations
-        //
-        int max_no_improvement;
-
-        bool print_forward_graph;
-        bool print_backward_graph;
-
-        int n_gradient_accumulation;
-
-        // ADAM parameters
-        struct {
-            int n_iter;
-
-            float sched; // schedule multiplier (fixed, decay or warmup)
-            float decay; // weight decay for AdamW, use 0.0f to disable
-            int   decay_min_ndim; // minimum number of tensor dimension to apply weight decay
-            float alpha; // learning rate
-            float beta1;
-            float beta2;
-            float eps;   // epsilon for numerical stability
-            float eps_f; // epsilon for convergence test
-            float eps_g; // epsilon for convergence test
-            float gclip; // gradient clipping
-        } adam;
-
-        // LBFGS parameters
-        struct {
-            int m; // number of corrections to approximate the inv. Hessian
-            int n_iter;
-            int max_linesearch;
-
-            float eps;      // convergence tolerance
-            float ftol;     // line search tolerance
-            float wolfe;
-            float min_step;
-            float max_step;
-
-            enum ggml_linesearch linesearch;
-        } lbfgs;
-    };
-
-    struct ggml_opt_context {
-        struct ggml_context * ctx;
-        struct ggml_opt_params params;
-
-        int iter;
-        int64_t nx; // number of parameter elements
-
-        bool just_initialized;
-
-        float loss_before;
-        float loss_after;
-
-        struct {
-            struct ggml_tensor * g;  // current gradient
-            struct ggml_tensor * m;  // first moment
-            struct ggml_tensor * v;  // second moment
-            struct ggml_tensor * pf; // past function values
-            float fx_best;
-            float fx_prev;
-            int n_no_improvement;
-        } adam;
-
-        struct {
-            struct ggml_tensor * x;    // current parameters
-            struct ggml_tensor * xp;   // previous parameters
-            struct ggml_tensor * g;    // current gradient
-            struct ggml_tensor * gp;   // previous gradient
-            struct ggml_tensor * d;    // search direction
-            struct ggml_tensor * pf;   // past function values
-            struct ggml_tensor * lmal; // the L-BFGS memory alpha
-            struct ggml_tensor * lmys; // the L-BFGS memory ys
-            struct ggml_tensor * lms;  // the L-BFGS memory s
-            struct ggml_tensor * lmy;  // the L-BFGS memory y
-            float fx_best;
-            float step;
-            int j;
-            int k;
-            int end;
-            int n_no_improvement;
-        } lbfgs;
-    };
-
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
 
-    GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
-
-    // optimize the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt(
-            struct ggml_context * ctx,
-            struct ggml_opt_params params,
-            struct ggml_tensor * f);
-
-    // initialize optimizer context
-    GGML_API void ggml_opt_init(
-            struct ggml_context     * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_opt_params    params,
-            int64_t                   nx);
-
-    // continue optimizing the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt_resume(
-            struct ggml_context * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_tensor * f);
-
-    // continue optimizing the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt_resume_g(
-            struct ggml_context * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_tensor * f,
-            struct ggml_cgraph * gf,
-            struct ggml_cgraph * gb,
-            ggml_opt_callback callback,
-            void * callback_data);
-
     //
     // quantization
     //
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index 71934c6791e32..ae7d3abc8de32 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -207,9 +207,11 @@ add_library(ggml-base
             ../include/ggml-alloc.h
             ../include/ggml-backend.h
             ../include/ggml-cpp.h
+            ../include/ggml-opt.h
             ggml.c
             ggml-alloc.c
             ggml-backend.cpp
+            ggml-opt.cpp
             ggml-threading.cpp
             ggml-threading.h
             ggml-quants.c
diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
index 041de9e3efc69..2b2240be85dd3 100644
--- a/ggml/src/ggml-alloc.c
+++ b/ggml/src/ggml-alloc.c
@@ -466,18 +466,12 @@ static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) {
     return ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
-static void ggml_gallocr_set_node_offset(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, size_t offset) {
-    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
-    hn->buffer_id = buffer_id;
-    hn->offset = offset;
-    hn->allocated = true;
-}
-
 static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) {
     return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
 static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
+    GGML_ASSERT(buffer_id >= 0);
     struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
 
     if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
@@ -816,7 +810,11 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
 }
 
 static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
-    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    size_t node_size = 0;
+    if (!node->data && !node->view_src) {
+        GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
+        node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    }
     return talloc->size_max >= node_size;
 }
 
diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index e48877ba8fa5a..634fe38eef871 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -279,7 +279,7 @@ void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, siz
     buf->iface.get_tensor(buf, tensor, data, offset, size);
 }
 
-GGML_API void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
index 61f53cd0145a6..df64879294981 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -12216,11 +12216,16 @@ static void ggml_compute_forward_opt_step_adamw_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
 
-    const struct ggml_tensor * src0        = dst->src[0];
-    const struct ggml_tensor * src0_grad   = dst->src[1];
-    const struct ggml_tensor * src0_grad_m = dst->src[2];
-    const struct ggml_tensor * src0_grad_v = dst->src[3];
+    const struct ggml_tensor * src0         = dst->src[0];
+    const struct ggml_tensor * src0_grad    = dst->src[1];
+    const struct ggml_tensor * src0_grad_m  = dst->src[2];
+    const struct ggml_tensor * src0_grad_v  = dst->src[3];
+    const struct ggml_tensor * adamw_params = dst->src[4];
+
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
+    GGML_ASSERT(ggml_nelements(adamw_params) == 7);
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -12237,16 +12242,14 @@ static void ggml_compute_forward_opt_step_adamw_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    /* const float   gnorm = 1.0f; */
-    int64_t       iter;   memcpy(&iter, &dst->op_params[0], sizeof(int64_t));
-    const float   alpha = ggml_get_op_params_f32(dst, 2);
-    const float   beta1 = ggml_get_op_params_f32(dst, 3);
-    const float   beta2 = ggml_get_op_params_f32(dst, 4);
-    const float   eps   = ggml_get_op_params_f32(dst, 5);
-    const float   wd    = ggml_get_op_params_f32(dst, 6);
-
-    const float beta1h  = alpha/(1.0f - powf(beta1, iter));
-    const float beta2h  =  1.0f/(1.0f - powf(beta2, iter));
+    const float * adamw_params_ptr = ggml_get_data_f32(adamw_params);
+    const float alpha  = adamw_params_ptr[0];
+    const float beta1  = adamw_params_ptr[1];
+    const float beta2  = adamw_params_ptr[2];
+    const float eps    = adamw_params_ptr[3];
+    const float wd     = adamw_params_ptr[4];
+    const float beta1h = adamw_params_ptr[5];
+    const float beta2h = adamw_params_ptr[6];
 
     for (int ir = ir0; ir < ir1; ++ir) {
         const int64_t i03 = ir/(ne02*ne01);
@@ -12270,17 +12273,9 @@ static void ggml_compute_forward_opt_step_adamw_f32(
             // The weight decay is applied independently of the Adam momenta m and v.
             // This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss.
             // See: https://arxiv.org/pdf/1711.05101v3.pdf
-            w[i00] = w[i00]*(1.0f - alpha*wd) - mh/vh;
+            w[i00] = w[i00]*(1.0f - alpha*wd) - alpha*mh/vh;
         }
     }
-
-    ggml_barrier(params->threadpool);
-    if (ith != 0) {
-        return;
-    }
-
-    iter++;
-    memcpy(&dst->op_params[0], &iter, sizeof(int64_t));
 }
 
 static void ggml_compute_forward_opt_step_adamw(
diff --git a/ggml/src/ggml-cuda/opt-step-adamw.cu b/ggml/src/ggml-cuda/opt-step-adamw.cu
index d6f13a9c62df2..35154f2996652 100644
--- a/ggml/src/ggml-cuda/opt-step-adamw.cu
+++ b/ggml/src/ggml-cuda/opt-step-adamw.cu
@@ -1,11 +1,11 @@
+#include "ggml-impl.h"
 #include "opt-step-adamw.cuh"
 
 #include <cstdint>
 
 static __global__ void opt_step_adamw_f32(
-    float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v, const int64_t k,
-    const float alpha, const float beta1, const float beta2, const float eps, const float wd,
-    const float beta1h, const float beta2h) {
+    float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v,
+    const float * __restrict__ pars, const int64_t k) {
 
     const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
 
@@ -13,6 +13,14 @@ static __global__ void opt_step_adamw_f32(
         return;
     }
 
+    const float alpha  = pars[0];
+    const float beta1  = pars[1];
+    const float beta2  = pars[2];
+    const float eps    = pars[3];
+    const float wd     = pars[4];
+    const float beta1h = pars[5];
+    const float beta2h = pars[6];
+
     const float gi = g[i];
     const float gmi = g_m[i]*beta1 +    gi*(1.0f - beta1);
     const float gvi = g_v[i]*beta2 + gi*gi*(1.0f - beta2);
@@ -23,58 +31,48 @@ static __global__ void opt_step_adamw_f32(
     const float mh =       gmi*beta1h;
     const float vh = sqrtf(gvi*beta2h) + eps;
 
-    x[i] = x[i]*(1.0f - alpha*wd) - mh/vh;
+    x[i] = x[i]*(1.0f - alpha*wd) - alpha*mh/vh;
 }
 
 static void opt_step_adamw_f32_cuda(
-    float * x, const float * g, float * g_m, float * g_v, const int64_t k,
-    const float alpha, const float beta1, const float beta2, const float eps, const float wd,
-    const float beta1h, const float beta2h, cudaStream_t stream) {
+    float * x, const float * g, float * g_m, float * g_v, const float * pars, const int64_t k, cudaStream_t stream) {
 
     const dim3 block_dims(CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
     const dim3 block_nums((k + CUDA_OPT_STEP_ADAMW_BLOCK_SIZE - 1) / CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
-    opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, k, alpha, beta1, beta2, eps, wd, beta1h, beta2h);
+    opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, pars, k);
 }
 
 void ggml_cuda_opt_step_adamw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0        = dst->src[0];
-    const ggml_tensor * src0_grad   = dst->src[1];
-    const ggml_tensor * src0_grad_m = dst->src[2];
-    const ggml_tensor * src0_grad_v = dst->src[3];
-
-    GGML_ASSERT(src0->type        == GGML_TYPE_F32);
-    GGML_ASSERT(src0_grad->type   == GGML_TYPE_F32);
-    GGML_ASSERT(src0_grad_m->type == GGML_TYPE_F32);
-    GGML_ASSERT(src0_grad_v->type == GGML_TYPE_F32);
+    const ggml_tensor * src0         = dst->src[0];
+    const ggml_tensor * src0_grad    = dst->src[1];
+    const ggml_tensor * src0_grad_m  = dst->src[2];
+    const ggml_tensor * src0_grad_v  = dst->src[3];
+    const ggml_tensor * adamw_params = dst->src[4];
+
+    GGML_ASSERT(src0->type         == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad->type    == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad_m->type  == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad_v->type  == GGML_TYPE_F32);
+    GGML_ASSERT(adamw_params->type == GGML_TYPE_F32);
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(src0_grad));
     GGML_ASSERT(ggml_is_contiguous(src0_grad_m));
     GGML_ASSERT(ggml_is_contiguous(src0_grad_v));
+    GGML_ASSERT(ggml_is_contiguous(adamw_params));
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
+    GGML_ASSERT(ggml_nelements(adamw_params) == 7);
 
-    float       * src0_d        = (float       *) src0->data;
-    const float * src0_grad_d   = (const float *) src0_grad->data;
-    float       * src0_grad_m_d = (float       *) src0_grad_m->data;
-    float       * src0_grad_v_d = (float       *) src0_grad_v->data;
+    float       * src0_d         = (float       *) src0->data;
+    const float * src0_grad_d    = (const float *) src0_grad->data;
+    float       * src0_grad_m_d  = (float       *) src0_grad_m->data;
+    float       * src0_grad_v_d  = (float       *) src0_grad_v->data;
+    const float * adamw_params_d = (const float *) adamw_params->data;
 
     cudaStream_t stream = ctx.stream();
 
     const int64_t ne = ggml_nelements(src0);
 
-    int64_t iter;  memcpy(&iter,  &dst->op_params[0], sizeof(int64_t));
-    float   alpha; memcpy(&alpha, &dst->op_params[2], sizeof(float));
-    float   beta1; memcpy(&beta1, &dst->op_params[3], sizeof(float));
-    float   beta2; memcpy(&beta2, &dst->op_params[4], sizeof(float));
-    float   eps;   memcpy(&eps,   &dst->op_params[5], sizeof(float));
-    float   wd;    memcpy(&wd,    &dst->op_params[6], sizeof(float));
-
-    const float beta1h  = alpha/(1.0f - powf(beta1, iter));
-    const float beta2h  =  1.0f/(1.0f - powf(beta2, iter));
-
-    opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, ne, alpha, beta1, beta2, eps, wd, beta1h, beta2h, stream);
-
-    iter++;
-    memcpy(&dst->op_params[0], &iter, sizeof(int64_t));
+    opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, adamw_params_d, ne, stream);
 }
diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
index aa4d2b85d7d7c..92a64fe5a5930 100644
--- a/ggml/src/ggml-impl.h
+++ b/ggml/src/ggml-impl.h
@@ -196,7 +196,7 @@ void ggml_hash_set_reset(struct ggml_hash_set * hash_set);
 static bool ggml_hash_contains(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
 
 // returns GGML_HASHSET_FULL if table is full, otherwise the current index of the key or where it should be inserted
-static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, const struct ggml_tensor * key);
 
 // returns GGML_HASHSET_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
 static size_t ggml_hash_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key);
@@ -210,7 +210,7 @@ static inline size_t ggml_hash(const struct ggml_tensor * p) {
     return (size_t)(uintptr_t)p >> 4;
 }
 
-static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, const struct ggml_tensor * key) {
     size_t h = ggml_hash(key) % hash_set->size;
 
     // linear probing
@@ -281,13 +281,14 @@ enum ggml_cgraph_eval_order {
 };
 
 struct ggml_cgraph {
-    int size;
-    int n_nodes;
-    int n_leafs;
-
-    struct ggml_tensor ** nodes;
-    struct ggml_tensor ** grads;
-    struct ggml_tensor ** leafs;
+    int size;    // maximum number of nodes/leafs/grads/grad_accs
+    int n_nodes; // number of nodes currently in use
+    int n_leafs; // number of leafs currently in use
+
+    struct ggml_tensor ** nodes;     // tensors with data that can change if the graph is evaluated
+    struct ggml_tensor ** grads;     // the outputs of these tensors are the gradients of the nodes
+    struct ggml_tensor ** grad_accs; // accumulators for node gradients
+    struct ggml_tensor ** leafs;     // tensors with constant data
 
     struct ggml_hash_set visited_hash_set;
 
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index b4b5cfd26891d..95b21fbf9c503 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -3639,6 +3639,12 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
     return ctx->all_data;
 }
 
+static void ggml_backend_metal_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    memset((char *)tensor->data + offset, value, size);
+
+    UNUSED(buffer);
+}
+
 static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     memcpy((char *)tensor->data + offset, data, size);
 
@@ -3671,7 +3677,7 @@ static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_
     /* .free_buffer     = */ ggml_backend_metal_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_metal_buffer_get_base,
     /* .init_tensor     = */ NULL,
-    /* .memset_tensor   = */ NULL,
+    /* .memset_tensor   = */ ggml_backend_metal_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_metal_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_metal_buffer_get_tensor,
     /* .cpy_tensor      = */ ggml_backend_metal_buffer_cpy_tensor,
diff --git a/ggml/src/ggml-opt.cpp b/ggml/src/ggml-opt.cpp
new file mode 100644
index 0000000000000..808aa0d023ab2
--- /dev/null
+++ b/ggml/src/ggml-opt.cpp
@@ -0,0 +1,867 @@
+#include "ggml-opt.h"
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <inttypes.h>
+#include <map>
+#include <random>
+#include <vector>
+
+struct ggml_opt_dataset {
+    struct ggml_context   * ctx;
+    ggml_backend_buffer_t   buf;
+    struct ggml_tensor    * data;
+    struct ggml_tensor    * labels;
+
+    int64_t ndata;
+    int64_t ndata_shard;
+    size_t  nbs_data;
+    size_t  nbs_labels;
+
+    std::vector<int64_t> permutation;
+};
+
+struct ggml_opt_context {
+    ggml_backend_sched_t    backend_sched;
+    ggml_cgraph           * allocated_graph;
+    ggml_cgraph           * allocated_graph_copy;
+    struct ggml_context   * ctx_static;
+    struct ggml_context   * ctx_static_cpu;
+    struct ggml_context   * ctx_compute;
+    struct ggml_context   * ctx_copy;
+    ggml_backend_buffer_t   buf_static;
+    ggml_backend_buffer_t   buf_static_cpu;
+    std::mt19937            rng;
+
+    struct ggml_tensor * inputs;
+    struct ggml_tensor * outputs;
+    struct ggml_tensor * labels;
+
+    struct ggml_tensor * loss;
+    struct ggml_tensor * pred;
+    struct ggml_tensor * ncorrect;
+
+    struct ggml_cgraph * gf;
+    struct ggml_cgraph * gb_grad;
+    struct ggml_cgraph * gb_opt;
+
+    int64_t iter;
+    int32_t opt_period;
+    int32_t opt_i;
+    bool    loss_per_datapoint;
+
+    ggml_opt_get_optimizer_params get_opt_pars;
+    void * get_opt_pars_ud;
+    struct ggml_tensor * adamw_params;
+};
+
+struct ggml_opt_result {
+    int64_t              ndata    = 0;
+    std::vector<float>   loss;
+    std::vector<int32_t> pred;
+    int64_t              ncorrect = 0;
+
+    bool loss_per_datapoint = false;
+    int64_t opt_period = -1;
+};
+
+// ====== Dataset ======
+
+ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label, int64_t ndata, int64_t ndata_shard) {
+    GGML_ASSERT(ne_datapoint >  0);
+    GGML_ASSERT(ne_label     >= 0);
+    GGML_ASSERT(ndata        >  0);
+    GGML_ASSERT(ndata_shard  >  0);
+
+    ggml_opt_dataset_t result = new ggml_opt_dataset;
+    result->ndata       = ndata;
+    result->ndata_shard = ndata_shard;
+
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ 2*ggml_tensor_overhead(),
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        result->ctx = ggml_init(params);
+    }
+
+    result->data = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_datapoint, ndata);
+    result->nbs_data = ggml_nbytes(result->data) * ndata_shard/ndata;
+
+    if (ne_label > 0) {
+        result->labels = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_label, ndata);
+        result->nbs_labels = ggml_nbytes(result->labels) * ndata_shard/ndata;
+    } else {
+        result->labels = nullptr;
+        result->nbs_labels = 0;
+    }
+
+    result->buf = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx, ggml_backend_cpu_buffer_type());
+
+    const int64_t nshards = ndata/ndata_shard;
+    result->permutation.resize(nshards);
+    for (int64_t i = 0; i < nshards; ++i) {
+        result->permutation[i] = i;
+    }
+    return result;
+}
+
+void ggml_opt_dataset_free(ggml_opt_dataset_t dataset) {
+    ggml_backend_buffer_free(dataset->buf);
+    ggml_free(dataset->ctx);
+    delete dataset;
+}
+
+struct ggml_tensor * ggml_opt_dataset_data(ggml_opt_dataset_t dataset) {
+    return dataset->data;
+}
+
+struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset) {
+    return dataset->labels;
+}
+
+void ggml_opt_dataset_shuffle(ggml_opt_context_t opt_ctx, ggml_opt_dataset_t dataset, int64_t idata) {
+    GGML_ASSERT(idata <= dataset->ndata);
+
+    if (idata < 0) {
+        std::shuffle(dataset->permutation.begin(), dataset->permutation.end(), opt_ctx->rng);
+        return;
+    }
+
+    GGML_ASSERT(idata % dataset->ndata_shard == 0);
+    const int64_t ishard_max = idata / dataset->ndata_shard;
+    std::shuffle(dataset->permutation.begin(), dataset->permutation.begin() + ishard_max, opt_ctx->rng);
+}
+
+void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor * data_batch, struct ggml_tensor * labels_batch, int64_t ibatch) {
+    GGML_ASSERT(   data_batch && ggml_is_contiguous(data_batch));
+    GGML_ASSERT(!labels_batch || ggml_is_contiguous(labels_batch));
+    GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr));
+
+    const size_t nb_data_batch = ggml_nbytes(data_batch);
+    GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0);
+    const int64_t shards_per_batch = nb_data_batch / dataset->nbs_data;
+
+    if (labels_batch) {
+        const size_t nb_labels_batch = ggml_nbytes(labels_batch);
+        GGML_ASSERT(nb_labels_batch == shards_per_batch*dataset->nbs_labels);
+    }
+
+    GGML_ASSERT((ibatch + 1)*shards_per_batch <= int64_t(dataset->permutation.size()));
+
+    for (int64_t ishard_batch = 0; ishard_batch < shards_per_batch; ++ishard_batch) {
+        const int64_t ishard = dataset->permutation[ibatch*shards_per_batch + ishard_batch];
+
+        const char * ptr_data = (const char *) dataset->data->data + ishard*dataset->nbs_data;
+        ggml_backend_tensor_set(data_batch, ptr_data, ishard_batch*dataset->nbs_data, dataset->nbs_data);
+
+        if (!labels_batch) {
+            continue;
+        }
+
+        const char * ptr_labels = (const char *) dataset->labels->data + ishard*dataset->nbs_labels;
+        ggml_backend_tensor_set(labels_batch, ptr_labels, ishard_batch*dataset->nbs_labels, dataset->nbs_labels);
+    }
+}
+
+// ====== Model / Context ======
+
+struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata) {
+    GGML_UNUSED(userdata);
+
+    ggml_opt_optimizer_params result;
+
+    result.adamw.alpha = 0.001f;
+    result.adamw.beta1 = 0.9f;
+    result.adamw.beta2 = 0.999f;
+    result.adamw.eps   = 1e-8f;
+    result.adamw.wd    = 0.0f;
+
+    return result;
+}
+
+struct ggml_opt_params ggml_opt_default_params(
+        ggml_backend_sched_t backend_sched,
+        struct ggml_context * ctx_compute,
+        struct ggml_tensor * inputs,
+        struct ggml_tensor * outputs,
+        enum ggml_opt_loss_type loss_type) {
+    return {
+        /*backend_sched   =*/ backend_sched,
+        /*ctx_compute     =*/ ctx_compute,
+        /*inputs          =*/ inputs,
+        /*logits          =*/ outputs,
+        /*loss_type       =*/ loss_type,
+        /*build_type      =*/ GGML_OPT_BUILD_TYPE_OPT,
+        /*opt_period      =*/ 1,
+        /*get_opt_pars    =*/ ggml_opt_get_default_optimizer_params,
+        /*get_opt_pars_ud =*/ nullptr,
+    };
+}
+
+static ggml_tensor * map_tensor(std::map<ggml_tensor *, ggml_tensor *> & tensor_map, ggml_context * ctx, ggml_tensor * tensor) {
+    if (!tensor) {
+        return nullptr;
+    }
+
+    if (tensor_map.find(tensor) != tensor_map.end()) {
+        return tensor_map[tensor];
+    }
+
+    ggml_tensor * new_tensor = ggml_dup_tensor(ctx, tensor);
+    tensor_map[tensor] = new_tensor;
+
+    new_tensor->op = tensor->op;
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        new_tensor->nb[i] = tensor->nb[i];
+    }
+    new_tensor->flags = tensor->flags;
+    memcpy(new_tensor->op_params, tensor->op_params, sizeof(tensor->op_params));
+    strcpy(new_tensor->name, tensor->name);
+    new_tensor->data = tensor->data;
+    new_tensor->buffer = tensor->buffer;
+    new_tensor->extra = tensor->extra;
+    new_tensor->view_offs = tensor->view_offs;
+    new_tensor->view_src = map_tensor(tensor_map, ctx, tensor->view_src);
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        new_tensor->src[i] = map_tensor(tensor_map, ctx, tensor->src[i]);
+    }
+
+    return new_tensor;
+}
+
+static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * graph) {
+    std::map<ggml_tensor *, ggml_tensor *> tensor_map;
+
+    ggml_cgraph * new_graph = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true);
+
+    for (int i = 0; i < graph->n_leafs; i++) {
+        ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->leafs[i]));
+    }
+    for (int i = 0; i < graph->n_nodes; i++) {
+        ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->nodes[i]));
+    }
+    for (int i = 0; i < graph->n_nodes; ++i) {
+        const size_t igrad_src = ggml_hash_find(&graph->visited_hash_set, graph->nodes[i]);
+        const size_t igrad_dst = ggml_hash_find(&new_graph->visited_hash_set, new_graph->nodes[i]);
+        graph->grads[igrad_dst]     = new_graph->grads[igrad_src];
+        graph->grad_accs[igrad_dst] = new_graph->grad_accs[igrad_src];
+    }
+
+    return new_graph;
+}
+
+static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) {
+    GGML_ASSERT(graph);
+    if (opt_ctx->allocated_graph == graph) {
+        return;
+    }
+
+    ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph
+
+    {
+        ggml_init_params params = {
+            /*.mem_size   =*/ ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE,
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        ggml_free(opt_ctx->ctx_copy);
+        opt_ctx->ctx_copy = ggml_init(params);
+    }
+
+    opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph);
+
+    ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
+    opt_ctx->allocated_graph = graph;
+}
+
+ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
+    ggml_opt_context_t result = new struct ggml_opt_context;
+    result->backend_sched        = params.backend_sched;
+    result->allocated_graph      = nullptr;
+    result->allocated_graph_copy = nullptr;
+    result->ctx_compute          = params.ctx_compute;
+    result->ctx_copy             = nullptr;
+    result->inputs               = params.inputs;
+    result->outputs              = params.outputs;
+    result->iter                 = 1;
+    result->opt_period           = params.opt_period;
+    result->opt_i                = 0;
+    result->get_opt_pars         = params.get_opt_pars;
+    result->get_opt_pars_ud      = params.get_opt_pars_ud;
+
+    GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically");
+    GGML_ASSERT(result->opt_period >= 1);
+
+    const bool accumulate = params.build_type == GGML_OPT_BUILD_TYPE_GRAD ||
+        (params.build_type == GGML_OPT_BUILD_TYPE_OPT && result->opt_period > 1);
+
+    ggml_set_input(result->inputs);
+    ggml_set_output(result->outputs);
+
+    result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass.
+    ggml_build_forward_expand(result->gf, result->outputs);
+
+    int n_param = 0;
+    for (int i = 0; i < result->gf->n_nodes; ++i) {
+        if (result->gf->nodes[i]->flags & GGML_TENSOR_FLAG_PARAM) {
+            n_param++;
+        }
+    }
+
+    {
+        // The static context is used for:
+        //   - gradients (1 tensor per param if using gradient accumulation)
+        //   - optimizer momenta (2 tensors per param)
+        //   - labels
+        //   - loss + its gradient (up to 5 tensors)
+        //   - pred
+        //   - ncorrect (2 tensors).
+        const size_t tensors_per_param = (accumulate ? 1 : 0) + (params.build_type == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
+        const size_t size_meta = (tensors_per_param*n_param + 9) * ggml_tensor_overhead();
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ size_meta,
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        result->ctx_static = ggml_init(params);
+    }
+    {
+        // The static cpu context is used for:
+        //   - optimizer parameters (1 for the entire context)
+        const size_t size_meta = 1 * ggml_tensor_overhead();
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ size_meta,
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        result->ctx_static_cpu = ggml_init(params);
+    }
+
+
+    switch (params.loss_type) {
+        case GGML_OPT_LOSS_TYPE_MEAN: {
+            result->labels = nullptr;
+            result->loss = ggml_sum(result->ctx_static, result->outputs);
+            ggml_set_name(result->loss, "loss_sum");
+            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
+            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
+            ggml_set_name(result->loss, "loss_mean");
+            result->loss_per_datapoint = true;
+            break;
+        }
+        case GGML_OPT_LOSS_TYPE_SUM: {
+            result->labels = nullptr;
+            result->loss = ggml_sum(result->ctx_static, result->outputs);
+            ggml_set_name(result->loss, "loss_sum");
+            result->loss_per_datapoint = false;
+            break;
+        }
+        case GGML_OPT_LOSS_TYPE_CROSS_ENTROPY: {
+            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
+            ggml_set_input(result->labels);
+            ggml_set_name(result->labels, "labels");
+            result->loss = ggml_cross_entropy_loss(result->ctx_static, result->outputs, result->labels);
+            ggml_set_name(result->loss, "loss_cross_entropy");
+            if (result->opt_period > 1) {
+                result->loss = ggml_scale(result->ctx_static, result->loss, 1.0f / result->opt_period);
+                ggml_set_name(result->loss, "loss_cross_entropy_scaled");
+            }
+            result->loss_per_datapoint = true;
+            break;
+        }
+        case GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR: {
+            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
+            ggml_set_input(result->labels);
+            ggml_set_name(result->labels, "labels");
+            result->loss = ggml_sub(result->ctx_static, result->outputs, result->labels);
+            ggml_set_name(result->loss, "loss_error");
+            result->loss = ggml_sqr(result->ctx_static, result->loss);
+            ggml_set_name(result->loss, "loss_squared_error");
+            result->loss = ggml_sum(result->ctx_static, result->loss);
+            ggml_set_name(result->loss, "loss_sum_squared_error");
+            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
+            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
+            ggml_set_name(result->loss, "loss_mean_squared_error");
+            result->loss_per_datapoint = true;
+            break;
+        }
+    }
+    ggml_set_output(result->loss);
+    ggml_set_loss(result->loss);
+    ggml_build_forward_expand(result->gf, result->loss);
+
+    result->pred = ggml_argmax(result->ctx_static, result->outputs);
+    ggml_set_name(result->pred, "pred");
+    ggml_set_output(result->pred);
+    ggml_build_forward_expand(result->gf, result->pred);
+
+    if (result->labels) {
+        result->ncorrect = ggml_count_equal(result->ctx_static, result->pred, ggml_argmax(result->ctx_static, result->labels));
+        ggml_set_name(result->ncorrect, "ncorrect");
+        ggml_set_output(result->ncorrect);
+        ggml_build_forward_expand(result->gf, result->ncorrect);
+    } else {
+        result->ncorrect = nullptr;
+    }
+
+    if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
+        result->gb_grad = nullptr;
+        result->gb_opt  = nullptr;
+
+        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
+        result->buf_static_cpu = nullptr;
+
+        ggml_opt_alloc_graph(result, result->gf);
+
+        return result;
+    }
+
+    // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients.
+    result->gb_grad = ggml_graph_dup(result->ctx_compute, result->gf);
+    ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate);
+
+    if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) {
+        result->gb_opt  = nullptr;
+
+        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
+        result->buf_static_cpu = nullptr;
+
+        ggml_opt_alloc_graph(result, result->gb_grad);
+        ggml_graph_reset(result->gb_grad);
+
+        return result;
+    }
+
+    GGML_ASSERT(params.build_type == GGML_OPT_BUILD_TYPE_OPT);
+
+    // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
+    result->gb_opt = ggml_graph_dup(result->ctx_compute, result->gb_grad);
+
+    result->adamw_params = ggml_new_tensor_1d(result->ctx_static_cpu, GGML_TYPE_F32, 7);
+    ggml_set_input(result->adamw_params);
+    ggml_set_name(result->adamw_params, "adamw_params");
+
+    for (int i = result->gf->n_nodes-1; i >= 0; --i) {
+        struct ggml_tensor * node = result->gb_opt->nodes[i];
+        struct ggml_tensor * grad = ggml_graph_get_grad(result->gb_opt, node);
+
+        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
+            struct ggml_tensor * m        = ggml_dup_tensor(result->ctx_static, node);
+            struct ggml_tensor * v        = ggml_dup_tensor(result->ctx_static, node);
+            struct ggml_tensor * opt_step = ggml_opt_step_adamw(result->ctx_compute, node, grad, m, v, result->adamw_params);
+            ggml_build_forward_expand(result->gb_opt, opt_step);
+        }
+    }
+
+    result->buf_static = ggml_backend_alloc_ctx_tensors(
+        result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
+
+    result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type());
+
+    ggml_opt_alloc_graph(result, result->gb_opt);
+    ggml_graph_reset(result->gb_opt);
+
+    return result;
+}
+
+void ggml_opt_free(ggml_opt_context_t opt_ctx) {
+    if (opt_ctx == nullptr) {
+        return;
+    }
+    ggml_backend_buffer_free(opt_ctx->buf_static);
+    ggml_backend_buffer_free(opt_ctx->buf_static_cpu);
+    ggml_free(opt_ctx->ctx_static);
+    ggml_free(opt_ctx->ctx_static_cpu);
+    delete opt_ctx;
+}
+
+void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer) {
+    if (optimizer) {
+        ggml_graph_reset(opt_ctx->gb_opt);
+        opt_ctx->iter = 1;
+    } else {
+        ggml_graph_reset(opt_ctx->gb_grad);
+    }
+}
+
+struct ggml_tensor * ggml_opt_inputs(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->inputs;
+}
+
+struct ggml_tensor * ggml_opt_outputs(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->outputs;
+}
+
+struct ggml_tensor * ggml_opt_labels(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->labels;
+}
+
+struct ggml_tensor * ggml_opt_loss(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->loss;
+}
+
+struct ggml_tensor * ggml_opt_pred(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->pred;
+}
+
+struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->ncorrect;
+}
+
+struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node) {
+    return ggml_graph_get_grad_acc(opt_ctx->gb_opt, node);
+}
+
+// ====== Optimization Result ======
+
+ggml_opt_result_t ggml_opt_result_init() {
+    return new ggml_opt_result;
+}
+
+void ggml_opt_result_free(ggml_opt_result_t result) {
+    delete result;
+}
+
+void ggml_opt_result_reset(ggml_opt_result_t result) {
+    result->ndata = 0;
+    result->loss.clear();
+    result->pred.clear();
+    result->ncorrect = 0;
+}
+
+void ggml_opt_result_ndata(ggml_opt_result_t result, int64_t * ndata) {
+    *ndata = result->ndata;
+}
+
+void ggml_opt_result_loss(ggml_opt_result_t result, double * loss, double * unc) {
+    const int64_t nbatches = result->loss.size(); // Number of physical batches.
+
+    if (nbatches == 0) {
+        *loss = 0.0;
+        *unc  = NAN;
+        return;
+    }
+
+    double sum         = 0.0;
+    double sum_squared = 0.0;
+
+    for (const float & loss : result->loss) {
+        // If the loss is per datapoint it was scaled by 1.0f/opt_period for each physical batch.
+        const float loss_scaled = result->loss_per_datapoint ? loss*result->opt_period : loss;
+        sum         += loss_scaled;
+        sum_squared += loss_scaled*loss_scaled;
+    }
+
+    const double mean = sum/nbatches;
+    *loss = result->loss_per_datapoint ? mean : sum;
+
+    if (!unc) {
+        return;
+    }
+
+    if (nbatches < 2) {
+        *unc = NAN;
+        return;
+    }
+
+    const double var_sum = sum_squared/nbatches - mean*mean; // variance without Bessel's correction, i.e. nbatches/(nbatches-1)
+    *unc = result->loss_per_datapoint ? sqrt(var_sum / (nbatches - 1)) : sqrt(var_sum * nbatches/(nbatches - 1));
+}
+
+void ggml_opt_result_pred(ggml_opt_result_t result, int32_t * pred) {
+    for (size_t i = 0; i < result->pred.size(); ++i) {
+        pred[i] = result->pred[i];
+    }
+}
+
+void ggml_opt_result_accuracy(ggml_opt_result_t result, double * accuracy, double * unc) {
+    *accuracy = result->ncorrect >= 0 ? double(result->ncorrect) / double(result->ndata) : NAN;
+
+    if (!unc) {
+        return;
+    }
+
+    *unc = result->ncorrect >= 0 && result->ndata >= 2 ?
+        sqrt((*accuracy) * (1.0 - (*accuracy)) / double(result->ndata - 1)) : NAN;
+}
+
+// ====== Computation ======
+
+static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph, ggml_opt_result * result) {
+    if (graph != opt_ctx->gf) {
+        struct ggml_opt_optimizer_params opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
+
+        GGML_ASSERT(opt_pars.adamw.alpha >  0.0f);
+        GGML_ASSERT(opt_pars.adamw.beta1 >= 0.0f);
+        GGML_ASSERT(opt_pars.adamw.beta1 <= 1.0f);
+        GGML_ASSERT(opt_pars.adamw.beta2 >= 0.0f);
+        GGML_ASSERT(opt_pars.adamw.beta2 <= 1.0f);
+        GGML_ASSERT(opt_pars.adamw.eps   >= 0.0f);
+        GGML_ASSERT(opt_pars.adamw.wd    >= 0.0f);
+        GGML_ASSERT(opt_pars.adamw.wd    <= 1.0f);
+
+        // beta1, beta2 after applying warmup
+        const float beta1h = 1.0f/(1.0f - powf(opt_pars.adamw.beta1, opt_ctx->iter));
+        const float beta2h = 1.0f/(1.0f - powf(opt_pars.adamw.beta2, opt_ctx->iter));
+
+        float * adamw_par_data = ggml_get_data_f32(opt_ctx->adamw_params);
+        adamw_par_data[0] = opt_pars.adamw.alpha;
+        adamw_par_data[1] = opt_pars.adamw.beta1;
+        adamw_par_data[2] = opt_pars.adamw.beta2;
+        adamw_par_data[3] = opt_pars.adamw.eps;
+        adamw_par_data[4] = opt_pars.adamw.wd;
+        adamw_par_data[5] = beta1h;
+        adamw_par_data[6] = beta2h;
+    }
+
+    ggml_opt_alloc_graph(opt_ctx, graph);
+    ggml_backend_sched_graph_compute(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
+    opt_ctx->iter += opt_ctx->allocated_graph == opt_ctx->gb_opt;
+
+    if (!result) {
+        return;
+    }
+
+    if (result->ndata == 0) {
+        result->loss_per_datapoint = opt_ctx->loss_per_datapoint;
+        result->opt_period         = opt_ctx->opt_period;
+    } else {
+        GGML_ASSERT(result->loss_per_datapoint == opt_ctx->loss_per_datapoint);
+        GGML_ASSERT(result->opt_period         == opt_ctx->opt_period);
+    }
+
+    const int64_t ndata = opt_ctx->outputs->ne[1];
+    GGML_ASSERT(result->ndata == ndata*int64_t(result->loss.size()) && "varying batch size not supported");
+    result->ndata += ndata;
+
+    GGML_ASSERT(ggml_is_scalar(opt_ctx->loss));
+    GGML_ASSERT(opt_ctx->loss->type == GGML_TYPE_F32);
+    float loss;
+    ggml_backend_tensor_get(opt_ctx->loss, &loss, 0, ggml_nbytes(opt_ctx->loss));
+    result->loss.push_back(loss);
+
+    GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32);
+    std::vector<int32_t> pred(ndata);
+    ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred));
+    result->pred.insert(result->pred.end(), pred.begin(), pred.end());
+
+    if (!opt_ctx->labels || result->ncorrect < 0) {
+        result->ncorrect = -1;
+        return;
+    }
+
+    GGML_ASSERT(ggml_is_scalar(opt_ctx->ncorrect));
+    GGML_ASSERT(opt_ctx->ncorrect->type == GGML_TYPE_I64);
+    int64_t ncorrect;
+    ggml_backend_tensor_get(opt_ctx->ncorrect, &ncorrect, 0, ggml_nbytes(opt_ctx->ncorrect));
+    result->ncorrect += ncorrect;
+}
+
+void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
+    ggml_opt_eval_graph(opt_ctx, opt_ctx->gf, result);
+}
+
+void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
+    if (opt_ctx->opt_period == 1) {
+        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
+        return;
+    }
+
+    const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
+    if (opt_i_next == 0) {
+        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
+        ggml_opt_reset(opt_ctx, /*optimizer =*/ false);
+    } else {
+        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_grad, result);
+    }
+    opt_ctx->opt_i = opt_i_next;
+}
+
+// ====== High-Level Functions ======
+
+void ggml_opt_epoch(
+        ggml_opt_context_t      opt_ctx,
+        ggml_opt_dataset_t      dataset,
+        ggml_opt_result_t       result_train,
+        ggml_opt_result_t       result_eval,
+        int64_t                 idata_split,
+        ggml_opt_epoch_callback callback_train,
+        ggml_opt_epoch_callback callback_eval) {
+    struct ggml_tensor * inputs = ggml_opt_inputs(opt_ctx);
+    struct ggml_tensor * labels = ggml_opt_labels(opt_ctx);
+    struct ggml_tensor * data   = ggml_opt_dataset_data(dataset);
+    GGML_ASSERT(data->ne[0] == inputs->ne[0]);
+
+    const int64_t ndata       =   data->ne[1];
+    const int64_t ndata_batch = inputs->ne[1];
+
+    GGML_ASSERT(data->ne[1] % inputs->ne[1] == 0);
+    const int64_t nbatches = ndata/ndata_batch;
+
+    idata_split = idata_split < 0 ? ndata : idata_split;
+    GGML_ASSERT(idata_split % ndata_batch == 0);
+    const int64_t ibatch_split = idata_split / ndata_batch;
+
+    int64_t ibatch = 0;
+    int64_t t_loop_start = ggml_time_us();
+    for (; ibatch < ibatch_split; ++ibatch) {
+        ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
+        ggml_opt_forward_backward(opt_ctx, result_train);
+        if (callback_train) {
+            callback_train(true, opt_ctx, dataset, result_train, ibatch+1, ibatch_split, t_loop_start);
+        }
+    }
+    t_loop_start = ggml_time_us();
+    for (; ibatch < nbatches; ++ibatch) {
+        ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
+        ggml_opt_forward(opt_ctx, result_eval);
+        if (callback_eval) {
+            callback_eval(false, opt_ctx, dataset, result_eval, ibatch+1-ibatch_split, nbatches-ibatch_split, t_loop_start);
+        }
+    }
+}
+
+void ggml_opt_epoch_callback_progress_bar(
+        bool               train,
+        ggml_opt_context_t opt_ctx,
+        ggml_opt_dataset_t dataset,
+        ggml_opt_result_t  result,
+        int64_t            ibatch,
+        int64_t            ibatch_max,
+        int64_t            t_start_us) {
+    fprintf(stderr, "%s[", train ? "train: " : "val:   ");
+
+    constexpr int64_t bar_length = 25;
+    for (int64_t j = 0; j < bar_length; ++j) {
+        const int64_t ibatch_j = ibatch_max * j/bar_length;
+        if (ibatch_j < ibatch) {
+            fprintf(stderr, "=");
+        } else if (ibatch_max * (j - 1)/bar_length < ibatch) {
+            fprintf(stderr, ">");
+        } else {
+            fprintf(stderr, " ");
+        }
+    }
+
+    const int64_t batch_size = ggml_opt_inputs(opt_ctx)->ne[1];
+    const int64_t idata      = ibatch*batch_size;
+    const int64_t idata_max  = ibatch_max*batch_size;
+
+    double loss;
+    double loss_unc;
+    ggml_opt_result_loss(result, &loss, &loss_unc);
+
+    double accuracy;
+    double accuracy_unc;
+    ggml_opt_result_accuracy(result, &accuracy, &accuracy_unc);
+
+    const int64_t t_ibatch_us = ggml_time_us() - t_start_us;
+    int64_t t_ibatch_s = t_ibatch_us / 1000000;
+    const int64_t t_ibatch_h = t_ibatch_s / 3600;
+    t_ibatch_s -= t_ibatch_h * 3600;
+    const int64_t t_ibatch_m = t_ibatch_s / 60;
+    t_ibatch_s -= t_ibatch_m * 60;
+
+    const int64_t t_eta_us = t_ibatch_us * (ibatch_max - ibatch)/ibatch;
+    int64_t t_eta_s = t_eta_us / 1000000;
+    const int64_t t_eta_h = t_eta_s / 3600;
+    t_eta_s -= t_eta_h * 3600;
+    const int64_t t_eta_m = t_eta_s / 60;
+    t_eta_s -= t_eta_m * 60;
+
+    fprintf(stderr, "| data=%06" PRId64 "/%06" PRId64 ", loss=%.6lf+-%.6lf, accuracy=%.2lf+-%.2lf%%, "
+            "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 ", ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 "]\r",
+            idata, idata_max, loss, loss_unc, 100.0*accuracy, 100.0*accuracy_unc,
+            t_ibatch_h, t_ibatch_m, t_ibatch_s, t_eta_h, t_eta_m, t_eta_s);
+    if (ibatch == ibatch_max) {
+        fprintf(stderr, "\n");
+    }
+    fflush(stderr);
+
+    GGML_UNUSED(dataset);
+}
+
+void ggml_opt_fit(
+        ggml_backend_sched_t            backend_sched,
+        ggml_context                  * ctx_compute,
+        ggml_tensor                   * inputs,
+        ggml_tensor                   * outputs,
+        ggml_opt_dataset_t              dataset,
+        enum ggml_opt_loss_type         loss_type,
+        ggml_opt_get_optimizer_params   get_opt_pars,
+        int64_t                         nepoch,
+        int64_t                         nbatch_logical,
+        float                           val_split,
+        bool                            silent) {
+    ggml_time_init();
+    const int64_t t_start_us = ggml_time_us();
+
+    const int64_t ndata           = ggml_opt_dataset_data(dataset)->ne[1];
+    const int64_t nbatch_physical = inputs->ne[1];
+    GGML_ASSERT(ndata          % nbatch_logical  == 0);
+    GGML_ASSERT(nbatch_logical % nbatch_physical == 0);
+
+    const int64_t opt_period       = nbatch_logical / nbatch_physical;
+    const int64_t nbatches_logical = ndata / nbatch_logical;
+
+    GGML_ASSERT(val_split >= 0.0f);
+    GGML_ASSERT(val_split <  1.0f);
+    const int64_t ibatch_split = int64_t(((1.0f - val_split) * nbatches_logical)) * opt_period; // train <-> val split index (physical)
+    const int64_t idata_split  = ibatch_split * nbatch_physical;
+
+    int64_t epoch = 1;
+
+    ggml_opt_params params = ggml_opt_default_params(backend_sched, ctx_compute, inputs, outputs, loss_type);
+    params.opt_period      = opt_period;
+    params.get_opt_pars    = get_opt_pars;
+    params.get_opt_pars_ud = &epoch;
+    ggml_opt_context_t opt_ctx = ggml_opt_init(params);
+
+    // Shuffling the data is generally useful but there is only a point if not all data is used in a single batch.
+    if (nbatch_logical < ndata) {
+        ggml_opt_dataset_shuffle(opt_ctx, dataset, -1); // Shuffle all data (train + validation).
+    }
+
+    ggml_opt_result_t result_train = ggml_opt_result_init();
+    ggml_opt_result_t result_val   = ggml_opt_result_init();
+
+    ggml_opt_epoch_callback epoch_callback = silent ? nullptr : ggml_opt_epoch_callback_progress_bar;
+
+    for (; epoch <= nepoch; ++epoch) {
+        if (nbatch_logical < idata_split) {
+            ggml_opt_dataset_shuffle(opt_ctx, dataset, idata_split);
+        }
+
+        ggml_opt_result_reset(result_train);
+        ggml_opt_result_reset(result_val);
+
+        if (!silent) {
+            fprintf(stderr, "%s: epoch %04" PRId64 "/%04" PRId64 ":\n", __func__, epoch, nepoch);
+        }
+        ggml_opt_epoch(opt_ctx, dataset, result_train, result_val, idata_split, epoch_callback, epoch_callback);
+        if (!silent) {
+            fprintf(stderr, "\n");
+        }
+    }
+
+    if (!silent) {
+        int64_t t_total_s = (ggml_time_us() - t_start_us) / 1000000;
+        const int64_t t_total_h = t_total_s / 3600;
+        t_total_s -= t_total_h * 3600;
+        const int64_t t_total_m = t_total_s / 60;
+        t_total_s -= t_total_m * 60;
+        fprintf(stderr, "%s: training took %02" PRId64 ":%02" PRId64 ":%02" PRId64 "\n", __func__, t_total_h, t_total_m, t_total_s);
+    }
+
+    ggml_opt_free(opt_ctx);
+    ggml_opt_result_free(result_train);
+    ggml_opt_result_free(result_val);
+}
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 5cdf59f2560a1..4a478fcaa2064 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1592,14 +1592,13 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         /*.op           =*/ GGML_OP_NONE,
         /*.op_params    =*/ { 0 },
         /*.flags        =*/ 0,
-        /*.grad         =*/ NULL,
         /*.src          =*/ { NULL },
         /*.view_src     =*/ view_src,
         /*.view_offs    =*/ view_offs,
         /*.data         =*/ obj_alloc_size > 0 ? (void *)(result + 1) : data,
         /*.name         =*/ { 0 },
         /*.extra        =*/ NULL,
-        ///*.padding      =*/ { 0 },
+        /*.padding      =*/ { 0 },
     };
 
 #ifdef __clang__
@@ -4194,8 +4193,6 @@ struct ggml_tensor * ggml_flash_attn_ext(
         GGML_ASSERT(mask);
     }
 
-    bool is_node = false;
-
     // permute(0, 2, 1, 3)
     int64_t ne[4] = { q->ne[0], q->ne[2], q->ne[1], q->ne[3] };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
@@ -4203,8 +4200,7 @@ struct ggml_tensor * ggml_flash_attn_ext(
     float params[] = { scale, max_bias, logit_softcap };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_FLASH_ATTN_EXT;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_FLASH_ATTN_EXT;
     result->src[0] = q;
     result->src[1] = k;
     result->src[2] = v;
@@ -4272,14 +4268,6 @@ struct ggml_tensor * ggml_flash_attn_back(
 
     GGML_ASSERT(ne2 % kvne2 == 0);
 
-    bool is_node = false;
-
-    if (q->grad || k->grad || v->grad) {
-        // when using this operation (in backwards pass) these grads are set.
-        // we don't want to create (big) grad of our result, so is_node is false.
-        is_node = false;
-    }
-
     // store gradients of q, k and v as continuous tensors concatenated in result.
     // note: v and gradv are actually transposed, i.e. v->ne[0] != D.
     const int64_t elem_q = ggml_nelements(q);
@@ -4302,8 +4290,7 @@ struct ggml_tensor * ggml_flash_attn_back(
     int32_t masked_i = masked ? 1 : 0;
     ggml_set_op_params(result, &masked_i, sizeof(masked_i));
 
-    result->op   = GGML_OP_FLASH_ATTN_BACK;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_FLASH_ATTN_BACK;
     result->src[0] = q;
     result->src[1] = k;
     result->src[2] = v;
@@ -4945,34 +4932,24 @@ struct ggml_tensor * ggml_opt_step_adamw(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * grad,
-        float                 alpha,
-        float                 beta1,
-        float                 beta2,
-        float                 eps,
-        float                 wd) {
+        struct ggml_tensor  * m,
+        struct ggml_tensor  * v,
+        struct ggml_tensor  * adamw_params) {
     GGML_ASSERT(a->flags & GGML_TENSOR_FLAG_PARAM);
     GGML_ASSERT(ggml_are_same_shape(a, grad));
-    GGML_ASSERT(alpha >  0.0f);
-    GGML_ASSERT(beta1 >= 0.0f && beta1 <= 1.0f);
-    GGML_ASSERT(beta2 >= 0.0f && beta2 <= 1.0f);
-    GGML_ASSERT(eps   >= 0.0f);
-    GGML_ASSERT(wd    >= 0.0f && wd    <= 1.0f);
+    GGML_ASSERT(ggml_are_same_shape(a, m));
+    GGML_ASSERT(ggml_are_same_shape(a, v));
+    GGML_ASSERT(adamw_params->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_nelements(adamw_params) == 7);
 
     struct ggml_tensor * result = ggml_view_tensor(ctx, a);
 
-    const int64_t iter = 1;
-    memcpy(&result->op_params[0], &iter, sizeof(int64_t));
-    ggml_set_op_params_f32(result, 2, alpha);
-    ggml_set_op_params_f32(result, 3, beta1);
-    ggml_set_op_params_f32(result, 4, beta2);
-    ggml_set_op_params_f32(result, 5, eps);
-    ggml_set_op_params_f32(result, 6, wd);
-
     result->op     = GGML_OP_OPT_STEP_ADAMW;
     result->src[0] = a;
     result->src[1] = grad;
-    result->src[2] = ggml_dup_tensor(ctx, grad);
-    result->src[3] = ggml_dup_tensor(ctx, grad);
+    result->src[2] = m;
+    result->src[3] = v;
+    result->src[4] = adamw_params;
 
     return result;
 }
@@ -5041,1112 +5018,514 @@ static void ggml_hash_map_free(struct hash_map * map) {
     GGML_FREE(map);
 }
 
-// gradient checkpointing
-
-static struct ggml_tensor * ggml_recompute_graph_node(
-        struct ggml_context * ctx,
-        struct ggml_cgraph  * graph,
-        struct hash_map     * replacements,
-        struct ggml_tensor  * node) {
-
-    if (node == NULL) {
-        return NULL;
-    }
-
-    if (node->flags & GGML_TENSOR_FLAG_PARAM) {
-        return node;
-    }
-
-    if (!ggml_hash_contains(&graph->visited_hash_set, node)) {
-        return node;
-    }
-
-    int count_children = 0;
-    for (int k = 0; k < GGML_MAX_SRC; ++k) {
-        if (node->src[k]) {
-            ++count_children;
-        }
-    }
-
-    if (count_children == 0) {
-        return node;
-    }
-
-    size_t i = ggml_hash_find(&replacements->set, node);
-    GGML_ASSERT(i != GGML_HASHSET_FULL); // assert that not full
-    if (replacements->set.keys[i] == node) {
-        return replacements->vals[i];
-    }
-
-    struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, GGML_MAX_DIMS, node->ne);
-
-    // insert clone into replacements
-    GGML_ASSERT(replacements->set.keys[i] == NULL); // assert that we don't overwrite
-    replacements->set.keys[i] = node;
-    replacements->vals[i] = clone;
-
-    clone->op       = node->op;
-    clone->grad     = node->grad;
-    clone->flags    = node->flags;
-    clone->extra    = node->extra;
-    for (int k = 0; k < GGML_MAX_DIMS; ++k) {
-        clone->nb[k] = node->nb[k];
-    }
-    for (int k = 0; k < GGML_MAX_SRC; ++k) {
-        clone->src[k] = ggml_recompute_graph_node(ctx, graph, replacements, node->src[k]);
-    }
-    if (node->view_src != NULL) {
-        clone->data = (node->view_src->data == NULL)
-                        ? NULL // view_src not yet allocated
-                        : (char *) node->view_src->data // view_src already allocated
-                                 + node->view_offs;
-        clone->view_src  = node->view_src;
-        clone->view_offs = node->view_offs;
-    }
-
-    GGML_ASSERT(sizeof(node->op_params) == sizeof(int32_t) * (GGML_MAX_OP_PARAMS / sizeof(int32_t)));
-    GGML_ASSERT(sizeof(node->name)      == GGML_MAX_NAME);
-    memcpy(clone->op_params, node->op_params, sizeof(node->op_params));
-    ggml_format_name(clone, "%s (clone)", ggml_get_name(node));
-
-    return clone;
-}
-
-void ggml_build_backward_gradient_checkpointing(
-        struct ggml_context   * ctx,
-        struct ggml_cgraph    * gf,
-        struct ggml_cgraph    * gb,
-        struct ggml_cgraph    * gb_tmp,
-        struct ggml_tensor  * * checkpoints,
-        int                     n_checkpoints) {
-    ggml_graph_cpy(gf, gb_tmp);
-    ggml_build_backward_expand(ctx, gf, gb_tmp, false);
-
-    if (n_checkpoints <= 0) {
-        ggml_graph_cpy(gb_tmp, gb);
-        return;
-    }
-
-    struct hash_map * replacements = ggml_new_hash_map(gf->n_nodes + gf->n_leafs + n_checkpoints);
-
-    // insert checkpoints in replacements
-    for (int i = 0; i < n_checkpoints; ++i) {
-        size_t k = ggml_hash_find(&replacements->set, checkpoints[i]);
-        GGML_ASSERT(k != GGML_HASHSET_FULL); // assert that not full
-        GGML_ASSERT(replacements->set.keys[k] == NULL); // assert that we don't overwrite
-        replacements->set.keys[k] = checkpoints[i];
-        replacements->vals[k]     = checkpoints[i];
-    }
-
-    ggml_graph_cpy(gf, gb);
-    // rewrite gb_tmp->nodes[gf->n_nodes:gb_tmp->n_nodes],
-    // replacing references to gb_tmp->nodes[0:gf->n_nodes] ( == gf->nodes[0:gf->n_nodes]),
-    // by recomputing them from checkpoints
-    for (int i = gf->n_nodes; i<gb_tmp->n_nodes; ++i) {
-        struct ggml_tensor * node = gb_tmp->nodes[i];
-        for (int k = 0; k < GGML_MAX_SRC; ++k) {
-            // insert new tensors recomputing src, reusing already made replacements,
-            // remember replacements: remember new tensors with mapping from corresponding gf nodes
-            // recurse for input tensors,
-            // unless (i.e. terminating when) input tensors are replacements (like checkpoints)
-            node->src[k] = ggml_recompute_graph_node(ctx, gf, replacements, node->src[k]);
-        }
-        // insert rewritten backward node with replacements made into resulting backward graph gb
-        ggml_build_forward_expand(gb, node);
-    }
-
-    ggml_hash_map_free(replacements);
-}
-
 // utility functions to change gradients
 // if a is in acc_table, modify gradients in-place and mark result as gradient accumulator
 // else if a is in zero_table, replace a
 // else, just add/subtract/etc. the gradients
 
-static struct ggml_tensor * ggml_add_or_set(
-        struct ggml_context  * ctx,
-        struct ggml_tensor   * a,
-        struct ggml_tensor   * b,
-        struct ggml_hash_set * zero_table,
-        struct ggml_hash_set * acc_table) {
-    if (ggml_hash_contains(acc_table, a)) {
-        struct ggml_tensor * ret = ggml_add_impl(ctx, a, b, true);
-        const size_t insert_result = ggml_hash_insert(acc_table, ret);
-        GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
-        GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
-        return ret;
-    }
-    if (ggml_hash_contains(zero_table, a)) {
-        return b;
+static void ggml_add_or_set(
+        struct ggml_context * ctx,
+        struct ggml_cgraph  * cgraph,
+        size_t                isrc,
+        struct ggml_tensor  * tensor) {
+    if (cgraph->grads[isrc]) {
+        cgraph->grads[isrc] = ggml_add_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]);
+    } else {
+        cgraph->grads[isrc] = tensor;
     }
-    return ggml_add_impl(ctx, a, b, false);
+    ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
 }
 
-static struct ggml_tensor * ggml_acc_or_set(
-        struct ggml_context  * ctx,
-        struct ggml_tensor   * a,
-        struct ggml_tensor   * b,
-        const  size_t          nb1,
-        const  size_t          nb2,
-        const  size_t          nb3,
-        const  size_t          offset,
-        struct ggml_hash_set * zero_table,
-        struct ggml_hash_set * acc_table) {
-    if (ggml_hash_contains(acc_table, a)) {
-        struct ggml_tensor * ret = ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, true);
-        const size_t insert_result = ggml_hash_insert(acc_table, ret);
-        GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
-        GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
-        return ret;
-    }
-    if (ggml_hash_contains(zero_table, a)) {
-        struct ggml_tensor * a_zero = ggml_scale(ctx, a, 0.0f); // FIXME this is going to produce NaN if a contains inf/NaN
-        return ggml_acc_impl(ctx, a_zero, b, nb1, nb2, nb3, offset, false);
+static void ggml_acc_or_set(
+        struct ggml_context * ctx,
+        struct ggml_cgraph  * cgraph,
+        size_t                isrc,
+        struct ggml_tensor  * src,
+        struct ggml_tensor  * tensor,
+        const  size_t         nb1,
+        const  size_t         nb2,
+        const  size_t         nb3,
+        const  size_t         offset) {
+    if (cgraph->grads[isrc]) {
+        cgraph->grads[isrc] = ggml_acc_impl(ctx, cgraph->grads[isrc], tensor, nb1, nb2, nb3, offset, cgraph->grad_accs[isrc]);
+    } else {
+        struct ggml_tensor * a_zero = ggml_scale(ctx, src, 0.0f); // FIXME this is going to produce NaN if a contains inf/NaN
+        cgraph->grads[isrc] = ggml_acc_impl(ctx, a_zero, tensor, nb1, nb2, nb3, offset, false);
     }
-    return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, false);
+    ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
 }
 
-static struct ggml_tensor * ggml_add1_or_set(
-        struct ggml_context  * ctx,
-        struct ggml_tensor   * a,
-        struct ggml_tensor   * b,
-        struct ggml_hash_set * zero_table,
-        struct ggml_hash_set * acc_table) {
-    if (ggml_hash_contains(acc_table, a)) {
-        struct ggml_tensor * ret = ggml_add1_impl(ctx, a, b, true);
-        const size_t insert_result = ggml_hash_insert(acc_table, ret);
-        GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
-        GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
-        return ret;
-    }
-    if (ggml_hash_contains(zero_table, a)) {
-        return ggml_repeat(ctx, b, a);
+static void ggml_add1_or_set(
+        struct ggml_context * ctx,
+        struct ggml_cgraph  * cgraph,
+        size_t                isrc,
+        struct ggml_tensor  * src,
+        struct ggml_tensor  * tensor) {
+    if (cgraph->grads[isrc]) {
+        cgraph->grads[isrc] = ggml_add1_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]);
+    } else {
+        cgraph->grads[isrc] = ggml_repeat(ctx, tensor, src);
     }
-    return ggml_add1_impl(ctx, a, b, false);
+    ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
 }
 
-static struct ggml_tensor * ggml_sub_or_set(
-        struct ggml_context  * ctx,
-        struct ggml_tensor   * a,
-        struct ggml_tensor   * b,
-        struct ggml_hash_set * zero_table,
-        struct ggml_hash_set * acc_table) {
-    if (ggml_hash_contains(acc_table, a)) {
-        struct ggml_tensor * ret = ggml_sub_impl(ctx, a, b, true);
-        const size_t insert_result = ggml_hash_insert(acc_table, ret);
-        GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
-        GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
-        return ret;
-    }
-    if (ggml_hash_contains(zero_table, a)) {
-        return ggml_neg(ctx, b);
+static void ggml_sub_or_set(
+        struct ggml_context * ctx,
+        struct ggml_cgraph  * cgraph,
+        size_t                isrc,
+        struct ggml_tensor  * tensor) {
+    if (cgraph->grads[isrc]) {
+        cgraph->grads[isrc] = ggml_sub_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]);
+    } else {
+        cgraph->grads[isrc] = ggml_neg(ctx, tensor);
     }
-    return ggml_sub_impl(ctx, a, b, false);
+    ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
 }
 
-static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, struct ggml_hash_set * zero_table, struct ggml_hash_set * acc_table) {
+static void ggml_compute_backward(
+        struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i, bool * grads_needed) {
+    struct ggml_tensor * tensor = cgraph->nodes[i];
+    struct ggml_tensor * grad   = ggml_graph_get_grad(cgraph, tensor);
+
+    if (!grad) {
+        return;
+    }
+
     struct ggml_tensor * src0 = tensor->src[0];
     struct ggml_tensor * src1 = tensor->src[1];
     struct ggml_tensor * src2 = tensor->src[2];
+    struct ggml_hash_set * hash_set = &cgraph->visited_hash_set;
+    const size_t isrc0 = ggml_hash_find(hash_set, src0);
+    const size_t isrc1 = ggml_hash_find(hash_set, src1);
+    const size_t isrc2 = ggml_hash_find(hash_set, src2);
+    const bool src0_needs_grads = isrc0 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc0) && grads_needed[isrc0];
+    const bool src1_needs_grads = isrc1 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc1) && grads_needed[isrc1];
+    const bool src2_needs_grads = isrc2 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc2) && grads_needed[isrc2];
 
     switch (tensor->op) {
-        case GGML_OP_DUP:
-            {
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_ADD:
-            {
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    if (ggml_are_same_shape(src0, src1)) {
-                        src1->grad = ggml_add_or_set(ctx, src1->grad,                       tensor->grad,        zero_table, acc_table);
-                    } else {
-                        src1->grad = ggml_add_or_set(ctx, src1->grad, ggml_repeat_back(ctx, tensor->grad, src1), zero_table, acc_table);
-                    }
-                }
-            } break;
-        case GGML_OP_ADD1:
-            {
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    src1->grad = ggml_add_or_set(ctx,
-                        src1->grad,
-                        ggml_mean(ctx, tensor->grad), // TODO: should probably be sum instead of mean
-                        zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_ACC:
-            {
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    const size_t nb1     = ((int32_t *) tensor->op_params)[0];
-                    const size_t nb2     = ((int32_t *) tensor->op_params)[1];
-                    const size_t nb3     = ((int32_t *) tensor->op_params)[2];
-                    const size_t offset  = ((int32_t *) tensor->op_params)[3];
-
-                    struct ggml_tensor * tensor_grad_view = ggml_view_4d(ctx,
-                        tensor->grad,
-                        src1->grad->ne[0],
-                        src1->grad->ne[1],
-                        src1->grad->ne[2],
-                        src1->grad->ne[3],
-                        nb1, nb2, nb3, offset);
-
-                    src1->grad =
-                        ggml_add_or_set(ctx,
-                            src1->grad,
-                            ggml_reshape(ctx,
-                                ggml_cont(ctx, tensor_grad_view),
-                                src1->grad),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SUB:
-            {
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    src1->grad = ggml_sub_or_set(ctx, src1->grad, tensor->grad, zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_MUL:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_mul(ctx, src1, tensor->grad),
-                                zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    src1->grad =
-                        ggml_add_or_set(ctx,
-                                src1->grad,
-                                ggml_mul(ctx, src0, tensor->grad),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_DIV:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_div(ctx, tensor->grad, src1),
-                                zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    src1->grad =
-                        ggml_sub_or_set(ctx,
-                                src1->grad,
-                                ggml_mul(ctx,
-                                    tensor->grad,
-                                    ggml_div(ctx, tensor, src1)),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SQR:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_scale(ctx,
-                                    ggml_mul(ctx, src0, tensor->grad),
-                                    2.0f),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SQRT:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_scale(ctx,
-                                    ggml_div(ctx,
-                                        tensor->grad,
-                                        tensor),
-                                    0.5f),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_LOG:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_div(ctx,
-                                    tensor->grad,
-                                    src0),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SIN:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_mul(ctx,
-                                    tensor->grad,
-                                    ggml_cos(ctx, src0)),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_COS:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_sub_or_set(ctx,
-                                src0->grad,
-                                ggml_mul(ctx,
-                                    tensor->grad,
-                                    ggml_sin(ctx, src0)),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SUM:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add1_or_set(ctx,
-                                src0->grad,
-                                tensor->grad,
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SUM_ROWS:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_repeat(ctx,
-                                    tensor->grad,
-                                    src0->grad),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_MEAN:
-        case GGML_OP_ARGMAX:
-        case GGML_OP_COUNT_EQUAL:
-            {
-                GGML_ABORT("fatal error"); // TODO: implement
-            }
-        case GGML_OP_REPEAT:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_repeat_back(ctx, tensor->grad, src0->grad),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_REPEAT_BACK:
-            {
-                if (src0->grad) {
-                    // TODO: test this
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_repeat(ctx, tensor->grad, src0->grad),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_CONCAT:
-            {
-                GGML_ABORT("fatal error"); // TODO: implement
-            }
-        case GGML_OP_SILU_BACK:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        case GGML_OP_DUP: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
             }
-        case GGML_OP_NORM:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_ADD: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
             }
-        case GGML_OP_RMS_NORM:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    float eps;
-                    memcpy(&eps, tensor->op_params, sizeof(float));
-
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_rms_norm_back(ctx, src0, tensor->grad, eps),
-                            zero_table, acc_table);
+            if (src1_needs_grads) {
+                struct ggml_tensor * tmp = grad;
+                if (!ggml_are_same_shape(src0, src1)) {
+                    tmp = ggml_repeat_back(ctx, tmp, src1);
                 }
-            } break;
-        case GGML_OP_RMS_NORM_BACK:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+                ggml_add_or_set(ctx, cgraph, isrc1, tmp);
             }
-        case GGML_OP_GROUP_NORM:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_ADD1: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
             }
-        case GGML_OP_MUL_MAT:
-            {
-                // https://cs231n.github.io/optimization-2/#staged
-                // # forward pass
-                // s0 = np.random.randn(5, 10)
-                // s1 = np.random.randn(10, 3)
-                // t = s0.dot(s1)
-
-                // # now suppose we had the gradient on t from above in the circuit
-                // dt = np.random.randn(*t.shape) # same shape as t
-                // ds0 = dt.dot(s1.T) #.T gives the transpose of the matrix
-                // ds1 = t.T.dot(dt)
-
-                // tensor.shape [m,p,qq,rr]
-                // src0.shape   [n,m,q1,r1]
-                // src1.shape   [n,p,qq,rr]
-
-                // necessary for llama
-                if (src0->grad) {
-                    struct ggml_tensor * s1_tg =
-                        ggml_out_prod(ctx, // [n,m,qq,rr]
-                            src1,          // [n,p,qq,rr]
-                            tensor->grad); // [m,p,qq,rr]
-                    const int64_t qq = s1_tg->ne[2];
-                    const int64_t rr = s1_tg->ne[3];
-                    const int64_t q1 = src0->ne[2];
-                    const int64_t r1 = src0->ne[3];
-                    const bool ne2_broadcasted = qq > q1;
-                    const bool ne3_broadcasted = rr > r1;
-                    if (ne2_broadcasted || ne3_broadcasted) {
-                        // sum broadcast repetitions of s1_tg into shape of src0
-                        s1_tg = ggml_repeat_back(ctx, s1_tg, src0);
-                    }
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad, // [n,m,q1,r1]
-                                s1_tg,      // [n,m,q1,r1]
-                                zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    src1->grad =
-                        ggml_add_or_set(ctx,
-                                src1->grad,                            // [n,p,qq,rr]
-                                // ggml_mul_mat(ctx,                   // [n,p,qq,rr]
-                                //     ggml_cont(ctx,                  // [m,n,q1,r1]
-                                //         ggml_transpose(ctx, src0)), // [m,n,q1,r1]
-                                //     tensor->grad),                  // [m,p,qq,rr]
-
-                                // // when src0 is bigger than tensor->grad (this is mostly the case in llama),
-                                // // avoid transpose of src0, rather transpose smaller tensor->grad
-                                // // and then use ggml_out_prod
-                                ggml_out_prod(ctx,                  // [n,p,qq,rr]
-                                    src0,                           // [n,m,q1,r1]
-                                    ggml_transpose(ctx,             // [p,m,qq,rr]
-                                        tensor->grad)),             // [m,p,qq,rr]
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_MUL_MAT_ID:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+            if (src1_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc1, ggml_mean(ctx, grad)); // TODO: should probably be sum instead of mean
             }
-        case GGML_OP_OUT_PROD:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_ACC: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
             }
-        case GGML_OP_SCALE:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    float s;
-                    memcpy(&s, tensor->op_params, sizeof(float));
-
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_scale_impl(ctx, tensor->grad, s, false),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SET:
-            {
-                const size_t nb1     = ((int32_t *) tensor->op_params)[0];
-                const size_t nb2     = ((int32_t *) tensor->op_params)[1];
-                const size_t nb3     = ((int32_t *) tensor->op_params)[2];
-                const size_t offset  = ((int32_t *) tensor->op_params)[3];
-
-                struct ggml_tensor * tensor_grad_view = NULL;
-
-                if (src0->grad || src1->grad) {
-                    GGML_ASSERT(src0->type == tensor->type);
-                    GGML_ASSERT(tensor->grad->type == tensor->type);
-                    GGML_ASSERT(!src1->grad || src1->grad->type == tensor->grad->type);
-
-                    tensor_grad_view = ggml_view_4d(ctx,
-                        tensor->grad, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
-                        nb1, nb2, nb3, offset);
-                }
+            if (src1_needs_grads) {
+                const size_t nb1    = ((int32_t *) tensor->op_params)[0];
+                const size_t nb2    = ((int32_t *) tensor->op_params)[1];
+                const size_t nb3    = ((int32_t *) tensor->op_params)[2];
+                const size_t offset = ((int32_t *) tensor->op_params)[3];
 
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx,
-                        src0->grad,
-                        ggml_acc_impl(ctx,
-                            tensor->grad,
-                            ggml_neg(ctx, tensor_grad_view),
-                            nb1, nb2, nb3, offset, false),
-                        zero_table, acc_table);
-                }
+                struct ggml_tensor * tensor_grad_view = ggml_view_4d(ctx,
+                    grad, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
+                    nb1, nb2, nb3, offset);
 
-                if (src1->grad) {
-                    src1->grad =
-                        ggml_add_or_set(ctx,
-                            src1->grad,
-                            ggml_reshape(ctx,
-                                ggml_cont(ctx, tensor_grad_view),
-                                src1->grad),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_CPY:
-            {
-                // necessary for llama
-                // cpy overwrites value of src1 by src0 and returns view(src1)
-                // the overwriting is mathematically equivalent to:
-                // tensor = src0 * 1 + src1 * 0
-                if (src0->grad) {
-                    // dsrc0 = dtensor * 1
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    // dsrc1 = dtensor * 0 -> noop
-                }
-            } break;
-        case GGML_OP_CONT:
-            {
-                // same as cpy
-                if (src0->grad) {
-                    GGML_ASSERT(ggml_is_contiguous(src0->grad));
-                    GGML_ASSERT(ggml_is_contiguous(tensor->grad));
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_RESHAPE:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            ggml_reshape(ctx,
-                                ggml_is_contiguous(tensor->grad)
-                                    ? tensor->grad
-                                    : ggml_cont(ctx, tensor->grad),
-                                src0->grad),
-                        zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_VIEW:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    size_t offset;
-
-                    memcpy(&offset, tensor->op_params, sizeof(offset));
-
-                    size_t nb1 = tensor->nb[1];
-                    size_t nb2 = tensor->nb[2];
-                    size_t nb3 = tensor->nb[3];
-
-                    if (src0->type != src0->grad->type) {
-                        // gradient is typically F32, but src0 could be other type
-                        size_t ng = ggml_element_size(src0->grad);
-                        size_t n0 = ggml_element_size(src0);
-                        GGML_ASSERT(offset % n0 == 0);
-                        GGML_ASSERT(nb1 % n0 == 0);
-                        GGML_ASSERT(nb2 % n0 == 0);
-                        GGML_ASSERT(nb3 % n0 == 0);
-                        offset = (offset / n0) * ng;
-                        nb1 = (nb1 / n0) * ng;
-                        nb2 = (nb2 / n0) * ng;
-                        nb3 = (nb3 / n0) * ng;
-                    }
-
-                    src0->grad = ggml_acc_or_set(ctx, src0->grad, tensor->grad, nb1, nb2, nb3, offset, zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_PERMUTE:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    int32_t * axes = (int32_t *) tensor->op_params;
-                    int axis0 = axes[0] & 0x3;
-                    int axis1 = axes[1] & 0x3;
-                    int axis2 = axes[2] & 0x3;
-                    int axis3 = axes[3] & 0x3;
-                    int axes_backward[4] = {0,0,0,0};
-                    axes_backward[axis0] = 0;
-                    axes_backward[axis1] = 1;
-                    axes_backward[axis2] = 2;
-                    axes_backward[axis3] = 3;
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            ggml_permute(ctx,
-                                tensor->grad,
-                                axes_backward[0],
-                                axes_backward[1],
-                                axes_backward[2],
-                                axes_backward[3]),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_TRANSPOSE:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            ggml_transpose(ctx, tensor->grad),
-                        zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_GET_ROWS:
-            {
-                // necessary for llama (only for tokenizer)
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            // last ggml_get_rows_back argument src0->grad is only
-                            // necessary to setup correct output shape
-                            ggml_get_rows_back(ctx, tensor->grad, src1, src0->grad),
-                        zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    // noop
-                }
-            } break;
-        case GGML_OP_GET_ROWS_BACK:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+                ggml_add_or_set(ctx, cgraph, isrc1, ggml_reshape(ctx, ggml_cont(ctx, tensor_grad_view), src1));
             }
-        case GGML_OP_DIAG:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_SUB: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
             }
-        case GGML_OP_DIAG_MASK_INF:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    const int n_past = ((int32_t *) tensor->op_params)[0];
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            /* ggml_diag_mask_inf_impl() shouldn't be here */
-                            /* ref:  https://github.com/ggerganov/llama.cpp/pull/4203#discussion_r1412377992 */
-                            ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false),
-                        zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_DIAG_MASK_ZERO:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    const int n_past = ((int32_t *) tensor->op_params)[0];
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false),
-                        zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SOFT_MAX:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            ggml_soft_max_back(ctx, tensor->grad, tensor),
-                        zero_table, acc_table);
-                }
-                GGML_ASSERT((!src1 || !src1->grad) && "backward pass for softmax mask not implemented");
-            } break;
-        case GGML_OP_SOFT_MAX_BACK:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+            if (src1_needs_grads) {
+                ggml_sub_or_set(ctx, cgraph, isrc1, grad);
             }
-        case GGML_OP_ROPE:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    //const int n_past = ((int32_t *) tensor->op_params)[0];
-                    const int n_dims     = ((int32_t *) tensor->op_params)[1];
-                    const int mode       = ((int32_t *) tensor->op_params)[2];
-                    //const int n_ctx      = ((int32_t *) tensor->op_params)[3];
-                    const int n_ctx_orig = ((int32_t *) tensor->op_params)[4];
-                    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-
-                    memcpy(&freq_base,   (int32_t *) tensor->op_params +  5, sizeof(float));
-                    memcpy(&freq_scale,  (int32_t *) tensor->op_params +  6, sizeof(float));
-                    memcpy(&ext_factor,  (int32_t *) tensor->op_params +  7, sizeof(float));
-                    memcpy(&attn_factor, (int32_t *) tensor->op_params +  8, sizeof(float));
-                    memcpy(&beta_fast,   (int32_t *) tensor->op_params +  9, sizeof(float));
-                    memcpy(&beta_slow,   (int32_t *) tensor->op_params + 10, sizeof(float));
-
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_rope_back(ctx,
-                                tensor->grad,
-                                src1,
-                                src2,
-                                n_dims,
-                                mode,
-                                n_ctx_orig,
-                                freq_base,
-                                freq_scale,
-                                ext_factor,
-                                attn_factor,
-                                beta_fast,
-                                beta_slow),
-                            zero_table, acc_table);
-                }
-                GGML_ASSERT((!src2 || !src2->grad) && "gradients for freq factors not implemented");
-            } break;
-        case GGML_OP_ROPE_BACK:
-            {
-                if (src0->grad) {
-                    //const int n_past = ((int32_t *) tensor->op_params)[0];
-                    const int n_dims     = ((int32_t *) tensor->op_params)[1];
-                    const int mode       = ((int32_t *) tensor->op_params)[2];
-                    //const int n_ctx      = ((int32_t *) tensor->op_params)[3];
-                    const int n_ctx_orig = ((int32_t *) tensor->op_params)[4];
-                    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-
-                    memcpy(&freq_base,   (int32_t *) tensor->op_params +  5, sizeof(float));
-                    memcpy(&freq_scale,  (int32_t *) tensor->op_params +  6, sizeof(float));
-                    memcpy(&ext_factor,  (int32_t *) tensor->op_params +  7, sizeof(float));
-                    memcpy(&attn_factor, (int32_t *) tensor->op_params +  8, sizeof(float));
-                    memcpy(&beta_fast,   (int32_t *) tensor->op_params +  9, sizeof(float));
-                    memcpy(&beta_slow,   (int32_t *) tensor->op_params + 10, sizeof(float));
-
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_rope_impl(ctx,
-                                tensor->grad,
-                                src1,
-                                src2,
-                                n_dims,
-                                mode,
-                                n_ctx_orig,
-                                freq_base,
-                                freq_scale,
-                                ext_factor,
-                                attn_factor,
-                                beta_fast,
-                                beta_slow,
-                                false),
-                            zero_table, acc_table);
+        } break;
+        case GGML_OP_MUL: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, src1, grad));
+            }
+            if (src1_needs_grads) {
+                struct ggml_tensor * tmp = ggml_mul(ctx, src0, grad);
+                if (!ggml_are_same_shape(src0, src1)) {
+                    tmp = ggml_repeat_back(ctx, tmp, src1);
                 }
-            } break;
-        case GGML_OP_CLAMP:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+                ggml_add_or_set(ctx, cgraph, isrc1, tmp);
             }
-        case GGML_OP_CONV_TRANSPOSE_1D:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_DIV: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_div(ctx, grad, src1));
             }
-        case GGML_OP_IM2COL:
-            {
-                if (src1->grad) {
-                    const int32_t s0    = ggml_get_op_params_i32(tensor, 0);
-                    const int32_t s1    = ggml_get_op_params_i32(tensor, 1);
-                    const int32_t p0    = ggml_get_op_params_i32(tensor, 2);
-                    const int32_t p1    = ggml_get_op_params_i32(tensor, 3);
-                    const int32_t d0    = ggml_get_op_params_i32(tensor, 4);
-                    const int32_t d1    = ggml_get_op_params_i32(tensor, 5);
-                    const bool    is_2D = ggml_get_op_params_i32(tensor, 6) == 1;
-
-                    src1->grad = ggml_add_or_set(ctx,
-                            src1->grad,
-                            ggml_im2col_back(ctx, src0, tensor->grad, src1->ne, s0, s1, p0, p1, d0, d1, is_2D),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_IM2COL_BACK:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+            if (src1_needs_grads) {
+                ggml_sub_or_set(ctx, cgraph, isrc1, ggml_mul(ctx, grad, ggml_div(ctx, tensor, src1)));
             }
-        case GGML_OP_CONV_TRANSPOSE_2D:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_SQR: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_scale(ctx, ggml_mul(ctx, src0, grad), 2.0f));
             }
-        case GGML_OP_POOL_1D:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_SQRT: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_scale(ctx, ggml_div(ctx, grad, tensor), 0.5f));
             }
-        case GGML_OP_POOL_2D:
-            {
-                if (src0->grad) {
-                    const enum ggml_op_pool op = ggml_get_op_params_i32(tensor, 0);
-                    const      int32_t      k0 = ggml_get_op_params_i32(tensor, 1);
-                    const      int32_t      k1 = ggml_get_op_params_i32(tensor, 2);
-                    const      int32_t      s0 = ggml_get_op_params_i32(tensor, 3);
-                    const      int32_t      s1 = ggml_get_op_params_i32(tensor, 4);
-                    const      int32_t      p0 = ggml_get_op_params_i32(tensor, 5);
-                    const      int32_t      p1 = ggml_get_op_params_i32(tensor, 6);
-
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_pool_2d_back(ctx, tensor->grad, src0, op, k0, k1, s0, s1, p0, p1),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_POOL_2D_BACK:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_LOG: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_div(ctx, grad, src0));
             }
-        case GGML_OP_UPSCALE:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_SIN: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, grad, ggml_cos(ctx, src0)));
             }
-        case GGML_OP_PAD:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_COS: {
+            if (src0_needs_grads) {
+                ggml_sub_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, grad, ggml_sin(ctx, src0)));
             }
-        case GGML_OP_ARANGE:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_SUM: {
+            if (src0_needs_grads) {
+                ggml_add1_or_set(ctx, cgraph, isrc0, src0, grad);
             }
-        case GGML_OP_TIMESTEP_EMBEDDING:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_SUM_ROWS: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_repeat(ctx, grad, src0));
             }
-        case GGML_OP_ARGSORT:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_MEAN: {
+            if (src0_needs_grads) {
+                ggml_add1_or_set(ctx, cgraph, isrc0, src0, ggml_scale_impl(ctx, grad, 1.0f/src0->ne[0], false));
             }
-        case GGML_OP_LEAKY_RELU:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_REPEAT: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_repeat_back(ctx, grad, src0));
             }
-        case GGML_OP_FLASH_ATTN_EXT:
-            {
-                GGML_ABORT("FA backward pass not adapted after rework");
-                struct ggml_tensor * flash_grad = NULL;
-                if (src0->grad || src1->grad || tensor->src[2]->grad) {
-                    int32_t t = ggml_get_op_params_i32(tensor, 0);
-                    GGML_ASSERT(t == 0 || t == 1);
-                    bool masked = t != 0;
-                    flash_grad =
-                        ggml_flash_attn_back(ctx,
-                            src0,
-                            src1,
-                            tensor->src[2],
-                            tensor->grad,
-                            masked);
+        } break;
+        case GGML_OP_REPEAT_BACK: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_repeat(ctx, grad, src0));
+            }
+        } break;
+        case GGML_OP_RMS_NORM: {
+            if (src0_needs_grads) {
+                float eps;
+                memcpy(&eps, tensor->op_params, sizeof(float));
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_rms_norm_back(ctx, src0, grad, eps));
+            }
+        } break;
+        case GGML_OP_MUL_MAT: {
+            // https://cs231n.github.io/optimization-2/#staged
+            // # forward pass
+            // s0 = np.random.randn(5, 10)
+            // s1 = np.random.randn(10, 3)
+            // t = s0.dot(s1)
+
+            // # now suppose we had the gradient on t from above in the circuit
+            // dt = np.random.randn(*t.shape) # same shape as t
+            // ds0 = dt.dot(s1.T) #.T gives the transpose of the matrix
+            // ds1 = t.T.dot(dt)
+
+            // tensor.shape [m,p,qq,rr]
+            // src0.shape   [n,m,q1,r1]
+            // src1.shape   [n,p,qq,rr]
+
+            if (src0_needs_grads) {
+                struct ggml_tensor * s1_tg =
+                    ggml_out_prod(ctx, // [n,m,qq,rr]
+                        src1,          // [n,p,qq,rr]
+                        grad);         // [m,p,qq,rr]
+                const int64_t qq = s1_tg->ne[2];
+                const int64_t rr = s1_tg->ne[3];
+                const int64_t q1 = src0->ne[2];
+                const int64_t r1 = src0->ne[3];
+                const bool ne2_broadcasted = qq > q1;
+                const bool ne3_broadcasted = rr > r1;
+                if (ne2_broadcasted || ne3_broadcasted) {
+                    // sum broadcast repetitions of s1_tg into shape of src0
+                    s1_tg = ggml_repeat_back(ctx, s1_tg, src0);
                 }
+                ggml_add_or_set(ctx, cgraph, isrc0, s1_tg /*= [n,m,q1,r1]*/);
+            }
+            if (src1_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc1,
+                        // ggml_mul_mat(ctx,                   // [n,p,qq,rr]
+                        //     ggml_cont(ctx,                  // [m,n,q1,r1]
+                        //         ggml_transpose(ctx, src0)), // [m,n,q1,r1]
+                        //     grad),                          // [m,p,qq,rr]
+
+                        // when src0 is bigger than tensor->grad (this is mostly the case in llama),
+                        // avoid transpose of src0, rather transpose smaller tensor->grad
+                        // and then use ggml_out_prod
+                        ggml_out_prod(ctx,      // [n,p,qq,rr]
+                            src0,               // [n,m,q1,r1]
+                            ggml_transpose(ctx, // [p,m,qq,rr]
+                                grad)));        // [m,p,qq,rr]
+            }
+        } break;
+        case GGML_OP_SCALE: {
+            if (src0_needs_grads) {
+                float s;
+                memcpy(&s, tensor->op_params, sizeof(float));
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_scale_impl(ctx, grad, s, false));
+            }
+        } break;
+        case GGML_OP_SET: {
+            const size_t nb1    = ((const int32_t *) tensor->op_params)[0];
+            const size_t nb2    = ((const int32_t *) tensor->op_params)[1];
+            const size_t nb3    = ((const int32_t *) tensor->op_params)[2];
+            const size_t offset = ((const int32_t *) tensor->op_params)[3];
+
+            struct ggml_tensor * tensor_grad_view = NULL;
+
+            if (src0_needs_grads || src1_needs_grads) {
+                GGML_ASSERT(src0->type == tensor->type);
+                GGML_ASSERT(!cgraph->grads[isrc0] ||                      cgraph->grads[isrc0]->type == grad->type);
+                GGML_ASSERT(!cgraph->grads[isrc1] || !src1_needs_grads || cgraph->grads[isrc1]->type == grad->type);
+
+                tensor_grad_view = ggml_view_4d(ctx,
+                    grad, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
+                    nb1, nb2, nb3, offset);
+            }
 
-                const int64_t elem_q = ggml_nelements(src0);
-                const int64_t elem_k = ggml_nelements(src1);
-                const int64_t elem_v = ggml_nelements(src2);
-
-                enum ggml_type result_type = flash_grad->type;
-                GGML_ASSERT(ggml_blck_size(result_type) == 1);
-                const size_t tsize = ggml_type_size(result_type);
-
-                const size_t offs_q = 0;
-                const size_t offs_k = offs_q + GGML_PAD(elem_q * tsize, GGML_MEM_ALIGN);
-                const size_t offs_v = offs_k + GGML_PAD(elem_k * tsize, GGML_MEM_ALIGN);
-
-                if (src0->grad) {
-                    struct ggml_tensor * view_q = ggml_view_1d(ctx, flash_grad, elem_q, offs_q);
-                    struct ggml_tensor * grad_q = ggml_reshape(ctx, view_q, src0);
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            grad_q,
-                            zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    struct ggml_tensor * view_k = ggml_view_1d(ctx, flash_grad, elem_k, offs_k);
-                    struct ggml_tensor * grad_k = ggml_reshape(ctx, view_k, src1);
-                    src1->grad = ggml_add_or_set(ctx,
-                            src1->grad,
-                            grad_k,
-                            zero_table, acc_table);
-                }
-                if (src2->grad) {
-                    struct ggml_tensor * view_v = ggml_view_1d(ctx, flash_grad, elem_v, offs_v);
-                    struct ggml_tensor * grad_v = ggml_reshape(ctx, view_v, src2);
-                    src2->grad = ggml_add_or_set(ctx,
-                            src2->grad,
-                            grad_v,
-                            zero_table, acc_table);
+            if (src0_needs_grads) {
+                struct ggml_tensor * tmp = ggml_neg(ctx, tensor_grad_view);
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_acc_impl(ctx, grad, tmp, nb1, nb2, nb3, offset, false));
+            }
+
+            if (src1_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc1, ggml_reshape(ctx, ggml_cont(ctx, tensor_grad_view), src1));
+            }
+        } break;
+        case GGML_OP_CPY: {
+            // cpy overwrites value of src1 by src0 and returns view(src1)
+            // the overwriting is mathematically equivalent to:
+            // tensor = src0 * 1 + src1 * 0
+            if (src0_needs_grads) {
+                // dsrc0 = dtensor * 1
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
+            }
+            if (src1_needs_grads) {
+                // dsrc1 = dtensor * 0 -> noop
+            }
+        } break;
+        case GGML_OP_CONT: {
+            // same as cpy
+            if (src0_needs_grads) {
+                GGML_ASSERT(!cgraph->grads[isrc0] || ggml_is_contiguous(cgraph->grads[isrc0]));
+                GGML_ASSERT(ggml_is_contiguous(grad));
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
+            }
+        } break;
+        case GGML_OP_RESHAPE: {
+            if (src0_needs_grads) {
+                struct ggml_tensor * grad_cont = ggml_is_contiguous(grad) ? grad : ggml_cont(ctx, grad);
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_reshape(ctx, grad_cont, src0));
+            }
+        } break;
+        case GGML_OP_VIEW: {
+            if (src0_needs_grads) {
+                size_t offset;
+
+                memcpy(&offset, tensor->op_params, sizeof(offset));
+
+                size_t nb1 = tensor->nb[1];
+                size_t nb2 = tensor->nb[2];
+                size_t nb3 = tensor->nb[3];
+
+                if (cgraph->grads[isrc0] && src0->type != cgraph->grads[isrc0]->type) {
+                    // gradient is typically F32, but src0 could be other type
+                    size_t ng = ggml_element_size(cgraph->grads[isrc0]);
+                    size_t n0 = ggml_element_size(src0);
+                    GGML_ASSERT(offset % n0 == 0);
+                    GGML_ASSERT(nb1 % n0 == 0);
+                    GGML_ASSERT(nb2 % n0 == 0);
+                    GGML_ASSERT(nb3 % n0 == 0);
+                    offset = (offset / n0) * ng;
+                    nb1 = (nb1 / n0) * ng;
+                    nb2 = (nb2 / n0) * ng;
+                    nb3 = (nb3 / n0) * ng;
                 }
-            } break;
-        case GGML_OP_FLASH_ATTN_BACK:
-            {
-                GGML_ABORT("fatal error"); // not supported
+
+                ggml_acc_or_set(ctx, cgraph, isrc0, src0, grad, nb1, nb2, nb3, offset);
             }
-        case GGML_OP_SSM_CONV:
-        case GGML_OP_SSM_SCAN:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_PERMUTE: {
+            if (src0_needs_grads) {
+                const int32_t * axes = (const int32_t *) tensor->op_params;
+                const int axis0 = axes[0] & 0x3;
+                const int axis1 = axes[1] & 0x3;
+                const int axis2 = axes[2] & 0x3;
+                const int axis3 = axes[3] & 0x3;
+                int axb[4] = {0,0,0,0}; // axes backward
+                axb[axis0] = 0;
+                axb[axis1] = 1;
+                axb[axis2] = 2;
+                axb[axis3] = 3;
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_permute(ctx, grad, axb[0], axb[1], axb[2], axb[3]));
             }
+        } break;
+        case GGML_OP_TRANSPOSE: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_transpose(ctx, grad));
+            }
+        } break;
+        case GGML_OP_GET_ROWS: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_get_rows_back(ctx, grad, src1, src0));
+            }
+            if (src1_needs_grads) {
+                // noop
+            }
+        } break;
+        case GGML_OP_DIAG_MASK_INF: {
+            if (src0_needs_grads) {
+                /* ggml_diag_mask_inf_impl() shouldn't be here */
+                /* ref:  https://github.com/ggerganov/llama.cpp/pull/4203#discussion_r1412377992 */
+                const int n_past = ((const int32_t *) tensor->op_params)[0];
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_diag_mask_zero_impl(ctx, grad, n_past, false));
+            }
+        } break;
+        case GGML_OP_DIAG_MASK_ZERO: {
+            if (src0_needs_grads) {
+                const int n_past = ((const int32_t *) tensor->op_params)[0];
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_diag_mask_zero_impl(ctx, grad, n_past, false));
+            }
+        } break;
+        case GGML_OP_SOFT_MAX: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_soft_max_back(ctx, grad, tensor));
+            }
+            GGML_ASSERT((!src1 || !src1_needs_grads) && "backward pass for softmax mask not implemented");
+        } break;
+        case GGML_OP_ROPE: {
+            if (src0_needs_grads) {
+                //const int n_past = ((int32_t *) tensor->op_params)[0];
+                const int n_dims     = ((const int32_t *) tensor->op_params)[1];
+                const int mode       = ((const int32_t *) tensor->op_params)[2];
+                //const int n_ctx      = ((int32_t *) tensor->op_params)[3];
+                const int n_ctx_orig = ((const int32_t *) tensor->op_params)[4];
+                float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+
+                memcpy(&freq_base,   (const float *) tensor->op_params +  5, sizeof(float));
+                memcpy(&freq_scale,  (const float *) tensor->op_params +  6, sizeof(float));
+                memcpy(&ext_factor,  (const float *) tensor->op_params +  7, sizeof(float));
+                memcpy(&attn_factor, (const float *) tensor->op_params +  8, sizeof(float));
+                memcpy(&beta_fast,   (const float *) tensor->op_params +  9, sizeof(float));
+                memcpy(&beta_slow,   (const float *) tensor->op_params + 10, sizeof(float));
+
+                ggml_add_or_set(ctx, cgraph, isrc0,
+                    ggml_rope_back(ctx, grad, src1, src2, n_dims, mode, n_ctx_orig, freq_base,
+                        freq_scale, ext_factor, attn_factor, beta_fast, beta_slow));
+            }
+            GGML_ASSERT((!src2 || !src2_needs_grads) && "gradients for freq factors not implemented");
+        } break;
+        case GGML_OP_IM2COL: {
+            if (src1_needs_grads) {
+                const int32_t s0    = ggml_get_op_params_i32(tensor, 0);
+                const int32_t s1    = ggml_get_op_params_i32(tensor, 1);
+                const int32_t p0    = ggml_get_op_params_i32(tensor, 2);
+                const int32_t p1    = ggml_get_op_params_i32(tensor, 3);
+                const int32_t d0    = ggml_get_op_params_i32(tensor, 4);
+                const int32_t d1    = ggml_get_op_params_i32(tensor, 5);
+                const bool    is_2D = ggml_get_op_params_i32(tensor, 6) == 1;
+
+                ggml_add_or_set(ctx, cgraph, isrc1, ggml_im2col_back(ctx, src0, grad, src1->ne, s0, s1, p0, p1, d0, d1, is_2D));
+            }
+        } break;
+        case GGML_OP_POOL_2D: {
+            if (src0_needs_grads) {
+                const enum ggml_op_pool op = ggml_get_op_params_i32(tensor, 0);
+                const      int32_t      k0 = ggml_get_op_params_i32(tensor, 1);
+                const      int32_t      k1 = ggml_get_op_params_i32(tensor, 2);
+                const      int32_t      s0 = ggml_get_op_params_i32(tensor, 3);
+                const      int32_t      s1 = ggml_get_op_params_i32(tensor, 4);
+                const      int32_t      p0 = ggml_get_op_params_i32(tensor, 5);
+                const      int32_t      p1 = ggml_get_op_params_i32(tensor, 6);
+
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_pool_2d_back(ctx, grad, src0, op, k0, k1, s0, s1, p0, p1));
+            }
+        } break;
         case GGML_OP_WIN_PART:
         case GGML_OP_WIN_UNPART:
-        case GGML_OP_UNARY:
-            {
-                switch (ggml_get_unary_op(tensor)) {
-                    case GGML_UNARY_OP_ABS:
-                        {
-                            if (src0->grad) {
-                                src0->grad =
-                                    ggml_add_or_set(ctx,
-                                            src0->grad,
-                                            ggml_mul(ctx,
-                                                ggml_sgn(ctx, src0),
-                                                tensor->grad),
-                                            zero_table, acc_table);
-                            }
-                        } break;
-                    case GGML_UNARY_OP_SGN:
-                        {
-                            if (src0->grad) {
-                                // noop
-                            }
-                        } break;
-                    case GGML_UNARY_OP_NEG:
-                        {
-                            if (src0->grad) {
-                                src0->grad = ggml_sub_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                            }
-                        } break;
-                    case GGML_UNARY_OP_STEP:
-                        {
-                            if (src0->grad) {
-                                // noop
-                            }
-                        } break;
-                    case GGML_UNARY_OP_TANH:
-                        {
-                            GGML_ABORT("fatal error"); // TODO: not implemented
-                        }
-                    case GGML_UNARY_OP_ELU:
-                        {
-                            GGML_ABORT("fatal error"); // TODO: not implemented
-                        }
-                    case GGML_UNARY_OP_RELU:
-                        {
-                            if (src0->grad) {
-                                src0->grad = ggml_add_or_set(ctx,
-                                        src0->grad,
-                                        ggml_mul(ctx,
-                                            ggml_step(ctx, src0),
-                                            tensor->grad),
-                                        zero_table, acc_table);
-                            }
-                        } break;
-                    case GGML_UNARY_OP_SIGMOID:
-                        {
-                            GGML_ABORT("fatal error"); // TODO: not implemented
-                        }
-                    case GGML_UNARY_OP_GELU:
-                        {
-                            GGML_ABORT("fatal error"); // TODO: not implemented
-                        }
-                    case GGML_UNARY_OP_GELU_QUICK:
-                        {
-                            GGML_ABORT("fatal error"); // TODO: not implemented
-                        }
-                    case GGML_UNARY_OP_SILU:
-                        {
-                            // necessary for llama
-                            if (src0->grad) {
-                                src0->grad = ggml_add_or_set(ctx,
-                                        src0->grad,
-                                        ggml_silu_back(ctx, src0, tensor->grad),
-                                        zero_table, acc_table);
-                            }
-                        } break;
-                    case GGML_UNARY_OP_EXP:
-                        {
-                            if (src0->grad) {
-                                src0->grad = ggml_add_or_set(ctx,
-                                        src0->grad,
-                                        ggml_mul(ctx, tensor, tensor->grad),
-                                        zero_table, acc_table);
-                            }
-                        } break;
-                    default:
-                        GGML_ABORT("fatal error");
-                }
-            } break;
-        case GGML_OP_GET_REL_POS:
-        case GGML_OP_ADD_REL_POS:
-        case GGML_OP_RWKV_WKV6:
-        case GGML_OP_MAP_UNARY:
-        case GGML_OP_MAP_BINARY:
-        case GGML_OP_MAP_CUSTOM1_F32:
-        case GGML_OP_MAP_CUSTOM2_F32:
-        case GGML_OP_MAP_CUSTOM3_F32:
-        case GGML_OP_MAP_CUSTOM1:
-        case GGML_OP_MAP_CUSTOM2:
-        case GGML_OP_MAP_CUSTOM3:
-            {
-                GGML_ABORT("fatal error"); // not supported
-            }
-        case GGML_OP_CROSS_ENTROPY_LOSS:
-            {
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_cross_entropy_loss_back(ctx,
-                                    src0,
-                                    src1,
-                                    tensor->grad),
-                                zero_table, acc_table);
-                }
-                GGML_ASSERT(!src1->grad && "backward pass for labels not implemented");
-            } break;
-        case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
-            {
-                GGML_ABORT("fatal error"); // not supported
+        case GGML_OP_UNARY: {
+            switch (ggml_get_unary_op(tensor)) {
+                case GGML_UNARY_OP_ABS: {
+                    if (src0_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, ggml_sgn(ctx, src0), grad));
+                    }
+                } break;
+                case GGML_UNARY_OP_SGN: {
+                    // noop
+                } break;
+                case GGML_UNARY_OP_NEG: {
+                    if (src0_needs_grads) {
+                        ggml_sub_or_set(ctx, cgraph, isrc0, grad);
+                    }
+                } break;
+                case GGML_UNARY_OP_STEP: {
+                    // noop
+                } break;
+                case GGML_UNARY_OP_RELU: {
+                    if (src0_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, ggml_step(ctx, src0), grad));
+                    }
+                } break;
+                case GGML_UNARY_OP_SILU: {
+                    if (src0_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_silu_back(ctx, src0, grad));
+                    }
+                } break;
+                case GGML_UNARY_OP_EXP: {
+                    if (src0_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, tensor, grad));
+                    }
+                } break;
+                default: {
+                    fprintf(stderr, "%s: unsupported unary op for backward pass: %s\n",
+                        __func__, ggml_unary_op_name(ggml_get_unary_op(tensor)));
+                    GGML_ABORT("fatal error");
+                } break;
             }
-        case GGML_OP_OPT_STEP_ADAMW:
-            {
-                GGML_ABORT("fatal error"); // not supported
+        } break;
+        case GGML_OP_CROSS_ENTROPY_LOSS: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_cross_entropy_loss_back(ctx, src0, src1, grad));
             }
-        case GGML_OP_NONE:
-            {
-                // nop
-            } break;
+            GGML_ASSERT(!src1_needs_grads && "backward pass for labels not implemented");
+        } break;
+        case GGML_OP_NONE: {
+            // noop
+        } break;
         case GGML_OP_COUNT:
-            {
-                GGML_ABORT("fatal error");
-            }
+        default: {
+            fprintf(stderr, "%s: unsupported ggml op for backward pass: %s\n", __func__, ggml_op_name(tensor->op));
+            GGML_ABORT("fatal error");
+        } break;
     }
 
-    for (int i = 0; i < GGML_MAX_SRC; ++i) {
-        if (tensor->src[i] && tensor->src[i]->grad) {
-            GGML_ASSERT(ggml_are_same_shape(tensor->src[i], tensor->src[i]->grad));
-        }
-    }
+    GGML_ASSERT(!src0_needs_grads || ggml_are_same_shape(src0, cgraph->grads[isrc0]));
+    GGML_ASSERT(!src1_needs_grads || ggml_are_same_shape(src1, cgraph->grads[isrc1]));
+    GGML_ASSERT(!src2_needs_grads || ggml_are_same_shape(src2, cgraph->grads[isrc2]));
 }
 
 static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
-    if (node->grad == NULL) {
-        // this usually happens when we generate intermediate nodes from constants in the backward pass
-        // it can also happen during forward pass, if the user performs computations with constants
-        if (node->op != GGML_OP_NONE) {
-            //GGML_PRINT_DEBUG("%s: warning: node %p has no grad, but op %d\n", __func__, (void *) node, node->op);
-        }
-    }
-
     // check if already visited
     if (ggml_hash_insert(&cgraph->visited_hash_set, node) == GGML_HASHSET_ALREADY_EXISTS) {
         return;
@@ -6207,18 +5586,42 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor *
     ggml_build_forward_impl(cgraph, tensor, true);
 }
 
-void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate) {
-    GGML_ASSERT(gf->n_nodes > 0);
-    GGML_ASSERT(gf->grads);
+void ggml_build_backward_expand(
+        struct ggml_context * ctx_static,
+        struct ggml_context * ctx_compute,
+        struct ggml_cgraph  * cgraph,
+        bool                  accumulate) {
+    GGML_ASSERT(cgraph->n_nodes > 0);
+    GGML_ASSERT(cgraph->grads);
+    GGML_ASSERT(cgraph->grad_accs);
+
+    const int n_nodes_f = cgraph->n_nodes;
 
-    for (int i = 0; i < gf->n_nodes; ++i) {
-        struct ggml_tensor * node = gf->nodes[i];
+    const size_t hash_size = ggml_hash_size(2*cgraph->size);
+    memset(cgraph->grads,     0, hash_size*sizeof(struct ggml_tensor *));
+    memset(cgraph->grad_accs, 0, hash_size*sizeof(struct ggml_tensor *));
+    bool * grads_needed = calloc(hash_size, sizeof(bool));
+
+    {
+        bool any_params = false;
+        bool any_loss   = false;
+        for (int i = 0; i < n_nodes_f; ++i) {
+            struct ggml_tensor * node = cgraph->nodes[i];
+            any_params = any_params || (node->flags & GGML_TENSOR_FLAG_PARAM);
+            any_loss   = any_loss   || (node->flags & GGML_TENSOR_FLAG_LOSS);
+        }
+        GGML_ASSERT(any_params && "no trainable parameters found, did you forget to call ggml_set_param?");
+        GGML_ASSERT(any_loss && "no training loss found, did you forget to call ggml_set_loss?");
+    }
+
+    for (int i = 0; i < n_nodes_f; ++i) {
+        struct ggml_tensor * node = cgraph->nodes[i];
 
         if (node->type == GGML_TYPE_I32) {
             continue;
         }
 
-        bool needs_grad = node->flags & GGML_TENSOR_FLAG_PARAM;
+        bool node_needs_grad = node->flags & GGML_TENSOR_FLAG_PARAM;
         bool ignore_src[GGML_MAX_SRC] = {false};
         switch (node->op) {
             // gradients in node->src[0] for one reason or another have no effect on output gradients
@@ -6246,14 +5649,14 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph *
                 break;
         }
         for (int j = 0; j < GGML_MAX_SRC; ++j) {
-            if (!node->src[j] || !node->src[j]->grad || ignore_src[j]) {
+            if (!node->src[j] || ignore_src[j] || !grads_needed[ggml_hash_find(&cgraph->visited_hash_set, node->src[j])]) {
                 continue;
             }
             GGML_ASSERT(node->src[j]->type == GGML_TYPE_F32 || node->src[j]->type == GGML_TYPE_F16);
-            needs_grad = true;
+            node_needs_grad = true;
             break;
         }
-        if (!needs_grad) {
+        if (!node_needs_grad) {
             continue;
         }
 
@@ -6261,73 +5664,21 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph *
         GGML_ASSERT(!node->view_src || node->op == GGML_OP_CPY || node->op == GGML_OP_VIEW ||
             node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE);
 
-        // create a new tensor with the same type and shape as the node and set it as grad
-        node->grad = ggml_dup_tensor(ctx, node);
-    }
-
-    // keep tables of original gradients for replacement/accumulation logic
-    struct ggml_hash_set zero_table = ggml_hash_set_new(gf->size);
-    struct ggml_hash_set acc_table  = ggml_hash_set_new(gf->size);
-    for (int i = 0; i < gf->n_nodes; i++) {
-        struct ggml_tensor * node = gf->nodes[i];
-
-        if (node->grad) {
-            {
-                const size_t insert_result = ggml_hash_insert(&zero_table, node->grad);
-                GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
-                GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
-            }
-
-            // only gradients of trainable parameters should be accumulated
-            if (accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) {
-                const size_t insert_result = ggml_hash_insert(&acc_table, node->grad);
-                GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
-                GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
-            }
+        const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
+        if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) {
+            cgraph->grads[igrad]     = ggml_dup_tensor(ctx_static, node);
+            cgraph->grad_accs[igrad] = cgraph->grads[igrad];
         }
+        grads_needed[igrad] = true;
     }
 
-    for (int i = gf->n_nodes - 1; i >= 0; i--) {
-        struct ggml_tensor * node = gf->nodes[i];
-
+    for (int i = n_nodes_f - 1; i >= 0; --i) {
         // inplace operations to add gradients are not created by ggml_compute_backward except for gradient accumulation
         // use allocator to automatically make inplace operations
-        if (node->grad) {
-            ggml_compute_backward(ctx, node, &zero_table, &acc_table);
-        }
+        ggml_compute_backward(ctx_compute, cgraph, i, grads_needed);
     }
 
-    for (int i = 0; i < gf->n_nodes; i++) {
-        struct ggml_tensor * node = gf->nodes[i];
-
-        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
-            GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node);
-            ggml_build_forward_expand(gb, node->grad);
-        }
-    }
-
-    ggml_hash_set_free(&zero_table);
-    ggml_hash_set_free(&acc_table);
-}
-
-void ggml_build_opt_adamw(
-        struct ggml_context * ctx,
-        struct ggml_cgraph  * gf,
-        struct ggml_cgraph  * gb,
-        float                 alpha,
-        float                 beta1,
-        float                 beta2,
-        float                 eps,
-        float                 wd) {
-    for (int i = 0; i < gf->n_nodes; i++) {
-        struct ggml_tensor * node = gf->nodes[i];
-
-        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
-            GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node);
-            struct ggml_tensor * opt_step = ggml_opt_step_adamw(ctx, node, node->grad, alpha, beta1, beta2, eps, wd);
-            ggml_build_forward_expand(gb, opt_step);
-        }
-    }
+    free(grads_needed);
 }
 
 static void * incr_ptr_aligned(void ** p, size_t size, size_t align) {
@@ -6345,7 +5696,8 @@ static size_t ggml_graph_nbytes(size_t size, bool grads) {
     incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // leafs
     incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // hash keys
     if (grads) {
-        incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // grads
+        incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // grads
+        incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // grad_accs
     }
     incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
 
@@ -6371,10 +5723,12 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
 
     void * p = cgraph + 1;
 
-    struct ggml_tensor ** nodes_ptr = incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
-    struct ggml_tensor ** leafs_ptr = incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
-    struct ggml_tensor ** hash_keys_ptr = incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
-    struct ggml_tensor ** grads_ptr = grads ? incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
+    struct ggml_tensor ** nodes_ptr     =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** leafs_ptr     =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** hash_keys_ptr =         incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** grads_ptr     = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
+    struct ggml_tensor ** grad_accs_ptr = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
+
     ggml_bitset_t * hash_used = incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
 
     // check that we allocated the correct amount of memory
@@ -6386,12 +5740,17 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
         /*.n_leafs      =*/ 0,
         /*.nodes        =*/ nodes_ptr,
         /*.grads        =*/ grads_ptr,
+        /*.grad_accs    =*/ grad_accs_ptr,
         /*.leafs        =*/ leafs_ptr,
         /*.hash_table   =*/ { hash_size, hash_used, hash_keys_ptr },
         /*.order        =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT,
     };
 
     ggml_hash_set_reset(&cgraph->visited_hash_set);
+    if (grads) {
+        memset(cgraph->grads,     0, hash_size*sizeof(struct ggml_tensor *));
+        memset(cgraph->grad_accs, 0, hash_size*sizeof(struct ggml_tensor *));
+    }
 
     return cgraph;
 }
@@ -6407,6 +5766,7 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1)
         /*.n_leafs      =*/ 0,
         /*.nodes        =*/ cgraph0->nodes + i0,
         /*.grads        =*/ cgraph0->grads ? cgraph0->grads + i0 : NULL,
+        /*.grad_accs    =*/ cgraph0->grad_accs ? cgraph0->grad_accs + i0 : NULL,
         /*.leafs        =*/ NULL,
         /*.hash_table   =*/ { 0, NULL, NULL },
         /*.order        =*/ cgraph0->order,
@@ -6432,19 +5792,23 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
         dst->nodes[i] = src->nodes[i];
     }
 
-    if (src->grads) {
-        GGML_ASSERT(dst->grads != NULL);
-        for (int i = 0; i < src->n_nodes; ++i) {
-            dst->grads[i] = src->grads[i];
-        }
-    }
-
     for (size_t i = 0; i < src->visited_hash_set.size; ++i) {
         // copy all hashset keys (tensors) that are in use
         if (ggml_bitset_get(src->visited_hash_set.used, i)) {
             ggml_hash_insert(&dst->visited_hash_set, src->visited_hash_set.keys[i]);
         }
     }
+
+    if (src->grads) {
+        GGML_ASSERT(dst->grads     != NULL);
+        GGML_ASSERT(dst->grad_accs != NULL);
+        for (int i = 0; i < src->n_nodes; ++i) {
+            const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]);
+            const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]);
+            dst->grads[igrad_dst]     = src->grads[igrad_src];
+            dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src];
+        }
+    }
 }
 
 struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph) {
@@ -6470,29 +5834,36 @@ void ggml_graph_reset(struct ggml_cgraph * cgraph) {
     GGML_ASSERT(cgraph->grads != NULL);
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        struct ggml_tensor * node = cgraph->nodes[i];
+        struct ggml_tensor * node     = cgraph->nodes[i];
+        struct ggml_tensor * grad_acc = ggml_graph_get_grad_acc(cgraph, node);
+
+        if (node->op == GGML_OP_OPT_STEP_ADAMW) {
+            // clear momenta
+            if (node->src[2]->data) {
+                ggml_set_zero(node->src[2]);
+            }
+            if (node->src[3]->data) {
+                ggml_set_zero(node->src[3]);
+            }
+        }
 
         // initial gradients of loss should be 1, 0 otherwise
-        if (node->grad) {
+        if (grad_acc) {
             if (node->flags & GGML_TENSOR_FLAG_LOSS) {
-                GGML_ASSERT(node->grad->buffer);
-                GGML_ASSERT(node->type == GGML_TYPE_F32);
-                GGML_ASSERT(ggml_is_scalar(node));
+                GGML_ASSERT(grad_acc->type == GGML_TYPE_F32);
+                GGML_ASSERT(ggml_is_scalar(grad_acc));
 
                 const float onef = 1.0f;
-                ggml_backend_tensor_set(node->grad, &onef, 0, ggml_nbytes(node->grad));
+                if (grad_acc->buffer) {
+                    ggml_backend_tensor_set(grad_acc, &onef, 0, sizeof(float));
+                } else {
+                    GGML_ASSERT(grad_acc->data);
+                    *((float *) grad_acc->data) = onef;
+                }
             } else {
-                ggml_set_zero(node->grad);
+                ggml_set_zero(grad_acc);
             }
         }
-
-        GGML_ASSERT(node);
-        if (node->op == GGML_OP_OPT_STEP_ADAMW) {
-            // set iteration to 1 and clear momenta
-            ggml_set_op_params_i32(node, 0, 1);
-            ggml_set_zero(node->src[2]);
-            ggml_set_zero(node->src[3]);
-        }
     }
 }
 
@@ -6530,7 +5901,7 @@ void ggml_graph_add_node(struct ggml_cgraph * cgraph, struct ggml_tensor * tenso
     cgraph->n_nodes++;
 }
 
-struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name) {
+struct ggml_tensor * ggml_graph_get_tensor(const struct ggml_cgraph * cgraph, const char * name) {
     for (int i = 0; i < cgraph->n_leafs; i++) {
         struct ggml_tensor * leaf = cgraph->leafs[i];
 
@@ -6550,6 +5921,16 @@ struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const ch
     return NULL;
 }
 
+struct ggml_tensor * ggml_graph_get_grad(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
+    const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
+    return igrad != GGML_HASHSET_FULL && ggml_bitset_get(cgraph->visited_hash_set.used, igrad) ? cgraph->grads[igrad] : NULL;
+}
+
+struct ggml_tensor * ggml_graph_get_grad_acc(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
+    const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
+    return igrad != GGML_HASHSET_FULL && ggml_bitset_get(cgraph->visited_hash_set.used, igrad) ? cgraph->grad_accs[igrad] : NULL;
+}
+
 void ggml_graph_print(const struct ggml_cgraph * cgraph) {
     GGML_LOG_INFO("=== GRAPH ===\n");
 
@@ -6560,7 +5941,8 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
         GGML_LOG_INFO(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s\n",
                 i,
                 node->ne[0], node->ne[1], node->ne[2],
-                ggml_op_name(node->op), (node->flags & GGML_TENSOR_FLAG_PARAM) ? "x" : node->grad ? "g" : " ");
+                ggml_op_name(node->op), (node->flags & GGML_TENSOR_FLAG_PARAM) ? "x" :
+                      ggml_graph_get_grad(cgraph, node) ? "g" : " ");
     }
 
     GGML_LOG_INFO("n_leafs = %d\n", cgraph->n_leafs);
@@ -6595,8 +5977,9 @@ static bool ggml_graph_find(const struct ggml_cgraph * cgraph, const struct ggml
 static struct ggml_tensor * ggml_graph_get_parent(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
     for (int i = 0; i < cgraph->n_nodes; i++) {
         struct ggml_tensor * parent = cgraph->nodes[i];
+        struct ggml_tensor * grad = ggml_graph_get_grad(cgraph, parent);
 
-        if (parent->grad == node) {
+        if (grad == node) {
             return parent;
         }
     }
@@ -6636,6 +6019,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
 
     for (int i = 0; i < gb->n_nodes; i++) {
         struct ggml_tensor * node = gb->nodes[i];
+        struct ggml_tensor * grad = ggml_graph_get_grad(gb, node);
 
         if (ggml_graph_get_parent(gb, node) != NULL) {
             continue;
@@ -6643,7 +6027,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
 
         if (node->flags & GGML_TENSOR_FLAG_PARAM) {
             snprintf(color, sizeof(color), "yellow");
-        } else if (node->grad) {
+        } else if (grad) {
             if (ggml_graph_find(gf, node)) {
                 snprintf(color, sizeof(color), "green");
             } else {
@@ -6670,8 +6054,8 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
             fprintf(fp, "%d [%" PRId64 ", %" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], node->ne[2], ggml_op_symbol(node->op));
         }
 
-        if (node->grad) {
-            fprintf(fp, " | <g>%s\"; ]\n", ggml_op_symbol(node->grad->op));
+        if (grad) {
+            fprintf(fp, " | <g>%s\"; ]\n", ggml_op_symbol(grad->op));
         } else {
             fprintf(fp, "\"; ]\n");
         }
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 6618d03d150a0..f8a59b6df175f 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -811,11 +811,11 @@ struct test_case {
 
         ggml_build_forward_expand(gf, out);
         ggml_graph_cpy(gf, gb);
-        ggml_build_backward_expand(ctx, gf, gb, false);
+        ggml_build_backward_expand(ctx, ctx, gb, false);
         if (expect.size() != 1 || expect[0] != 0.0f) {
             GGML_ASSERT(ggml_graph_n_nodes(gb) > ggml_graph_n_nodes(gf));
             for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-                GGML_ASSERT(!(t->flags & GGML_TENSOR_FLAG_PARAM) || t->grad->op != GGML_OP_NONE);
+                GGML_ASSERT(!(t->flags & GGML_TENSOR_FLAG_PARAM) || ggml_graph_get_grad(gb, t)->op != GGML_OP_NONE);
             }
         }
 
@@ -862,7 +862,13 @@ struct test_case {
             const char * bn = ggml_backend_name(backend);
             const int64_t ne = ggml_nelements(t);
 
-            std::vector<float> ga = tensor_to_float(t->grad);
+            std::vector<float> ga;
+            struct ggml_tensor * grad = ggml_graph_get_grad(gb, t);
+            if (grad) {
+                ga = tensor_to_float(grad);
+            } else {
+                ga.resize(ne); // default value is 0.0f
+            }
 
             for (int64_t i = 0; i < ne; ++i) { // gradient algebraic
                 // check for nans
@@ -2500,6 +2506,35 @@ struct test_sum_rows : public test_case {
     }
 };
 
+// GGML_OP_MEAN
+struct test_mean : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_mean(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {10, 5, 4, 3})
+        : type(type), ne(ne) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_param(ctx, a);
+        ggml_set_name(a, "a");
+
+        ggml_tensor * out = ggml_mean(ctx, a);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    float grad_eps() override {
+        return 0.1f * ne[0]*ne[1]*ne[2]*ne[3];
+    }
+};
+
 // GGML_OP_UPSCALE
 struct test_upscale : public test_case {
     const ggml_type type;
@@ -2834,24 +2869,14 @@ struct test_cross_entropy_loss : public test_case {
 struct test_opt_step_adamw : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
-    const float alpha;
-    const float beta1;
-    const float beta2;
-    const float eps;
-    const float wd;
 
     std::string vars() override {
-        return VARS_TO_STR7(type, ne, alpha, beta1, beta2, eps, wd);
+        return VARS_TO_STR2(type, ne);
     }
 
     test_opt_step_adamw(ggml_type type = GGML_TYPE_F32,
-            std::array<int64_t, 4> ne = {10, 5, 4, 3},
-            float alpha = 1e-3f,
-            float beta1 = 0.9f,
-            float beta2 = 0.999f,
-            float eps = 1e-8f,
-            float wd = 0.0f)
-        : type(type), ne(ne), alpha(alpha), beta1(beta1), beta2(beta2), eps(eps), wd(wd) {}
+            std::array<int64_t, 4> ne = {10, 5, 4, 3})
+        : type(type), ne(ne) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
@@ -2861,7 +2886,16 @@ struct test_opt_step_adamw : public test_case {
         ggml_tensor * grad = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
         ggml_set_name(grad, "grad");
 
-        ggml_tensor * out = ggml_opt_step_adamw(ctx, a, grad, alpha, beta1, beta2, eps, wd);
+        ggml_tensor * grad_m = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
+        ggml_set_name(grad_m, "grad_m");
+
+        ggml_tensor * grad_v = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
+        ggml_set_name(grad_v, "grad_v");
+
+        ggml_tensor * adamw_params = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 7);
+        ggml_set_name(adamw_params, "adamw_params");
+
+        ggml_tensor * out = ggml_opt_step_adamw(ctx, a, grad, grad_m, grad_v, adamw_params);
         ggml_set_name(out, "out");
 
         return out;
@@ -2869,7 +2903,7 @@ struct test_opt_step_adamw : public test_case {
 
     void initialize_tensors(ggml_context * ctx) override {
         for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-            init_tensor_uniform(t, 0.0f, 1.0f); // grad_v needs non-negative values.
+            init_tensor_uniform(t, 0.0f, 1.0f); // grad_v and adamw_params need non-negative values.
         }
     }
 
@@ -3735,6 +3769,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
     test_cases.emplace_back(new test_sum());
     test_cases.emplace_back(new test_sum_rows());
+    test_cases.emplace_back(new test_mean());
     test_cases.emplace_back(new test_upscale());
     test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, { 512, 512, 3, 1 }, 2, true));
     test_cases.emplace_back(new test_upscale_ext());
@@ -3766,9 +3801,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     }
 
     test_cases.emplace_back(new test_cross_entropy_loss());
-    for (float wd : {0.0f, 1e-2f}) {
-        test_cases.emplace_back(new test_opt_step_adamw(GGML_TYPE_F32, {10, 5, 4, 3}, 1.0f, 1e-3f, 0.9f, 0.999f, wd));
-    }
+    test_cases.emplace_back(new test_opt_step_adamw(GGML_TYPE_F32, {10, 5, 4, 3}));
 
     // these tests are disabled to save execution time, but they can be handy for debugging
 #if 0
@@ -3938,6 +3971,8 @@ int main(int argc, char ** argv) {
         ggml_backend_free(backend);
     }
 
+    ggml_quantize_free();
+
     printf("%zu/%zu backends passed\n", n_ok, ggml_backend_dev_count());
 
     if (n_ok != ggml_backend_dev_count()) {
@@ -3945,8 +3980,6 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    ggml_quantize_free();
-
     printf("\033[1;32mOK\033[0m\n");
     return 0;
 }
diff --git a/tests/test-opt.cpp b/tests/test-opt.cpp
index 546ca230ba417..4abe85c74a4f6 100644
--- a/tests/test-opt.cpp
+++ b/tests/test-opt.cpp
@@ -1,181 +1,892 @@
 #include "ggml.h"
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include "ggml-cpu.h"
+#include "ggml-opt.h"
 
 #include <cmath>
-#include <cstdio>
-#include <cstdlib>
-#include <cassert>
-
-#define MAX_NARGS 2
-
-#if defined(__GNUC__)
-#pragma GCC diagnostic ignored "-Wdouble-promotion"
-#endif
-
-//
-// logging
-//
-#define GGML_DEBUG 0
-#if (GGML_DEBUG >= 1)
-#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG(...)
-#endif
-
-#if (GGML_DEBUG >= 5)
-#define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG_5(...)
-#endif
-
-#if (GGML_DEBUG >= 10)
-#define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG_10(...)
-#endif
-
-#define GGML_PRINT(...) printf(__VA_ARGS__)
-
-
-static float frand(void) {
-    return (float)rand()/(float)RAND_MAX;
-}
-
-static struct ggml_tensor * get_random_tensor(
-    struct ggml_context * ctx0, int ndims, int64_t ne[], float fmin, float fmax
-) {
-    struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F32, ndims, ne);
-
-    switch (ndims) {
-        case 1:
-            for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((float *)result->data)[i0] = frand()*(fmax - fmin) + fmin;
+#include <inttypes.h>
+#include <random>
+#include <string>
+#include <thread>
+#include <vector>
+
+static bool almost_equal(const double a, const double b, const double atol) {
+    return fabs(a - b) < atol;
+}
+
+constexpr int64_t ne_datapoint = 2;
+constexpr int64_t ne_label     = 1;
+constexpr int64_t ndata        = 6;
+
+struct helper_ctx_data {
+    std::vector<ggml_opt_dataset_t>   datasets_supervised;
+    std::vector<struct ggml_tensor *> data_batch;
+    std::vector<struct ggml_tensor *> labels_batch;
+
+    ggml_opt_dataset_t       dataset_unsupervised;
+    struct ggml_context    * ctx_static;
+    struct ggml_context    * ctx_compute;
+    struct ggml_opt_params   opt_params;
+    ggml_opt_context_t       opt_ctx;
+    struct ggml_tensor     * inputs;
+    struct ggml_tensor     * weights;
+    struct ggml_tensor     * outputs;
+    ggml_backend_buffer_t    buf;
+    ggml_opt_result_t        result;
+    ggml_opt_result_t        result2;
+};
+
+// These default values make it easier to check optimization results vs. expected values.
+static ggml_opt_optimizer_params helper_get_test_opt_pars(void * userdata) {
+    ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(userdata);
+    result.adamw.alpha = 1.0f;
+    result.adamw.beta1 = 0.0f;
+    result.adamw.beta2 = 0.0f;
+    result.adamw.eps   = 0.0f;
+    return result;
+}
+
+static helper_ctx_data helper_get_ctx_data(
+        ggml_backend_sched_t    backend_sched,
+        ggml_backend_t          backend,
+        const bool              init_opt_ctx       = true,
+        const bool              optimizer_defaults = true,
+        int64_t                 nbatch_logical     = 1,
+        int64_t                 nbatch_physical    = 1,
+        enum ggml_opt_loss_type loss_type          = GGML_OPT_LOSS_TYPE_SUM) {
+    std::vector<ggml_opt_dataset_t> datasets(ndata);
+    for (int64_t ndata_shard = 1; ndata_shard <= ndata; ++ndata_shard) {
+        ggml_opt_dataset_t dataset = ggml_opt_dataset_init(ne_datapoint, ne_label, ndata, ndata_shard);
+
+        float * data   = ggml_get_data_f32(ggml_opt_dataset_data(  dataset));
+        float * labels = ggml_get_data_f32(ggml_opt_dataset_labels(dataset));
+
+        for (int64_t idata = 0; idata < ndata; ++idata) {
+            for (int64_t id = 0; id < ne_datapoint; ++id) {
+                data[  idata*ne_datapoint + id] =     16*idata + id;
             }
-            break;
-        case 2:
-            for (int i1 = 0; i1 < ne[1]; i1++) {
-                for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((float *)result->data)[i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
-                }
+            for (int64_t il = 0; il < ne_label;     ++il) {
+                labels[idata*ne_label     + il] = 16*(16*idata + il);
             }
-            break;
-        case 3:
-            for (int i2 = 0; i2 < ne[2]; i2++) {
-                for (int i1 = 0; i1 < ne[1]; i1++) {
-                    for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((float *)result->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
+        }
+
+        datasets[ndata_shard-1] = dataset;
+    }
+
+    ggml_opt_dataset_t dataset_unsupervised = ggml_opt_dataset_init(1, 0, ndata, /*ndata_shard =*/ 1);
+
+    float * data = ggml_get_data_f32(ggml_opt_dataset_data(dataset_unsupervised));
+
+    for (int64_t idata = 0; idata < ndata; ++idata) {
+        data[idata] = idata;
+    }
+
+    struct ggml_context * ctx_static;
+    struct ggml_context * ctx_compute;
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ (2*ndata + 2)*ggml_tensor_overhead(),
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        ctx_static = ggml_init(params);
+    }
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ GGML_DEFAULT_GRAPH_SIZE*ggml_tensor_overhead() + 3*ggml_graph_overhead(),
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        ctx_compute = ggml_init(params);
+    }
+
+    std::vector<struct ggml_tensor *>   data_batch(ndata);
+    std::vector<struct ggml_tensor *> labels_batch(ndata);
+    for (int64_t ndata_batch = 1; ndata_batch <= ndata; ++ndata_batch) {
+        data_batch[ndata_batch-1]   = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, ndata_batch*ne_datapoint);
+        labels_batch[ndata_batch-1] = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, ndata_batch*ne_label);
+    }
+
+    struct ggml_tensor * inputs = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, nbatch_physical);
+    ggml_set_name(inputs, "inputs");
+
+    struct ggml_tensor * weights = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, 1);
+    ggml_set_name(weights, "weights");
+    ggml_set_param(ctx_static, weights);
+
+    struct ggml_tensor * intermediary = ggml_add(ctx_compute, inputs, weights);
+
+    struct ggml_tensor * outputs = ggml_scale(ctx_compute, intermediary, 1.0f);
+    ggml_set_name(outputs, "outputs");
+
+    ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx_static, backend);
+    const float w0 = float(ndata)/2;
+    ggml_backend_tensor_set(weights, &w0, 0, sizeof(float));
+
+    GGML_ASSERT(nbatch_logical % nbatch_physical == 0);
+    const int32_t opt_period = nbatch_logical / nbatch_physical;
+
+    struct ggml_opt_params opt_params = ggml_opt_default_params(backend_sched, ctx_compute, inputs, outputs, loss_type);
+    opt_params.opt_period = opt_period;
+    if (!optimizer_defaults) {
+        opt_params.get_opt_pars = helper_get_test_opt_pars;
+    }
+    ggml_opt_context_t opt_ctx = init_opt_ctx ? ggml_opt_init(opt_params) : nullptr;
+
+    ggml_opt_result_t result  = ggml_opt_result_init();
+    ggml_opt_result_t result2 = ggml_opt_result_init();
+
+    return {datasets, data_batch, labels_batch, dataset_unsupervised, ctx_static, ctx_compute, opt_params, opt_ctx, inputs, weights, outputs, buf, result, result2};
+}
+
+static void helper_free_ctx_data(struct helper_ctx_data ctx_data) {
+    ggml_opt_result_free(ctx_data.result);
+    ggml_opt_result_free(ctx_data.result2);
+    ggml_opt_free(ctx_data.opt_ctx);
+    ggml_backend_buffer_free(ctx_data.buf);
+    ggml_free(ctx_data.ctx_static);
+    ggml_free(ctx_data.ctx_compute);
+    for (ggml_opt_dataset_t dataset : ctx_data.datasets_supervised) {
+        ggml_opt_dataset_free(dataset);
+    }
+    ggml_opt_dataset_free(ctx_data.dataset_unsupervised);
+}
+
+static void helper_after_test(
+        const char * func, const bool high_level, const std::string options,
+        const std::string subtest, const bool subtest_ok, int & ntest, int & npass) {
+    printf("  %s(high_level=%s%s, subtest=%s): ",
+           func, high_level ? "yes" : "no", options.c_str(), subtest.c_str());
+    if (subtest_ok) {
+        printf("\033[1;32mOK\033[0m\n");
+        npass++;
+    } else {
+        printf("\033[1;31mFAIL\033[0m\n");
+    }
+    ntest++;
+}
+
+static std::pair<int, int> test_dataset(ggml_backend_sched_t backend_sched, ggml_backend_t backend, const bool shuffle) {
+    int ntest = 0;
+    int npass = 0;
+
+    struct helper_ctx_data cd = helper_get_ctx_data(backend_sched, backend);
+
+    for (int64_t ndata_shard = 1; ndata_shard <= ndata; ++ndata_shard) {
+        ggml_opt_dataset_t dataset = cd.datasets_supervised[ndata_shard-1];
+
+        if (shuffle) {
+            ggml_opt_dataset_shuffle(cd.opt_ctx, dataset, -1);
+        }
+
+        for (int64_t ndata_batch = 1; ndata_batch <= ndata; ++ndata_batch) {
+            if (ndata_batch % ndata_shard != 0) {
+                continue;
+            }
+            bool subtest_ok = true;
+
+            struct ggml_tensor *   data_batch =   cd.data_batch[ndata_batch-1];
+            struct ggml_tensor * labels_batch = cd.labels_batch[ndata_batch-1];
+
+            std::vector<float>   data(ggml_nelements(  data_batch));
+            std::vector<float> labels(ggml_nelements(labels_batch));
+
+            std::vector<int64_t> idata_shuffled;
+            const int64_t nbatches = ndata / ndata_batch;
+            for (int64_t ibatch = 0; ibatch < nbatches; ++ibatch) {
+                ggml_opt_dataset_get_batch(dataset, data_batch, labels_batch, ibatch);
+
+                ggml_backend_tensor_get(  data_batch,   data.data(), 0, ggml_nbytes(  data_batch));
+                ggml_backend_tensor_get(labels_batch, labels.data(), 0, ggml_nbytes(labels_batch));
+
+                for (int64_t idata_batch = 0; idata_batch < ndata_batch; ++idata_batch) {
+                    const int64_t idata = ibatch*ndata_batch + idata_batch;
+                    const int64_t idata_found = data[idata_batch*ne_datapoint] / 16;
+                    subtest_ok = subtest_ok && (shuffle || idata_found == idata);
+                    idata_shuffled.push_back(idata_found);
+
+                    for (int64_t id = 0; id < ne_datapoint; ++id) {
+                        if (data[  idata_batch*ne_datapoint + id] != 16*idata_found + id) {
+                            subtest_ok = false;
+                        }
+                    }
+                    for (int64_t il = 0; il < ne_label;     ++il) {
+                        if (labels[idata_batch*ne_label     + il] != 16*(16*idata_found + il)) {
+                            subtest_ok = false;
+                        }
                     }
                 }
             }
-            break;
-        case 4:
-            for (int i3 = 0; i3 < ne[3]; i3++) {
-                for (int i2 = 0; i2 < ne[2]; i2++) {
-                    for (int i1 = 0; i1 < ne[1]; i1++) {
-                        for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((float *)result->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
-                        }
+
+            if (!shuffle || ndata % ndata_batch == 0) {
+                const int ndata_max = (ndata / ndata_batch) * ndata_batch;
+
+                for (int64_t idata = 0; subtest_ok && idata < ndata_max; ++idata) {
+                    int ninstances = 0;
+                    for (int64_t id : idata_shuffled) {
+                        ninstances += id == idata;
                     }
+                    if (ninstances != 1) {
+                        subtest_ok = false;
+                    }
+                }
+            }
+
+            printf("  %s(shuffle=%s, ndata_shard=%" PRId64 ", ndata_batch=%" PRId64 "): ",
+                   __func__, shuffle ? "yes" : "no", ndata_shard, ndata_batch);
+            if (subtest_ok) {
+                printf("\033[1;32mOK\033[0m\n");
+                npass++;
+            } else {
+                printf("\033[1;31mFAIL\033[0m\n");
+            }
+            ntest++;
+        }
+    }
+
+    helper_free_ctx_data(cd);
+
+    return std::make_pair(npass, ntest);
+}
+
+static std::pair<int, int> test_grad(ggml_backend_sched_t backend_sched, ggml_backend_t backend) {
+    int ntest = 0;
+    int npass = 0;
+
+    struct helper_ctx_data cd = helper_get_ctx_data(backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false,
+    /*nbatch_logical =*/ 999999, /*nbatch_physical =*/ 1);
+
+    std::vector<float> grad_history(ndata);
+    for (int64_t idata = 0; idata < ndata; ++idata) {
+        grad_history[idata] = NAN;
+    }
+
+    for (int idata = 0; idata < ndata; ++idata) {
+        const float idataf = idata;
+        ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
+        ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+        ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata, 0, sizeof(float));
+    }
+
+    {
+        bool subtest_ok = true;
+        for (int idata = 0; idata < ndata; ++idata) {
+            if (grad_history[idata] != idata + 1) {
+                subtest_ok = false;
+            }
+        }
+        printf("  %s(): ", __func__);
+        if (subtest_ok) {
+            printf("\033[1;32mOK\033[0m\n");
+            npass++;
+        } else {
+            printf("\033[1;31mFAIL\033[0m\n");
+        }
+        ntest++;
+    }
+
+    helper_free_ctx_data(cd);
+
+    return std::make_pair(npass, ntest);
+}
+
+static void helper_after_test_forward_backward(
+        const char * func, const bool high_level, const bool shuffle,
+        const std::string subtest, const bool subtest_ok, int & ntest, int & npass) {
+    std::string options = ", shuffle=";
+    options += shuffle ? "yes" : "no";
+    helper_after_test(func, high_level, options, subtest, subtest_ok, ntest, npass);
+}
+
+static std::pair<int, int> test_forward_backward(
+        ggml_backend_sched_t backend_sched, ggml_backend_t backend, const bool high_level, const bool shuffle) {
+    int ntest = 0;
+    int npass = 0;
+
+    struct helper_ctx_data cd = helper_get_ctx_data(backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false);
+    struct ggml_tensor * loss = ggml_opt_loss(cd.opt_ctx);
+
+    std::vector<float> loss_history(ndata);
+    for (int64_t idata = 0; idata < ndata; ++idata) {
+        loss_history[idata] = NAN;
+    }
+
+    {
+        int64_t ndata;
+        ggml_opt_result_ndata(cd.result, &ndata);
+        double loss;
+        double loss_unc;
+        ggml_opt_result_loss(cd.result, &loss, &loss_unc);
+        double accuracy;
+        double accuracy_unc;
+        ggml_opt_result_accuracy(cd.result, &accuracy, &accuracy_unc);
+        const bool subtest_ok = ndata == 0 && loss == 0.0 && std::isnan(loss_unc) && std::isnan(accuracy) && std::isnan(accuracy_unc);
+        helper_after_test_forward_backward(__func__, high_level, shuffle, "results_initial", subtest_ok, ntest, npass);
+    }
+
+    if (high_level) {
+        ggml_opt_dataset_t dataset = cd.dataset_unsupervised;
+        if (shuffle) {
+            ggml_opt_dataset_shuffle(cd.opt_ctx, dataset, -1);
+        }
+        ggml_opt_epoch(cd.opt_ctx, dataset, nullptr, cd.result, 0, nullptr, nullptr);
+    } else {
+        for (int idata = 0; idata < ndata; ++idata) {
+            const float idataf = idata;
+            ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
+            ggml_opt_forward(cd.opt_ctx, cd.result);
+            ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float));
+        }
+    }
+
+    {
+        float weights;
+        ggml_backend_tensor_get(cd.weights, &weights, 0, sizeof(float));
+        const bool subtest_ok = weights == ndata/2;
+        helper_after_test_forward_backward(__func__, high_level, shuffle, "weights_after_forward", subtest_ok, ntest, npass);
+    }
+    {
+        int64_t ndata;
+        ggml_opt_result_ndata(cd.result, &ndata);
+        bool subtest_ok = ndata == 6;
+
+        double loss;
+        double loss_unc;
+        ggml_opt_result_loss(cd.result, &loss, &loss_unc);
+        subtest_ok = subtest_ok && loss == 33.0 && almost_equal(loss_unc, sqrt(3.5), 1e-10);
+
+        double accuracy;
+        double accuracy_unc;
+        ggml_opt_result_accuracy(cd.result, &accuracy, &accuracy_unc);
+        subtest_ok = subtest_ok && std::isnan(accuracy) && std::isnan(accuracy_unc);
+
+        helper_after_test_forward_backward(__func__, high_level, shuffle, "results_after_forward", subtest_ok, ntest, npass);
+    }
+
+    float w0;
+    ggml_backend_tensor_get(cd.weights, &w0, 0, sizeof(float));
+    for (int i = 0; i < 10; ++i) {
+        ggml_opt_forward_backward(cd.opt_ctx, nullptr);
+    }
+    ggml_backend_tensor_set(cd.weights, &w0, 0, sizeof(float));
+
+    ggml_opt_reset(cd.opt_ctx, /*optimizer =*/ false);
+    ggml_opt_result_reset(cd.result);
+
+    for (int64_t idata = 0; idata < ndata; ++idata) {
+        loss_history[idata] = NAN;
+    }
+
+    if (high_level) {
+        ggml_opt_dataset_t dataset = cd.dataset_unsupervised;
+        if (shuffle) {
+            ggml_opt_dataset_shuffle(cd.opt_ctx, dataset, -1);
+        }
+        ggml_opt_epoch(cd.opt_ctx, dataset, cd.result, nullptr, ndata, nullptr, nullptr);
+    } else {
+        for (int idata = 0; idata < ndata; ++idata) {
+            const float idataf = idata;
+            ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
+            ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+            ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float));
+        }
+    }
+
+    {
+        float weights;
+        ggml_backend_tensor_get(cd.weights, &weights, 0, sizeof(float));
+        const bool subtest_ok = weights == -ndata/2;
+        helper_after_test_forward_backward(__func__, high_level, shuffle, "weights_after_forward_backward", subtest_ok, ntest, npass);
+    }
+    {
+        int64_t ndata;
+        ggml_opt_result_ndata(cd.result, &ndata);
+        bool subtest_ok = ndata == 6;
+
+        double loss;
+        double loss_unc;
+        ggml_opt_result_loss(cd.result, &loss, &loss_unc);
+        subtest_ok = subtest_ok && loss == 18.0 && (shuffle || loss_unc == 0.0);
+
+        double accuracy;
+        double accuracy_unc;
+        ggml_opt_result_accuracy(cd.result, &accuracy, &accuracy_unc);
+        subtest_ok = subtest_ok && std::isnan(accuracy) && std::isnan(accuracy_unc);
+
+        helper_after_test_forward_backward(__func__, high_level, shuffle, "result_after_forward_backward", subtest_ok, ntest, npass);
+    }
+
+    helper_free_ctx_data(cd);
+
+    return std::make_pair(npass, ntest);
+}
+
+static std::pair<int, int> test_epoch_vs_fit(ggml_backend_sched_t backend_sched, ggml_backend_t backend) {
+    int ntest = 0;
+    int npass = 0;
+
+    float weights_epoch;
+    float weights_fit;
+
+    {
+        struct helper_ctx_data cd = helper_get_ctx_data(backend_sched, backend, /*init_opt_ctx =*/ true);
+        ggml_opt_dataset_t dataset = cd.dataset_unsupervised;
+
+        ggml_opt_dataset_shuffle(cd.opt_ctx, dataset, -1);
+        ggml_opt_epoch(cd.opt_ctx, dataset, cd.result, nullptr, ndata, nullptr, nullptr);
+
+        ggml_backend_tensor_get(cd.weights, &weights_epoch, 0, ggml_nbytes(cd.weights));
+        helper_free_ctx_data(cd);
+    }
+    {
+        struct helper_ctx_data cd = helper_get_ctx_data(backend_sched, backend, /*init_opt_ctx =*/ false);
+        ggml_opt_dataset_t dataset = cd.dataset_unsupervised;
+
+        ggml_opt_fit(backend_sched, cd.ctx_compute, cd.inputs, cd.outputs, dataset,
+            GGML_OPT_LOSS_TYPE_SUM, ggml_opt_get_default_optimizer_params, 1, 1, 0.0f, true);
+
+        ggml_backend_tensor_get(cd.weights, &weights_fit, 0, ggml_nbytes(cd.weights));
+        helper_free_ctx_data(cd);
+    }
+
+    const bool subtest_ok = weights_epoch == weights_fit;
+
+    printf("  %s(): ", __func__);
+    if (subtest_ok) {
+        printf("\033[1;32mOK\033[0m\n");
+        npass++;
+    } else {
+        printf("\033[1;31mFAIL\033[0m\n");
+    }
+    ntest++;
+
+    return std::make_pair(npass, ntest);
+}
+
+static void helper_after_test_idata_split(
+        const char * func, const bool high_level, const int epoch,
+        const std::string subtest, const bool subtest_ok, int & ntest, int & npass) {
+    std::string options = ", epoch=";
+    options += std::to_string(epoch);
+    helper_after_test(func, high_level, options, subtest, subtest_ok, ntest, npass);
+}
+
+static std::pair<int, int> test_idata_split(ggml_backend_sched_t backend_sched, ggml_backend_t backend, const bool high_level) {
+    int ntest = 0;
+    int npass = 0;
+
+    struct helper_ctx_data cd = helper_get_ctx_data(backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false);
+    struct ggml_tensor * loss = ggml_opt_loss(cd.opt_ctx);
+    const int idata_split = ndata * 2/3;
+
+    std::vector<float> loss_history(ndata);
+    for (int64_t idata = 0; idata < ndata; ++idata) {
+        loss_history[idata] = NAN;
+    }
+
+    for (int epoch = 1; epoch <= 4; ++epoch) {
+        if (high_level) {
+            ggml_opt_epoch(cd.opt_ctx, cd.dataset_unsupervised, cd.result, cd.result2, idata_split, nullptr, nullptr);
+        } else {
+            int idata = 0;
+            for (; idata < idata_split; ++idata) {
+                const float idataf = idata;
+                ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
+                ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+                ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float));
+            }
+            for (; idata < ndata; ++idata) {
+                const float idataf = idata;
+                ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
+                ggml_opt_forward(cd.opt_ctx, cd.result2);
+                ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float));
+            }
+        }
+
+        {
+            float weights;
+            ggml_backend_tensor_get(cd.weights, &weights, 0, sizeof(float));
+            const bool subtest_ok = weights == ndata/2 - epoch*idata_split;
+            helper_after_test_idata_split(__func__, high_level, epoch, "weights", subtest_ok, ntest, npass);
+        }
+        {
+            int64_t ndata_result;
+            ggml_opt_result_ndata(cd.result, &ndata_result);
+            bool subtest_ok = ndata_result == idata_split;
+
+            double loss;
+            double loss_unc;
+            ggml_opt_result_loss(cd.result, &loss, &loss_unc);
+            subtest_ok = subtest_ok && loss == 28.0 - epoch*16.0 && loss_unc == 0.0;
+
+            double accuracy;
+            double accuracy_unc;
+            ggml_opt_result_accuracy(cd.result, &accuracy, &accuracy_unc);
+            subtest_ok = subtest_ok && std::isnan(accuracy) && std::isnan(accuracy_unc);
+
+            helper_after_test_idata_split(__func__, high_level, epoch, "results_backward", subtest_ok, ntest, npass);
+        }
+        {
+            int64_t ndata_result;
+            ggml_opt_result_ndata(cd.result2, &ndata_result);
+            bool subtest_ok = ndata_result == ndata - idata_split;
+
+            double loss;
+            double loss_unc;
+            ggml_opt_result_loss(cd.result2, &loss, &loss_unc);
+            subtest_ok = subtest_ok && loss == 15.0 - epoch*8 && almost_equal(loss_unc, sqrt(0.5), 1e-10);
+
+            double accuracy;
+            double accuracy_unc;
+            ggml_opt_result_accuracy(cd.result2, &accuracy, &accuracy_unc);
+            subtest_ok = subtest_ok && std::isnan(accuracy) && std::isnan(accuracy_unc);
+
+            helper_after_test_idata_split(__func__, high_level, epoch, "results_forward", subtest_ok, ntest, npass);
+        }
+
+        ggml_opt_result_reset(cd.result);
+        ggml_opt_result_reset(cd.result2);
+    }
+
+    helper_free_ctx_data(cd);
+
+    return std::make_pair(npass, ntest);
+}
+
+static void helper_after_test_gradient_accumulation(
+        const char * func, const int nbatch_physical, const enum ggml_opt_loss_type loss_type, const int epoch,
+        const std::string subtest, const bool subtest_ok, int & ntest, int & npass) {
+    std::string options = ", nbatch_physical=";
+    options += std::to_string(nbatch_physical);
+    options += ", loss_type=";
+    options += loss_type == GGML_OPT_LOSS_TYPE_MEAN ? "mean" : "sum";
+    options += ", epoch=";
+    options += std::to_string(epoch);
+    helper_after_test(func, false, options, subtest, subtest_ok, ntest, npass);
+}
+
+static std::pair<int, int> test_gradient_accumulation(
+        ggml_backend_sched_t backend_sched, ggml_backend_t backend, const int32_t nbatch_physical, const enum ggml_opt_loss_type loss_type) {
+    int ntest = 0;
+    int npass = 0;
+
+    struct helper_ctx_data cd = helper_get_ctx_data(
+        backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false, /*nbatch_logical =*/ 6, nbatch_physical, loss_type);
+    struct ggml_tensor * loss = ggml_opt_loss(cd.opt_ctx);
+
+    std::vector<float> grad_history(ndata);
+    for (int64_t idata = 0; idata < ndata; ++idata) {
+        grad_history[idata] = NAN;
+    }
+
+    for (int epoch = 1; epoch <= 4; ++epoch) {
+        if (nbatch_physical == 1) {
+            for (int idata = 0; idata < ndata; ++idata) {
+                const float idataf = idata;
+                ggml_backend_tensor_set(cd.inputs, &idataf, 0, 1*sizeof(float));
+                ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+                ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata, 0, 1*sizeof(float));
+            }
+        } else if (nbatch_physical == 2) {
+            for (int idata = 0; idata < ndata; idata += 2) {
+                const float idataf[2] = {float(idata + 0), float(idata + 1)};
+                ggml_backend_tensor_set(cd.inputs, idataf, 0, 2*sizeof(float));
+                ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+
+                grad_history[idata + 0] = 0.0f;
+                ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata + 1, 0, 1*sizeof(float));
+            }
+        } else {
+            GGML_ASSERT(false);
+        }
+
+        {
+            GGML_ASSERT(ndata == 6);
+            constexpr double atol = 1e-6;
+            bool subtest_ok = true;
+            if (loss_type == GGML_OPT_LOSS_TYPE_SUM) {
+                if (nbatch_physical == 1) {
+                    subtest_ok = subtest_ok && almost_equal(grad_history[0], 1.0, atol);
+                    subtest_ok = subtest_ok && almost_equal(grad_history[2], 3.0, atol);
+                    subtest_ok = subtest_ok && almost_equal(grad_history[4], 5.0, atol);
+                } else {
+                    subtest_ok = subtest_ok && almost_equal(grad_history[0], 0.0, atol);
+                    subtest_ok = subtest_ok && almost_equal(grad_history[2], 0.0, atol);
+                    subtest_ok = subtest_ok && almost_equal(grad_history[4], 0.0, atol);
                 }
+                subtest_ok = subtest_ok && almost_equal(grad_history[1], 2.0, atol);
+                subtest_ok = subtest_ok && almost_equal(grad_history[3], 4.0, atol);
+                subtest_ok = subtest_ok && almost_equal(grad_history[5], 0.0, atol);
+            } else if (loss_type == GGML_OPT_LOSS_TYPE_MEAN) {
+                if (nbatch_physical == 1) {
+                    subtest_ok = subtest_ok && almost_equal(grad_history[0], 1.0/ndata, atol);
+                    subtest_ok = subtest_ok && almost_equal(grad_history[2], 3.0/ndata, atol);
+                    subtest_ok = subtest_ok && almost_equal(grad_history[4], 5.0/ndata, atol);
+                } else {
+                    subtest_ok = subtest_ok && almost_equal(grad_history[0], 0.0/ndata, atol);
+                    subtest_ok = subtest_ok && almost_equal(grad_history[2], 0.0/ndata, atol);
+                    subtest_ok = subtest_ok && almost_equal(grad_history[4], 0.0/ndata, atol);
+                }
+                subtest_ok = subtest_ok && almost_equal(grad_history[1], 2.0/ndata, atol);
+                subtest_ok = subtest_ok && almost_equal(grad_history[3], 4.0/ndata, atol);
+                subtest_ok = subtest_ok && almost_equal(grad_history[5], 0.0/ndata, atol);
+            } else {
+                GGML_ASSERT(false);
+            }
+            helper_after_test_gradient_accumulation(__func__, nbatch_physical, loss_type, epoch, "grads", subtest_ok, ntest, npass);
+        }
+        {
+            float weights;
+            ggml_backend_tensor_get(cd.weights, &weights, 0, sizeof(float));
+            const bool subtest_ok = weights == (ndata/2) - epoch;
+            helper_after_test_gradient_accumulation(__func__, nbatch_physical, loss_type, epoch, "weights", subtest_ok, ntest, npass);
+        }
+        {
+            int64_t ndata_result;
+            ggml_opt_result_ndata(cd.result, &ndata_result);
+            bool subtest_ok = ndata_result == ndata/nbatch_physical;
+
+            double loss;
+            ggml_opt_result_loss(cd.result, &loss, /*loss_unc =*/ nullptr);
+            if (loss_type == GGML_OPT_LOSS_TYPE_SUM) {
+                subtest_ok = subtest_ok && loss == (39.0 - epoch*6.0);
+            } else if (loss_type == GGML_OPT_LOSS_TYPE_MEAN) {
+                subtest_ok = subtest_ok && almost_equal(loss, (39.0 - epoch*6.0) / ndata, 1e-6);
+            } else {
+                GGML_ASSERT(false);
             }
-            break;
-        default:
-            assert(false);
+
+            double accuracy;
+            double accuracy_unc;
+            ggml_opt_result_accuracy(cd.result, &accuracy, &accuracy_unc);
+            subtest_ok = subtest_ok && std::isnan(accuracy) && std::isnan(accuracy_unc);
+
+            helper_after_test_gradient_accumulation(__func__, nbatch_physical, loss_type, epoch, "results", subtest_ok, ntest, npass);
+        }
+
+        ggml_opt_result_reset(cd.result);
     }
 
+    helper_free_ctx_data(cd);
+
+    return std::make_pair(npass, ntest);
+}
+
+static ggml_opt_optimizer_params helper_get_regression_opt_pars(void * userdata) {
+    ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(userdata);
+    result.adamw.alpha = 0.1f;
     return result;
 }
 
-int main(void) {
-    struct ggml_init_params params = {
-        /* .mem_size   = */ 1024*1024*1024,
-        /* .mem_buffer = */ NULL,
-        /* .no_alloc   = */ false,
-    };
+static std::pair<int, int> test_regression(ggml_backend_sched_t backend_sched, ggml_backend_t backend) {
+    int ntest = 0;
+    int npass = 0;
 
-    struct ggml_context * ctx = ggml_init(params);
+    // Test for simple regression with f(x) = a*x + b
 
-    int64_t ne1[4] = {4, 128, 1, 1};
-    int64_t ne2[4] = {4, 256, 1, 1};
-    int64_t ne3[4] = {128, 256, 1, 1};
+    constexpr int64_t ndata_regression = 201;
+    constexpr float a_true = 1.2f;
+    constexpr float b_true = 3.4f;
 
-    struct ggml_tensor * a = get_random_tensor(ctx, 2, ne1, -1, +1);
-    struct ggml_tensor * b = get_random_tensor(ctx, 2, ne2, -1, +1);
-    ggml_set_param(ctx, a);
-    ggml_set_param(ctx, b);
+    std::mt19937 gen(12345);
+    std::normal_distribution<float> nd{0.0f, 0.1f};
 
-    struct ggml_tensor * c = get_random_tensor(ctx, 2, ne3, -1, +1);
+    ggml_opt_dataset_t dataset = ggml_opt_dataset_init(1, 1, ndata_regression, ndata_regression);
 
-    struct ggml_tensor * ab = ggml_mul_mat(ctx, a, b);
-    struct ggml_tensor * d  = ggml_sub(ctx, c, ab);
-    struct ggml_tensor * e  = ggml_sum(ctx, ggml_sqr(ctx, d));
+    float * data   = ggml_get_data_f32(ggml_opt_dataset_data(  dataset));
+    float * labels = ggml_get_data_f32(ggml_opt_dataset_labels(dataset));
 
-    struct ggml_cgraph * ge = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, true);
-    ggml_build_forward_expand(ge, e);
-    ggml_graph_reset(ge);
+    constexpr float x_min = -100.0f;
+    constexpr float x_max =  100.0f;
 
-    ggml_graph_compute_with_ctx(ctx, ge, /*n_threads*/ 1);
+    for (int64_t idata = 0; idata < ndata_regression; ++idata) {
+        const float x = x_min + (x_max - x_min) * idata/(ndata_regression-1);
+        const float y = a_true*x + b_true + nd(gen);
 
-    const float fe = ggml_get_f32_1d(e, 0);
-    printf("%s: e = %.4f\n", __func__, fe);
+        data[idata]   = x;
+        labels[idata] = y;
+    }
 
-    struct ggml_opt_params opt_params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
+    struct ggml_context * ctx_static;
+    struct ggml_context * ctx_compute;
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ 3*ggml_tensor_overhead(),
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        ctx_static = ggml_init(params);
+    }
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ GGML_DEFAULT_GRAPH_SIZE*ggml_tensor_overhead() + 3*ggml_graph_overhead(),
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        ctx_compute = ggml_init(params);
+    }
 
-    ggml_opt(ctx, opt_params, e);
+    // The first dimension is the dimension of the datapoints, the second dimension is the number of datapoints.
+    struct ggml_tensor * x = ggml_new_tensor_2d(ctx_static, GGML_TYPE_F32, 1, ndata_regression);
+    ggml_set_name(x, "x");
+
+    struct ggml_tensor * a = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, 1);
+    ggml_set_name(a, "a");
+    ggml_set_param(ctx_static, a);
+
+    struct ggml_tensor * b = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, 1);
+    ggml_set_name(b, "b");
+    ggml_set_param(ctx_static, b);
+
+    struct ggml_tensor * f = ggml_add(ctx_compute, ggml_mul(ctx_compute, x, a), b);
+    ggml_set_name(f, "f");
+    ggml_set_param(ctx_static, f);
+
+    ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx_static, backend);
+    const float a0 = 1.0f;
+    const float b0 = 3.0f;
+    ggml_backend_tensor_set(a, &a0, 0, sizeof(float));
+    ggml_backend_tensor_set(b, &b0, 0, sizeof(float));
+
+    ggml_opt_fit(backend_sched, ctx_compute, x, f, dataset, GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR,
+        helper_get_regression_opt_pars, 100, ndata_regression, 0.0f, true);
+
+    {
+        float a_fit;
+        ggml_backend_tensor_get(a, &a_fit, 0, sizeof(float));
+        float b_fit;
+        ggml_backend_tensor_get(b, &b_fit, 0, sizeof(float));
+        const bool subtest_ok = almost_equal(a_fit, a_true, 1e-2) && almost_equal(b_fit, b_true, 1e-2);
+        printf("  %s(subtest=weights): ", __func__);
+        if (subtest_ok) {
+            printf("\033[1;32mOK\033[0m\n");
+            npass++;
+        } else {
+            printf("\033[1;31mFAIL\033[0m\n");
+        }
+        ntest++;
+    }
 
-    ggml_graph_reset(ge);
+    ggml_backend_buffer_free(buf);
+    ggml_free(ctx_static);
+    ggml_opt_dataset_free(dataset);
 
-    ggml_graph_compute_with_ctx(ctx, ge, /*n_threads*/ 1);
+    return std::make_pair(npass, ntest);
+}
 
-    const float fe_opt = ggml_get_f32_1d(e, 0);
-    printf("%s: original  e = %.4f\n", __func__, fe);
-    printf("%s: optimized e = %.4f\n", __func__, fe_opt);
+static std::pair<int, int> test_backend(ggml_backend_sched_t backend_sched, ggml_backend_t backend) {
+    int npass = 0;
+    int ntest = 0;
 
-    const bool success = (fe_opt <= fe);
-    assert(success);
+    for (bool shuffle : {false, true}) {
+        std::pair<int, int> partial = test_dataset(backend_sched, backend, shuffle);
+        npass += partial.first;
+        ntest += partial.second;
+    }
+    {
+        std::pair<int, int> partial = test_grad(backend_sched, backend);
+        npass += partial.first;
+        ntest += partial.second;
+    }
+    for (bool high_level : {false, true}){
+        for (bool shuffle : {false, true}) {
+            if (!high_level && shuffle) {
+                continue;
+            }
 
-    ggml_free(ctx);
-    return success ? 0 : -1;
+            std::pair<int, int> partial = test_forward_backward(backend_sched, backend, high_level, shuffle);
+            npass += partial.first;
+            ntest += partial.second;
+        }
+    }
+    {
+        std::pair<int, int> partial = test_epoch_vs_fit(backend_sched, backend);
+        npass += partial.first;
+        ntest += partial.second;
+    }
+    for (bool high_level : {false, true}){
+        std::pair<int, int> partial = test_idata_split(backend_sched, backend, high_level);
+        npass += partial.first;
+        ntest += partial.second;
+    }
+    for (int32_t nbatch_physical : {2, 1}) {
+        for (enum ggml_opt_loss_type loss_type : {GGML_OPT_LOSS_TYPE_SUM, GGML_OPT_LOSS_TYPE_MEAN}) {
+            std::pair<int, int> partial = test_gradient_accumulation(backend_sched, backend, nbatch_physical, loss_type);
+            npass += partial.first;
+            ntest += partial.second;
+        }
+    }
+    {
+        std::pair<int, int> partial = test_regression(backend_sched, backend);
+        npass += partial.first;
+        ntest += partial.second;
+    }
+
+    return std::make_pair(npass, ntest);
 }
-// int64_t ne1[4] = {4, 128, 1, 1};
-// int64_t ne2[4] = {4, 256, 1, 1};;
-// int64_t ne3[4] = {128, 256, 1, 1};
-// main: original  e = 25890.9375
-// main: optimized e = 10094.7031
 
-// int64_t ne1[4] = {8, 128, 1, 1};
-// int64_t ne2[4] = {8, 256, 1, 1};;
-// int64_t ne3[4] = {128, 256, 1, 1};
-// main: original  e = 39429.5078
-// main: optimized e = 9275.8936
+int main(void) {
+    const size_t dev_count = ggml_backend_dev_count();
+    printf("Testing %zu devices\n\n", dev_count);
+    size_t n_ok = 0;
+
+    std::vector<ggml_backend_dev_t> devs;
+    std::vector<ggml_backend_t>     backends;
 
-// int64_t ne1[4] = {16, 128, 1, 1};
-// int64_t ne2[4] = {16, 256, 1, 1};;
-// int64_t ne3[4] = {128, 256, 1, 1};
-// main: original  e = 68371.1328
-// main: optimized e = 7854.4502
+    for (size_t i = 0; i < dev_count; ++i) {
+        devs.push_back(ggml_backend_dev_get(i));
 
+        ggml_backend_t backend = ggml_backend_dev_init(devs[i], NULL);
+        GGML_ASSERT(backend != NULL);
 
-// int64_t ne1[4] = {32, 128, 1, 1};
-// int64_t ne2[4] = {32, 256, 1, 1};;
-// int64_t ne3[4] = {128, 256, 1, 1};
-// main: original  e = 126061.1953
-// main: optimized e = 5451.0166
+        if (ggml_backend_is_cpu(backend)) {
+            ggml_backend_cpu_set_n_threads(backend, std::thread::hardware_concurrency() / 2);
+        }
+
+        backends.push_back(backend);
+    }
 
-// int64_t ne1[4] = {4, 1024, 1, 1};
-// int64_t ne2[4] = {4, 2048, 1, 1};;
-// int64_t ne3[4] = {1024, 2048, 1, 1};
-// main: original  e = 1620817.8750
-// main: optimized e = 698387.6875
+    for (size_t i = 0; i < dev_count; ++i) {
+        // Put the backend to be tested in front so that it's prioritized:
+        std::vector<ggml_backend_t> backends_modded = {backends[i]};
+        backends_modded.insert(backends_modded.end(), backends.begin(), backends.end());
 
-// another run on M1
-// int64_t ne1[4] = {4, 1024, 1, 1};
-// int64_t ne2[4] = {4, 2048, 1, 1};;
-// int64_t ne3[4] = {1024, 2048, 1, 1};
-// main: original  e = 1629595.6250
-// main: optimized e = 698169.1250
+        ggml_backend_sched_t backend_sched = ggml_backend_sched_new(
+            backends_modded.data(), nullptr, backends_modded.size(), GGML_DEFAULT_GRAPH_SIZE, false);
 
-// int64_t ne1[4] = {32, 1024, 1, 1};
-// int64_t ne2[4] = {32, 2048, 1, 1};;
-// int64_t ne3[4] = {1024, 2048, 1, 1};
-// main: original  e = 8146770.5000
-// main: optimized e = 651119.1250
+        printf("Backend %zu/%zu: %s\n", i + 1, dev_count, ggml_backend_dev_name(devs[i]));
+        printf("  Device description: %s\n", ggml_backend_dev_description(devs[i]));
+        size_t free, total; // NOLINT
+        ggml_backend_dev_memory(devs[i], &free, &total);
+        printf("  Device memory: %zu MB (%zu MB free)\n", total / 1024 / 1024, free / 1024 / 1024);
+        printf("\n");
+
+        std::pair<int, int> result = test_backend(backend_sched, backends[i]);
+
+        printf("  %d/%d tests passed\n", result.first, result.second);
+        printf("  Backend %s: ", ggml_backend_name(backends[i]));
+        if (result.first == result.second) {
+            printf("\033[1;32mOK\033[0m\n");
+            n_ok++;
+        } else {
+            printf("\033[1;31mFAIL\033[0m\n");
+        }
+
+        printf("\n");
+
+        ggml_backend_sched_free(backend_sched);
+    }
+
+    for (ggml_backend_t backend : backends) {
+        ggml_backend_free(backend);
+    }
+
+    printf("%zu/%zu backends passed\n", n_ok, dev_count);
+    if (n_ok != dev_count) {
+        printf("\033[1;31mFAIL\033[0m\n");
+        return 1;
+    }
+    printf("\033[1;32mOK\033[0m\n");
+    return 0;
+}

From 68fcb4759c0ca2874b59d9c1e6a35ceca1cc04ce Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sat, 16 Nov 2024 21:32:41 +0200
Subject: [PATCH 009/102] ggml : fix compile warnings (#0)

ggml-ci
---
 ggml/src/ggml.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 4a478fcaa2064..ee72a173e07fd 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -5501,7 +5501,7 @@ static void ggml_compute_backward(
                     fprintf(stderr, "%s: unsupported unary op for backward pass: %s\n",
                         __func__, ggml_unary_op_name(ggml_get_unary_op(tensor)));
                     GGML_ABORT("fatal error");
-                } break;
+                } //break;
             }
         } break;
         case GGML_OP_CROSS_ENTROPY_LOSS: {
@@ -5517,7 +5517,7 @@ static void ggml_compute_backward(
         default: {
             fprintf(stderr, "%s: unsupported ggml op for backward pass: %s\n", __func__, ggml_op_name(tensor->op));
             GGML_ABORT("fatal error");
-        } break;
+        } //break;
     }
 
     GGML_ASSERT(!src0_needs_grads || ggml_are_same_shape(src0, cgraph->grads[isrc0]));

From 84274a10c399d843cff39feb2b86dc8224f613d8 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sat, 16 Nov 2024 21:34:03 +0200
Subject: [PATCH 010/102] tests : remove test-grad0

---
 Makefile             |    6 -
 tests/CMakeLists.txt |    3 +-
 tests/test-grad0.cpp | 1684 ------------------------------------------
 3 files changed, 1 insertion(+), 1692 deletions(-)
 delete mode 100644 tests/test-grad0.cpp

diff --git a/Makefile b/Makefile
index 539370e0639f4..ae9a9c6323a01 100644
--- a/Makefile
+++ b/Makefile
@@ -48,7 +48,6 @@ TEST_TARGETS = \
 	tests/test-backend-ops \
 	tests/test-chat-template \
 	tests/test-double-float \
-	tests/test-grad0 \
 	tests/test-grammar-integration \
 	tests/test-grammar-parser \
 	tests/test-json-schema-to-grammar \
@@ -1499,11 +1498,6 @@ tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp \
 	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-grad0: tests/test-grad0.cpp \
-	$(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
 tests/test-opt: tests/test-opt.cpp \
 	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 08ad66b49fdd4..b06f122e89873 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -116,9 +116,8 @@ llama_target_and_test(test-sampling.cpp)
 llama_target_and_test(test-chat-template.cpp)
 
 llama_target_and_test(test-grammar-parser.cpp)
-llama_target_and_test(test-llama-grammar.cpp)
 llama_target_and_test(test-grammar-integration.cpp)
-llama_target_and_test(test-grad0.cpp)
+llama_target_and_test(test-llama-grammar.cpp)
 llama_target_and_test(test-barrier.cpp)
 # llama_target_and_test(test-opt.cpp) # SLOW
 llama_target_and_test(test-backend-ops.cpp)
diff --git a/tests/test-grad0.cpp b/tests/test-grad0.cpp
deleted file mode 100644
index c712dba7fc8d4..0000000000000
--- a/tests/test-grad0.cpp
+++ /dev/null
@@ -1,1684 +0,0 @@
-#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnings on Windows
-#include "ggml.h"
-#include "ggml-cpu.h"
-
-#include <cfloat>
-#include <cmath>
-#include <cstdint>
-#include <cstdio>
-#include <cstdlib>
-#include <cassert>
-#include <initializer_list>
-#include <vector>
-
-#if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
-#endif
-
-#if defined(__GNUC__)
-#pragma GCC diagnostic ignored "-Wdouble-promotion"
-#endif
-
-#define MAX_NARGS 3
-
-#undef MIN
-#undef MAX
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-
-#define GGML_SILU_FP16
-
-//
-// logging
-//
-
-#if (GGML_DEBUG >= 1)
-#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG(...)
-#endif
-
-#if (GGML_DEBUG >= 5)
-#define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG_5(...)
-#endif
-
-#if (GGML_DEBUG >= 10)
-#define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG_10(...)
-#endif
-
-#define GGML_PRINT(...) printf(__VA_ARGS__)
-
-static float frand(void) {
-    return (float)rand()/(float)RAND_MAX;
-}
-
-static int irand(int n) {
-    if (n == 0) return 0;
-    return rand()%n;
-}
-
-static void get_random_dims(int64_t * dims, int ndims) {
-    dims[0] = dims[1] = dims[2] = dims[3] = 1;
-
-    for (int i = 0; i < ndims; i++) {
-        dims[i] = 1 + irand(4);
-    }
-}
-
-static struct ggml_tensor * get_random_tensor_f32(
-        struct ggml_context * ctx0,
-        int ndims,
-        int64_t ne[],
-        float fmin,
-        float fmax) {
-    struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F32, ndims, ne);
-
-    switch (ndims) {
-        case 1:
-            for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((float *)result->data)[i0] = frand()*(fmax - fmin) + fmin;
-            }
-            break;
-        case 2:
-            for (int i1 = 0; i1 < ne[1]; i1++) {
-                for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((float *)result->data)[i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
-                }
-            }
-            break;
-        case 3:
-            for (int i2 = 0; i2 < ne[2]; i2++) {
-                for (int i1 = 0; i1 < ne[1]; i1++) {
-                    for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((float *)result->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
-                    }
-                }
-            }
-            break;
-        case 4:
-            for (int i3 = 0; i3 < ne[3]; i3++) {
-                for (int i2 = 0; i2 < ne[2]; i2++) {
-                    for (int i1 = 0; i1 < ne[1]; i1++) {
-                        for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((float *)result->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
-                        }
-                    }
-                }
-            }
-            break;
-        default:
-            assert(false);
-    }
-
-    return result;
-}
-
-static struct ggml_tensor * get_random_tensor_f16(
-        struct ggml_context * ctx0,
-        int ndims,
-        int64_t ne[],
-        float fmin,
-        float fmax) {
-    struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F16, ndims, ne);
-
-    switch (ndims) {
-        case 1:
-            for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((ggml_fp16_t *)result->data)[i0] = ggml_fp32_to_fp16(frand()*(fmax - fmin) + fmin);
-            }
-            break;
-        case 2:
-            for (int i1 = 0; i1 < ne[1]; i1++) {
-                for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((ggml_fp16_t *)result->data)[i1*ne[0] + i0] = ggml_fp32_to_fp16(frand()*(fmax - fmin) + fmin);
-                }
-            }
-            break;
-        case 3:
-            for (int i2 = 0; i2 < ne[2]; i2++) {
-                for (int i1 = 0; i1 < ne[1]; i1++) {
-                    for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((ggml_fp16_t *)result->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = ggml_fp32_to_fp16(frand()*(fmax - fmin) + fmin);
-                    }
-                }
-            }
-            break;
-        case 4:
-            for (int i3 = 0; i3 < ne[3]; i3++) {
-                for (int i2 = 0; i2 < ne[2]; i2++) {
-                    for (int i1 = 0; i1 < ne[1]; i1++) {
-                        for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((ggml_fp16_t *)result->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = ggml_fp32_to_fp16(frand()*(fmax - fmin) + fmin);
-                        }
-                    }
-                }
-            }
-            break;
-        default:
-            assert(false);
-    }
-
-    return result;
-}
-
-static struct ggml_tensor * get_random_tensor_i32(
-        struct ggml_context * ctx0,
-        int ndims,
-        int64_t ne[],
-        int32_t imin,
-        int32_t imax) {
-    struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_I32, ndims, ne);
-
-    switch (ndims) {
-        case 1:
-            for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((int32_t *)result->data)[i0] = irand(imax - imin) + imin;
-            }
-            break;
-        case 2:
-            for (int i1 = 0; i1 < ne[1]; i1++) {
-                for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((int32_t *)result->data)[i1*ne[0] + i0] = irand(imax - imin) + imin;
-                }
-            }
-            break;
-        case 3:
-            for (int i2 = 0; i2 < ne[2]; i2++) {
-                for (int i1 = 0; i1 < ne[1]; i1++) {
-                    for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((int32_t *)result->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = irand(imax - imin) + imin;
-                    }
-                }
-            }
-            break;
-        case 4:
-            for (int i3 = 0; i3 < ne[3]; i3++) {
-                for (int i2 = 0; i2 < ne[2]; i2++) {
-                    for (int i1 = 0; i1 < ne[1]; i1++) {
-                        for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((int32_t *)result->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = irand(imax - imin) + imin;
-                        }
-                    }
-                }
-            }
-            break;
-        default:
-            assert(false);
-    }
-
-    return result;
-}
-
-static bool check_gradient(
-        const char * op_name,
-        struct ggml_context * ctx0,
-        struct ggml_tensor * x[],
-        struct ggml_tensor * f,
-        int ndims,
-        int nargs,
-        float eps,
-        float max_error_abs,
-        float max_error_rel,
-        std::vector<double> expected_vals) {
-
-    static int n_threads = -1;
-    if (n_threads < 0) {
-        n_threads = GGML_DEFAULT_N_THREADS;
-
-        const char *env = getenv("GGML_N_THREADS");
-        if (env) {
-            n_threads = atoi(env);
-        }
-
-        printf("GGML_N_THREADS = %d\n", n_threads);
-    }
-
-    struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, GGML_DEFAULT_GRAPH_SIZE, true);
-    struct ggml_cgraph * gb = ggml_new_graph_custom(ctx0, GGML_DEFAULT_GRAPH_SIZE, true);
-    ggml_build_forward_expand(gf, f);
-    ggml_graph_cpy(gf, gb);
-    ggml_build_backward_expand(ctx0, gf, gb, false);
-
-    ggml_graph_compute_with_ctx(ctx0, gf, n_threads);
-
-    ggml_graph_reset(gb);
-    if (f->grad) {
-        ggml_set_f32(f->grad, 1.0f);
-    }
-
-    ggml_graph_compute_with_ctx(ctx0, gb, n_threads);
-
-    // ggml_graph_dump_dot(gf, NULL, "test-grad0-forward.dot");
-    // ggml_graph_dump_dot(gb, gf,  "test-grad0-backward.dot");
-
-    for (int i = 0; i < nargs; ++i) {
-        bool all_g0_bad = true;
-        const int nelements = ggml_nelements(x[i]);
-        for (int k = 0; k < nelements; ++k) {
-            // Calculate gradient numerically:
-            const float x0 = ggml_get_f32_1d(x[i], k);
-            const float xm = x0 - eps;
-            const float xp = x0 + eps;
-            ggml_set_f32_1d(x[i], k, xp);
-
-            ggml_graph_compute_with_ctx(ctx0, gf, n_threads);
-
-            const double f0 = ggml_get_f32_1d(f, 0);
-
-            ggml_set_f32_1d(x[i], k, xm);
-
-            ggml_graph_compute_with_ctx(ctx0, gf, n_threads);
-
-            const double f1 = ggml_get_f32_1d(f, 0);
-            const double g0 = (f0 - f1)/(2.0*(double) eps);
-
-            // The numerical calculation of the gradient fails around noncontinuities (e.g. 0 for ReLU).
-            // In such cases, provide a vector of expected values and skip the comparison for failed calculations.
-            if (!expected_vals.empty()) {
-                bool matches_any = false;
-                for (const double & ev : expected_vals) {
-                    const double error_abs = std::fabs(g0 - ev);
-                    if (error_abs > max_error_abs) {
-                        continue;
-                    }
-                    const double error_rel = g0 != 0.0 ? fabs(g0 - ev)/fabs(g0) : 0.0;
-                    if (error_rel > max_error_rel) {
-                        continue;
-                    }
-                    matches_any = true;
-                    break;
-                }
-                if (!matches_any) {
-                    continue;
-                }
-            }
-            all_g0_bad = false;
-
-            ggml_set_f32_1d(x[i], k, x0);
-
-            // compute gradient using backward graph
-            ggml_graph_reset(gb);
-            if (f->grad) {
-                ggml_set_f32(f->grad, 1.0f);
-            }
-
-            ggml_graph_compute_with_ctx(ctx0, gb, n_threads);
-
-            const double g1 = ggml_get_f32_1d(x[i]->grad, k);
-
-            const double error_abs = fabs(g0 - g1);
-            const double error_rel = g0 != 0.0 ? fabs(g0 - g1)/fabs(g0) : 0.0;
-
-            if (error_abs > max_error_abs || error_rel > max_error_rel) {
-                printf("%s: ndims=%d, i=%d, k=%d, x0=%f, xm=%f, xp=%f, f0=%f, f1=%f, g0=%f, g1=%f, eps=%f, error_abs=%f, error_rel=%f\n",
-                            op_name, ndims, i, k, x0, xm, xp, f0, f1, g0, g1, eps, error_abs, error_rel);
-                //assert(false);
-                return false;
-            }
-        }
-        if (all_g0_bad) {
-            printf("%s: numerical calculation of the gradient failed for all values\n", op_name);
-            return false;
-        }
-    }
-
-    return true;
-}
-
-// TODO: clean-up this ..
-static bool check_mat_mul(
-        const struct ggml_tensor * y,
-        const struct ggml_tensor * x0,
-        const struct ggml_tensor * x1) {
-    float * dst  = (float *) y->data;
-    float * src0 = (float *) x0->data;
-    float * src1 = (float *) x1->data;
-
-    const int nc = x0->ne[1];
-    const int nr = x1->ne[1];
-    const int nk = x0->ne[0];
-
-    GGML_PRINT_DEBUG("check_mat_mul: nc=%d, nr=%d, nk=%d\n", nc, nr, nk);
-
-    GGML_PRINT_DEBUG("x0:\n");
-    for (int j = 0; j < x0->ne[1]; ++j) {
-        for (int i = 0; i < x0->ne[0]; ++i) {
-            GGML_PRINT_DEBUG("%6.3f ", src0[j*nk + i]);
-        }
-        GGML_PRINT_DEBUG("\n");
-    }
-    GGML_PRINT_DEBUG("\n");
-
-    GGML_PRINT_DEBUG("x1:\n");
-    for (int j = 0; j < x1->ne[1]; ++j) {
-        for (int i = 0; i < x1->ne[0]; ++i) {
-            GGML_PRINT_DEBUG("%6.3f ", src1[j*nk + i]);
-        }
-        GGML_PRINT_DEBUG("\n");
-    }
-    GGML_PRINT_DEBUG("\n");
-
-    GGML_PRINT_DEBUG("y: n_dims = %d, (%lld, %lld)\n", y->n_dims, y->ne[0], y->ne[1]);
-    for (int j = 0; j < y->ne[1]; ++j) {
-        for (int i = 0; i < y->ne[0]; ++i) {
-            GGML_PRINT_DEBUG("%6.3f ", dst[j*nr + i]);
-        }
-        GGML_PRINT_DEBUG("\n");
-    }
-
-    for (int i = 0; i < nr; ++i) {
-        for (int j = 0; j < nc; ++j) {
-            float sum = 0.0f;
-
-            for (int k = 0; k < nk; ++k) {
-                sum += src0[j*nk + k]*src1[i*nk + k];
-            }
-
-            if (fabsf(dst[i*nc + j] - sum) > 1e-5f) {
-                fprintf(stderr, "check_mat_mul: dst[%d] = %f, sum = %f\n", i*nc + j, dst[i*nc + j], sum);
-                assert(false);
-                return false;
-            }
-        }
-    }
-
-    return true;
-}
-
-#define NUM_PERMUTATIONS (4*3*2*1)
-
-int main(int argc, const char ** argv) {
-    struct ggml_init_params params = {
-        /* .mem_size   = */ 256*1024*1024,
-        /* .mem_buffer = */ NULL,
-        /* .no_alloc   = */ false,
-    };
-
-    int64_t ne[4];
-
-    int all_permutations[4 * NUM_PERMUTATIONS];
-    {
-        int count = 0;
-        for (int ax0=0; ax0<4; ++ax0) {
-            for (int ax1=0; ax1<4; ++ax1) {
-                if (ax1 == ax0) continue;
-                for (int ax2=0; ax2<4; ++ax2) {
-                    if (ax2 == ax0) continue;
-                    if (ax2 == ax1) continue;
-                    for (int ax3=0; ax3<4; ++ax3) {
-                        if (ax3 == ax0) continue;
-                        if (ax3 == ax1) continue;
-                        if (ax3 == ax2) continue;
-                        assert(count < NUM_PERMUTATIONS);
-                        all_permutations[count*4+0] = ax0;
-                        all_permutations[count*4+1] = ax1;
-                        all_permutations[count*4+2] = ax2;
-                        all_permutations[count*4+3] = ax3;
-                        ++count;
-                    }
-                }
-            }
-        }
-    }
-
-    unsigned seed_iter = 1;
-
-    // original loop: 1000
-    int niter = 4;
-    const char *env = getenv("GGML_NLOOP");
-    if (env != NULL) {
-        niter = atoi(env);
-    }
-    if (argc > 1) {
-        niter = atoi(argv[1]);
-    }
-    for (int iter = 0; iter < niter; ++iter) {
-        srand(seed_iter);
-        seed_iter = rand();
-        unsigned seed = rand();
-
-        printf("test-grad0: iter:%d/%d\n", (iter+1), niter);
-        struct ggml_context * ctx0 = ggml_init(params);
-
-        get_random_dims(ne, 4);
-
-        struct ggml_tensor * x[MAX_NARGS];
-
-        // add f32
-        {
-            srand(seed);
-            const int nargs = 2;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_add(ctx0, x[0], x[1]));
-
-                check_gradient("add f32", ctx0, x, f, ndims, nargs, 1e-3f, 2e-3f, 2e-3f, {});
-            }
-        }
-
-        // add f16
-        {
-            srand(seed);
-            const int nargs = 2;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f16(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_add(ctx0, x[0], x[1]));
-
-                check_gradient("add f16", ctx0, x, f, ndims, nargs, 1e-1f, 2e-1f, 2e-1f, {});
-            }
-        }
-
-        // sub
-        {
-            srand(seed);
-            const int nargs = 2;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_sub(ctx0, x[0], x[1]));
-
-                check_gradient("sub", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
-            }
-        }
-
-        // mul
-        {
-            srand(seed);
-            const int nargs = 2;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_mul(ctx0, x[0], x[1]));
-
-                check_gradient("mul", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // div
-        {
-            srand(seed);
-            const int nargs = 2;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, 0.5f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_div(ctx0, x[0], x[1]));
-
-                check_gradient("div", ctx0, x, f, ndims, nargs, 1e-3f, 1e-1f, 1e-1f, {});
-            }
-        }
-
-        // sqr
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqr(ctx0, x[0]));
-
-                check_gradient("sqr", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // sqrt
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, 2.0f*1e-3f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqrt(ctx0, x[0]));
-
-                check_gradient("sqrt", ctx0, x, f, ndims, nargs, 1e-3f, 2e-2f, 1e-1f, {});
-            }
-        }
-
-        // log
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, 2.0f*1e-3f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_log(ctx0, x[0]));
-
-                check_gradient("log", ctx0, x, f, ndims, nargs, 1e-3f, INFINITY, 1e-1f, {});
-            }
-        }
-
-        // sum
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, x[0]);
-
-                check_gradient("sum", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
-            }
-        }
-
-
-        // sum_rows
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqr(ctx0, ggml_sum_rows(ctx0, x[0])));
-
-                check_gradient("sum_rows", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY, {});
-            }
-        }
-
-        // mean, not yet fully implemented
-        if(0)
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_mean(ctx0, x[0]));
-
-                check_gradient("mean", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
-            }
-        }
-
-        // argmax
-        if (0)
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_argmax(ctx0, x[0]));
-
-                check_gradient("argmax", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
-            }
-        }
-
-        // repeat
-        {
-            srand(seed);
-            int64_t ne2[4];
-            get_random_dims(ne2, 4);
-
-            ne2[0] = ne[0] * ne2[0];
-            ne2[1] = ne[1] * ne2[1];
-            ne2[2] = 1;
-            ne2[3] = 1;
-
-            const int nargs = 1;
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[0]);
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x[1], ggml_repeat(ctx0, x[0], x[1]))));
-
-                check_gradient("repeat", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY, {});
-            }
-        }
-
-        // repeat back
-        {
-            srand(seed);
-            int64_t ne2[4];
-            get_random_dims(ne2, 4);
-
-            ne2[0] = ne[0] * ne2[0];
-            ne2[1] = ne[1] * ne2[1];
-            ne2[2] = 1;
-            ne2[3] = 1;
-
-            const int nargs = 1;
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[0]);
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x[0], ggml_repeat_back(ctx0, x[1], x[0]))));
-
-                check_gradient("repeat back", ctx0, x, f, ndims, nargs, 1e-3f, 1e-2f, INFINITY, {});
-            }
-        }
-
-        // abs
-        {
-           const int nargs = 1;
-
-           for (int ndims = 1; ndims <= 4; ++ndims) {
-               for (int i = 0; i < nargs; ++i) {
-                   x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                   ggml_set_param(ctx0, x[i]);
-               }
-
-               struct ggml_tensor * f = ggml_sum(ctx0, ggml_abs(ctx0, x[0]));
-
-               check_gradient("abs", ctx0, x, f, ndims, nargs, 1e-3f, INFINITY, 1e-3f, {-1.0, 1.0});
-           }
-        }
-
-        // sgn
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor* f = ggml_sum(ctx0, ggml_sgn(ctx0, x[0]));
-
-                check_gradient("sgn", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {0.0});
-            }
-        }
-
-        // neg
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor* f = ggml_sum(ctx0, ggml_neg(ctx0, x[0]));
-
-                check_gradient("neg", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
-            }
-        }
-
-        // step
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor* f = ggml_sum(ctx0, ggml_step(ctx0, x[0]));
-
-                check_gradient("step", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {0.0});
-            }
-        }
-
-        // tanh, not yet fully implemented
-        if(0)
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor* f = ggml_sum(ctx0, ggml_tanh(ctx0, x[0]));
-
-                check_gradient("tanh", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
-            }
-        }
-
-        // mul_mat
-        {
-            srand(seed);
-            const int nargs = 2;
-
-            for (int ndims = 2; ndims <= 4; ++ndims) {
-                int max_nrep = (ndims >= 3) ? 2 : 1;
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                for (int nrep2 = 1; nrep2 < max_nrep; ++nrep2) {
-                    for (int nrep3 = 1; nrep3 < max_nrep; ++nrep3) {
-                        {
-                            int64_t ne2[4];
-                            get_random_dims(ne2, 4);
-                            ne2[0] = ne[0];
-                            ne2[2] = nrep2 * ne[2];
-                            ne2[3] = nrep3 * ne[3];
-                            x[1] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
-                        }
-
-                        ggml_set_param(ctx0, x[0]);
-                        ggml_set_param(ctx0, x[1]);
-
-                        struct ggml_tensor * m = ggml_mul_mat(ctx0, x[1], x[0]);
-                        struct ggml_tensor * f = ggml_sum(ctx0, m);
-
-                        GGML_PRINT_DEBUG("testing: mul_mat, [%lld, %lld] (%d) * [%lld, %lld] (%d)\n", x[1]->ne[0], x[1]->ne[1], x[1]->n_dims, x[0]->ne[0], x[0]->ne[1], x[0]->n_dims);
-
-                        check_gradient("mul_mat", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-                        if (ndims == 2) {
-                            // check_mat_mul does not support ndims > 2
-                            check_mat_mul(m, x[1], x[0]);
-                        }
-                    }
-                }
-            }
-        }
-
-        // elu, not yet fully implemented
-        if(0)
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor* f = ggml_sum(ctx0, ggml_elu(ctx0, x[0]));
-
-                check_gradient("elu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
-            }
-        }
-
-        // relu
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor* f = ggml_sum(ctx0, ggml_relu(ctx0, x[0]));
-
-                check_gradient("relu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {0.0, 1.0});
-            }
-        }
-
-        // gelu, not yet fully implemented
-        if(0)
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor* f = ggml_sum(ctx0, ggml_gelu(ctx0, x[0]));
-
-                check_gradient("gelu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, 1e-3f, {});
-            }
-        }
-
-        // silu
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_silu(ctx0, x[0]));
-
-#ifdef GGML_SILU_FP16
-                // due to GGML_SILU_FP16 the finite difference method will be slightly wrong -> increase error bounds.
-                check_gradient("silu", ctx0, x, f, ndims, nargs, 1e-3f, 0.5, INFINITY, {});
-#else
-                check_gradient("silu", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-#endif
-            }
-        }
-
-        // rms_norm
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_rms_norm(ctx0, x[0], 1e-6f));
-
-                check_gradient("rms_norm", ctx0, x, f, ndims, nargs, 1e-4f, 1.0f, INFINITY, {});
-            }
-        }
-
-        // scale
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-
-                const float s = -1.0f + 2.0f*frand();
-
-                ggml_set_param(ctx0, x[0]);
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_scale(ctx0, x[0], s));
-
-                check_gradient("scale", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // cpy f32
-        {
-            srand(seed);
-            const int nargs = 2;
-
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-                // x[1] is overwritten by x[0], so the gradients don't propagate to x[1]
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_cpy(ctx0, x[0], x[1]));
-
-                check_gradient("cpy f32", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // cpy f16
-        {
-            srand(seed);
-            const int nargs = 2;
-
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                for (int i = 0; i < nargs; ++i) {
-                    x[i] = get_random_tensor_f16(ctx0, ndims, ne, -1.0f, 1.0f);
-                    ggml_set_param(ctx0, x[i]);
-                }
-                // x[1] is overwritten by x[0], so the gradients don't propagate to x[1]
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_cpy(ctx0, x[0], x[1]));
-
-                check_gradient("cpy f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY, {});
-            }
-        }
-
-        // reshape (1d->nd)
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                int64_t ne2[4];
-                ne2[0] = 1;
-                ne2[1] = 1;
-                ne2[2] = 1;
-                ne2[3] = 1;
-                for (int i = 0; i < ndims; ++i) {
-                    ne2[0] *= ne[i];
-                }
-                x[0] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
-                x[1] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[0]);
-
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_reshape(ctx0, x[0], x[1]));
-                check_gradient("reshape", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // reshape (nd->1d)
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            for (int ndims = 1; ndims <= 2; ++ndims) {
-                int64_t ne2[4];
-                ne2[0] = 1;
-                ne2[1] = 1;
-                ne2[2] = 1;
-                ne2[3] = 1;
-                for (int i = 0; i < ndims; ++i) {
-                    ne2[0] *= ne[i];
-                }
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[0]);
-
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_reshape(ctx0, x[0], x[1]));
-                check_gradient("reshape", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // acc 1d
-        {
-            srand(seed);
-            int64_t ne2[4] = { 1, 1, 1, 1 };
-
-            const int nargs = 2;
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[0]);
-
-                get_random_dims(ne2, 1);
-                while ((ne2[0] > ne[0]) || (ne2[0] > ggml_nelements(x[0]))) {
-                    get_random_dims(ne2, 1);
-                }
-
-                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[1]);
-
-                const int max_offset = MAX(0, ggml_nelements(x[0]) - ggml_nelements(x[1]));
-                const int offset = irand(max_offset) * ggml_element_size(x[0]);
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_acc(ctx0, x[0], x[1], x[0]->nb[1], x[0]->nb[2], x[0]->nb[3], offset));
-
-                check_gradient("acc 1d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // acc 2d
-        {
-            srand(seed);
-            int64_t ne2[4]         = { 1, 1, 1, 1 };
-            int64_t max_offsets[4] = { 0, 0, 0, 0 };
-            int64_t offsets[4]     = { 0, 0, 0, 0 };
-
-            const int nargs = 2;
-            for (int ndims = 2; ndims <= 4; ++ndims) {
-
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[0]);
-
-                get_random_dims(ne2, 2);
-                while ((ne2[0] > ne[0]) || (ne2[1] > ne[1]) || (ne2[0]*ne2[1] > ggml_nelements(x[0]))) {
-                    get_random_dims(ne2, 2);
-                }
-
-                x[1] = get_random_tensor_f32(ctx0, 2, ne2, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[1]);
-
-                max_offsets[0] = MAX(0, x[0]->ne[0] - x[1]->ne[0]);
-                max_offsets[1] = MAX(0, x[0]->ne[1] - x[1]->ne[1]);
-                offsets[0] = irand(max_offsets[0]) * x[0]->nb[0];
-                offsets[1] = irand(max_offsets[1]) * x[0]->nb[1];
-                const int offset = offsets[0] + offsets[1];
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_acc(ctx0, x[0], x[1], x[0]->nb[1], x[0]->nb[2], x[0]->nb[3], offset));
-
-                check_gradient("acc 2d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // acc 3d
-        {
-            srand(seed);
-            int64_t ne2[4]         = { 1, 1, 1, 1 };
-            int64_t max_offsets[4] = { 0, 0, 0, 0 };
-            int64_t offsets[4]     = { 0, 0, 0, 0 };
-
-            const int nargs = 2;
-            for (int ndims = 3; ndims <= 4; ++ndims) {
-
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[0]);
-
-                get_random_dims(ne2, 3);
-                while ((ne2[0] > ne[0]) || (ne2[1] > ne[1]) || (ne2[2] > ne[2]) || (ne2[0]*ne2[1]*ne2[2] > ggml_nelements(x[0]))) {
-                    get_random_dims(ne2, 3);
-                }
-
-                x[1] = get_random_tensor_f32(ctx0, 3, ne2, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[1]);
-
-                max_offsets[0] = MAX(0, x[0]->ne[0] - x[1]->ne[0]);
-                max_offsets[1] = MAX(0, x[0]->ne[1] - x[1]->ne[1]);
-                max_offsets[2] = MAX(0, x[0]->ne[2] - x[1]->ne[2]);
-                offsets[0] = irand(max_offsets[0]) * x[0]->nb[0];
-                offsets[1] = irand(max_offsets[1]) * x[0]->nb[1];
-                offsets[2] = irand(max_offsets[2]) * x[0]->nb[2];
-                const int offset = offsets[0] + offsets[1] + offsets[2];
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_acc(ctx0, x[0], x[1], x[0]->nb[1], x[0]->nb[2], x[0]->nb[3], offset));
-
-                check_gradient("acc 3d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // acc 4d
-        {
-            srand(seed);
-            int64_t ne2[4]         = { 1, 1, 1, 1 };
-            int64_t max_offsets[4] = { 0, 0, 0, 0 };
-            int64_t offsets[4]     = { 0, 0, 0, 0 };
-
-            const int nargs = 2;
-            for (int ndims = 4; ndims <= 4; ++ndims) {
-
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[0]);
-
-                get_random_dims(ne2, 4);
-                while ((ne2[0] > ne[0]) || (ne2[1] > ne[1]) || (ne2[2] > ne[2]) || (ne2[3] > ne[3]) || (ne2[0]*ne2[1]*ne2[2]*ne2[3] > ggml_nelements(x[0]))) {
-                    get_random_dims(ne2, 4);
-                }
-
-                x[1] = get_random_tensor_f32(ctx0, 4, ne2, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[1]);
-
-                max_offsets[0] = MAX(0, x[0]->ne[0] - x[1]->ne[0]);
-                max_offsets[1] = MAX(0, x[0]->ne[1] - x[1]->ne[1]);
-                max_offsets[2] = MAX(0, x[0]->ne[2] - x[1]->ne[2]);
-                max_offsets[3] = MAX(0, x[0]->ne[3] - x[1]->ne[3]);
-                offsets[0] = irand(max_offsets[0]) * x[0]->nb[0];
-                offsets[1] = irand(max_offsets[1]) * x[0]->nb[1];
-                offsets[2] = irand(max_offsets[2]) * x[0]->nb[2];
-                offsets[3] = irand(max_offsets[3]) * x[0]->nb[3];
-                const int offset = offsets[0] + offsets[1] + offsets[2] + offsets[3];
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_acc(ctx0, x[0], x[1], x[0]->nb[1], x[0]->nb[2], x[0]->nb[3], offset));
-
-                check_gradient("acc 4d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // set_1d
-        {
-            srand(seed);
-            int64_t ne2[4];
-
-            const int nargs = 2;
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[0]);
-
-                get_random_dims(ne2, 1);
-                while ((ne2[0] > ne[0]) || (ne2[0] > ggml_nelements(x[0]))) {
-                    get_random_dims(ne2, 1);
-                }
-
-                x[1] = get_random_tensor_f32(ctx0, 1, ne2, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[1]);
-
-                const int max_offset = MAX(0, ggml_nelements(x[0]) - ggml_nelements(x[1]));
-                const int offset = irand(max_offset) * ggml_element_size(x[0]);
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_set_1d(ctx0, x[0], x[1], offset));
-
-                check_gradient("set_1d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // set_2d
-        {
-            srand(seed);
-            int64_t ne2[4];
-            int64_t max_offsets[4] = { 0, 0, 0, 0 };
-            int64_t offsets[4]     = { 0, 0, 0, 0 };
-
-            const int nargs = 1;
-            for (int ndims = 2; ndims <= 4; ++ndims) {
-
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[0]);
-
-                get_random_dims(ne2, 2);
-                while ((ne2[0] > ne[0]) || (ne2[1] > ne[1]) || (ne2[0]*ne2[1] > ggml_nelements(x[0]))) {
-                    get_random_dims(ne2, 2);
-                }
-
-                x[1] = get_random_tensor_f32(ctx0, 2, ne2, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[1]);
-
-                max_offsets[0] = MAX(0, x[0]->ne[0] - x[1]->ne[0]);
-                max_offsets[1] = MAX(0, x[0]->ne[1] - x[1]->ne[1]);
-                offsets[0] = irand(max_offsets[0]) * x[0]->nb[0];
-                offsets[1] = irand(max_offsets[1]) * x[0]->nb[1];
-                const int offset = offsets[0] + offsets[1];
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_set_2d(ctx0, x[0], x[1], x[1]->nb[1], offset));
-
-                check_gradient("set_2d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // view_1d
-        {
-            srand(seed);
-            const int nargs = 1;
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-
-                ggml_set_param(ctx0, x[0]);
-
-                const int k0 = irand(ggml_nelements(x[0]));
-                const int k1 = irand(ggml_nelements(x[0]));
-                const int i0 = MIN(k0, k1);
-                const int i1 = MAX(k0, k1);
-
-                const int offset = i0 * sizeof(float);
-                const int nelem  = i1 - i0;
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_view_1d(ctx0, x[0], nelem, offset));
-
-                check_gradient("view_1d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // view_2d
-        {
-            srand(seed);
-            int64_t ne2[4];
-            int64_t nb2[4];
-
-            const int nargs = 1;
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-
-                get_random_dims(ne2, 2);
-                while (ne2[0]*ne2[1] > ggml_nelements(x[0])) {
-                    get_random_dims(ne2, 2);
-                }
-                const int count = ne2[0]*ne2[1];
-
-                nb2[0] = sizeof(float);
-                nb2[1] = nb2[0]*ne2[0];
-
-                ggml_set_param(ctx0, x[0]);
-
-                const int max_offset = ggml_nelements(x[0]) - count;
-                const int offset = irand(max_offset+1) * sizeof(float);
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_view_2d(ctx0, x[0], ne2[0], ne2[1], nb2[1], offset));
-
-                check_gradient("view_2d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // view_3d
-        {
-            srand(seed);
-            int64_t ne2[4] = {1,1,1,1};
-            int64_t nb2[4] = {0,0,0,0};
-
-            const int nargs = 1;
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-
-                get_random_dims(ne2, 3);
-                while (ne2[0]*ne2[1]*ne2[2] > ggml_nelements(x[0])) {
-                    get_random_dims(ne2, 3);
-                }
-                const int count = ne2[0]*ne2[1]*ne2[2];
-
-                nb2[0] = sizeof(float);
-                nb2[1] = nb2[0]*ne2[0];
-                nb2[2] = nb2[1]*ne2[1];
-
-                ggml_set_param(ctx0, x[0]);
-
-                const int max_offset = ggml_nelements(x[0]) - count;
-                const int offset = irand(max_offset+1) * sizeof(float);
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_view_3d(ctx0, x[0], ne2[0], ne2[1], ne2[2], nb2[1], nb2[2], offset));
-
-                check_gradient("view_3d", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // permute
-        {
-            srand(seed);
-            int64_t ne2[4];
-
-            const int nargs = 1;
-            for (int ndims = 1; ndims <= 4; ++ndims)
-            {
-                // ggml_permute will set axes of dimensions below n_dims to 1.
-                // to make ggml_permute work correctly on all axes,
-                // the input tensor needs maximal n_dim of 4.
-                for (int i=0; i<ndims; ++i) {
-                    ne2[i] = ne[i];
-                }
-                for (int i=ndims; i<4; ++i) {
-                    ne2[i] = 1;
-                }
-                x[0] = get_random_tensor_f32(ctx0, 4, ne2, -1.0f, 1.0f);
-
-                ggml_set_param(ctx0, x[0]);
-
-                const int p = irand(NUM_PERMUTATIONS);
-                const int ax0 = all_permutations[p*4+0];
-                const int ax1 = all_permutations[p*4+1];
-                const int ax2 = all_permutations[p*4+2];
-                const int ax3 = all_permutations[p*4+3];
-
-                // sum requires contiguous tensor rows
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_cont(ctx0, ggml_permute(ctx0, x[0], ax0, ax1, ax2, ax3)));
-
-                check_gradient("permute", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // transpose
-        {
-            srand(seed);
-            int64_t ne2[4];
-
-            const int nargs = 1;
-            for (int ndims = 1; ndims <= 4; ++ndims)
-            {
-                // ggml_transpose will set axes of dimensions below n_dims to 1.
-                // to make ggml_transpose work correctly on all axes,
-                // the input tensor needs maximal n_dim of 4.
-                for (int i=0; i<ndims; ++i) {
-                    ne2[i] = ne[i];
-                }
-                for (int i=ndims; i<4; ++i) {
-                    ne2[i] = 1;
-                }
-                x[0] = get_random_tensor_f32(ctx0, 4, ne2, -1.0f, 1.0f);
-
-                ggml_set_param(ctx0, x[0]);
-
-                // sum requires contiguous tensor rows
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, x[0])));
-
-                check_gradient("transpose", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // get_rows
-        {
-            srand(seed);
-            int64_t ne2[4] = {ne[0], ne[1], 1, 1};
-            int64_t ne3[4] = {1+irand(ne[1]), 1, 1, 1};
-            const int nargs = 1;
-            const int ndims = 2;
-            x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
-            x[1] = get_random_tensor_i32(ctx0, 1, ne3, 0, ne2[1]);
-
-            ggml_set_param(ctx0, x[0]);
-
-            struct ggml_tensor * f = ggml_sum(ctx0, ggml_get_rows(ctx0, x[0], x[1]));
-
-            check_gradient("get_rows", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-        }
-
-        // diag_mask_inf
-        {
-            srand(seed);
-            const int nargs = 1;
-            const int ndims = 2;
-
-            x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-            ggml_set_param(ctx0, x[0]);
-
-            int n_past = irand(ne[0]);
-
-            struct ggml_tensor * f = ggml_sum(ctx0, ggml_diag_mask_inf(ctx0, x[0], n_past));
-
-            check_gradient("diag_mask_inf", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-        }
-
-        // diag_mask_zero
-        {
-            srand(seed);
-            const int nargs = 1;
-            const int ndims = 2;
-
-            x[0] = get_random_tensor_f32(ctx0, ndims, ne, -1.0f, 1.0f);
-            ggml_set_param(ctx0, x[0]);
-
-            int n_past = irand(ne[0]);
-
-            struct ggml_tensor * f = ggml_sum(ctx0, ggml_diag_mask_zero(ctx0, x[0], n_past));
-
-            check_gradient("diag_mask_zero", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-        }
-
-        // softmax
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            int64_t ne2[4];
-            get_random_dims(ne2, 4);
-
-            for (int ndims = 1; ndims <= 3; ++ndims) {
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
-                ggml_set_param(ctx0, x[0]);
-
-                float eps = 1e-6f;
-                // dont use only sum as aggregation, because sum of softmax is always 1 -> finite differences should not work
-                // instead use sum(log(soft_max()*(1-eps)+eps)); use eps to avoid log(0)
-                struct ggml_tensor * f = ggml_sum(ctx0,
-                                            ggml_log(ctx0,
-                                                ggml_add1(ctx0,
-                                                    ggml_scale(ctx0,
-                                                        ggml_soft_max(ctx0, x[0]),
-                                                        1.0f - eps),
-                                                    ggml_new_f32(ctx0, eps))));
-
-                check_gradient("softmax", ctx0, x, f, ndims, nargs, 1e-3f, 2e-1f, INFINITY, {});
-                // NOTE: softmax forward is computed using f16 table lookup instead of using actual expf, but backward assumes actual expf.
-                // this may result in different gradients too finite differences.
-                // when this test reports errors, first try to replace the table lookup with actual expf and test again to see if just that was the cause.
-                // if only the table lookup causes gradients to differ this is acceptable.
-            }
-        }
-
-        // cross_entropy_loss
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            int64_t ne2[4];
-            get_random_dims(ne2, 4);
-
-            for (int ndims = 1; ndims <= 4; ++ndims) {
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
-                x[1] = get_random_tensor_f32(ctx0, ndims, ne2, 0.0f, 1.0f);
-                // the second argument to cross_entropy_loss must sum up to 1 for each row
-                int nr = ggml_nrows(x[1]);
-                int nc = ggml_nelements(x[1]) / nr;
-                for (int ir = 0; ir < nr; ++ir) {
-                    float sum = 0;
-                    for (int ic = 0; ic < nc; ++ic) {
-                        sum += ((float *) x[1]->data)[ic + ir*nc];
-                    }
-                    for (int ic = 0; ic < nc; ++ic) {
-                        ((float *) x[1]->data)[ic + ir*nc] /= sum;
-                    }
-                }
-                ggml_set_param(ctx0, x[0]);
-
-                struct ggml_tensor * f = ggml_cross_entropy_loss(ctx0, x[0], x[1]);
-
-                check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // rope f32
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            int64_t ne2[4];
-            get_random_dims(ne2, 4);
-            ne2[0] += ne2[0] % 2;
-            int n_rot = ne2[0];
-
-            for (int ndims = 3; ndims <= 4; ++ndims) {
-                for (int mode = 0; mode < 4; ++mode) {
-                    for (int n_past = 1; n_past < ne2[2]; ++n_past) {
-                        x[0] = get_random_tensor_f32(ctx0, ndims, ne2, -1.0f, 1.0f);
-
-                        struct ggml_tensor * p = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne2[2]);
-                        for (int i = 0; i < ne2[2]; ++i) {
-                            ((int32_t *) p->data)[i] = n_past + i;
-                        }
-
-                        ggml_set_param(ctx0, x[0]);
-
-                        const bool skip_past = (mode & 1);
-                        if (skip_past) {
-                            // we have no past, so this would have to work on uninitialized memory.
-                            // we only test the gradients here;
-                            // skip_past should have no influence on gradient computation.
-                            // so when other modes work, we assume that this does as well.
-                            continue;
-                        }
-
-                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], p, n_rot, mode));
-
-                        GGML_PRINT_DEBUG("rope f32: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
-                        check_gradient("rope f32", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY, {});
-                    }
-                }
-            }
-        }
-
-        // rope f16
-        {
-            srand(seed);
-            const int nargs = 1;
-
-            int64_t ne2[4];
-            get_random_dims(ne2, 4);
-            ne2[0] += ne2[0] % 2;
-            int n_rot = ne2[0];
-
-            for (int ndims = 3; ndims <= 4; ++ndims) {
-                for (int mode = 0; mode < 4; ++mode) {
-                    for (int n_past = 1; n_past < ne2[2]; ++n_past) {
-                        x[0] = get_random_tensor_f16(ctx0, ndims, ne2, -1.0f, 1.0f);
-
-                        struct ggml_tensor * p = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ne2[2]);
-                        for (int i = 0; i < ne2[2]; ++i) {
-                            ((int32_t *) p->data)[i] = n_past + i;
-                        }
-
-                        ggml_set_param(ctx0, x[0]);
-
-                        const bool skip_past = (mode & 1);
-                        if (skip_past) {
-                            // we have no past, so this would have to work on uninitialized memory.
-                            // we only test the gradients here;
-                            // skip_past should have no influence on gradient computation.
-                            // so when other modes work, we assume that this does as well.
-                            continue;
-                        }
-
-                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_rope(ctx0, x[0], p, n_rot, mode));
-
-                        GGML_PRINT_DEBUG("rope f16: n_past: %d n_rot: %d mode: %d\n", n_past, n_rot, mode);
-                        check_gradient("rope f16", ctx0, x, f, ndims, nargs, 1e-1f, 1e-1f, INFINITY, {});
-                    }
-                }
-            }
-        }
-
-        // im2col f32
-        {
-            srand(seed);
-            const int nargs = 1;
-            const int ndims = 4;
-
-            for (const bool is_2D : {false, true}) {
-                int64_t ne0[ndims];
-                int64_t ne1[ndims];
-                get_random_dims(ne0, ndims);
-                get_random_dims(ne1, ndims);
-
-                // // Ensure that the output is not zero-sized:
-                ne1[0] += 8;
-                ne1[1] += 8;
-
-                if (is_2D) {
-                    ne1[2] = ne0[2];
-                } else {
-                    ne1[1] = ne0[1];
-                    ne0[3] = 1;
-                    ne1[3] = 1;
-                }
-
-                // The order of arguments is swapped because the first tensor is only used for its shape.
-                x[1] = get_random_tensor_f16(ctx0, ndims, ne0, -1.0f, 1.0f);
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne1, -1.0f, 1.0f);
-
-                ggml_set_param(ctx0, x[0]);
-
-                const int s0 =         1 + irand(2);
-                const int s1 = is_2D ? 1 + irand(2) : 0;
-                const int p0 =         0 + irand(2);
-                const int p1 = is_2D ? 0 + irand(2) : 0;
-                const int d0 =         1 + irand(2);
-                const int d1 = is_2D ? 1 + irand(2) : 0;
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_im2col(ctx0, x[1], x[0], s0, s1, p0, p1, d0, d1, is_2D, GGML_TYPE_F32));
-
-                GGML_PRINT_DEBUG("im2col f32: is_2D=%s, s0=%d, s1=%d, p0=%d, p1=%d, d0=%d, d1=%d\n", is_2D ? "yes" : "no", s0, s1, p0, p1, d0, d1);
-                check_gradient("im2col f32", ctx0, x, f, ndims, nargs, 1e-2f, 1e-3f, INFINITY, {});
-            }
-        }
-
-        // pool_2d f32
-        {
-            srand(seed);
-            const int nargs = 1;
-            const int ndims = 4;
-
-            for (const enum ggml_op_pool op : {GGML_OP_POOL_AVG, GGML_OP_POOL_MAX}) {
-                int64_t ne0[ndims];
-                get_random_dims(ne0, ndims);
-
-                ne0[0] += 8;
-                ne0[1] += 8;
-
-                x[0] = get_random_tensor_f32(ctx0, ndims, ne0, -1.0f, 1.0f);
-
-                ggml_set_param(ctx0, x[0]);
-
-                const int k0 = 2 + irand(2);
-                const int k1 = 2 + irand(2);
-                const int s0 = 2 + irand(2);
-                const int s1 = 2 + irand(2);
-                const int p0 = 0 + irand(2);
-                const int p1 = 0 + irand(2);
-
-                struct ggml_tensor * f = ggml_sum(ctx0, ggml_pool_2d(ctx0, x[0], op, k0, k1, s0, s1, p0, p1));
-
-                GGML_PRINT_DEBUG("ggml_pool_2d f32: op=%s k0=%d, k1=%d, s0=%d, s1=%d, p0=%d, p1=%d\n",
-                                 op == GGML_OP_POOL_MAX ? "max" : "avg", k0, k1, s0, s1, p0, p1);
-                std::vector<double> expected_vals;
-                if (op == GGML_OP_POOL_MAX) {
-                    expected_vals.push_back(0.0);
-                    expected_vals.push_back(1.0);
-                }
-                check_gradient("ggml_pool_2d f32", ctx0, x, f, ndims, nargs, 1e-3f, 1e-3f, INFINITY, expected_vals);
-            }
-        }
-
-        // flash_attn f32
-        // TODO: adapt to ggml_flash_attn_ext() changes
-        //{
-        //    srand(seed);
-        //    const int nargs = 3;
-
-        //    int64_t ne2[4];
-
-        //    get_random_dims(ne2, 4);
-        //    int64_t D = ne2[0];
-        //    int64_t N = ne2[1];
-        //    int64_t M = ne2[2] + N;
-        //    int64_t B = ne2[3];
-
-        //    for (int masked = 0; masked <= 1; ++masked) {
-        //        for (int ndims = 2; ndims <= 4; ++ndims) {
-        //            int max_nrep = (ndims >= 3) ? 2 : 1;
-        //            for (int nrep = 1; nrep < max_nrep; ++nrep) {
-        //                int64_t neq[4] = { D, N, B*nrep, ne[3] };
-        //                int64_t nek[4] = { D, M, B, ne[3] };
-        //                int64_t nev[4] = { M, D, B, ne[3] };
-        //                if (ndims == 2) {
-        //                    neq[2] = 1; neq[3] = 1;
-        //                    nek[2] = 1; nek[3] = 1;
-        //                    nev[2] = 1; nev[3] = 1;
-        //                } else if (ndims == 3) {
-        //                    neq[3] = 1;
-        //                    nek[3] = 1;
-        //                    nev[3] = 1;
-        //                }
-        //                x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
-        //                x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
-        //                x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
-        //                ggml_set_param(ctx0, x[0]);
-        //                ggml_set_param(ctx0, x[1]);
-        //                ggml_set_param(ctx0, x[2]);
-
-        //                struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
-
-        //                check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY, {});
-        //            }
-        //        }
-        //    }
-        //}
-
-        ggml_free(ctx0);
-    }
-
-    return 0;
-}

From a4200cafadebb7576a9a3905039858f5e73ce4cc Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sat, 16 Nov 2024 21:35:31 +0200
Subject: [PATCH 011/102] make : add ggml-opt (#0)

ggml-ci
---
 Makefile | 1 +
 1 file changed, 1 insertion(+)

diff --git a/Makefile b/Makefile
index ae9a9c6323a01..fecf1f693e195 100644
--- a/Makefile
+++ b/Makefile
@@ -996,6 +996,7 @@ OBJ_GGML = \
 	$(DIR_GGML)/src/ggml-alloc.o \
 	$(DIR_GGML)/src/ggml-backend.o \
 	$(DIR_GGML)/src/ggml-backend-reg.o \
+	$(DIR_GGML)/src/ggml-opt.o \
 	$(DIR_GGML)/src/ggml-quants.o \
 	$(DIR_GGML)/src/ggml-threading.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \

From 5d9e59979c2fd91c99d03c23e8df9c50c9d55a85 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sat, 16 Nov 2024 21:38:01 +0200
Subject: [PATCH 012/102] ggml : adapt AMX to tensor->grad removal (#0)

ggml-ci
---
 ggml/src/ggml-amx/ggml-amx.cpp | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/ggml/src/ggml-amx/ggml-amx.cpp b/ggml/src/ggml-amx/ggml-amx.cpp
index 37da985399c58..8568e7965fd2e 100644
--- a/ggml/src/ggml-amx/ggml-amx.cpp
+++ b/ggml/src/ggml-amx/ggml-amx.cpp
@@ -317,8 +317,6 @@ static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const st
             const enum ggml_type type = src0->type;
             const int64_t ne0 = op->ne[0];
 
-            bool is_training = src0->grad || src1->grad;
-
             // amx kernels enables for Q4_0, Q4_1, Q8_0, F16
             // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
             bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
@@ -326,7 +324,6 @@ static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const st
             bool can_use_amx =
                 is_contiguous_2d(src0) &&       // src0 must be contiguous
                 is_contiguous_2d(src1) &&       // src1 must be contiguous
-                !is_training &&                 // inference only
                 src1->type == GGML_TYPE_F32 &&  // src1 must be float32
                 has_amx_kernels &&              // with amx kernel impls
                 ne0 % (TILE_N * 2) == 0;        // out_features is 32x

From 24203e9dd7355a4a10bc32d959fd0148d37bf666 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sat, 16 Nov 2024 23:40:39 +0200
Subject: [PATCH 013/102] ggml : inttypes.h -> cinttypes (#0)

ggml-ci
---
 ggml/src/ggml-opt.cpp        | 2 +-
 tests/test-opt.cpp           | 2 +-
 tests/test-quantize-perf.cpp | 1 -
 3 files changed, 2 insertions(+), 3 deletions(-)

diff --git a/ggml/src/ggml-opt.cpp b/ggml/src/ggml-opt.cpp
index 808aa0d023ab2..040205a31bab0 100644
--- a/ggml/src/ggml-opt.cpp
+++ b/ggml/src/ggml-opt.cpp
@@ -8,7 +8,7 @@
 #include <algorithm>
 #include <cmath>
 #include <cstdint>
-#include <inttypes.h>
+#include <cinttypes>
 #include <map>
 #include <random>
 #include <vector>
diff --git a/tests/test-opt.cpp b/tests/test-opt.cpp
index 4abe85c74a4f6..f90c92b4b8ecf 100644
--- a/tests/test-opt.cpp
+++ b/tests/test-opt.cpp
@@ -5,7 +5,7 @@
 #include "ggml-opt.h"
 
 #include <cmath>
-#include <inttypes.h>
+#include <cinttypes>
 #include <random>
 #include <string>
 #include <thread>
diff --git a/tests/test-quantize-perf.cpp b/tests/test-quantize-perf.cpp
index ac0d1271430b1..2882884938388 100644
--- a/tests/test-quantize-perf.cpp
+++ b/tests/test-quantize-perf.cpp
@@ -7,7 +7,6 @@
 #include <algorithm>
 #include <assert.h>
 #include <functional>
-#include <inttypes.h>
 #include <math.h>
 #include <memory>
 #include <stdio.h>

From eda7e1d4f54711de1c9b40502d6c88bbc217da60 Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Sun, 17 Nov 2024 07:31:17 +0100
Subject: [PATCH 014/102] ggml : fix possible buffer use after free in sched
 reserve (#9930)

---
 ggml/src/ggml-backend.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index 634fe38eef871..9a6010d361224 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -1538,12 +1538,13 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 
     ggml_backend_sched_split_graph(sched, measure_graph);
 
+    ggml_backend_sched_synchronize(sched);
+
     if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
         return false;
     }
 
     ggml_backend_sched_reset(sched);
-    ggml_backend_sched_synchronize(sched);
 
     return true;
 }

From 467576b6cc7d2b9220f55bc635aa51469cf26fb5 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Sun, 17 Nov 2024 09:06:34 +0100
Subject: [PATCH 015/102] CMake: default to -arch=native for CUDA build
 (#10320)

---
 README.md                         |  4 ++--
 ggml/src/ggml-cuda/CMakeLists.txt | 15 +++++++++------
 2 files changed, 11 insertions(+), 8 deletions(-)

diff --git a/README.md b/README.md
index 6ab6acf1282c6..5f7933c132dc3 100644
--- a/README.md
+++ b/README.md
@@ -459,14 +459,14 @@ To learn more how to measure perplexity using llama.cpp, [read this documentatio
 - Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205)
 - A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532)
 
-## Other documentations
+## Other documentation
 
 - [main (cli)](./examples/main/README.md)
 - [server](./examples/server/README.md)
 - [jeopardy](./examples/jeopardy/README.md)
 - [GBNF grammars](./grammars/README.md)
 
-**Development documentations**
+**Development documentation**
 
 - [How to build](./docs/build.md)
 - [Running on Docker](./docs/docker.md)
diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt
index 40ed2bdf375e9..860552f3a96df 100644
--- a/ggml/src/ggml-cuda/CMakeLists.txt
+++ b/ggml/src/ggml-cuda/CMakeLists.txt
@@ -6,15 +6,18 @@ if (CUDAToolkit_FOUND)
     message(STATUS "CUDA Toolkit found")
 
     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-        # 52 == lowest CUDA 12 standard
-        # 60 == FP16 CUDA intrinsics
-        # 61 == integer CUDA intrinsics
-        # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
-        if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+        # native == GPUs available at build time
+        # 52     == Maxwell, lowest CUDA 12 standard
+        # 60     == P100, FP16 CUDA intrinsics
+        # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
+        # 70     == V100, FP16 tensor cores
+        # 75     == Turing, int6 tensor cores
+        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6")
+            set(CMAKE_CUDA_ARCHITECTURES "native")
+        elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
             set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
         else()
             set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
-            #set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")

From c3ea58aca406911eb4d409cdbfc76683393442b6 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Sun, 17 Nov 2024 09:09:55 +0100
Subject: [PATCH 016/102] CUDA: remove DMMV, consolidate F16 mult mat vec
 (#10318)

---
 Makefile                                 |  57 --
 docs/build.md                            |  11 -
 ggml/CMakeLists.txt                      |   5 -
 ggml/src/ggml-cuda/CMakeLists.txt        |  13 -
 ggml/src/ggml-cuda/dmmv.cu               | 699 -----------------------
 ggml/src/ggml-cuda/ggml-cuda.cu          | 204 +------
 ggml/src/ggml-cuda/mmv.cu                | 223 ++++++++
 ggml/src/ggml-cuda/{dmmv.cuh => mmv.cuh} |  16 +-
 ggml/src/ggml-hip/CMakeLists.txt         |   7 -
 ggml/src/ggml-musa/CMakeLists.txt        |  11 -
 10 files changed, 246 insertions(+), 1000 deletions(-)
 delete mode 100644 ggml/src/ggml-cuda/dmmv.cu
 create mode 100644 ggml/src/ggml-cuda/mmv.cu
 rename ggml/src/ggml-cuda/{dmmv.cuh => mmv.cuh} (55%)

diff --git a/Makefile b/Makefile
index fecf1f693e195..bd0bad160a1f2 100644
--- a/Makefile
+++ b/Makefile
@@ -635,10 +635,6 @@ else ifndef CUDA_POWER_ARCH
 	MK_NVCCFLAGS += -arch=native
 endif # CUDA_DOCKER_ARCH
 
-ifdef GGML_CUDA_FORCE_DMMV
-	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
-endif # GGML_CUDA_FORCE_DMMV
-
 ifdef GGML_CUDA_FORCE_MMQ
 	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
 endif # GGML_CUDA_FORCE_MMQ
@@ -647,20 +643,6 @@ ifdef GGML_CUDA_FORCE_CUBLAS
 	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_CUBLAS
 endif # GGML_CUDA_FORCE_CUBLAS
 
-ifdef GGML_CUDA_DMMV_X
-	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
-else
-	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
-endif # GGML_CUDA_DMMV_X
-
-ifdef GGML_CUDA_MMV_Y
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
-else ifdef GGML_CUDA_DMMV_Y
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_DMMV_Y) # for backwards compatibility
-else
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
-endif # GGML_CUDA_MMV_Y
-
 ifdef GGML_CUDA_F16
 	MK_NVCCFLAGS += -DGGML_CUDA_F16
 endif # GGML_CUDA_F16
@@ -669,12 +651,6 @@ ifdef GGML_CUDA_DMMV_F16
 	MK_NVCCFLAGS += -DGGML_CUDA_F16
 endif # GGML_CUDA_DMMV_F16
 
-ifdef GGML_CUDA_KQUANTS_ITER
-	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
-else
-	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
-endif
-
 ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
 	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
 else
@@ -783,10 +759,6 @@ ifdef GGML_HIPBLAS
 		AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
 	endif
 
-	GGML_CUDA_DMMV_X       ?= 32
-	GGML_CUDA_MMV_Y        ?= 1
-	GGML_CUDA_KQUANTS_ITER ?= 2
-
 	MK_CPPFLAGS += -DGGML_USE_HIP -DGGML_USE_CUDA
 
 ifdef GGML_HIP_UMA
@@ -800,13 +772,6 @@ endif # GGML_HIP_UMA
 	HIPCC ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
 
 	HIPFLAGS += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
-	HIPFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
-	HIPFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
-	HIPFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
-
-ifdef GGML_CUDA_FORCE_DMMV
-	HIPFLAGS += -DGGML_CUDA_FORCE_DMMV
-endif # GGML_CUDA_FORCE_DMMV
 
 ifdef GGML_CUDA_FORCE_MMQ
 	HIPFLAGS += -DGGML_CUDA_FORCE_MMQ
@@ -869,10 +834,6 @@ ifdef GGML_MUSA
 
 	MUSAFLAGS += $(addprefix --cuda-gpu-arch=, $(MTGPU_TARGETS))
 
-ifdef GGML_CUDA_FORCE_DMMV
-	MUSAFLAGS += -DGGML_CUDA_FORCE_DMMV
-endif # GGML_CUDA_FORCE_DMMV
-
 ifdef GGML_CUDA_FORCE_MMQ
 	MUSAFLAGS += -DGGML_CUDA_FORCE_MMQ
 endif # GGML_CUDA_FORCE_MMQ
@@ -881,18 +842,6 @@ ifdef GGML_CUDA_FORCE_CUBLAS
 	MUSAFLAGS += -DGGML_CUDA_FORCE_CUBLAS
 endif # GGML_CUDA_FORCE_CUBLAS
 
-ifdef GGML_CUDA_DMMV_X
-	MUSAFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
-else
-	MUSAFLAGS += -DGGML_CUDA_DMMV_X=32
-endif # GGML_CUDA_DMMV_X
-
-ifdef GGML_CUDA_MMV_Y
-	MUSAFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
-else
-	MUSAFLAGS += -DGGML_CUDA_MMV_Y=1
-endif # GGML_CUDA_MMV_Y
-
 ifdef GGML_CUDA_F16
 	MUSAFLAGS += -DGGML_CUDA_F16
 endif # GGML_CUDA_F16
@@ -901,12 +850,6 @@ ifdef GGML_CUDA_DMMV_F16
 	MUSAFLAGS += -DGGML_CUDA_F16
 endif # GGML_CUDA_DMMV_F16
 
-ifdef GGML_CUDA_KQUANTS_ITER
-	MUSAFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
-else
-	MUSAFLAGS += -DK_QUANTS_PER_ITERATION=2
-endif
-
 ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
 	MUSAFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
 else
diff --git a/docs/build.md b/docs/build.md
index 811bbb409c250..359952b30ff93 100644
--- a/docs/build.md
+++ b/docs/build.md
@@ -186,13 +186,9 @@ The following compilation options are also available to tweak performance:
 
 | Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
 |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| GGML_CUDA_FORCE_DMMV          | Boolean                | false   | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
-| GGML_CUDA_DMMV_X              | Positive integer >= 32 | 32      | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants.                                         |
-| GGML_CUDA_MMV_Y               | Positive integer       | 1       | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended.                                                                                                                                         |
 | GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
 | GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
 | GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
-| GGML_CUDA_KQUANTS_ITER        | 1 or 2                 | 2       | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                                                                     |
 | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
 | GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
 
@@ -268,13 +264,6 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 
 The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
 If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
-The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
-
-| Option                 | Legal values           | Default | Description                                                                                                                                                                                                                                    |
-|------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| GGML_CUDA_DMMV_X       | Positive integer >= 32 | 32      | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-| GGML_CUDA_MMV_Y        | Positive integer       | 1       | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants.                                                                       |
-| GGML_CUDA_KQUANTS_ITER | 1 or 2                 | 2       | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                             |
 
 ### Vulkan
 
diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index fd94998264547..a82818d607f5e 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -128,14 +128,9 @@ option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"
 
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
-option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
-set   (GGML_CUDA_DMMV_X   "32" CACHE STRING "ggml: x stride for dmmv CUDA kernels")
-set   (GGML_CUDA_MMV_Y     "1" CACHE STRING "ggml: y block size for mmv CUDA kernels")
 option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
-set   (GGML_CUDA_KQUANTS_ITER "2" CACHE STRING
-                                            "ggml: iters./thread per block for Q2_K/Q6_K")
 set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt
index 860552f3a96df..3dde0f3666187 100644
--- a/ggml/src/ggml-cuda/CMakeLists.txt
+++ b/ggml/src/ggml-cuda/CMakeLists.txt
@@ -54,21 +54,12 @@ if (CUDAToolkit_FOUND)
     target_link_libraries(ggml-cuda PRIVATE ggml-base)
     target_include_directories(ggml-cuda PRIVATE . ..)
 
-    # TODO: change the definitions to this target only
-
-    add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
-    add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
-    add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
     add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
 
     if (GGML_CUDA_GRAPHS)
         add_compile_definitions(GGML_CUDA_USE_GRAPHS)
     endif()
 
-    if (GGML_CUDA_FORCE_DMMV)
-        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
-    endif()
-
     if (GGML_CUDA_FORCE_MMQ)
         add_compile_definitions(GGML_CUDA_FORCE_MMQ)
     endif()
@@ -81,10 +72,6 @@ if (CUDAToolkit_FOUND)
         add_compile_definitions(GGML_CUDA_NO_VMM)
     endif()
 
-    if (DEFINED GGML_CUDA_DMMV_Y)
-        add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_DMMV_Y}) # for backwards compatibility
-    endif()
-
     if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
         add_compile_definitions(GGML_CUDA_F16)
     endif()
diff --git a/ggml/src/ggml-cuda/dmmv.cu b/ggml/src/ggml-cuda/dmmv.cu
deleted file mode 100644
index 00e21b5d77e3c..0000000000000
--- a/ggml/src/ggml-cuda/dmmv.cu
+++ /dev/null
@@ -1,699 +0,0 @@
-#include "dmmv.cuh"
-#include "dequantize.cuh"
-#include "convert.cuh"
-
-#ifndef K_QUANTS_PER_ITERATION
-#define K_QUANTS_PER_ITERATION 2
-#else
-static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
-#endif
-
-static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q2_K * x = (const block_q2_K *)vx + ib0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int step = 16/K_QUANTS_PER_ITERATION;
-
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0...15 or 0...7
-
-    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15 or 0...14 in steps of 2
-    const int q_offset = 32*im + l0;
-    const int s_offset = 8*im;
-    const int y_offset = 128*im + l0;
-
-    uint32_t aux[4];
-    const uint8_t * d = (const uint8_t *)aux;
-    const uint8_t * m = (const uint8_t *)(aux + 2);
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + y_offset;
-        const uint8_t * q = x[i].qs + q_offset;
-
-        const float dall = __low2half(x[i].dm);
-        const float dmin = __high2half(x[i].dm);
-
-        const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
-        aux[0] = a[0] & 0x0f0f0f0f;
-        aux[1] = a[1] & 0x0f0f0f0f;
-        aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
-        aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
-
-        float sum1 = 0, sum2 = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
-                  + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
-                  + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
-                  + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
-                  + y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
-                  + y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
-                  + y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
-                  +y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
-            sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
-                  + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
-
-        }
-        tmp += dall * sum1 - dmin * sum2;
-
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q3_K * x = (const block_q3_K *)vx + ib0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    const uint16_t kmask1 = 0x0303;
-    const uint16_t kmask2 = 0x0f0f;
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int n  = K_QUANTS_PER_ITERATION;               // iterations in the inner loop
-    const int step = 16/K_QUANTS_PER_ITERATION;
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0....15 or 0...7
-
-    const uint8_t m = 1 << (4*im);
-
-    const int l0 = n*in;                                 // 0...15 or 0...14 in steps of 2
-    const int q_offset =  32*im + l0;
-    const int y_offset = 128*im + l0;
-
-    uint16_t utmp[4];
-    const int8_t * s = (const int8_t *)utmp;
-
-    const uint16_t s_shift = 4*im;
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + y_offset;
-        const uint8_t * q = x[i].qs + q_offset;
-        const uint8_t * h = x[i].hmask + l0;
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
-        utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
-        utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
-        utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
-
-        const float d = x[i].d;
-
-        float sum = 0;
-        for (int l = 0; l < n; ++l) {
-            sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
-                 + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
-                 + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
-                 + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
-            sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
-                 + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
-                 + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
-                + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
-        }
-        tmp += d * sum;
-
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q4_K * x = (const block_q4_K *)vx + ib0;
-
-    const uint16_t kmask1 = 0x3f3f;
-    const uint16_t kmask2 = 0x0f0f;
-    const uint16_t kmask3 = 0xc0c0;
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int step = 8/K_QUANTS_PER_ITERATION;           // 8 or 4
-
-    const int il  = tid/step;                            // 0...3
-    const int ir  = tid - step*il;                       // 0...7 or 0...3
-    const int n   = 2 * K_QUANTS_PER_ITERATION;          // 2 or 4
-
-    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const int in = il%2;
-
-    const int l0 = n*(2*ir + in);
-    const int q_offset = 32*im + l0;
-    const int y_offset = 64*im + l0;
-
-    uint16_t aux[4];
-    const uint8_t * sc = (const uint8_t *)aux;
-
-#if K_QUANTS_PER_ITERATION == 2
-    uint32_t q32[4];
-    const uint8_t * q4 = (const uint8_t *)q32;
-#else
-    uint16_t q16[4];
-    const uint8_t * q4 = (const uint8_t *)q16;
-#endif
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y1 = yy + i*QK_K + y_offset;
-        const float   * y2 = y1 + 128;
-
-        const float dall = __low2half(x[i].dm);
-        const float dmin = __high2half(x[i].dm);
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux[0] = a[im+0] & kmask1;
-        aux[1] = a[im+2] & kmask1;
-        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
-        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
-
-#if K_QUANTS_PER_ITERATION == 2
-        const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset);
-        const uint32_t * q2 = q1 + 16;
-
-        q32[0] = q1[0] & 0x0f0f0f0f;
-        q32[1] = q1[0] & 0xf0f0f0f0;
-        q32[2] = q2[0] & 0x0f0f0f0f;
-        q32[3] = q2[0] & 0xf0f0f0f0;
-
-        float4 s = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        for (int l = 0; l < 4; ++l) {
-            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+ 4];
-            s.z += y2[l] * q4[l+8]; s.w += y2[l+32] * q4[l+12];
-            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
-        }
-        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
-#else
-        const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset);
-        const uint16_t * q2 = q1 + 32;
-
-        q16[0] = q1[0] & 0x0f0f;
-        q16[1] = q1[0] & 0xf0f0;
-        q16[2] = q2[0] & 0x0f0f;
-        q16[3] = q2[0] & 0xf0f0;
-
-        float4 s = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        for (int l = 0; l < 2; ++l) {
-            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
-            s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
-            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
-        }
-        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
-#endif
-
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (tid == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols) {
-
-    const int row = blockIdx.x;
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q5_K * x = (const block_q5_K *)vx + ib0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    const uint16_t kmask1 = 0x3f3f;
-    const uint16_t kmask2 = 0x0f0f;
-    const uint16_t kmask3 = 0xc0c0;
-
-    const int tid = threadIdx.x/2;  // 0...15
-    const int ix  = threadIdx.x%2;
-
-    const int il  = tid/4;     // 0...3
-    const int ir  = tid - 4*il;// 0...3
-    const int n   = 2;
-
-    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const int in = il%2;
-
-    const int l0 = n*(2*ir + in);
-    const int q_offset = 32*im + l0;
-    const int y_offset = 64*im + l0;
-
-    const uint8_t hm1  = 1 << (2*im);
-    const uint8_t hm2  = hm1 << 4;
-
-    uint16_t aux[4];
-    const uint8_t * sc = (const uint8_t *)aux;
-
-    uint16_t q16[8];
-    const uint8_t * q4 = (const uint8_t *)q16;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2) {
-
-        const uint8_t * ql1 = x[i].qs + q_offset;
-        const uint8_t * qh  = x[i].qh + l0;
-        const float   * y1  = yy + i*QK_K + y_offset;
-        const float   * y2  = y1 + 128;
-
-        const float dall = __low2half(x[i].dm);
-        const float dmin = __high2half(x[i].dm);
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux[0] = a[im+0] & kmask1;
-        aux[1] = a[im+2] & kmask1;
-        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
-        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
-
-        float4 sum = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        const uint16_t * q1 = (const uint16_t *)ql1;
-        const uint16_t * q2 = q1 + 32;
-        q16[0] = q1[0] & 0x0f0f;
-        q16[1] = q1[8] & 0x0f0f;
-        q16[2] = (q1[0] >> 4) & 0x0f0f;
-        q16[3] = (q1[8] >> 4) & 0x0f0f;
-        q16[4] = q2[0] & 0x0f0f;
-        q16[5] = q2[8] & 0x0f0f;
-        q16[6] = (q2[0] >> 4) & 0x0f0f;
-        q16[7] = (q2[8] >> 4) & 0x0f0f;
-        for (int l = 0; l < n; ++l) {
-            sum.x += y1[l+ 0] * (q4[l +0] + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
-                   + y1[l+16] * (q4[l +2] + (qh[l+16] & (hm1 << 0) ? 16 : 0));
-            sum.y += y1[l+32] * (q4[l +4] + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
-                   + y1[l+48] * (q4[l +6] + (qh[l+16] & (hm1 << 1) ? 16 : 0));
-            sum.z += y2[l+ 0] * (q4[l +8] + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
-                   + y2[l+16] * (q4[l+10] + (qh[l+16] & (hm2 << 0) ? 16 : 0));
-            sum.w += y2[l+32] * (q4[l+12] + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
-                   + y2[l+48] * (q4[l+14] + (qh[l+16] & (hm2 << 1) ? 16 : 0));
-            smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
-                  + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
-        }
-        tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q6_K * x = (const block_q6_K *)vx + ib0;
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
-
-    const int step = 16/K_QUANTS_PER_ITERATION;          // 16 or 8
-
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0...15 or 0...7
-
-#if K_QUANTS_PER_ITERATION == 1
-    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15
-    const int is = 0;
-#else
-    const int l0 = 4 * in;                               // 0, 4, 8, ..., 28
-    const int is = in / 4;
-#endif
-    const int ql_offset = 64*im + l0;
-    const int qh_offset = 32*im + l0;
-    const int s_offset  =  8*im + is;
-    const int y_offset = 128*im + l0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + y_offset;
-        const uint8_t * ql = x[i].ql + ql_offset;
-        const uint8_t * qh = x[i].qh + qh_offset;
-        const int8_t  * s  = x[i].scales + s_offset;
-
-        const float d = x[i].d;
-
-#if K_QUANTS_PER_ITERATION == 1
-        float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
-                  + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
-                  + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
-                  + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
-                  + y[64] * s[4] * d * ((int8_t)((ql[ 0]  >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
-                  + y[80] * s[5] * d * ((int8_t)((ql[16]  >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
-                  + y[96] * s[6] * d * ((int8_t)((ql[32]  >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
-                  +y[112] * s[7] * d * ((int8_t)((ql[48]  >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
-        tmp += sum;
-#else
-        float sum = 0;
-        for (int l = 0; l < 4; ++l) {
-            sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
-                 + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
-                 + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
-                 + y[l+96] * s[6] * d * ((int8_t)((ql[l+32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
-        }
-        tmp += sum;
-#endif
-
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (tid == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __device__ void convert_f16(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
-    const half * x = (const half *) vx;
-    // load 2 halfs into register in a single instruction
-    const half2 x_reg = *((half2 *) &(x[ib + iqs]));
-    // automatic half -> float type cast if dfloat == float
-    v.x = __low2float(x_reg);
-    v.y = __high2float(x_reg);
-}
-
-static constexpr __device__ dequantize_kernel_t get_dequantize_kernel(ggml_type type) {
-    return type == GGML_TYPE_Q4_0 ? dequantize_q4_0 :
-        type == GGML_TYPE_Q4_1 ? dequantize_q4_1 :
-        type == GGML_TYPE_Q5_0 ? dequantize_q5_0 :
-        type == GGML_TYPE_Q5_1 ? dequantize_q5_1 :
-        type == GGML_TYPE_Q8_0 ? dequantize_q8_0 :
-        type == GGML_TYPE_F16 ? convert_f16 :
-        nullptr;
-}
-
-template <ggml_type type>
-static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) {
-    constexpr int qk = ggml_cuda_type_traits<type>::qk; // quantized weights per x block
-    constexpr int qr = ggml_cuda_type_traits<type>::qr; // number of quantized weights per data value in x block
-    constexpr dequantize_kernel_t dequantize_kernel = get_dequantize_kernel(type);
-
-    const int64_t row = (int64_t)blockIdx.x*blockDim.y + threadIdx.y;
-
-    if (row >= nrows) {
-        return;
-    }
-
-    const int tid = threadIdx.x;
-
-    const int iter_stride = 2*GGML_CUDA_DMMV_X;
-    const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
-    const int y_offset = qr == 1 ? 1 : qk/2;
-
-// partial sum for each thread
-#ifdef GGML_CUDA_F16
-    half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics
-#else
-    float tmp = 0.0f;
-#endif // GGML_CUDA_F16
-
-    for (int i = 0; i < ncols; i += iter_stride) {
-        const int col = i + vals_per_iter*tid;
-        const int64_t ib = ((int64_t)row*ncols + col)/qk; // x block index
-        const int iqs = (col%qk)/qr; // x quant index
-        const int iybs = col - col%qk; // y block start index
-
-// processing >2 values per i iter is faster for fast GPUs
-#pragma unroll
-        for (int j = 0; j < vals_per_iter; j += 2) {
-            // process 2 vals per j iter
-
-            // dequantize
-            // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
-            dfloat2 v;
-            dequantize_kernel(vx, ib, iqs + j/qr, v);
-
-            // matrix multiplication
-            // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
-#ifdef GGML_CUDA_F16
-            if ( y_offset == 1 ) {
-                // load 2 dfloats into register in a single instruction
-                const dfloat2 y_reg = *((dfloat2 *) &(y[iybs + iqs + j/qr]));
-                tmp += __hmul2(v, y_reg);
-            }
-            else {
-                tmp += __hmul2(v, {
-                        y[iybs + iqs + j/qr + 0],
-                        y[iybs + iqs + j/qr + y_offset]
-                    });
-            }
-#else
-            if ( y_offset == 1 ) {
-                // load 2 dfloats into register in a single instruction
-                const dfloat2 y_reg = *((dfloat2 *) &(y[iybs + iqs + j/qr]));
-                tmp += v.x * y_reg.x;
-                tmp += v.y * y_reg.y;
-            }
-            else {
-                tmp += v.x * y[iybs + iqs + j/qr + 0];
-                tmp += v.y * y[iybs + iqs + j/qr + y_offset];
-            }
-#endif // GGML_CUDA_F16
-        }
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (tid == 0) {
-#ifdef GGML_CUDA_F16
-        dst[row] = tmp.x + tmp.y;
-#else
-        dst[row] = tmp;
-#endif // GGML_CUDA_F16
-    }
-}
-
-static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q4_0>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q4_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q5_0>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q5_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q8_0>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q2_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2 / K_QUANTS_PER_ITERATION;
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q3_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2 / K_QUANTS_PER_ITERATION;
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q4_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const dim3 block_dims(32, 1, 1);
-    dequantize_mul_mat_vec_q5_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
-}
-
-static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2 / K_QUANTS_PER_ITERATION;
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_F16>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-void ggml_cuda_op_dequantize_mul_mat_vec(
-    ggml_backend_cuda_context & ctx,
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream) {
-    GGML_UNUSED(ctx);
-    const int64_t ne00 = src0->ne[0];
-    const int64_t row_diff = row_high - row_low;
-
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
-#ifdef GGML_CUDA_F16
-    ggml_cuda_pool_alloc<half> src1_dfloat_a(ctx.pool());
-    half * src1_dfloat = nullptr; // dfloat == half
-
-    bool src1_convert_f16 =
-        src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
-        src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
-        src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
-
-    if (src1_convert_f16) {
-        src1_dfloat = src1_dfloat_a.alloc(ne00);
-        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
-        GGML_ASSERT(to_fp16_cuda != nullptr);
-        to_fp16_cuda(src1_ddf_i, src1_dfloat, ne00, stream);
-    }
-#else
-    const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion
-#endif // GGML_CUDA_F16
-
-    switch (src0->type) {
-        case GGML_TYPE_Q4_0:
-            dequantize_mul_mat_vec_q4_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            dequantize_mul_mat_vec_q4_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            dequantize_mul_mat_vec_q5_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            dequantize_mul_mat_vec_q5_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            dequantize_mul_mat_vec_q8_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q2_K:
-            dequantize_mul_mat_vec_q2_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q3_K:
-            dequantize_mul_mat_vec_q3_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q4_K:
-            dequantize_mul_mat_vec_q4_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_K:
-            dequantize_mul_mat_vec_q5_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q6_K:
-            dequantize_mul_mat_vec_q6_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_F16:
-            convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-            break;
-    }
-
-    GGML_UNUSED(src1);
-    GGML_UNUSED(dst);
-    GGML_UNUSED(src1_ddq_i);
-    GGML_UNUSED(src1_ncols);
-    GGML_UNUSED(src1_padded_row_size);
-}
-
-bool ggml_cuda_dmmv_type_supported(ggml_type src0_type) {
-    return src0_type == GGML_TYPE_Q4_0 || src0_type == GGML_TYPE_Q4_1 ||
-        src0_type == GGML_TYPE_Q5_0 || src0_type == GGML_TYPE_Q5_1 ||
-        src0_type == GGML_TYPE_Q8_0 || src0_type == GGML_TYPE_Q2_K ||
-        src0_type == GGML_TYPE_Q3_K || src0_type == GGML_TYPE_Q4_K ||
-        src0_type == GGML_TYPE_Q5_K || src0_type == GGML_TYPE_Q6_K ||
-        src0_type == GGML_TYPE_F16;
-}
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 07f043328a864..ef56e944d01ab 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -16,11 +16,11 @@
 #include "ggml-cuda/cpy.cuh"
 #include "ggml-cuda/cross-entropy-loss.cuh"
 #include "ggml-cuda/diagmask.cuh"
-#include "ggml-cuda/dmmv.cuh"
 #include "ggml-cuda/fattn.cuh"
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
 #include "ggml-cuda/mmq.cuh"
+#include "ggml-cuda/mmv.cuh"
 #include "ggml-cuda/mmvq.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
@@ -1020,114 +1020,6 @@ typedef void (*ggml_cuda_op_mul_mat_t)(
 
 #define MUL_MAT_SRC1_COL_STRIDE 128
 
-static __global__ void mul_mat_p021_f16_f32(
-    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y) {
-
-    const half * x = (const half *) vx;
-
-    const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
-    const int channel = blockDim.z*blockIdx.z + threadIdx.z;
-    const int channel_x = channel / (nchannels_y / nchannels_x);
-
-    const int nrows_y = ncols_x;
-    const int nrows_dst = nrows_x;
-    const int row_dst = row_x;
-
-    float tmp = 0.0f;
-
-    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
-        const int col_x = col_x0 + threadIdx.x;
-
-        if (col_x >= ncols_x) {
-            break;
-        }
-
-        // x is transposed and permuted
-        const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x;
-        const float xi = __half2float(x[ix]);
-
-        const int row_y = col_x;
-
-        // y is not transposed but permuted
-        const int iy = channel*nrows_y + row_y;
-
-        tmp += xi * y[iy];
-    }
-
-    // dst is not transposed and not permuted
-    const int idst = channel*nrows_dst + row_dst;
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[idst] = tmp;
-    }
-}
-
-static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
-    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
-    const int row_stride_x, const int channel_stride_x, const int channel_x_divisor) {
-
-    const half * x = (const half *) vx;
-
-    const int row_x     = blockDim.y*blockIdx.y + threadIdx.y;
-    const int channel   = blockDim.z*blockIdx.z + threadIdx.z;
-    const int channel_x = channel / channel_x_divisor;
-
-    const int nrows_y   = ncols_x;
-    const int nrows_dst = nrows_x;
-    const int row_dst   = row_x;
-
-    const int idst = channel*nrows_dst + row_dst;
-
-    float tmp = 0.0f;
-
-    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
-        const int col_x = col_x0 + threadIdx.x;
-
-        if (col_x >= ncols_x) {
-            break;
-        }
-
-        const int row_y = col_x;
-
-        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
-        const int iy = channel*nrows_y + row_y;
-
-        const float xi = __half2float(x[ix]);
-
-        tmp += xi * y[iy];
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[idst] = tmp;
-    }
-}
-
-static void ggml_mul_mat_p021_f16_f32_cuda(
-    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
-    const int nchannels_x, const int nchannels_y, cudaStream_t stream) {
-
-    const dim3 block_nums(1, nrows_x, nchannels_y);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y);
-}
-
-static void ggml_mul_mat_vec_nc_f16_f32_cuda(
-    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int row_stride_x,
-    const int nchannels_x, const int nchannels_y, const int channel_stride_x, cudaStream_t stream) {
-
-    const dim3 block_nums(1, nrows_x, nchannels_y);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
-        (vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y/nchannels_x);
-}
-
 static cudaError_t ggml_cuda_cpy_tensor_2d(
     void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
 
@@ -1654,58 +1546,6 @@ static void ggml_cuda_op_mul_mat(
     }
 }
 
-static void ggml_cuda_mul_mat_vec_p021(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
-    GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
-    GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-
-    const int64_t ne12 = src1->ne[2];
-
-    cudaStream_t main_stream = ctx.stream();
-
-    void  * src0_ddq = src0->data;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf  = (float *) dst->data;
-
-    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
-}
-
-static void ggml_cuda_mul_mat_vec_nc(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(!ggml_is_transposed(src0));
-    GGML_ASSERT(!ggml_is_transposed(src1));
-    GGML_ASSERT(!ggml_is_permuted(src0));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-
-    const int64_t nb01 = src0->nb[1];
-    const int64_t nb02 = src0->nb[2];
-
-    const int64_t ne12 = src1->ne[2];
-
-    cudaStream_t main_stream = ctx.stream();
-
-    void  * src0_ddq = src0->data;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf  = (float *) dst->data;
-
-    const int64_t row_stride_x = nb01 / sizeof(half);
-    const int64_t channel_stride_x = nb02 / sizeof(half);
-
-    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
-}
-
 static __global__ void k_compute_batched_ptrs(
         const half * src0_as_f16, const half * src1_as_f16, char * dst,
         const void ** ptrs_src, void ** ptrs_dst,
@@ -1879,21 +1719,17 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
-    bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
+    bool use_mul_mat_vec   = src0->type == GGML_TYPE_F16
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % (GGML_CUDA_DMMV_X*2) == 0 && src1->ne[1] == 1;
-    bool          use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
+        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
+    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
-    bool              use_mul_mat_q =  ggml_is_quantized(src0->type)
+    bool use_mul_mat_q     = ggml_is_quantized(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
-    // if mmvq is available it's a better choice than dmmv:
-#ifndef GGML_CUDA_FORCE_DMMV
-    use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
-#endif // GGML_CUDA_FORCE_DMMV
-
-    bool any_gpus_with_slow_fp16 = false;
+    bool any_gpus_with_slow_fp16   = false;
+    bool any_gpus_without_fp16_mma = false;
 
     if (split) {
         ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
@@ -1904,14 +1740,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
                 continue;
             }
 
-            const int cc            = ggml_cuda_info().devices[id].cc;
-            use_mul_mat_q           = use_mul_mat_q           && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
+            const int cc              = ggml_cuda_info().devices[id].cc;
+            use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+            any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_available(cc);
+            any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc);
         }
     } else {
-        const int cc            = ggml_cuda_info().devices[ctx.device].cc;
-        use_mul_mat_q           = use_mul_mat_q           && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
+        const int cc              = ggml_cuda_info().devices[ctx.device].cc;
+        use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+        any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_available(cc);
+        any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc);
     }
 
     // debug helpers
@@ -1922,18 +1760,14 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
-        // FP32 precision KQ single-batch for batch size 1 without FlashAttention
-        ggml_cuda_mul_mat_vec_p021(ctx, src0, src1, dst);
-    } else if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
-        // FP32 precision KQV single-batch for batch size 1 without FlashAttention
-        ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst);
+    if (!split && src0->type == GGML_TYPE_F16 && src1->ne[1] == 1 && dst->ne[3] == 1 && (src0->ne[1] < MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+        ggml_cuda_mul_mat_vec(ctx, src0, src1, dst);
     } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
                && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
-    } else if (use_dequantize_mul_mat_vec) {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, nullptr);
+    } else if (use_mul_mat_vec) {
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec, nullptr);
     } else if (use_mul_mat_vec_q) {
         ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, quantize_row_q8_1_cuda);
     } else if (use_mul_mat_q) {
diff --git a/ggml/src/ggml-cuda/mmv.cu b/ggml/src/ggml-cuda/mmv.cu
new file mode 100644
index 0000000000000..cfe91f4283fa6
--- /dev/null
+++ b/ggml/src/ggml-cuda/mmv.cu
@@ -0,0 +1,223 @@
+#include "common.cuh"
+#include "mmv.cuh"
+
+template <typename type_acc, int block_size>
+static __global__ void mul_mat_vec(
+        const half * __restrict__ x, const float * __restrict__ y, float * __restrict__ dst, const int64_t ncols2, const int64_t stride_row,
+        const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst) {
+    const int64_t row     = blockIdx.x;
+    const int64_t channel = blockIdx.z;
+    const int     tid     = threadIdx.x;
+
+    x   += (channel/channel_ratio)*stride_channel_x + row*stride_row;
+    y   +=  channel               *stride_channel_y;
+    dst +=  channel               *stride_channel_dst;
+
+    const half2  * x2 = (const half2  *) x;
+    const float2 * y2 = (const float2 *) y;
+
+    extern __shared__ char data_mmv[];
+    float * buf_iw = (float *) data_mmv;
+
+    if (block_size > WARP_SIZE) {
+        if (tid < WARP_SIZE) {
+            buf_iw[tid] = 0.0f;
+        }
+        __syncthreads();
+    }
+
+    float sumf;
+
+    if (std::is_same<type_acc, float>::value) {
+        sumf = 0.0f;
+
+        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmpx = __half22float2(x2[col2]);
+            const float2 tmpy = y2[col2];
+            sumf += tmpx.x * tmpy.x;
+            sumf += tmpx.y * tmpy.y;
+        }
+    } else {
+#ifdef FP16_AVAILABLE
+        half2 sumh2 = make_half2(0.0f, 0.0f);
+
+        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmp = y2[col2];
+            sumh2 += x2[col2] * make_half2(tmp.x, tmp.y);
+        }
+
+        sumf = __low2float(sumh2) + __high2float(sumh2);
+#else
+        NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+    }
+
+    sumf = warp_reduce_sum(sumf);
+
+    if (block_size > WARP_SIZE) {
+        buf_iw[tid/WARP_SIZE] = sumf;
+        __syncthreads();
+        if (tid > WARP_SIZE) {
+            return;
+        }
+        sumf = buf_iw[tid];
+        sumf = warp_reduce_sum(sumf);
+    }
+
+    if (tid != 0) {
+        return;
+    }
+
+    dst[row] = sumf;
+}
+
+template <typename type_acc>
+static void launch_mul_mat_vec_cuda(
+        const half * x, const float * y, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
+        cudaStream_t stream) {
+    GGML_ASSERT(ncols      % 2 == 0);
+    GGML_ASSERT(stride_row % 2 == 0);
+    GGML_ASSERT(nchannels_y % nchannels_x == 0);
+    const int64_t channel_ratio = nchannels_y / nchannels_x;
+
+    int64_t block_size_best = WARP_SIZE;
+    int64_t niter_best      = (ncols + 2*WARP_SIZE - 1) / (2*WARP_SIZE);
+    for (int64_t block_size = 2*WARP_SIZE; block_size <= 256; block_size += WARP_SIZE) {
+        const int64_t niter = (ncols + 2*block_size - 1) / (2*block_size);
+        if (niter < niter_best) {
+            niter_best      = niter;
+            block_size_best = block_size;
+        }
+    }
+
+    const int smem = WARP_SIZE*sizeof(float);
+    const dim3 block_nums(nrows, 1, nchannels_y);
+    const dim3 block_dims(block_size_best, 1, 1);
+    switch (block_size_best) {
+        case   32: {
+            mul_mat_vec<type_acc,  32><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case   64: {
+            mul_mat_vec<type_acc,  64><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case   96: {
+            mul_mat_vec<type_acc,  96><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  128: {
+            mul_mat_vec<type_acc, 128><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  160: {
+            mul_mat_vec<type_acc, 160><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  192: {
+            mul_mat_vec<type_acc, 192><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  224: {
+            mul_mat_vec<type_acc, 224><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  256: {
+            mul_mat_vec<type_acc, 256><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        default: {
+            GGML_ABORT("fatal error");
+        } break;
+    }
+}
+
+static void mul_mat_vec_cuda(
+        const half * x, const float * y, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
+        enum ggml_prec prec, cudaStream_t stream) {
+    switch (prec) {
+        case GGML_PREC_DEFAULT: {
+            launch_mul_mat_vec_cuda<half>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
+                stride_channel_x, stride_channel_y, stride_channel_dst, stream);
+        } break;
+        case GGML_PREC_F32: {
+            launch_mul_mat_vec_cuda<float>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
+                stride_channel_x, stride_channel_y, stride_channel_dst, stream);
+        } break;
+    }
+}
+
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+
+    GGML_ASSERT(src1->ne[1] == 1);
+
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+
+    const half  * src0_d = (const half  *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       *  dst_d = (float       *)  dst->data;
+
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne12 = src1->ne[2];
+    GGML_ASSERT(dst->ne[2] == ne12);
+
+    GGML_ASSERT(src0->ne[3] == 1);
+    GGML_ASSERT(src1->ne[3] == 1);
+    GGML_ASSERT( dst->ne[3] == 1);
+
+    const int64_t stride_row         = src0->nb[1] / ggml_type_size(src0->type);
+    const int64_t channel_stride_x   = src0->nb[2] / ggml_type_size(src0->type);
+    const int64_t channel_stride_y   = src1->nb[2] / ggml_type_size(src1->type);
+    const int64_t channel_stride_dst =  dst->nb[2] / ggml_type_size( dst->type);
+
+    mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, stride_row, ne02, ne12, channel_stride_x, channel_stride_y, channel_stride_dst, prec, ctx.stream());
+}
+
+void ggml_cuda_op_mul_mat_vec(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    GGML_ASSERT(src1_ncols == 1);
+
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+
+
+    // ggml_cuda_op provides single, contiguous matrices
+    const int64_t stride_row         = ne00;
+    const int64_t nchannels_x        = 1;
+    const int64_t nchannels_y        = 1;
+    const int64_t channel_stride_x   = 0;
+    const int64_t channel_stride_y   = 0;
+    const int64_t channel_stride_dst = 0;
+
+    mul_mat_vec_cuda((const half *) src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
+        nchannels_x, nchannels_y, channel_stride_x, channel_stride_y, channel_stride_dst, prec, stream);
+
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddq_i);
+    GGML_UNUSED(src1_ncols);
+    GGML_UNUSED(src1_padded_row_size);
+}
diff --git a/ggml/src/ggml-cuda/dmmv.cuh b/ggml/src/ggml-cuda/mmv.cuh
similarity index 55%
rename from ggml/src/ggml-cuda/dmmv.cuh
rename to ggml/src/ggml-cuda/mmv.cuh
index e727eb97f6aad..78a1cd4a6906a 100644
--- a/ggml/src/ggml-cuda/dmmv.cuh
+++ b/ggml/src/ggml-cuda/mmv.cuh
@@ -1,20 +1,12 @@
 #include "common.cuh"
 
-// dmmv = dequantize_mul_mat_vec
+// maximum number of src0 rows with which to use mul_mat_vec over cuBLAS if FP16 tensor cores are available
+#define MMV_MAX_ROWS 512
 
-// TODO: remove this?
-#ifndef GGML_CUDA_DMMV_X
-#define GGML_CUDA_DMMV_X 32
-#endif
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
 
-#ifndef GGML_CUDA_MMV_Y
-#define GGML_CUDA_MMV_Y 1
-#endif
-
-void ggml_cuda_op_dequantize_mul_mat_vec(
+void ggml_cuda_op_mul_mat_vec(
     ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, cudaStream_t stream);
-
-bool ggml_cuda_dmmv_type_supported(ggml_type src0_type);
diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt
index 5ed186ded16be..fccf8eb8440b8 100644
--- a/ggml/src/ggml-hip/CMakeLists.txt
+++ b/ggml/src/ggml-hip/CMakeLists.txt
@@ -75,18 +75,11 @@ target_include_directories(ggml-hip PRIVATE . ..)
 target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
 
 add_compile_definitions(GGML_USE_HIP)
-add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
-add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
-add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
 
 if (GGML_HIP_UMA)
     add_compile_definitions(GGML_HIP_UMA)
 endif()
 
-if (GGML_CUDA_FORCE_DMMV)
-    add_compile_definitions(GGML_CUDA_FORCE_DMMV)
-endif()
-
 if (GGML_CUDA_FORCE_MMQ)
     add_compile_definitions(GGML_CUDA_FORCE_MMQ)
 endif()
diff --git a/ggml/src/ggml-musa/CMakeLists.txt b/ggml/src/ggml-musa/CMakeLists.txt
index 8edc75cc5a9c7..f3c0136920540 100644
--- a/ggml/src/ggml-musa/CMakeLists.txt
+++ b/ggml/src/ggml-musa/CMakeLists.txt
@@ -58,19 +58,12 @@ if (MUSAToolkit_FOUND)
     target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
 
     add_compile_definitions(GGML_USE_MUSA)
-    add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
-    add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
-    add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
     add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
 
     if (GGML_CUDA_GRAPHS)
         add_compile_definitions(GGML_CUDA_USE_GRAPHS)
     endif()
 
-    if (GGML_CUDA_FORCE_DMMV)
-        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
-    endif()
-
     if (GGML_CUDA_FORCE_MMQ)
         add_compile_definitions(GGML_CUDA_FORCE_MMQ)
     endif()
@@ -83,10 +76,6 @@ if (MUSAToolkit_FOUND)
         add_compile_definitions(GGML_CUDA_NO_VMM)
     endif()
 
-    if (DEFINED GGML_CUDA_DMMV_Y)
-        add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_DMMV_Y}) # for backwards compatibility
-    endif()
-
     if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
         add_compile_definitions(GGML_CUDA_F16)
     endif()

From a43178299c2f74f14bcfc659bfd9fd32d931d1f4 Mon Sep 17 00:00:00 2001
From: FirstTimeEZ <179362031+FirstTimeEZ@users.noreply.github.com>
Date: Sun, 17 Nov 2024 21:39:22 +1300
Subject: [PATCH 017/102] ggml : fix undefined reference to 'getcpu' (#10354)

https://github.com/ggerganov/llama.cpp/issues/10352
---
 ggml/src/ggml-cpu/ggml-cpu.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
index df64879294981..0d23669c25bea 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -2369,7 +2369,7 @@ void ggml_numa_init(enum ggml_numa_strategy numa_flag) {
     // figure out which node we're on
     uint current_cpu;
     int getcpu_ret = 0;
-#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 28) || defined(__COSMOPOLITAN__)
+#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 33) || defined(__COSMOPOLITAN__)
     getcpu_ret = getcpu(&current_cpu, &g_state.numa.current_node);
 #else
     // old glibc doesn't have a wrapper for this call. Fall back on direct syscall

From cf32a9b93ad859ea592c31785b6bd3b4b2121463 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sun, 17 Nov 2024 11:23:01 +0200
Subject: [PATCH 018/102] metal : refactor kernel args into structs (#10238)

* metal : add kernel arg structs (wip)

* metal : fattn args

ggml-ci

* metal : cont + avoid potential int overflow [no ci]

* metal : mul mat struct (wip)

* cont : mul mat vec

* cont : pass by reference

* cont : args is first argument

* cont : use char ptr

* cont : shmem style

* cont : thread counters style

* cont : mul mm id

ggml-ci

* cont : int safety + register optimizations

ggml-ci

* metal : GGML_OP_CONCAT

ggml-ci

* metal : GGML_OP_ADD, GGML_OP_SUB, GGML_OP_MUL, GGML_OP_DIV

* metal : GGML_OP_REPEAT

* metal : GGML_OP_CPY

* metal : GGML_OP_RMS_NORM

* metal : GGML_OP_NORM

* metal : add TODOs for rest of ops

* ggml : add ggml-metal-impl.h

ggml-ci
---
 Makefile                              |    5 +-
 ggml/src/ggml-metal/CMakeLists.txt    |   18 +-
 ggml/src/ggml-metal/ggml-metal-impl.h |  249 ++
 ggml/src/ggml-metal/ggml-metal.m      |  681 +++--
 ggml/src/ggml-metal/ggml-metal.metal  | 3592 +++++++++----------------
 5 files changed, 1949 insertions(+), 2596 deletions(-)
 create mode 100644 ggml/src/ggml-metal/ggml-metal-impl.h

diff --git a/Makefile b/Makefile
index bd0bad160a1f2..95110d4eb7d8b 100644
--- a/Makefile
+++ b/Makefile
@@ -906,6 +906,7 @@ endif # GGML_METAL
 ifdef GGML_METAL
 ggml/src/ggml-metal/ggml-metal.o: \
 	ggml/src/ggml-metal/ggml-metal.m \
+	ggml/src/ggml-metal/ggml-metal-impl.h \
 	ggml/include/ggml-metal.h \
 	ggml/include/ggml.h
 	$(CC) $(CFLAGS) -c $< -o $@
@@ -913,9 +914,11 @@ ggml/src/ggml-metal/ggml-metal.o: \
 ifdef GGML_METAL_EMBED_LIBRARY
 ggml/src/ggml-metal-embed.o: \
 	ggml/src/ggml-metal/ggml-metal.metal \
+	ggml/src/ggml-metal/ggml-metal-impl.h \
 	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
-	@sed -e '/__embed_ggml-common.h__/r ggml/src/ggml-common.h' -e '/__embed_ggml-common.h__/d' < ggml/src/ggml-metal/ggml-metal.metal > ggml/src/ggml-metal/ggml-metal-embed.metal
+	@sed -e '/__embed_ggml-common.h__/r      ggml/src/ggml-common.h'                -e '/__embed_ggml-common.h__/d'      < ggml/src/ggml-metal/ggml-metal.metal           > ggml/src/ggml-metal/ggml-metal-embed.metal.tmp
+	@sed -e '/#include "ggml-metal-impl.h"/r ggml/src/ggml-metal/ggml-metal-impl.h' -e '/#include "ggml-metal-impl.h"/d' < ggml/src/ggml-metal/ggml-metal-embed.metal.tmp > ggml/src/ggml-metal/ggml-metal-embed.metal
 	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
 	@echo ".section __DATA, __ggml_metallib"                       >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
 	@echo ".globl _ggml_metallib_start"                            >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
diff --git a/ggml/src/ggml-metal/CMakeLists.txt b/ggml/src/ggml-metal/CMakeLists.txt
index e0992c7449b62..b237d79f47ddb 100644
--- a/ggml/src/ggml-metal/CMakeLists.txt
+++ b/ggml/src/ggml-metal/CMakeLists.txt
@@ -25,9 +25,10 @@ if (GGML_METAL_USE_BF16)
     add_compile_definitions(GGML_METAL_USE_BF16)
 endif()
 
-# copy ggml-common.h and ggml-metal.metal to bin directory
-configure_file(../ggml-common.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h    COPYONLY)
-configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
+# copy metal files to bin directory
+configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
+configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
+configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)
 
 if (GGML_METAL_EMBED_LIBRARY)
     enable_language(ASM)
@@ -36,24 +37,27 @@ if (GGML_METAL_EMBED_LIBRARY)
 
     set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h")
     set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+    set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
 
     file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
 
     # merge ggml-common.h and ggml-metal.metal into a single file
-    set(METALLIB_EMBED_ASM    "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-    set(METALLIB_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
+    set(METALLIB_EMBED_ASM        "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
+    set(METALLIB_SOURCE_EMBED     "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
+    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
 
     add_custom_command(
         OUTPUT ${METALLIB_EMBED_ASM}
         COMMAND echo "Embedding Metal library"
-        COMMAND sed -e '/__embed_ggml-common.h__/r ${METALLIB_COMMON}' -e '/__embed_ggml-common.h__/d' < ${METALLIB_SOURCE} > ${METALLIB_SOURCE_EMBED}
+        COMMAND sed -e '/__embed_ggml-common.h__/r         ${METALLIB_COMMON}' -e '/__embed_ggml-common.h__/d'         < ${METALLIB_SOURCE}           > ${METALLIB_SOURCE_EMBED_TMP}
+        COMMAND sed -e '/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}'   -e '/\#include \"ggml-metal-impl.h\"/d' < ${METALLIB_SOURCE_EMBED_TMP} > ${METALLIB_SOURCE_EMBED}
         COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
         COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
         COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
         COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
         COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
         COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
-        DEPENDS ggml-metal.metal ../ggml-common.h
+        DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
         COMMENT "Generate assembly for embedded Metal library"
     )
 
diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h
new file mode 100644
index 0000000000000..53c13549650c8
--- /dev/null
+++ b/ggml/src/ggml-metal/ggml-metal-impl.h
@@ -0,0 +1,249 @@
+#ifndef GGML_METAL_IMPL
+#define GGML_METAL_IMPL
+
+// kernel argument structs
+//
+// - element counters (e.g. ne00) typically use int32_t to reduce register usage
+//   however, be careful from int overflows when using those in the kernel implementation
+//
+// - strides (e.g. nb00) use uint64_t
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  dim;
+} ggml_metal_kargs_concat;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    uint64_t offs;
+} ggml_metal_kargs_bin;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_repeat;
+
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int64_t  ne0;
+    int64_t  ne1;
+    int64_t  ne2;
+    int64_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_cpy;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  n_past;
+    int32_t  n_dims;
+    int32_t  n_ctx_orig;
+    float    freq_base;
+    float    freq_scale;
+    float    ext_factor;
+    float    attn_factor;
+    float    beta_fast;
+    float    beta_slow;
+} ggml_metal_kargs_rope;
+
+typedef struct {
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne11;
+    int32_t  ne_12_2; // assume K and V are same shape
+    int32_t  ne_12_3;
+    uint64_t nb_12_1;
+    uint64_t nb_12_2;
+    uint64_t nb_12_3;
+    uint64_t nb31;
+    int32_t  ne1;
+    int32_t  ne2;
+    float    scale;
+    float    max_bias;
+    float    m0;
+    float    m1;
+    uint16_t n_head_log2;
+    float    logit_softcap;
+} ggml_metal_kargs_flash_attn_ext;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne02;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne12;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int16_t  r2;
+    int16_t  r3;
+} ggml_metal_kargs_mul_mm;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int16_t  r2;
+    int16_t  r3;
+} ggml_metal_kargs_mul_mv;
+
+typedef struct {
+    int32_t  nei0;
+    int32_t  nei1;
+    uint64_t nbi1;
+    int32_t  ne00;
+    int32_t  ne02;
+    uint64_t nb01;
+    uint64_t nb02;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    int32_t  ne0;
+    int32_t  ne1;
+} ggml_metal_kargs_mul_mm_id;
+
+typedef struct {
+    int32_t  nei0;
+    int32_t  nei1;
+    uint64_t nbi1;
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    int32_t  ne0;
+    int32_t  ne1;
+    uint64_t nb1;
+} ggml_metal_kargs_mul_mv_id;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne00_4;
+    uint64_t nb01;
+    float    eps;
+} ggml_metal_kargs_norm;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne00_4;
+    uint64_t nb01;
+    float    eps;
+} ggml_metal_kargs_rms_norm;
+
+#endif // GGML_METAL_IMPL
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index 95b21fbf9c503..58fee4bfd1296 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -2,6 +2,7 @@
 
 #import "ggml-impl.h"
 #import "ggml-backend-impl.h"
+#import "ggml-metal-impl.h"
 
 #import <Foundation/Foundation.h>
 
@@ -1193,35 +1194,39 @@ static void ggml_metal_encode_node(
 
                 const int32_t dim = ((const int32_t *) dst->op_params)[0];
 
+                ggml_metal_kargs_concat args = {
+                    /*.ne00 =*/ ne00,
+                    /*.ne01 =*/ ne01,
+                    /*.ne02 =*/ ne02,
+                    /*.ne03 =*/ ne03,
+                    /*.nb00 =*/ nb00,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb03 =*/ nb03,
+                    /*.ne10 =*/ ne10,
+                    /*.ne11 =*/ ne11,
+                    /*.ne12 =*/ ne12,
+                    /*.ne13 =*/ ne13,
+                    /*.nb10 =*/ nb10,
+                    /*.nb11 =*/ nb11,
+                    /*.nb12 =*/ nb12,
+                    /*.nb13 =*/ nb13,
+                    /*.ne0  =*/ ne0,
+                    /*.ne1  =*/ ne1,
+                    /*.ne2  =*/ ne2,
+                    /*.ne3  =*/ ne3,
+                    /*.nb0  =*/ nb0,
+                    /*.nb1  =*/ nb1,
+                    /*.nb2  =*/ nb2,
+                    /*.nb3  =*/ nb3,
+                    /*.dim  =*/ dim,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
-                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
-                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
-                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
-                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
-                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
-                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
-                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
-                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
-                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
-                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
-                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
-                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
-                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
-                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
-                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
-                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
-                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
-                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
-                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
-                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
-                [encoder setBytes:&dim  length:sizeof(dim)  atIndex:27];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
 
                 const int nth = MIN(1024, ne0);
 
@@ -1239,8 +1244,6 @@ static void ggml_metal_encode_node(
 
                 bool bcast_row = false;
 
-                int64_t nb = ne00; // used by the "row" kernels
-
                 id<MTLComputePipelineState> pipeline = nil;
 
                 if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
@@ -1249,7 +1252,6 @@ static void ggml_metal_encode_node(
                     // src1 is a row
                     GGML_ASSERT(ne11 == 1);
 
-                    nb = ne00 / 4;
                     switch (dst->op) {
                         case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break;
                         case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB_ROW].pipeline; break;
@@ -1269,36 +1271,39 @@ static void ggml_metal_encode_node(
                     }
                 }
 
+                ggml_metal_kargs_bin args = {
+                    /*.ne00 =*/ ne00,
+                    /*.ne01 =*/ ne01,
+                    /*.ne02 =*/ ne02,
+                    /*.ne03 =*/ ne03,
+                    /*.nb00 =*/ nb00,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb03 =*/ nb03,
+                    /*.ne10 =*/ ne10,
+                    /*.ne11 =*/ ne11,
+                    /*.ne12 =*/ ne12,
+                    /*.ne13 =*/ ne13,
+                    /*.nb10 =*/ nb10,
+                    /*.nb11 =*/ nb11,
+                    /*.nb12 =*/ nb12,
+                    /*.nb13 =*/ nb13,
+                    /*.ne0  =*/ ne0,
+                    /*.ne1  =*/ ne1,
+                    /*.ne2  =*/ ne2,
+                    /*.ne3  =*/ ne3,
+                    /*.nb0  =*/ nb0,
+                    /*.nb1  =*/ nb1,
+                    /*.nb2  =*/ nb2,
+                    /*.nb3  =*/ nb3,
+                    /*.offs =*/ offs,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
-                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
-                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
-                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
-                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
-                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
-                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
-                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
-                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
-                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
-                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
-                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
-                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
-                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
-                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
-                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
-                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
-                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
-                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
-                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
-                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
-                [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
-                [encoder setBytes:&nb   length:sizeof(nb)   atIndex:28];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
 
                 if (bcast_row) {
                     const int64_t n = ggml_nelements(dst)/4;
@@ -1322,25 +1327,29 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("fatal error");
                 }
 
+                ggml_metal_kargs_repeat args = {
+                    /*.ne00 =*/ ne00,
+                    /*.ne01 =*/ ne01,
+                    /*.ne02 =*/ ne02,
+                    /*.ne03 =*/ ne03,
+                    /*.nb00 =*/ nb00,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb03 =*/ nb03,
+                    /*.ne0  =*/ ne0,
+                    /*.ne1  =*/ ne1,
+                    /*.ne2  =*/ ne2,
+                    /*.ne3  =*/ ne3,
+                    /*.nb0  =*/ nb0,
+                    /*.nb1  =*/ nb1,
+                    /*.nb2  =*/ nb2,
+                    /*.nb3  =*/ nb3,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
-                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
-                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
-                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
-                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
-                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
-                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
-                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
-                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
-                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
-                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
-                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
-                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
-                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
-                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
 
                 const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
 
@@ -1369,25 +1378,29 @@ static void ggml_metal_encode_node(
 
                     const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline;
 
+                    ggml_metal_kargs_cpy args = {
+                        /*.ne00 =*/ ne00,
+                        /*.ne01 =*/ ne01,
+                        /*.ne02 =*/ ne02,
+                        /*.ne03 =*/ ne03,
+                        /*.nb00 =*/ nb00,
+                        /*.nb01 =*/ nb01,
+                        /*.nb02 =*/ nb02,
+                        /*.nb03 =*/ nb03,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                        /*.ne2  =*/ ne2,
+                        /*.ne3  =*/ ne3,
+                        /*.nb0  =*/ nb0,
+                        /*.nb1  =*/ nb1,
+                        /*.nb2  =*/ nb2,
+                        /*.nb3  =*/ nb3,
+                    };
+
                     [encoder setComputePipelineState:pipeline];
-                    [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                    [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                    [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                    [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
-                    [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
-                    [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
-                    [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
-                    [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
-                    [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
-                    [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
-                    [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
-                    [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
-                    [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
-                    [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
-                    [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
-                    [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
-                    [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
-                    [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
+                    [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
 
                     const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
 
@@ -1396,35 +1409,39 @@ static void ggml_metal_encode_node(
 
                 const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline;
 
+                ggml_metal_kargs_bin args = {
+                    /*.ne00 =*/ ne00,
+                    /*.ne01 =*/ ne01,
+                    /*.ne02 =*/ ne02,
+                    /*.ne03 =*/ ne03,
+                    /*.nb00 =*/ nb00,
+                    /*.nb01 =*/ pnb1,
+                    /*.nb02 =*/ pnb2,
+                    /*.nb03 =*/ pnb3,
+                    /*.ne10 =*/ ne10,
+                    /*.ne11 =*/ ne11,
+                    /*.ne12 =*/ ne12,
+                    /*.ne13 =*/ ne13,
+                    /*.nb10 =*/ nb10,
+                    /*.nb11 =*/ nb11,
+                    /*.nb12 =*/ nb12,
+                    /*.nb13 =*/ nb13,
+                    /*.ne0  =*/ ne0,
+                    /*.ne1  =*/ ne1,
+                    /*.ne2  =*/ ne2,
+                    /*.ne3  =*/ ne3,
+                    /*.nb0  =*/ nb0,
+                    /*.nb1  =*/ pnb1,
+                    /*.nb2  =*/ pnb2,
+                    /*.nb3  =*/ pnb3,
+                    /*.offs =*/ offs,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
-                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
-                [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:8];
-                [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:9];
-                [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:10];
-                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
-                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
-                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
-                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
-                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
-                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
-                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
-                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
-                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
-                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
-                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
-                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
-                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
-                [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:24];
-                [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:25];
-                [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:26];
-                [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
 
                 const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
 
@@ -1465,10 +1482,10 @@ static void ggml_metal_encode_node(
                 memcpy(&max, ((const int32_t *) dst->op_params) + 1, sizeof(float));
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
-                [encoder setBytes:&min length:sizeof(min) atIndex:2];
-                [encoder setBytes:&max length:sizeof(max) atIndex:3];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&min   length:sizeof(min) atIndex:2];
+                [encoder setBytes:&max   length:sizeof(max) atIndex:3];
 
                 const int64_t n = ggml_nelements(dst);
 
@@ -1640,6 +1657,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -1715,6 +1733,8 @@ static void ggml_metal_encode_node(
                 const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
                 const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
+                // TODO: add ggml_metal_kargs struct
+                // TODO: optimize (see https://github.com/ggerganov/llama.cpp/pull/10238/commits/7941b6b9ec29a2866fec6fa6c51612515ca509f6)
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
                 if (id_src1) {
@@ -1731,6 +1751,7 @@ static void ggml_metal_encode_node(
                 [encoder setBytes:&m0          length:sizeof(m0)          atIndex:8];
                 [encoder setBytes:&m1          length:sizeof(m1)          atIndex:9];
                 [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:10];
+
                 [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                 [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
@@ -1747,6 +1768,7 @@ static void ggml_metal_encode_node(
                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF].pipeline;
                 }
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -1771,6 +1793,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
                 [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
@@ -1841,6 +1864,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
@@ -1959,24 +1983,29 @@ static void ggml_metal_encode_node(
                                 default: GGML_ABORT("MUL MAT-MAT not implemented");
                             }
 
+                            ggml_metal_kargs_mul_mm args = {
+                                /*.ne00 =*/ ne00,
+                                /*.ne02 =*/ ne02,
+                                /*.nb01 =*/ nb01,
+                                /*.nb02 =*/ nb02,
+                                /*.nb03 =*/ nb03,
+                                /*.ne12 =*/ ne12,
+                                /*.nb10 =*/ nb10,
+                                /*.nb11 =*/ nb11,
+                                /*.nb12 =*/ nb12,
+                                /*.nb13 =*/ nb13,
+                                /*.ne0  =*/ ne0,
+                                /*.ne1  =*/ ne1,
+                                /*.r2   =*/ r2,
+                                /*.r3   =*/ r3,
+                            };
+
                             [encoder setComputePipelineState:pipeline];
-                            [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
-                            [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
-                            [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
-                            [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
-                            [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:4];
-                            [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
-                            [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
-                            [encoder setBytes:&nb03    length:sizeof(nb03) atIndex:7];
-                            [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:8];
-                            [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:9];
-                            [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:10];
-                            [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:11];
-                            [encoder setBytes:&nb13    length:sizeof(nb13) atIndex:12];
-                            [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:13];
-                            [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
-                            [encoder setBytes:&r2      length:sizeof(r2)   atIndex:15];
-                            [encoder setBytes:&r3      length:sizeof(r3)   atIndex:16];
+                            [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                            [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                            [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                            [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+
                             [encoder setThreadgroupMemoryLength:8192 atIndex:0];
                             [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
                         } else {
@@ -2154,28 +2183,32 @@ static void ggml_metal_encode_node(
                                     }
                             };
 
+                            ggml_metal_kargs_mul_mv args = {
+                                /*.ne00 =*/ ne00,
+                                /*.ne01 =*/ ne01,
+                                /*.ne02 =*/ ne02,
+                                /*.nb00 =*/ nb00,
+                                /*.nb01 =*/ nb01,
+                                /*.nb02 =*/ nb02,
+                                /*.nb03 =*/ nb03,
+                                /*.ne10 =*/ ne10,
+                                /*.ne11 =*/ ne11,
+                                /*.ne12 =*/ ne12,
+                                /*.nb10 =*/ nb10,
+                                /*.nb11 =*/ nb11,
+                                /*.nb12 =*/ nb12,
+                                /*.nb13 =*/ nb13,
+                                /*.ne0  =*/ ne0,
+                                /*.ne1  =*/ ne1,
+                                /*.r2   =*/ r2,
+                                /*.r3   =*/ r3,
+                            };
+
                             [encoder setComputePipelineState:pipeline];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
-                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
-                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
-                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
-                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10];
-                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
-                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
-                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:13];
-                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:14];
-                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:15];
-                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:16];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:17];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:18];
-                            [encoder setBytes:&r2   length:sizeof(r2)   atIndex:19];
-                            [encoder setBytes:&r3   length:sizeof(r3)   atIndex:20];
+                            [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
 
                             if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
                                 src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
@@ -2288,27 +2321,30 @@ static void ggml_metal_encode_node(
                         default: GGML_ABORT("MUL_MAT_ID not implemented");
                     }
 
+                    ggml_metal_kargs_mul_mm_id args = {
+                        /*.nei0 =*/ ne20,
+                        /*.nei1 =*/ ne21,
+                        /*.nbi1 =*/ nb21,
+                        /*.ne00 =*/ ne00,
+                        /*.ne02 =*/ ne02,
+                        /*.nb01 =*/ nb01,
+                        /*.nb02 =*/ nb02,
+                        /*.ne11 =*/ ne11,
+                        /*.ne12 =*/ ne12,
+                        /*.ne13 =*/ ne13,
+                        /*.nb10 =*/ nb10,
+                        /*.nb11 =*/ nb11,
+                        /*.nb12 =*/ nb12,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                    };
+
                     [encoder setComputePipelineState:pipeline];
-                    [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
-                    [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
-                    [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
-                    [encoder setBuffer:id_src2 offset:offs_src2    atIndex:3];
-                    [encoder setBytes:&ne20    length:sizeof(ne20) atIndex:4];
-                    [encoder setBytes:&ne21    length:sizeof(ne21) atIndex:5];
-                    [encoder setBytes:&nb21    length:sizeof(nb21) atIndex:6];
-                    [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:7];
-                    [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:8];
-                    [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:9];
-                    [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:10];
-                    [encoder setBytes:&ne11    length:sizeof(ne11) atIndex:11];
-                    [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:12];
-                    [encoder setBytes:&ne13    length:sizeof(ne13) atIndex:13];
-                    [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:14];
-                    [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:15];
-                    [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:16];
-                    [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:17];
-                    [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:18];
-                    [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:19];
+                    [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+                    [encoder setBuffer:id_src2 offset:offs_src2    atIndex:4];
 
                     [encoder setThreadgroupMemoryLength:GGML_PAD(8192 + dst_rows*4/*sizeof(ushort2)*/, 16) atIndex:0];
 
@@ -2467,30 +2503,34 @@ static void ggml_metal_encode_node(
                         GGML_ASSERT(ne00 >= nth0*nth1);
                     }
 
+                    ggml_metal_kargs_mul_mv_id args = {
+                        /*.nei0 =*/ ne20,
+                        /*.nei1 =*/ ne21,
+                        /*.nbi1 =*/ nb21,
+                        /*.ne00 =*/ ne00,
+                        /*.ne01 =*/ ne01,
+                        /*.ne02 =*/ ne02,
+                        /*.nb00 =*/ nb00,
+                        /*.nb01 =*/ nb01,
+                        /*.nb02 =*/ nb02,
+                        /*.ne10 =*/ ne10,
+                        /*.ne11 =*/ ne11,
+                        /*.ne12 =*/ ne12,
+                        /*.ne13 =*/ ne13,
+                        /*.nb10 =*/ nb10,
+                        /*.nb11 =*/ nb11,
+                        /*.nb12 =*/ nb12,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                        /*.nb1  =*/ nb1,
+                    };
+
                     [encoder setComputePipelineState:pipeline];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                    [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
-                    [encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
-                    [encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
-                    [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6];
-                    [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7];
-                    [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:8];
-                    [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:9];
-                    [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:10];
-                    [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:11];
-                    [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:12];
-                    [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:13];
-                    [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:14];
-                    [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:15];
-                    [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:16];
-                    [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:17];
-                    [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:18];
-                    [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:19];
-                    [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:20];
-                    [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:21];
-                    [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:22];
+                    [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
+                    [encoder setBuffer:id_src2 offset:offs_src2 atIndex:4];
 
                     const int64_t _ne1 = 1;
                     const int tgz = dst_rows;
@@ -2563,6 +2603,7 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("not implemented");
                 }
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0     offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_src1     offset:offs_src1 atIndex:1];
@@ -2586,20 +2627,28 @@ static void ggml_metal_encode_node(
                 float eps;
                 memcpy(&eps, dst->op_params, sizeof(float));
 
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM].pipeline;
+
                 int nth = 32; // SIMD width
 
-                while (nth < ne00/4 && nth < 1024) {
+                while (nth < ne00/4 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
                     nth *= 2;
                 }
 
-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM].pipeline;
+                nth = MIN(nth, ne00/4);
+
+                ggml_metal_kargs_rms_norm args = {
+                    /*.ne00   =*/ ne00,
+                    /*.ne00_4 =*/ ne00/4,
+                    /*.nb01   =*/ nb01,
+                    /*.eps    =*/ eps,
+                };
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
-                [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+
                 [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                 const int64_t nrows = ggml_nrows(src0);
@@ -2624,6 +2673,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GROUP_NORM].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0  offset:offs_src0        atIndex:0];
                 [encoder setBuffer:id_dst   offset:offs_dst         atIndex:1];
@@ -2641,22 +2691,35 @@ static void ggml_metal_encode_node(
             } break;
         case GGML_OP_NORM:
             {
+                GGML_ASSERT(ne00 % 4 == 0);
                 GGML_ASSERT(ggml_is_contiguous_1(src0));
 
                 float eps;
                 memcpy(&eps, dst->op_params, sizeof(float));
 
-                const int nth = MIN(256, ne00);
-
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_NORM].pipeline;
 
+                int nth = 32; // SIMD width
+
+                while (nth < ne00/4 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                    nth *= 2;
+                }
+
+                nth = MIN(nth, ne00/4);
+
+                ggml_metal_kargs_norm args = {
+                    /*.ne00   =*/ ne00,
+                    /*.ne00_4 =*/ ne00/4,
+                    /*.nb01   =*/ nb01,
+                    /*.eps    =*/ eps,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
-                [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
-                [encoder setThreadgroupMemoryLength:GGML_PAD(nth*sizeof(float), 16) atIndex:0];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+
+                [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                 const int64_t nrows = ggml_nrows(src0);
 
@@ -2706,40 +2769,44 @@ static void ggml_metal_encode_node(
                     };
                 }
 
+                ggml_metal_kargs_rope args = {
+                    /*.ne00        =*/ ne00,
+                    /*.ne01        =*/ ne01,
+                    /*.ne02        =*/ ne02,
+                    /*.ne03        =*/ ne03,
+                    /*.nb00        =*/ nb00,
+                    /*.nb01        =*/ nb01,
+                    /*.nb02        =*/ nb02,
+                    /*.nb03        =*/ nb03,
+                    /*.ne0         =*/ ne0,
+                    /*.ne1         =*/ ne1,
+                    /*.ne2         =*/ ne2,
+                    /*.ne3         =*/ ne3,
+                    /*.nb0         =*/ nb0,
+                    /*.nb1         =*/ nb1,
+                    /*.nb2         =*/ nb2,
+                    /*.nb3         =*/ nb3,
+                    /*.n_past      =*/ n_past,
+                    /*.n_dims      =*/ n_dims,
+                    /*.n_ctx_orig  =*/ n_ctx_orig,
+                    /*.freq_base   =*/ freq_base,
+                    /*.freq_scale  =*/ freq_scale,
+                    /*.ext_factor  =*/ ext_factor,
+                    /*.attn_factor =*/ attn_factor,
+                    /*.beta_fast   =*/ beta_fast,
+                    /*.beta_slow   =*/ beta_slow,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
-                [encoder setBuffer:id_src1     offset:offs_src1        atIndex:1];
+                [encoder setBytes:&args length:sizeof(args)     atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0     atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1     atIndex:2];
                 if (id_src2 != nil) {
-                    [encoder setBuffer:id_src2 offset:offs_src2        atIndex:2];
+                    [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
                 } else {
-                    [encoder setBuffer:id_src0 offset:offs_src0        atIndex:2];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:3];
                 }
-                [encoder setBuffer:id_dst      offset:offs_dst         atIndex:3];
-                [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:4];
-                [encoder setBytes:&ne01        length:sizeof( int64_t) atIndex:5];
-                [encoder setBytes:&ne02        length:sizeof( int64_t) atIndex:6];
-                [encoder setBytes:&ne03        length:sizeof( int64_t) atIndex:7];
-                [encoder setBytes:&nb00        length:sizeof(uint64_t) atIndex:8];
-                [encoder setBytes:&nb01        length:sizeof(uint64_t) atIndex:9];
-                [encoder setBytes:&nb02        length:sizeof(uint64_t) atIndex:10];
-                [encoder setBytes:&nb03        length:sizeof(uint64_t) atIndex:11];
-                [encoder setBytes:&ne0         length:sizeof( int64_t) atIndex:12];
-                [encoder setBytes:&ne1         length:sizeof( int64_t) atIndex:13];
-                [encoder setBytes:&ne2         length:sizeof( int64_t) atIndex:14];
-                [encoder setBytes:&ne3         length:sizeof( int64_t) atIndex:15];
-                [encoder setBytes:&nb0         length:sizeof(uint64_t) atIndex:16];
-                [encoder setBytes:&nb1         length:sizeof(uint64_t) atIndex:17];
-                [encoder setBytes:&nb2         length:sizeof(uint64_t) atIndex:18];
-                [encoder setBytes:&nb3         length:sizeof(uint64_t) atIndex:19];
-                [encoder setBytes:&n_past      length:sizeof(     int) atIndex:20];
-                [encoder setBytes:&n_dims      length:sizeof(     int) atIndex:21];
-                [encoder setBytes:&n_ctx_orig  length:sizeof(     int) atIndex:22];
-                [encoder setBytes:&freq_base   length:sizeof(   float) atIndex:23];
-                [encoder setBytes:&freq_scale  length:sizeof(   float) atIndex:24];
-                [encoder setBytes:&ext_factor  length:sizeof(   float) atIndex:25];
-                [encoder setBytes:&attn_factor length:sizeof(   float) atIndex:26];
-                [encoder setBytes:&beta_fast   length:sizeof(   float) atIndex:27];
-                [encoder setBytes:&beta_slow   length:sizeof(   float) atIndex:28];
+                [encoder setBuffer:id_dst  offset:offs_dst      atIndex:4];
 
                 [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
@@ -2796,6 +2863,7 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("fatal error");
                 };
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src1 offset:offs_src1       atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst        atIndex:1];
@@ -2836,6 +2904,7 @@ static void ggml_metal_encode_node(
 
                 const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UPSCALE_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -2870,6 +2939,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -2906,6 +2976,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARANGE_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_dst  offset:offs_dst    atIndex:0];
                 [encoder setBytes:&ne0   length:sizeof(ne0)   atIndex:1];
@@ -2927,6 +2998,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -2965,6 +3037,7 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("fatal error");
                 };
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
                 [encoder setBuffer:id_dst      offset:offs_dst         atIndex:1];
@@ -2983,6 +3056,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst    atIndex:1];
@@ -3224,37 +3298,41 @@ static void ggml_metal_encode_node(
                     }
                 }
 
+                ggml_metal_kargs_flash_attn_ext args = {
+                    /*.ne01          =*/ ne01,
+                    /*.ne02          =*/ ne02,
+                    /*.ne03          =*/ ne03,
+                    /*.nb01          =*/ nb01,
+                    /*.nb02          =*/ nb02,
+                    /*.nb03          =*/ nb03,
+                    /*.ne11          =*/ ne11,
+                    /*.ne_12_2       =*/ ne12,
+                    /*.ne_12_3       =*/ ne13,
+                    /*.nb_12_1       =*/ nb11,
+                    /*.nb_12_2       =*/ nb12,
+                    /*.nb_12_3       =*/ nb13,
+                    /*.nb31          =*/ nb31,
+                    /*.ne1           =*/ ne1,
+                    /*.ne2           =*/ ne2,
+                    /*.scale         =*/ scale,
+                    /*.max_bias      =*/ max_bias,
+                    /*.m0            =*/ m0,
+                    /*.m1            =*/ m1,
+                    /*.n_head_log2   =*/ n_head_log2,
+                    /*.logit_softcap =*/ logit_softcap,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0     offset:offs_src0           atIndex:0];
-                [encoder setBuffer:id_src1     offset:offs_src1           atIndex:1];
-                [encoder setBuffer:id_src2     offset:offs_src2           atIndex:2];
+                [encoder setBytes:&args length:sizeof(args)     atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0     atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1     atIndex:2];
+                [encoder setBuffer:id_src2 offset:offs_src2     atIndex:3];
                 if (id_src3) {
-                    [encoder setBuffer:id_src3     offset:offs_src3           atIndex:3];
+                    [encoder setBuffer:id_src3 offset:offs_src3 atIndex:4];
                 } else {
-                    [encoder setBuffer:id_src0     offset:offs_src0           atIndex:3];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:4];
                 }
-                [encoder setBuffer:id_dst        offset:offs_dst              atIndex:4];
-                [encoder setBytes:&ne01          length:sizeof( int64_t)      atIndex:5];
-                [encoder setBytes:&ne02          length:sizeof( int64_t)      atIndex:6];
-                [encoder setBytes:&ne03          length:sizeof( int64_t)      atIndex:7];
-                [encoder setBytes:&nb01          length:sizeof(uint64_t)      atIndex:8];
-                [encoder setBytes:&nb02          length:sizeof(uint64_t)      atIndex:9];
-                [encoder setBytes:&nb03          length:sizeof(uint64_t)      atIndex:10];
-                [encoder setBytes:&ne11          length:sizeof( int64_t)      atIndex:11];
-                [encoder setBytes:&ne12          length:sizeof( int64_t)      atIndex:12];
-                [encoder setBytes:&ne13          length:sizeof( int64_t)      atIndex:13];
-                [encoder setBytes:&nb11          length:sizeof(uint64_t)      atIndex:14];
-                [encoder setBytes:&nb12          length:sizeof(uint64_t)      atIndex:15];
-                [encoder setBytes:&nb13          length:sizeof(uint64_t)      atIndex:16];
-                [encoder setBytes:&nb31          length:sizeof(uint64_t)      atIndex:17];
-                [encoder setBytes:&ne1           length:sizeof( int64_t)      atIndex:18];
-                [encoder setBytes:&ne2           length:sizeof( int64_t)      atIndex:19];
-                [encoder setBytes:&scale         length:sizeof(   float)      atIndex:20];
-                [encoder setBytes:&max_bias      length:sizeof(   float)      atIndex:21];
-                [encoder setBytes:&m0            length:sizeof(m0)            atIndex:22];
-                [encoder setBytes:&m1            length:sizeof(m1)            atIndex:23];
-                [encoder setBytes:&n_head_log2   length:sizeof(n_head_log2)   atIndex:24];
-                [encoder setBytes:&logit_softcap length:sizeof(logit_softcap) atIndex:25];
+                [encoder setBuffer:id_dst offset:offs_dst       atIndex:5];
 
                 if (!use_vec_kernel) {
                     // half8x8 kernel
@@ -3389,25 +3467,29 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("not implemented");
                 }
 
+                ggml_metal_kargs_cpy args = {
+                    /*.ne00 =*/ ne00,
+                    /*.ne01 =*/ ne01,
+                    /*.ne02 =*/ ne02,
+                    /*.ne03 =*/ ne03,
+                    /*.nb00 =*/ nb00,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb03 =*/ nb03,
+                    /*.ne0  =*/ ne0,
+                    /*.ne1  =*/ ne1,
+                    /*.ne2  =*/ ne2,
+                    /*.ne3  =*/ ne3,
+                    /*.nb0  =*/ nb0,
+                    /*.nb1  =*/ nb1,
+                    /*.nb2  =*/ nb2,
+                    /*.nb3  =*/ nb3,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
-                [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
-                [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
-                [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
-                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
-                [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
-                [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
-                [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
-                [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
-                [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
-                [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
-                [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
-                [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
-                [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
-                [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
 
                 [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
@@ -3452,6 +3534,7 @@ static void ggml_metal_encode_node(
                 const int64_t n_threads = MIN((int64_t)[pipeline maxTotalThreadsPerThreadgroup], parallel_elements);
                 const int64_t n_tg = (parallel_elements + n_threads - 1) / n_threads;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0       atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst        atIndex:1];
diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
index 8c7fcb11303b8..86fdf1c18cfb6 100644
--- a/ggml/src/ggml-metal/ggml-metal.metal
+++ b/ggml/src/ggml-metal/ggml-metal.metal
@@ -6,6 +6,7 @@ __embed_ggml-common.h__
 // TODO: this should not be a relative path, but can't figure out how to set Metal include paths in Package.swift
 #include "../ggml-common.h"
 #endif
+#include "ggml-metal-impl.h"
 
 #include <metal_stdlib>
 
@@ -497,240 +498,131 @@ enum ggml_sort_order {
 // pros: works for non-contiguous tensors, supports broadcast across all dims
 // cons: not very efficient
 kernel void kernel_add(
+        constant ggml_metal_kargs_bin & args,
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne10,
-        constant  int64_t & ne11,
-        constant  int64_t & ne12,
-        constant  int64_t & ne13,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb12,
-        constant uint64_t & nb13,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        constant  int64_t & offs,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig.z;
-    const int64_t i02 = tgpig.y;
-    const int64_t i01 = tgpig.x;
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
 
-    const int64_t i13 = i03 % ne13;
-    const int64_t i12 = i02 % ne12;
-    const int64_t i11 = i01 % ne11;
+    const int i13 = i03%args.ne13;
+    const int i12 = i02%args.ne12;
+    const int i11 = i01%args.ne11;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
-    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + offs;
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs;
+    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11;
+    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
 
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i10 = i0 % ne10;
-        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) + *((device float *)(src1_ptr + i10*nb10));
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i10 = i0%args.ne10;
+        *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) + *((device float *)(src1_ptr + i10*args.nb10));
     }
 }
 
 kernel void kernel_sub(
+        constant ggml_metal_kargs_bin & args,
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne10,
-        constant  int64_t & ne11,
-        constant  int64_t & ne12,
-        constant  int64_t & ne13,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb12,
-        constant uint64_t & nb13,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        constant  int64_t & offs,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig.z;
-    const int64_t i02 = tgpig.y;
-    const int64_t i01 = tgpig.x;
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
 
-    const int64_t i13 = i03 % ne13;
-    const int64_t i12 = i02 % ne12;
-    const int64_t i11 = i01 % ne11;
+    const int i13 = i03%args.ne13;
+    const int i12 = i02%args.ne12;
+    const int i11 = i01%args.ne11;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
-    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + offs;
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs;
+    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11;
+    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
 
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i10 = i0 % ne10;
-        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) - *((device float *)(src1_ptr + i10*nb10));
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i10 = i0%args.ne10;
+        *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) - *((device float *)(src1_ptr + i10*args.nb10));
     }
 }
 
 kernel void kernel_mul(
+        constant ggml_metal_kargs_bin & args,
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne10,
-        constant  int64_t & ne11,
-        constant  int64_t & ne12,
-        constant  int64_t & ne13,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb12,
-        constant uint64_t & nb13,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig.z;
-    const int64_t i02 = tgpig.y;
-    const int64_t i01 = tgpig.x;
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
 
-    const int64_t i13 = i03 % ne13;
-    const int64_t i12 = i02 % ne12;
-    const int64_t i11 = i01 % ne11;
+    const int i13 = i03%args.ne13;
+    const int i12 = i02%args.ne12;
+    const int i11 = i01%args.ne11;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
-    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11;
+    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1;
 
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i10 = i0 % ne10;
-        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) * *((device float *)(src1_ptr + i10*nb10));
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i10 = i0%args.ne10;
+        *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) * *((device float *)(src1_ptr + i10*args.nb10));
     }
 }
 
 kernel void kernel_div(
+        constant ggml_metal_kargs_bin & args,
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne10,
-        constant  int64_t & ne11,
-        constant  int64_t & ne12,
-        constant  int64_t & ne13,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb12,
-        constant uint64_t & nb13,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig.z;
-    const int64_t i02 = tgpig.y;
-    const int64_t i01 = tgpig.x;
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
 
-    const int64_t i13 = i03 % ne13;
-    const int64_t i12 = i02 % ne12;
-    const int64_t i11 = i01 % ne11;
+    const int i13 = i03%args.ne13;
+    const int i12 = i02%args.ne12;
+    const int i11 = i01%args.ne11;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
-    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11;
+    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1;
 
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i10 = i0 % ne10;
-        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) / *((device float *)(src1_ptr + i10*nb10));
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i10 = i0%args.ne10;
+        *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) / *((device float *)(src1_ptr + i10*args.nb10));
     }
 }
 
 template<typename T>
 kernel void kernel_repeat(
+        constant ggml_metal_kargs_repeat & args,
         device const char * src0,
         device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i3 = tgpig.z;
-    const int64_t i2 = tgpig.y;
-    const int64_t i1 = tgpig.x;
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
 
-    const int64_t i03 = i3 % ne03;
-    const int64_t i02 = i2 % ne02;
-    const int64_t i01 = i1 % ne01;
+    const int i03 = i3%args.ne03;
+    const int i02 = i2%args.ne02;
+    const int i01 = i1%args.ne01;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
-    device       char * dst_ptr  = dst  +  i3*nb3  +  i2*nb2  +  i1*nb1 ;
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device       char * dst_ptr  = dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1;
 
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i00 = i0 % ne00;
-        *((device T *)(dst_ptr + i0*nb0)) = *((device T *)(src0_ptr + i00*nb00));
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i00 = i0%args.ne00;
+        *((device T *)(dst_ptr + i0*args.nb0)) = *((device T *)(src0_ptr + i00*args.nb00));
     }
 }
 
@@ -744,38 +636,42 @@ template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat
 // assumption: src1 is a row
 // broadcast src1 into src0
 kernel void kernel_add_row(
+        constant ggml_metal_kargs_bin & args,
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant   uint64_t & nb [[buffer(28)]],
         uint tpig[[thread_position_in_grid]]) {
+    const uint nb = args.ne00/4;
     dst[tpig] = src0[tpig] + src1[tpig % nb];
 }
 
 kernel void kernel_sub_row(
+        constant ggml_metal_kargs_bin & args,
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant   uint64_t & nb [[buffer(28)]],
         uint tpig[[thread_position_in_grid]]) {
+    const uint nb = args.ne00/4;
     dst[tpig] = src0[tpig] - src1[tpig % nb];
 }
 
 kernel void kernel_mul_row(
+        constant ggml_metal_kargs_bin & args,
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant   uint64_t & nb  [[buffer(28)]],
         uint tpig[[thread_position_in_grid]]) {
+    const uint nb = args.ne00/4;
     dst[tpig] = src0[tpig] * src1[tpig % nb];
 }
 
 kernel void kernel_div_row(
+        constant ggml_metal_kargs_bin & args,
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant   uint64_t & nb  [[buffer(28)]],
         uint tpig[[thread_position_in_grid]]) {
+    const uint nb = args.ne00/4;
     dst[tpig] = src0[tpig] / src1[tpig % nb];
 }
 
@@ -1345,102 +1241,112 @@ kernel void kernel_ssm_scan_f32(
 }
 
 kernel void kernel_norm(
-        device const  void * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant     float & eps,
-        threadgroup float  * sum [[threadgroup(0)]],
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const float * x = (device const float *) ((device const char *) src0 + tgpig*nb01);
-    // MEAN
-    // parallel sum
-    sum[tpitg] = 0.0f;
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        sum[tpitg] += x[i00];
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint   tgpig[[threadgroup_position_in_grid]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    device const float4 * x = (device const float4 *) (src0 + tgpig*args.nb01);
+
+    float4 sumf4(0.0f);
+
+    float sumf = 0.0f;
+
+    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
+        sumf4 += x[i00];
     }
-    // reduce
+    sumf = sumf4[0] + sumf4[1] + sumf4[2] + sumf4[3];
+    sumf = simd_sum(sumf);
+
     threadgroup_barrier(mem_flags::mem_threadgroup);
-    for (uint i = ntg/2; i > 0; i /= 2) {
-        if (tpitg < i) {
-            sum[tpitg] += sum[tpitg + i];
-        }
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
     }
-    const float mean  = sum[0] / ne00;
 
-    // recenter and VARIANCE
     threadgroup_barrier(mem_flags::mem_threadgroup);
-    device float * y = dst + tgpig*ne00;
-    sum[tpitg] = 0.0f;
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean = sumf/args.ne00;
+
+    device float4 * y = (device float4 *) dst + tgpig*args.ne00_4;
+
+    sumf = 0.0f;
+    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
         y[i00] = x[i00] - mean;
-        sum[tpitg] += y[i00] * y[i00];
+        sumf += dot(y[i00], y[i00]);
     }
+    sumf = simd_sum(sumf);
 
-    // reduce
     threadgroup_barrier(mem_flags::mem_threadgroup);
-    for (uint i = ntg/2; i > 0; i /= 2) {
-        if (tpitg < i) {
-            sum[tpitg] += sum[tpitg + i];
-        }
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
     }
-    const float variance = sum[0] / ne00;
 
-    const float scale = 1.0f/sqrt(variance + eps);
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float variance = sumf/args.ne00;
+
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
         y[i00] = y[i00] * scale;
     }
 }
 
 kernel void kernel_rms_norm(
-        device const  void * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant     float & eps,
-        threadgroup float  * buf [[threadgroup(0)]],
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const float4 * x = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
+        constant ggml_metal_kargs_rms_norm & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint   tgpig[[threadgroup_position_in_grid]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    device const float4 * x = (device const float4 *) (src0 + tgpig*args.nb01);
 
-    float4 sumf = 0;
-    float all_sum = 0;
+    float sumf = 0.0f;
 
     // parallel sum
-    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        sumf += x[i00] * x[i00];
+    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
+        sumf += dot(x[i00], x[i00]);
     }
-    all_sum = sumf[0] + sumf[1] + sumf[2] + sumf[3];
-    all_sum = simd_sum(all_sum);
-    if (ntg > N_SIMDWIDTH) {
-        if (sgitg == 0) {
-            buf[tiisg] = 0.0f;
-        }
+    sumf = simd_sum(sumf);
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+    threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        if (tiisg == 0) {
-            buf[sgitg] = all_sum;
-        }
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+    threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        all_sum = buf[tiisg];
-        all_sum = simd_sum(all_sum);
-    }
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
 
-    const float mean  = all_sum/ne00;
-    const float scale = 1.0f/sqrt(mean + eps);
+    const float mean  = sumf/args.ne00;
+    const float scale = 1.0f/sqrt(mean + args.eps);
 
-    device float4 * y = (device float4 *) (dst + tgpig*ne00);
-    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
+    device float4 * y = (device float4 *) dst + tgpig*args.ne00_4;
+    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
         y[i00] = x[i00] * scale;
     }
 }
@@ -1628,31 +1534,17 @@ inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thre
 //      quantizations where the block size is 32. It also does not
 //      guard against the number of rows not being divisible by
 //      N_DST, so this is another explicit assumption of the implementation.
-template<typename block_q_type, int nr, int nsg, int nw>
+template<typename block_q_type, int nr, int nsg, int nw, typename args_t>
 void mul_vec_q_n_f32_impl(
-        device const void  * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                     uint3   tgpig,
-                     uint    tiisg,
-                     uint    sgitg) {
-    const int nb = ne00/QK4_0;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+    const int nb = args.ne00/QK4_0;
 
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
@@ -1660,19 +1552,19 @@ void mul_vec_q_n_f32_impl(
 
     const int first_row = (r0 * nsg + sgitg) * nr;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-  //const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+  //const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-  //device const block_q_type * x = (device const block_q_type *) ((device char *) src0 + offset0);
-    device const float        * y = (device const float        *) ((device char *) src1 + offset1);
+  //device const block_q_type * x = (device const block_q_type *) (src0 + offset0);
+    device const float        * y = (device const float        *) (src1 + offset1);
 
     // pointers to src0 rows
     device const block_q_type * ax[nr];
     for (int row = 0; row < nr; ++row) {
-        const uint offset0 = (first_row + row)*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+        const uint64_t offset0 = (first_row + row)*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
 
         ax[row] = (device const block_q_type *) ((device char *) src0 + offset0);
     }
@@ -1680,10 +1572,10 @@ void mul_vec_q_n_f32_impl(
     float yl[16]; // src1 vector cache
     float sumf[nr] = {0.f};
 
-    const int ix = (tiisg/2);
-    const int il = (tiisg%2)*8;
+    const short ix = (tiisg/2);
+    const short il = (tiisg%2)*8;
 
-    device const float * yb = y + ix * QK4_0 + il;
+    device const float * yb = y + ix*QK4_0 + il;
 
     // each thread in a SIMD group deals with half a block.
     for (int ib = ix; ib < nb; ib += nw/2) {
@@ -1708,324 +1600,216 @@ void mul_vec_q_n_f32_impl(
         yb += QK4_0 * 16;
     }
 
+    device float * dst_f32 = (device float *) dst + im*args.ne0*args.ne1 + r1*args.ne0;
+
     for (int row = 0; row < nr; ++row) {
         const float tot = simd_sum(sumf[row]);
-        if (tiisg == 0 && first_row + row < ne01) {
-            dst[im*ne0*ne1 + r1*ne0 + first_row + row] = tot;
+
+        if (tiisg == 0 && first_row + row < args.ne01) {
+            dst_f32[first_row + row] = tot;
         }
     }
 }
 
 kernel void kernel_mul_mv_q4_0_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,nb01,nb02,nb03,ne10,ne12,nb11,nb12,nb13,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 kernel void kernel_mul_mv_q4_1_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
-     mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,nb01,nb02,nb03,ne10,ne12,nb11,nb12,nb13,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+     mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 kernel void kernel_mul_mv_q5_0_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,nb01,nb02,nb03,ne10,ne12,nb11,nb12,nb13,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 kernel void kernel_mul_mv_q5_1_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,nb01,nb02,nb03,ne10,ne12,nb11,nb12,nb13,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
-
 #define NB_Q8_0 8
 
+template<typename args_t>
 void kernel_mul_mv_q8_0_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
     const int nr  = N_DST;
     const int nsg = N_SIMDGROUP;
     const int nw  = N_SIMDWIDTH;
 
-    const int nb = ne00/QK8_0;
+    const int nb = args.ne00/QK8_0;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
-    const int first_row = (r0 * nsg + sgitg) * nr;
+    const int first_row = (r0*nsg + sgitg)*nr;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-  //const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+  //const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-  //device const block_q8_0 * x = (device const block_q8_0 *) ((device char *) src0 + offset0);
-    device const float      * y = (device const float      *) ((device char *) src1 + offset1);
+  //device const block_q8_0 * x = (device const block_q8_0 *) (src0 + offset0);
+    device const float      * y = (device const float      *) (src1 + offset1);
 
     // pointers to src0 rows
     device const block_q8_0 * ax[nr];
     for (int row = 0; row < nr; ++row) {
-        const uint offset0 = (first_row + row)*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+        const uint64_t offset0 = (first_row + row)*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
 
         ax[row] = (device const block_q8_0 *) ((device char *) src0 + offset0);
     }
 
     float yl[NB_Q8_0];
-    float sumf[nr]={0.f};
+    float sumf[nr] = { 0.f };
 
-    const int ix = tiisg/4;
-    const int il = tiisg%4;
+    const short ix = tiisg/4;
+    const short il = tiisg%4;
 
-    device const float * yb = y + ix * QK8_0 + NB_Q8_0*il;
+    device const float * yb = y + ix*QK8_0 + il*NB_Q8_0;
 
     // each thread in a SIMD group deals with NB_Q8_0 quants at a time
     for (int ib = ix; ib < nb; ib += nw/4) {
-        for (int i = 0; i < NB_Q8_0; ++i) {
+        for (short i = 0; i < NB_Q8_0; ++i) {
             yl[i] = yb[i];
         }
 
         for (int row = 0; row < nr; row++) {
-            device const int8_t * qs = ax[row][ib].qs + NB_Q8_0*il;
+            device const int8_t * qs = ax[row][ib].qs + il*NB_Q8_0;
             float sumq = 0.f;
-            for (int iq = 0; iq < NB_Q8_0; ++iq) {
+            for (short iq = 0; iq < NB_Q8_0; ++iq) {
                 sumq += qs[iq] * yl[iq];
             }
             sumf[row] += sumq*ax[row][ib].d;
         }
 
-        yb += NB_Q8_0 * nw;
+        yb += nw*NB_Q8_0;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < nr; ++row) {
         const float tot = simd_sum(sumf[row]);
-        if (tiisg == 0 && first_row + row < ne01) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+
+        if (tiisg == 0 && first_row + row < args.ne01) {
+            dst_f32[first_row + row] = tot;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_q8_0_f32")]]
 kernel void kernel_mul_mv_q8_0_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    kernel_mul_mv_q8_0_f32_impl(src0,src1,dst,ne00,ne01,ne02,nb01,nb02,nb03,ne10,ne12,nb11,nb12,nb13,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    kernel_mul_mv_q8_0_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 #define N_MV_T_T 4
 
-template<typename T0, typename T04, typename T1, typename T14>
+template<typename T0, typename T04, typename T1, typename T14, typename args_t>
 void kernel_mul_mv_impl(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb00,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne11,
-                   int64_t   ne12,
-                  uint64_t   nb10,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-                   uint3     tgpig,
-                   uint      tiisg) {
-    const int64_t r0 = tgpig.x;
-    const int64_t rb = tgpig.y*N_MV_T_T;
-    const int64_t im = tgpig.z;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig,
+        ushort tiisg) {
+    const int r0 = tgpig.x;
+    const int rb = tgpig.y*N_MV_T_T;
+    const int im = tgpig.z;
+
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
+
+    const uint64_t offset0 = r0*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
 
     device const T0 * x = (device const T0 *) (src0 + offset0);
 
-    if (ne00 < 128) {
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1;
+
+    if (args.ne00 < 128) {
         for (int row = 0; row < N_MV_T_T; ++row) {
             int r1 = rb + row;
-            if (r1 >= ne11) {
+            if (r1 >= args.ne11) {
                 break;
             }
 
-            const uint offset1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+            const uint64_t offset1 = r1*args.nb11 + (i12   )*args.nb12 + (i13   )*args.nb13;
 
             device const T1 * y = (device const T1 *) (src1 + offset1);
 
             float sumf = 0;
-            for (int i = tiisg; i < ne00; i += 32) {
+            for (int i = tiisg; i < args.ne00; i += 32) {
                 sumf += (T0) x[i] * (T1) y[i];
             }
 
             float all_sum = simd_sum(sumf);
             if (tiisg == 0) {
-                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+                dst_f32[(uint64_t)r1*args.ne0 + r0] = all_sum;
             }
         }
     } else {
         device const T04 * x4 = (device const T04 *) x;
         for (int row = 0; row < N_MV_T_T; ++row) {
             int r1 = rb + row;
-            if (r1 >= ne11) {
+            if (r1 >= args.ne11) {
                 break;
             }
 
-            const uint offset1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+            const uint64_t offset1 = r1*args.nb11 + (i12   )*args.nb12 + (i13   )*args.nb13;
 
             device const T1  * y  = (device const T1  *) (src1 + offset1);
             device const T14 * y4 = (device const T14 *) y;
 
             float sumf = 0;
-            for (int i = tiisg; i < ne00/4; i += 32) {
-                for (int k = 0; k < 4; ++k) sumf += (float) (x4[i][k] * y4[i][k]);
+            for (int i = tiisg; i < args.ne00/4; i += 32) {
+                sumf += dot((float4) x4[i], (float4) y4[i]);
             }
 
             float all_sum = simd_sum(sumf);
             if (tiisg == 0) {
-                for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) (x[i] * y[i]);
-                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+                for (int i = 4*(args.ne00/4); i < args.ne00; ++i) all_sum += (float) (x[i] * y[i]);
+                dst_f32[(uint64_t)r1*args.ne0 + r0] = all_sum;
             }
         }
     }
@@ -2033,51 +1817,17 @@ void kernel_mul_mv_impl(
 
 template<typename T0, typename T04, typename T1, typename T14>
 kernel void kernel_mul_mv(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]]) {
-    kernel_mul_mv_impl<T0, T04, T1, T14>(
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]]) {
+    kernel_mul_mv_impl<T0, T04, T1, T14, constant ggml_metal_kargs_mul_mv &>(
+        args,
         src0,
         src1,
         dst,
-        ne00,
-        ne01,
-        ne02,
-        nb00,
-        nb01,
-        nb02,
-        nb03,
-        ne10,
-        ne11,
-        ne12,
-        nb10,
-        nb11,
-        nb12,
-        nb13,
-        ne0,
-        ne1,
-        r2,
-        r3,
         tgpig,
         tiisg);
 }
@@ -2094,65 +1844,50 @@ template [[host_name("kernel_mul_mv_bf16_bf16")]] kernel mul_mv_t kernel_mul_mv<
 
 template<typename T, typename T4>
 kernel void kernel_mul_mv_1row(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]]) {
 
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int64_t im = tgpig.z;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = r0*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 = r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
     device const T     * x = (device const T     *) (src0 + offset0);
     device const float * y = (device const float *) (src1 + offset1);
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     float sumf = 0;
-    if (ne00 < 128) {
-        for (int i = tiisg; i < ne00; i += 32) {
+    if (args.ne00 < 128) {
+        for (int i = tiisg; i < args.ne00; i += 32) {
             sumf += (float) x[i] * (float) y[i];
         }
         float all_sum = simd_sum(sumf);
         if (tiisg == 0) {
-            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            dst_f32[r0] = all_sum;
         }
     } else {
         device const T4     * x4 = (device const T4     *) x;
         device const float4 * y4 = (device const float4 *) y;
 
-        for (int i = tiisg; i < ne00/4; i += 32) {
-            for (int k = 0; k < 4; ++k) sumf += (float) (x4[i][k] * y4[i][k]);
+        for (int i = tiisg; i < args.ne00/4; i += 32) {
+            sumf += dot((float4) x4[i], y4[i]);
         }
 
         float all_sum = simd_sum(sumf);
 
         if (tiisg == 0) {
-            for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) (x[i] * y[i]);
-            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            for (int i = 4*(args.ne00/4); i < args.ne00; ++i) all_sum += (float) (x[i] * y[i]);
+            dst_f32[r0] = all_sum;
         }
     }
 }
@@ -2167,54 +1902,39 @@ template [[host_name("kernel_mul_mv_bf16_f32_1row")]] kernel mul_mv_1row_t kerne
 // Assumes row size (ne00) is a multiple of 4
 template<typename T, typename T4>
 kernel void kernel_mul_mv_l4(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]]) {
 
-    const int nrows = ne11;
-    const int64_t r0 = tgpig.x;
-    const int64_t im = tgpig.z;
+    const int nrows = args.ne11;
+    const int r0 = tgpig.x;
+    const int im = tgpig.z;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    const uint64_t offset0 = r0*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
 
     device const T4 * x4 = (device const T4 *) (src0 + offset0);
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1;
+
     for (int r1 = 0; r1 < nrows; ++r1) {
-        const uint offset1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+        const uint64_t offset1 = r1*args.nb11 + (i12   )*args.nb12 + (i13   )*args.nb13;
 
         device const float4 * y4 = (device const float4 *) (src1 + offset1);
 
         float sumf = 0;
-        for (int i = tiisg; i < ne00/4; i += 32) {
-            for (int k = 0; k < 4; ++k) sumf += (float) (x4[i][k] * y4[i][k]);
+        for (int i = tiisg; i < args.ne00/4; i += 32) {
+            sumf += dot((float4) x4[i], y4[i]);
         }
 
         float all_sum = simd_sum(sumf);
         if (tiisg == 0) {
-            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            dst_f32[(uint64_t)r1*args.ne0 + r0] = all_sum;
         }
     }
 }
@@ -2234,7 +1954,7 @@ static float rope_yarn_ramp(const float low, const float high, const int i0) {
 // YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
 // MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
 static void rope_yarn(
-    float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
+    float theta_extrap, float freq_scale, float corr_dims[2], int i0, float ext_factor, float mscale,
     thread float * cos_theta, thread float * sin_theta) {
     // Get n-d rotational scaling corrected for extrapolation
     float theta_interp = freq_scale * theta_extrap;
@@ -2266,65 +1986,41 @@ static void rope_yarn_corr_dims(
 
 template<typename T>
 kernel void kernel_rope_norm(
-        device const    void * src0,
-        device const int32_t * src1,
-        device const   float * src2,
-        device         float * dst,
-        constant     int64_t & ne00,
-        constant     int64_t & ne01,
-        constant     int64_t & ne02,
-        constant     int64_t & ne03,
-        constant    uint64_t & nb00,
-        constant    uint64_t & nb01,
-        constant    uint64_t & nb02,
-        constant    uint64_t & nb03,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant     int64_t & ne2,
-        constant     int64_t & ne3,
-        constant    uint64_t & nb0,
-        constant    uint64_t & nb1,
-        constant    uint64_t & nb2,
-        constant    uint64_t & nb3,
-        constant         int & n_past,
-        constant         int & n_dims,
-        constant         int & n_ctx_orig,
-        constant       float & freq_base,
-        constant       float & freq_scale,
-        constant       float & ext_factor,
-        constant       float & attn_factor,
-        constant       float & beta_fast,
-        constant       float & beta_slow,
-        uint  tiitg[[thread_index_in_threadgroup]],
-        uint3 tptg[[threads_per_threadgroup]],
-        uint3 tgpig[[threadgroup_position_in_grid]]) {
-    const int64_t i3 = tgpig[2];
-    const int64_t i2 = tgpig[1];
-    const int64_t i1 = tgpig[0];
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
 
     float corr_dims[2];
-    rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
 
-    device const int32_t * pos = src1;
+    device const int32_t * pos = (device const int32_t *) src1;
 
     const float theta_base = (float) pos[i2];
-    const float inv_ndims = -1.f/n_dims;
+    const float inv_ndims = -1.f/args.n_dims;
 
     float cos_theta;
     float sin_theta;
 
-    for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
-        if (i0 < n_dims) {
-            const int64_t ic = i0/2;
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
 
-            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
 
-            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
 
-            rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
 
-            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
 
             const float x0 = src[0];
             const float x1 = src[1];
@@ -2332,8 +2028,8 @@ kernel void kernel_rope_norm(
             dst_data[0] = x0*cos_theta - x1*sin_theta;
             dst_data[1] = x0*sin_theta + x1*cos_theta;
         } else {
-            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
 
             dst_data[0] = src[0];
             dst_data[1] = src[1];
@@ -2343,74 +2039,50 @@ kernel void kernel_rope_norm(
 
 template<typename T>
 kernel void kernel_rope_neox(
-        device const    void * src0,
-        device const int32_t * src1,
-        device const   float * src2,
-        device         float * dst,
-        constant     int64_t & ne00,
-        constant     int64_t & ne01,
-        constant     int64_t & ne02,
-        constant     int64_t & ne03,
-        constant    uint64_t & nb00,
-        constant    uint64_t & nb01,
-        constant    uint64_t & nb02,
-        constant    uint64_t & nb03,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant     int64_t & ne2,
-        constant     int64_t & ne3,
-        constant    uint64_t & nb0,
-        constant    uint64_t & nb1,
-        constant    uint64_t & nb2,
-        constant    uint64_t & nb3,
-        constant         int & n_past,
-        constant         int & n_dims,
-        constant         int & n_ctx_orig,
-        constant       float & freq_base,
-        constant       float & freq_scale,
-        constant       float & ext_factor,
-        constant       float & attn_factor,
-        constant       float & beta_fast,
-        constant       float & beta_slow,
-        uint  tiitg[[thread_index_in_threadgroup]],
-        uint3 tptg[[threads_per_threadgroup]],
-        uint3 tgpig[[threadgroup_position_in_grid]]) {
-    const int64_t i3 = tgpig[2];
-    const int64_t i2 = tgpig[1];
-    const int64_t i1 = tgpig[0];
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
 
     float corr_dims[2];
-    rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
 
-    device const int32_t * pos = src1;
+    device const int32_t * pos = (device const int32_t *) src1;
 
     const float theta_base = (float) pos[i2];
-    const float inv_ndims = -1.f/n_dims;
+    const float inv_ndims = -1.f/args.n_dims;
 
     float cos_theta;
     float sin_theta;
 
-    for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
-        if (i0 < n_dims) {
-            const int64_t ic = i0/2;
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
 
-            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
 
-            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
 
-            rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
 
-            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
 
             const float x0 = src[0];
-            const float x1 = src[n_dims/2];
+            const float x1 = src[args.n_dims/2];
 
-            dst_data[0]        = x0*cos_theta - x1*sin_theta;
-            dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
         } else {
-            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
 
             dst_data[0] = src[0];
             dst_data[1] = src[1];
@@ -2808,37 +2480,17 @@ template<
     short KV = 8,    // key/value processed per each simdgroup
     short C  = 32>   // cache items per threadgroup
 kernel void kernel_flash_attn_ext(
-        device const  char * q,
-        device const  char * k,
-        device const  char * v,
-        device const  char * mask,
-        device       float * dst,
-        constant   int32_t & ne01,
-        constant   int32_t & ne02,
-        constant   int32_t & ne03,
-        constant  uint32_t & nb01,
-        constant  uint32_t & nb02,
-        constant  uint32_t & nb03,
-        constant   int32_t & ne11,
-        constant   int32_t & ne_12_2, // assume K and V are same shape
-        constant   int32_t & ne_12_3,
-        constant  uint32_t & nb_12_1,
-        constant  uint32_t & nb_12_2,
-        constant  uint32_t & nb_12_3,
-        constant  uint32_t & nb31,
-        constant   int32_t & ne1,
-        constant   int32_t & ne2,
-        constant     float & scale,
-        constant     float & max_bias,
-        constant     float & m0,
-        constant     float & m1,
-        constant  uint16_t & n_head_log2,
-        constant     float & logit_softcap,
-        threadgroup   half * shared [[threadgroup(0)]],
-        ushort3  tgpig[[threadgroup_position_in_grid]],
-        ushort3    ntg[[threads_per_threadgroup]],
-        ushort   tiisg[[thread_index_in_simdgroup]],
-        ushort   sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_flash_attn_ext & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * mask,
+        device       char * dst,
+        threadgroup  half * shmem_f16 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3   ntg[[threads_per_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
     const short nsg = ntg.y; // number of simdgroups
 
     const int iq3 = tgpig[2];
@@ -2854,27 +2506,27 @@ kernel void kernel_flash_attn_ext(
     const short TS = nsg*SH;   // shared memory size per query in (s_t == float)
     const short T  = D + 2*TS; // shared memory size per query in (half)
 
-    threadgroup q_t  * sq  = (threadgroup q_t  *) (shared +              0*D); // holds the query data
-    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shared +              0*D); // same as above but in q4_t
-    threadgroup o_t  * so  = (threadgroup o_t  *) (shared +              0*D); // reuse query data for accumulation
-    threadgroup o4_t * so4 = (threadgroup o4_t *) (shared +              0*D); // same as above but in o4_t
-    threadgroup s_t  * ss  = (threadgroup s_t  *) (shared + 2*sgitg*SH + Q*D); // scratch buffer for attention, mask and diagonal matrix
+    threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +              0*D); // holds the query data
+    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +              0*D); // same as above but in q4_t
+    threadgroup o_t  * so  = (threadgroup o_t  *) (shmem_f16 +              0*D); // reuse query data for accumulation
+    threadgroup o4_t * so4 = (threadgroup o4_t *) (shmem_f16 +              0*D); // same as above but in o4_t
+    threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + Q*D); // scratch buffer for attention, mask and diagonal matrix
 
-    threadgroup k_t    * sk    = (threadgroup k_t    *) (shared + sgitg*(4*16*KV) + Q*T); // scratch buffer to load K in shared memory
-    threadgroup k4x4_t * sk4x4 = (threadgroup k4x4_t *) (shared + sgitg*(4*16*KV) + Q*T); // same as above but in k4x4_t
+    threadgroup k_t    * sk    = (threadgroup k_t    *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // scratch buffer to load K in shared memory
+    threadgroup k4x4_t * sk4x4 = (threadgroup k4x4_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // same as above but in k4x4_t
 
-    threadgroup v_t    * sv    = (threadgroup v_t    *) (shared + sgitg*(4*16*KV) + Q*T); // scratch buffer to load V in shared memory
-    threadgroup v4x4_t * sv4x4 = (threadgroup v4x4_t *) (shared + sgitg*(4*16*KV) + Q*T); // same as above but in v4x4_t
+    threadgroup v_t    * sv    = (threadgroup v_t    *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // scratch buffer to load V in shared memory
+    threadgroup v4x4_t * sv4x4 = (threadgroup v4x4_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // same as above but in v4x4_t
 
     // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper)
     o8x8_t lo[D8];
 
     // load heads from Q to shared memory
     for (short j = sgitg; j < Q; j += nsg) {
-        device const float4 * q4 = (device const float4 *) ((device const char *) q + ((iq1 + j)*nb01 + iq2*nb02 + iq3*nb03));
+        device const float4 * q4 = (device const float4 *) ((device const char *) q + ((iq1 + j)*args.nb01 + iq2*args.nb02 + iq3*args.nb03));
 
         for (short i = tiisg; i < D4; i += NW) {
-            if (iq1 + j < ne01) {
+            if (iq1 + j < args.ne01) {
                 sq4[j*D4 + i] = (q4_t) q4[i];
             } else {
                 sq4[j*D4 + i] = (q4_t) 0.0f;
@@ -2907,11 +2559,11 @@ kernel void kernel_flash_attn_ext(
         const short ty = tiisg/4;
 
         // broadcast kv
-        //const short rk2 = ne02/ne12;
-        //const short rk3 = ne03/ne13;
+        //const short rk2 = args.ne02/args.ne12;
+        //const short rk3 = args.ne03/args.ne13;
 
-        const short ikv2 = iq2/(ne02/ne_12_2);
-        const short ikv3 = iq3/(ne03/ne_12_3);
+        const short ikv2 = iq2/(args.ne02/args.ne_12_2);
+        const short ikv3 = iq3/(args.ne03/args.ne_12_3);
 
         // load the queries from shared memory into local memory
         q8x8_t mq[D8];
@@ -2925,20 +2577,20 @@ kernel void kernel_flash_attn_ext(
         half slope = 1.0f;
 
         // ALiBi
-        if (max_bias > 0.0f) {
+        if (args.max_bias > 0.0f) {
             const short h = iq2;
 
-            const half  base = h < n_head_log2 ? m0 : m1;
-            const short exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+            const half  base = h < args.n_head_log2 ? args.m0 : args.m1;
+            const short exph = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
 
             slope = pow(base, exph);
         }
 
         // loop over the KV cache
         // each simdgroup handles blocks of Q rows and C columns
-        for (int ic0 = 0; ic0 < ne11; ic0 += C*nsg) {
+        for (int ic0 = 0; ic0 < args.ne11; ic0 += C*nsg) {
             const int ic = ic0 + C*sgitg;
-            if (ic >= ne11) {
+            if (ic >= args.ne11) {
                 break;
             }
 
@@ -2949,7 +2601,7 @@ kernel void kernel_flash_attn_ext(
                 // load the mask in shared memory
                 #pragma unroll(Q)
                 for (short j = 0; j < Q; ++j) {
-                    device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*nb31);
+                    device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31);
 
                     const half m = pm[ic + tiisg];
 
@@ -2972,18 +2624,18 @@ kernel void kernel_flash_attn_ext(
                     // this is compile-time check, so it does not have runtime overhead
                     if (is_same<kd4x4_t, k4x4_t>::value) {
                         // we can read directly from global memory
-                        device const k_t * pk = (device const k_t *) ((device const char *) k + ((ic + 8*cc)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
+                        device const k_t * pk = (device const k_t *) ((device const char *) k + ((ic + 8*cc)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
 
                         #pragma unroll(D8)
                         for (short i = 0; i < D8; ++i) {
                             k8x8_t mk;
-                            simdgroup_load(mk, pk + i*8, nb_12_1/sizeof(k_t), 0, true); // transpose // TODO: use ne10
+                            simdgroup_load(mk, pk + i*8, args.nb_12_1/sizeof(k_t), 0, true); // transpose // TODO: use ne10
 
                             simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk);
                         }
                     } else {
                         for (short ii = 0; ii < D16; ii += 4) {
-                            device const kd4x4_t * pk4x4 = (device const kd4x4_t *) ((device const char *) k + ((ic + 8*cc + ty)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
+                            device const kd4x4_t * pk4x4 = (device const kd4x4_t *) ((device const char *) k + ((ic + 8*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
 
                             if (D16%4 == 0) {
                                 // the head is evenly divisible by 4*16 = 64, so no need for bound checks
@@ -3042,10 +2694,10 @@ kernel void kernel_flash_attn_ext(
                     const half m = M[j];
 
                     // scale and apply the logitcap / mask
-                    half s = ss[j*TS + tiisg]*scale;
+                    half s = ss[j*TS + tiisg]*args.scale;
 
-                    if (logit_softcap != 0.0f) {
-                        s = logit_softcap*precise::tanh(s);
+                    if (args.logit_softcap != 0.0f) {
+                        s = args.logit_softcap*precise::tanh(s);
                     }
 
                     // mqk = mqk + mask*slope
@@ -3087,18 +2739,18 @@ kernel void kernel_flash_attn_ext(
 
                     if (is_same<vd4x4_t, v4x4_t>::value) {
                         // we can read directly from global memory
-                        device const v_t * pv = (device const v_t *) ((device const char *) v + ((ic + 8*cc)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
+                        device const v_t * pv = (device const v_t *) ((device const char *) v + ((ic + 8*cc)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
 
                         #pragma unroll(D8)
                         for (short i = 0; i < D8; ++i) {
                             v8x8_t mv;
-                            simdgroup_load(mv, pv + i*8, nb_12_1/sizeof(v_t), 0, false); // TODO: use ne20
+                            simdgroup_load(mv, pv + i*8, args.nb_12_1/sizeof(v_t), 0, false); // TODO: use ne20
 
                             simdgroup_multiply_accumulate(lo[i], ms, mv, lo[i]);
                         }
                     } else {
                         for (short ii = 0; ii < D16; ii += 4) {
-                            device const vd4x4_t * pv4x4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 8*cc + ty)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
+                            device const vd4x4_t * pv4x4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 8*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
 
                             if (D16%4 == 0) {
                                 // no need for bound checks
@@ -3227,11 +2879,11 @@ kernel void kernel_flash_attn_ext(
 
     // final rescale with 1/S and store to global memory
     if (sgitg == 0) {
-        for (short j = 0; j < Q && iq1 + j < ne01; ++j) {
+        for (short j = 0; j < Q && iq1 + j < args.ne01; ++j) {
             const float S = ss[j*TS + 0];
 
             for (short i = tiisg; i < D4; i += NW) {
-                dst4[((int64_t)iq3*ne2*ne1 + iq2 + (iq1 + j)*ne1)*D4 + i] = (float4) so4[j*D4 + i]/S;
+                dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*D4 + i] = (float4) so4[j*D4 + i]/S;
             }
         }
     }
@@ -3323,38 +2975,17 @@ template<
     short Q  = 1,    // queries per threadgroup
     short C  = 32>   // cache items per threadgroup
 kernel void kernel_flash_attn_ext_vec(
-        device const  char * q,
-        device const  char * k,
-        device const  char * v,
-        device const  char * mask,
-        device       float * dst,
-        constant   int32_t & ne01,
-        constant   int32_t & ne02,
-        constant   int32_t & ne03,
-        constant  uint32_t & nb01,
-        constant  uint32_t & nb02,
-        constant  uint32_t & nb03,
-        constant   int32_t & ne11,
-        constant   int32_t & ne_12_2, // assume K and V are same shape
-        constant   int32_t & ne_12_3,
-        constant  uint32_t & nb_12_1,
-        constant  uint32_t & nb_12_2,
-        constant  uint32_t & nb_12_3,
-        constant  uint32_t & nb31,
-        constant   int32_t & ne1,
-        constant   int32_t & ne2,
-        constant     float & scale,
-        constant     float & max_bias,
-        constant     float & m0,
-        constant     float & m1,
-        constant  uint16_t & n_head_log2,
-        constant     float & logit_softcap,
-        threadgroup   half * shared [[threadgroup(0)]],
-        ushort3  tgpig[[threadgroup_position_in_grid]],
-        ushort3  tpitg[[thread_position_in_threadgroup]],
-        ushort3    ntg[[threads_per_threadgroup]],
-        ushort   tiisg[[thread_index_in_simdgroup]],
-        ushort   sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_flash_attn_ext & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * mask,
+        device       char * dst,
+        threadgroup  half * shmem_f16 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3   ntg[[threads_per_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
     const short nsg = ntg.y; // number of simdgroups
 
     const int iq3 = tgpig[2];
@@ -3369,22 +3000,22 @@ kernel void kernel_flash_attn_ext_vec(
 
     const short T = D + nsg*SH; // shared memory size per query in (half)
 
-  //threadgroup q_t    * sq    = (threadgroup q_t    *) (shared +                0*D); // holds the query data
-    threadgroup q4_t   * sq4   = (threadgroup q4_t   *) (shared +                0*D); // same as above but in q4_t
-    threadgroup q4x4_t * sq4x4 = (threadgroup q4x4_t *) (shared +                0*D); // same as above but in q4x4_t
-    threadgroup s_t    * ss    = (threadgroup s_t    *) (shared + sgitg*SH     + Q*D); // scratch buffer for attention
-    threadgroup s4_t   * ss4   = (threadgroup s4_t   *) (shared + sgitg*SH     + Q*D); // same as above but in s4_t
-    threadgroup half   * sm    = (threadgroup half   *) (shared + sgitg*SH + C + Q*D); // scratch buffer for mask
-    threadgroup o4x4_t * sr4x4 = (threadgroup o4x4_t *) (shared + sgitg*D      + Q*T); // scratch buffer for the results
+  //threadgroup q_t    * sq    = (threadgroup q_t    *) (shmem_f16 +                0*D); // holds the query data
+    threadgroup q4_t   * sq4   = (threadgroup q4_t   *) (shmem_f16 +                0*D); // same as above but in q4_t
+    threadgroup q4x4_t * sq4x4 = (threadgroup q4x4_t *) (shmem_f16 +                0*D); // same as above but in q4x4_t
+    threadgroup s_t    * ss    = (threadgroup s_t    *) (shmem_f16 + sgitg*SH     + Q*D); // scratch buffer for attention
+    threadgroup s4_t   * ss4   = (threadgroup s4_t   *) (shmem_f16 + sgitg*SH     + Q*D); // same as above but in s4_t
+    threadgroup half   * sm    = (threadgroup half   *) (shmem_f16 + sgitg*SH + C + Q*D); // scratch buffer for mask
+    threadgroup o4x4_t * sr4x4 = (threadgroup o4x4_t *) (shmem_f16 + sgitg*D      + Q*T); // scratch buffer for the results
 
     // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper)
     o4x4_t lo[D16/NL];
 
     // load heads from Q to shared memory
-    device const float4 * q4 = (device const float4 *) ((device const char *) q + (iq1*nb01 + iq2*nb02 + iq3*nb03));
+    device const float4 * q4 = (device const float4 *) ((device const char *) q + (iq1*args.nb01 + iq2*args.nb02 + iq3*args.nb03));
 
     for (short i = tiisg; i < D4; i += NW) {
-        if (iq1 < ne01) {
+        if (iq1 < args.ne01) {
             sq4[i] = (q4_t) q4[i];
         } else {
             sq4[i] = (q4_t) 0.0f;
@@ -3412,11 +3043,11 @@ kernel void kernel_flash_attn_ext_vec(
         const short ty = tiisg/NL;
 
         // broadcast kv
-        //const short rk2 = ne02/ne12;
-        //const short rk3 = ne03/ne13;
+        //const short rk2 = args.ne02/args.ne12;
+        //const short rk3 = args.ne03/args.ne13;
 
-        const short ikv2 = iq2/(ne02/ne_12_2);
-        const short ikv3 = iq3/(ne03/ne_12_3);
+        const short ikv2 = iq2/(args.ne02/args.ne_12_2);
+        const short ikv3 = iq3/(args.ne03/args.ne_12_3);
 
         // load the queries from shared memory into local memory
         q4x4_t mq[D16/NL];
@@ -3429,25 +3060,25 @@ kernel void kernel_flash_attn_ext_vec(
         const bool has_mask = mask != q;
 
         // pointer to the mask
-        device const half * pm = (device const half *) (mask + iq1*nb31);
+        device const half * pm = (device const half *) (mask + iq1*args.nb31);
 
         half slope = 1.0f;
 
         // ALiBi
-        if (max_bias > 0.0f) {
+        if (args.max_bias > 0.0f) {
             const short h = iq2;
 
-            const half  base = h < n_head_log2 ? m0 : m1;
-            const short exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+            const half  base = h < args.n_head_log2 ? args.m0 : args.m1;
+            const short exph = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
 
             slope = pow(base, exph);
         }
 
         // loop over the KV cache
         // each simdgroup handles blocks of Q rows and C columns
-        for (int ic0 = 0; ic0 < ne11; ic0 += C*nsg) {
+        for (int ic0 = 0; ic0 < args.ne11; ic0 += C*nsg) {
             const int ic = ic0 + C*sgitg;
-            if (ic >= ne11) {
+            if (ic >= args.ne11) {
                 break;
             }
 
@@ -3461,7 +3092,7 @@ kernel void kernel_flash_attn_ext_vec(
                 for (short cc = 0; cc < C/4; ++cc) {
                     qk_t mqka[4] = { 0.0, 0.0, 0.0, 0.0 };
 
-                    device const kd4x4_t * pk = (device const kd4x4_t *) ((device const char *) k + ((ic + 4*cc + ty)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
+                    device const kd4x4_t * pk = (device const kd4x4_t *) ((device const char *) k + ((ic + 4*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
 
                     #pragma unroll(D16/NL)
                     for (short ii = 0; ii < D16; ii += NL) {
@@ -3497,10 +3128,10 @@ kernel void kernel_flash_attn_ext_vec(
 
                     // mqk = mqk*scale + mask*slope
                     if (tx == 0) {
-                        mqk *= scale;
+                        mqk *= args.scale;
 
-                        if (logit_softcap != 0.0f) {
-                            mqk = logit_softcap*precise::tanh(mqk);
+                        if (args.logit_softcap != 0.0f) {
+                            mqk = args.logit_softcap*precise::tanh(mqk);
                         }
 
                         mqk += sm[4*cc + ty]*slope;
@@ -3539,7 +3170,7 @@ kernel void kernel_flash_attn_ext_vec(
             // O = O + (Q*K^T)*V
             {
                 for (short cc = 0; cc < C/4; ++cc) {
-                    device const vd4x4_t * pv4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 4*cc + ty)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
+                    device const vd4x4_t * pv4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 4*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
 
                     const s4x4_t ms(ss[4*cc + ty]);
 
@@ -3644,7 +3275,7 @@ kernel void kernel_flash_attn_ext_vec(
         const float S = ss[0];
 
         for (short i = tiisg; i < D16; i += NW) {
-            dst44[((int64_t)iq3*ne2*ne1 + iq2 + (iq1)*ne1)*D16 + i] = (float4x4) sr4x4[i]/S;
+            dst44[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)iq1*args.ne1)*D16 + i] = (float4x4) sr4x4[i]/S;
         }
     }
 }
@@ -3686,42 +3317,27 @@ template [[host_name("kernel_flash_attn_ext_vec_q8_0_h256")]] kernel flash_attn_
 
 template<typename T0, typename T1>
 kernel void kernel_cpy(
-        device  const void * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+    const int64_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
 
-    device T1 * dst_data = (device T1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
-        device const T0 * src = (device T0 *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
         dst_data[i00] = (T1) src[0];
     }
 }
@@ -3741,42 +3357,27 @@ template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy<bf
 #endif
 
 kernel void kernel_cpy_f32_q8_0(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK8_0;
+    const int64_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK8_0;
 
-    device block_q8_0 * dst_data = (device block_q8_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device block_q8_0 * dst_data = (device block_q8_0 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x*QK8_0; i00 < ne00; i00 += ntg.x*QK8_0) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x*QK8_0; i00 < args.ne00; i00 += ntg.x*QK8_0) {
+        device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
         float amax = 0.0f; // absolute max
 
@@ -3799,42 +3400,27 @@ kernel void kernel_cpy_f32_q8_0(
 }
 
 kernel void kernel_cpy_f32_q4_0(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_0;
+    const int64_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK4_0;
 
-    device block_q4_0 * dst_data = (device block_q4_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device block_q4_0 * dst_data = (device block_q4_0 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x*QK4_0; i00 < ne00; i00 += ntg.x*QK4_0) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x*QK4_0; i00 < args.ne00; i00 += ntg.x*QK4_0) {
+        device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
         float amax = 0.0f; // absolute max
         float max  = 0.0f;
@@ -3866,42 +3452,27 @@ kernel void kernel_cpy_f32_q4_0(
 }
 
 kernel void kernel_cpy_f32_q4_1(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_1;
+    const int64_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK4_1;
 
-    device block_q4_1 * dst_data = (device block_q4_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device block_q4_1 * dst_data = (device block_q4_1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x*QK4_1; i00 < ne00; i00 += ntg.x*QK4_1) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x*QK4_1; i00 < args.ne00; i00 += ntg.x*QK4_1) {
+        device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
         float min = FLT_MAX;
         float max = -FLT_MAX;
@@ -3932,42 +3503,27 @@ kernel void kernel_cpy_f32_q4_1(
 }
 
 kernel void kernel_cpy_f32_q5_0(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_0;
+    const int64_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK5_0;
 
-    device block_q5_0 * dst_data = (device block_q5_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device block_q5_0 * dst_data = (device block_q5_0 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x*QK5_0; i00 < ne00; i00 += ntg.x*QK5_0) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x*QK5_0; i00 < args.ne00; i00 += ntg.x*QK5_0) {
+        device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
         float amax = 0.0f; // absolute max
         float max  = 0.0f;
@@ -4005,42 +3561,27 @@ kernel void kernel_cpy_f32_q5_0(
 }
 
 kernel void kernel_cpy_f32_q5_1(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_1;
+    const int64_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK5_1;
 
-    device block_q5_1 * dst_data = (device block_q5_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device block_q5_1 * dst_data = (device block_q5_1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x*QK5_1; i00 < ne00; i00 += ntg.x*QK5_1) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x*QK5_1; i00 < args.ne00; i00 += ntg.x*QK5_1) {
+        device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
         float max = src[0];
         float min = src[0];
@@ -4088,42 +3629,27 @@ static inline int best_index_int8(int n, constant float * val, float x) {
 }
 
 kernel void kernel_cpy_f32_iq4_nl(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_NL;
+    const int64_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK4_NL;
 
-    device block_iq4_nl * dst_data = (device block_iq4_nl *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device block_iq4_nl * dst_data = (device block_iq4_nl *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x*QK4_NL; i00 < ne00; i00 += ntg.x*QK4_NL) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x*QK4_NL; i00 < args.ne00; i00 += ntg.x*QK4_NL) {
+        device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
         float amax = 0.0f; // absolute max
         float max  = 0.0f;
@@ -4159,104 +3685,66 @@ kernel void kernel_cpy_f32_iq4_nl(
         }
 
         dst_data[i00/QK4_NL].d = sumq2 > 0 ? sumqx/sumq2 : d;
-
     }
 }
 
 kernel void kernel_concat(
+    constant ggml_metal_kargs_concat & args,
     device  const char * src0,
     device  const char * src1,
     device        char * dst,
-    constant   int64_t & ne00,
-    constant   int64_t & ne01,
-    constant   int64_t & ne02,
-    constant   int64_t & ne03,
-    constant  uint64_t & nb00,
-    constant  uint64_t & nb01,
-    constant  uint64_t & nb02,
-    constant  uint64_t & nb03,
-    constant   int64_t & ne10,
-    constant   int64_t & ne11,
-    constant   int64_t & ne12,
-    constant   int64_t & ne13,
-    constant  uint64_t & nb10,
-    constant  uint64_t & nb11,
-    constant  uint64_t & nb12,
-    constant  uint64_t & nb13,
-    constant   int64_t & ne0,
-    constant   int64_t & ne1,
-    constant   int64_t & ne2,
-    constant   int64_t & ne3,
-    constant  uint64_t & nb0,
-    constant  uint64_t & nb1,
-    constant  uint64_t & nb2,
-    constant  uint64_t & nb3,
-    constant   int32_t & dim,
-    uint3 tgpig[[threadgroup_position_in_grid]],
-    uint3 tpitg[[thread_position_in_threadgroup]],
-    uint3   ntg[[threads_per_threadgroup]]) {
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    ushort3 tpitg[[thread_position_in_threadgroup]],
+    ushort3   ntg[[threads_per_threadgroup]]) {
 
-    const int64_t i3 = tgpig.z;
-    const int64_t i2 = tgpig.y;
-    const int64_t i1 = tgpig.x;
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
 
-    int64_t o[4] = {0, 0, 0, 0};
-    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+    int o[4] = {0, 0, 0, 0};
+    o[args.dim] = args.dim == 0 ? args.ne00 : (args.dim == 1 ? args.ne01 : (args.dim == 2 ? args.ne02 : args.ne03));
 
     device const float * x;
 
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-            x = (device const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            x = (device const float *)(src0 + (i3       )*args.nb03 + (i2       )*args.nb02 + (i1       )*args.nb01 + (i0       )*args.nb00);
         } else {
-            x = (device const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
+            x = (device const float *)(src1 + (i3 - o[3])*args.nb13 + (i2 - o[2])*args.nb12 + (i1 - o[1])*args.nb11 + (i0 - o[0])*args.nb10);
         }
 
-        device float * y = (device float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+        device float * y = (device float *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
         *y = *x;
     }
 }
 
+template<typename args_t>
 void kernel_mul_mv_q2_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_q2_K * x = (device const block_q2_K *) ((device char *) src0 + offset0);
-    device const float      * y = (device const float      *) ((device char *) src1 + offset1);
+    device const block_q2_K * x = (device const block_q2_K *) (src0 + offset0);
+    device const float      * y = (device const float      *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
@@ -4305,92 +3793,64 @@ void kernel_mul_mv_q2_K_f32_impl(
                                  (acc1[3] + 1.f/256.f * acc2[3]) * (sc[6] & 0xF) * 1.f/64.f) -
                          dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) + sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));
 
-            qs += nb01/2;
-            sc += nb01;
-            dh += nb01/2;
+            qs += args.nb01/2;
+            sc += args.nb01;
+            dh += args.nb01/2;
         }
 
         y4 += 4 * QK_K;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_q2_K_f32")]]
 kernel void kernel_mul_mv_q2_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_q2_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_q2_K_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
+template<typename args_t>
 void kernel_mul_mv_q3_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int64_t im = tgpig.z;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
+
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_q3_K * x = (device const block_q3_K *) ((device char *) src0 + offset0);
-    device const float     * yy = (device const float      *) ((device char *) src1 + offset1);
+    device const block_q3_K * x = (device const block_q3_K *) (src0 + offset0);
+    device const float     * yy = (device const float      *) (src1 + offset1);
 
     float yl[32];
 
@@ -4420,9 +3880,10 @@ void kernel_mul_mv_q3_K_f32_impl(
 
     const ushort4 hm = mm[2*ip + il/2];
 
-    const int shift = 2*il;
-    const float    v1 = il == 0 ? 4.f : 64.f;
-    const float    v2 = 4.f * v1;
+    const short shift = 2*il;
+
+    const float v1 = il == 0 ? 4.f : 64.f;
+    const float v2 = 4.f * v1;
 
     const uint16_t s_shift1 = 4*ip;
     const uint16_t s_shift2 = s_shift1 + il;
@@ -4491,10 +3952,10 @@ void kernel_mul_mv_q3_K_f32_impl(
             sumf1[row] += d1 * (scales[1] - 32);
             sumf2[row] += d2 * (scales[3] - 32);
 
-            q  += nb01/2;
-            h  += nb01/2;
-            a  += nb01/2;
-            dh += nb01/2;
+            q  += args.nb01/2;
+            h  += args.nb01/2;
+            a  += args.nb01/2;
+            dh += args.nb01/2;
         }
 
         y1 += 4 * QK_K;
@@ -4504,66 +3965,39 @@ void kernel_mul_mv_q3_K_f32_impl(
         const float sumf = (sumf1[row] + 0.25f * sumf2[row]) / (1 << shift);
         sumf1[row] = simd_sum(sumf);
     }
+
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     if (tiisg == 0) {
         for (int row = 0; row < 2; ++row) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = sumf1[row];
+            dst_f32[first_row + row] = sumf1[row];
         }
     }
 }
 
 [[host_name("kernel_mul_mv_q3_K_f32")]]
 kernel void kernel_mul_mv_q3_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_q3_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_q3_K_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
+template<typename args_t>
 void kernel_mul_mv_q4_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
 
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
@@ -4574,21 +4008,21 @@ void kernel_mul_mv_q4_K_f32_impl(
     const int iq = it/4;     // 0 or 1
     const int ir = it%4;     // 0...3
 
-    const int nb = ne00/QK_K;
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
     //const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
     const int first_row = r0 * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_q4_K * x = (device const block_q4_K *) ((device char *) src0 + offset0);
-    device const float      * y = (device const float      *) ((device char *) src1 + offset1);
+    device const block_q4_K * x = (device const block_q4_K *) (src0 + offset0);
+    device const float      * y = (device const float      *) (src1 + offset1);
 
     float yl[16];
     float yh[16];
@@ -4641,92 +4075,64 @@ void kernel_mul_mv_q4_K_f32_impl(
                                  (acc2[2] + 1.f/256.f * acc2[3]) * sc8[5] * 1.f/16.f) -
                          dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
 
-            q1 += nb01/2;
-            sc += nb01/2;
-            dh += nb01/2;
+            q1 += args.nb01/2;
+            sc += args.nb01/2;
+            dh += args.nb01/2;
         }
 
         y4 += 4 * QK_K;
     }
 
+    device float * dst_f32 = (device float *) dst + (int64_t)im*args.ne0*args.ne1 + (int64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_q4_K_f32")]]
 kernel void kernel_mul_mv_q4_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_q4_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_q4_K_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
+template<typename args_t>
 void kernel_mul_mv_q5_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
+
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_q5_K * x = (device const block_q5_K *) ((device char *) src0 + offset0);
-    device const float     * yy = (device const float      *) ((device char *) src1 + offset1);
+    device const block_q5_K * x = (device const block_q5_K *) (src0 + offset0);
+    device const float     * yy = (device const float      *) (src1 + offset1);
 
     float sumf[2]={0.f};
 
@@ -4800,98 +4206,70 @@ void kernel_mul_mv_q5_K_f32_impl(
                                  sc8[5] * (acc1[3]/16.f + 16.f*acc2[3])) -
                          dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
 
-            q1 += nb01;
-            qh += nb01;
-            dh += nb01/2;
-            a  += nb01/2;
+            q1 += args.nb01;
+            qh += args.nb01;
+            dh += args.nb01/2;
+            a  += args.nb01/2;
         }
 
         y1 += 4 * QK_K;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < 2; ++row) {
         const float tot = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+            dst_f32[first_row + row] = tot;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_q5_K_f32")]]
 kernel void kernel_mul_mv_q5_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_q5_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_q5_K_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
+template <typename args_t>
 void kernel_mul_mv_q6_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
 
     const uint8_t kmask1 = 0x03;
     const uint8_t kmask2 = 0x0C;
     const uint8_t kmask3 = 0x30;
     const uint8_t kmask4 = 0xC0;
 
-    const int nb = ne00/QK_K;
+    const int nb = args.ne00/QK_K;
 
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int     im = tgpig.z;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
 
-    const int row = 2 * r0 + sgitg;
+    const int row = 2*r0 + sgitg;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =  r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =  r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_q6_K * x = (device const block_q6_K *) ((device char *) src0 + offset0);
-    device const float     * yy = (device const float      *) ((device char *) src1 + offset1);
+    device const block_q6_K * x = (device const block_q6_K *) (src0 + offset0);
+    device const float     * yy = (device const float      *) (src1 + offset1);
 
     float sumf = 0;
 
@@ -4908,7 +4286,6 @@ void kernel_mul_mv_q6_K_f32_impl(
     const int q_offset_h = 32*ip + l0;
 
     for (int i = ix; i < nb; i += 2) {
-
         device const uint8_t * q1 = x[i].ql + q_offset_l;
         device const uint8_t * q2 = q1 + 32;
         device const uint8_t * qh = x[i].qh + q_offset_h;
@@ -4930,98 +4307,70 @@ void kernel_mul_mv_q6_K_f32_impl(
 
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     const float tot = simd_sum(sumf);
     if (tiisg == 0) {
-        dst[r1*ne0 + im*ne0*ne1 + row] = tot;
+        dst_f32[row] = tot;
     }
 }
 
 [[host_name("kernel_mul_mv_q6_K_f32")]]
 kernel void kernel_mul_mv_q6_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_q6_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_q6_K_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 // ======================= "True" 2-bit
 
+template<typename args_t>
 void kernel_mul_mv_iq2_xxs_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq2_xxs * x = (device const block_iq2_xxs *) ((device char *) src0 + offset0);
-    device const float         * y = (device const float         *) ((device char *) src1 + offset1);
+    device const block_iq2_xxs * x = (device const block_iq2_xxs *) (src0 + offset0);
+    device const float         * y = (device const float         *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
 
     const int nb32 = nb * (QK_K / 32);
 
-    threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
-    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 256);
+    threadgroup uint64_t * svalues = (threadgroup uint64_t *)(shmem);
+    threadgroup uint8_t  * ssigns  = (threadgroup uint8_t  *)(svalues + 256);
     {
         int nval = 4;
         int pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) values[pos + i] = iq2xxs_grid[pos + i];
+        for (int i = 0; i < nval; ++i) svalues[pos + i] = iq2xxs_grid[pos + i];
         nval = 2;
         pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        for (int i = 0; i < nval; ++i) ssigns[pos+i] = ksigns_iq2xs[pos+i];
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
@@ -5051,114 +4400,85 @@ void kernel_mul_mv_iq2_xxs_f32_impl(
 
             float sum = 0;
             for (int l = 0; l < 4; ++l) {
-                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + aux8[l]);
-                const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(svalues + aux8[l]);
+                const uint8_t signs = ssigns[(aux32 >> 7*l) & 127];
                 for (int j = 0; j < 8; ++j) {
                     sum += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
                 }
             }
             sumf[row] += d * sum;
 
-            dh += nb01/2;
-            q2 += nb01/2;
+            dh += args.nb01/2;
+            q2 += args.nb01/2;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+            dst_f32[first_row + row] = all_sum * 0.25f;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_iq2_xxs_f32")]]
 kernel void kernel_mul_mv_iq2_xxs_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_iq2_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    kernel_mul_mv_iq2_xxs_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
+template<typename args_t>
 void kernel_mul_mv_iq2_xs_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq2_xs * x = (device const block_iq2_xs *) ((device char *) src0 + offset0);
-    device const float        * y = (device const float        *) ((device char *) src1 + offset1);
+    device const block_iq2_xs * x = (device const block_iq2_xs *) (src0 + offset0);
+    device const float        * y = (device const float        *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
 
     const int nb32 = nb * (QK_K / 32);
 
-    threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
-    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 512);
+    threadgroup uint64_t * svalues = (threadgroup uint64_t *)(shmem);
+    threadgroup uint8_t  * ssigns  = (threadgroup uint8_t  *)(svalues + 512);
     {
         int nval = 8;
         int pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) values[pos + i] = iq2xs_grid[pos + i];
+        for (int i = 0; i < nval; ++i) svalues[pos + i] = iq2xs_grid[pos + i];
         nval = 2;
         pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        for (int i = 0; i < nval; ++i) ssigns[pos+i] = ksigns_iq2xs[pos+i];
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
@@ -5190,122 +4510,94 @@ void kernel_mul_mv_iq2_xs_f32_impl(
 
             float sum1 = 0, sum2 = 0;
             for (int l = 0; l < 2; ++l) {
-                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
-                const uint8_t signs = shared_signs[(q2[l] >> 9)];
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(svalues + (q2[l] & 511));
+                const uint8_t signs = ssigns[(q2[l] >> 9)];
                 for (int j = 0; j < 8; ++j) {
                     sum1 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
                 }
             }
             for (int l = 2; l < 4; ++l) {
-                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
-                const uint8_t signs = shared_signs[(q2[l] >> 9)];
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(svalues + (q2[l] & 511));
+                const uint8_t signs = ssigns[(q2[l] >> 9)];
                 for (int j = 0; j < 8; ++j) {
                     sum2 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
                 }
             }
             sumf[row] += d1 * sum1 + d2 * sum2;
 
-            dh += nb01/2;
-            q2 += nb01/2;
-            sc += nb01;
+            dh += args.nb01/2;
+            q2 += args.nb01/2;
+            sc += args.nb01;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+            dst_f32[first_row + row] = all_sum * 0.25f;
         }
-    }
-}
-
-[[host_name("kernel_mul_mv_iq2_xs_f32")]]
-kernel void kernel_mul_mv_iq2_xs_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    }
+}
+
+[[host_name("kernel_mul_mv_iq2_xs_f32")]]
+kernel void kernel_mul_mv_iq2_xs_f32(
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq2_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq2_xs_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
+template <typename args_t>
 void kernel_mul_mv_iq3_xxs_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq3_xxs * x = (device const block_iq3_xxs *) ((device char *) src0 + offset0);
-    device const float         * y = (device const float         *) ((device char *) src1 + offset1);
+    device const block_iq3_xxs * x = (device const block_iq3_xxs *) (src0 + offset0);
+    device const float         * y = (device const float         *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
 
     const int nb32 = nb * (QK_K / 32);
 
-    threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
-    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 256);
+    threadgroup uint32_t * svalues = (threadgroup uint32_t *)(shmem);
+    threadgroup uint8_t  * ssigns  = (threadgroup uint8_t  *)(svalues + 256);
     {
         int nval = 4;
         int pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) values[pos + i] = iq3xxs_grid[pos + i];
+        for (int i = 0; i < nval; ++i) svalues[pos + i] = iq3xxs_grid[pos + i];
         nval = 2;
         pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        for (int i = 0; i < nval; ++i) ssigns[pos+i] = ksigns_iq2xs[pos+i];
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
@@ -5314,7 +4606,6 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
     device const float * y4 = y + 32 * ix;
 
     for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
-
         for (int i = 0; i < 32; ++i) {
             yl[i] = y4[i];
         }
@@ -5328,16 +4619,15 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
         device const half * dh = &xr->d;
 
         for (int row = 0; row < N_DST; row++) {
-
             const float db = dh[0];
             const uint32_t aux32 = gas[0] | (gas[1] << 16);
             const float d = db * (0.5f + (aux32 >> 28));
 
             float2 sum = {0};
             for (int l = 0; l < 4; ++l) {
-                const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + q3[2*l+0]);
-                const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + q3[2*l+1]);
-                const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
+                const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(svalues + q3[2*l+0]);
+                const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(svalues + q3[2*l+1]);
+                const uint8_t signs = ssigns[(aux32 >> 7*l) & 127];
                 for (int j = 0; j < 4; ++j) {
                     sum[0] += yl[8*l + j + 0] * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
                     sum[1] += yl[8*l + j + 4] * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
@@ -5345,103 +4635,75 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
             }
             sumf[row] += d * (sum[0] + sum[1]);
 
-            dh  += nb01/2;
-            q3  += nb01;
-            gas += nb01/2;
+            dh  += args.nb01/2;
+            q3  += args.nb01;
+            gas += args.nb01/2;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.5f;
+            dst_f32[first_row + row] = all_sum * 0.5f;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_iq3_xxs_f32")]]
 kernel void kernel_mul_mv_iq3_xxs_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq3_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq3_xxs_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
+template<typename args_t>
 void kernel_mul_mv_iq3_s_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq3_s * x = (device const block_iq3_s *) ((device char *) src0 + offset0);
-    device const float       * y = (device const float       *) ((device char *) src1 + offset1);
+    device const block_iq3_s * x = (device const block_iq3_s *) (src0 + offset0);
+    device const float       * y = (device const float       *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
 
     const int nb32 = nb * (QK_K / 32);
 
-    threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
+    threadgroup uint32_t * svalues = (threadgroup uint32_t *) shmem;
     {
         int nval = 8;
         int pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) values[pos + i] = iq3s_grid[pos + i];
+        for (int i = 0; i < nval; ++i) svalues[pos + i] = iq3s_grid[pos + i];
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
@@ -5472,8 +4734,8 @@ void kernel_mul_mv_iq3_s_f32_impl(
 
             float2 sum = {0};
             for (int l = 0; l < 4; ++l) {
-                const threadgroup uint32_t * table1 = qh[0] & kmask_iq2xs[2*l+0] ? values + 256 : values;
-                const threadgroup uint32_t * table2 = qh[0] & kmask_iq2xs[2*l+1] ? values + 256 : values;
+                const threadgroup uint32_t * table1 = qh[0] & kmask_iq2xs[2*l+0] ? svalues + 256 : svalues;
+                const threadgroup uint32_t * table2 = qh[0] & kmask_iq2xs[2*l+1] ? svalues + 256 : svalues;
                 const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(table1 + qs[2*l+0]);
                 const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(table2 + qs[2*l+1]);
                 for (int j = 0; j < 4; ++j) {
@@ -5483,105 +4745,77 @@ void kernel_mul_mv_iq3_s_f32_impl(
             }
             sumf[row] += d * (sum[0] + sum[1]);
 
-            dh    += nb01/2;
-            qs    += nb01;
-            qh    += nb01;
-            sc    += nb01;
-            signs += nb01;
+            dh    += args.nb01/2;
+            qs    += args.nb01;
+            qh    += args.nb01;
+            sc    += args.nb01;
+            signs += args.nb01;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_iq3_s_f32")]]
 kernel void kernel_mul_mv_iq3_s_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq3_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq3_s_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
+template <typename args_t>
 void kernel_mul_mv_iq2_s_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq2_s * x = (device const block_iq2_s *) ((device char *) src0 + offset0);
-    device const float       * y = (device const float       *) ((device char *) src1 + offset1);
+    device const block_iq2_s * x = (device const block_iq2_s *) (src0 + offset0);
+    device const float       * y = (device const float       *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
 
     const int nb32 = nb * (QK_K / 32);
 
-    //threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
+    //threadgroup uint64_t * svalues = (threadgroup uint64_t *) shmem;
     //{
     //    int nval = 32;
     //    int pos  = (32*sgitg + tiisg)*nval;
-    //    for (int i = 0; i < nval; ++i) values[pos + i] = iq2s_grid[pos + i];
+    //    for (int i = 0; i < nval; ++i) svalues[pos + i] = iq2s_grid[pos + i];
     //    threadgroup_barrier(mem_flags::mem_threadgroup);
     //}
 
@@ -5613,8 +4847,8 @@ void kernel_mul_mv_iq2_s_f32_impl(
 
             float2 sum = {0};
             for (int l = 0; l < 2; ++l) {
-                //const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
-                //const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
+                //const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(svalues + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
+                //const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(svalues + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
                 constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
                 constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
                 for (int j = 0; j < 8; ++j) {
@@ -5624,94 +4858,66 @@ void kernel_mul_mv_iq2_s_f32_impl(
             }
             sumf[row] += d1 * sum[0] + d2 * sum[1];
 
-            dh    += nb01/2;
-            qs    += nb01;
-            qh    += nb01;
-            sc    += nb01;
-            signs += nb01;
+            dh    += args.nb01/2;
+            qs    += args.nb01;
+            qh    += args.nb01;
+            sc    += args.nb01;
+            signs += args.nb01;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+            dst_f32[first_row + row] = all_sum * 0.25f;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_iq2_s_f32")]]
 kernel void kernel_mul_mv_iq2_s_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq2_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq2_s_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
+template<typename args_t>
 void kernel_mul_mv_iq1_s_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_value,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq1_s * x = (device const block_iq1_s *) ((device char *) src0 + offset0);
-    device const float       * y = (device const float       *) ((device char *) src1 + offset1);
+    device const block_iq1_s * x = (device const block_iq1_s *) (src0 + offset0);
+    device const float       * y = (device const float       *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
@@ -5754,61 +4960,50 @@ void kernel_mul_mv_iq1_s_f32_impl(
             }
             sumf[row] += (float)dh[0] * (sum + sumy * (qh[0] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA)) * (2*((qh[0] >> 12) & 7) + 1);
 
-            dh += nb01/2;
-            qs += nb01;
-            qh += nb01/2;
+            dh += args.nb01/2;
+            qs += args.nb01;
+            qh += args.nb01/2;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
+template <typename args_t>
 void kernel_mul_mv_iq1_m_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_value,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq1_m * x = (device const block_iq1_m *) ((device char *) src0 + offset0);
-    device const float       * y = (device const float       *) ((device char *) src1 + offset1);
+    device const block_iq1_m * x = (device const block_iq1_m *) (src0 + offset0);
+    device const float       * y = (device const float       *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
@@ -5860,66 +5055,55 @@ void kernel_mul_mv_iq1_m_f32_impl(
             sumf[row] += (float)scale.f16 * ((sum[0] + delta1) * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 7) + 1) +
                                              (sum[1] + delta2) * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 7) + 1));
 
-            sc += nb01/2;
-            qs += nb01;
-            qh += nb01;
+            sc += args.nb01/2;
+            qs += args.nb01;
+            qh += args.nb01;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
+template<typename args_t>
 void kernel_mul_mv_iq4_nl_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values_i8,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    threadgroup float * shared_values = (threadgroup float *)shared_values_i8;
-    const int nb = ne00/QK4_NL;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+    const int nb = args.ne00/QK4_NL;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
     const int first_row = (r0 * 2 + sgitg) * 2;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq4_nl * x = (device const block_iq4_nl *) ((device char *) src0 + offset0);
-    device const float        * y = (device const float        *) ((device char *) src1 + offset1);
+    device const block_iq4_nl * x = (device const block_iq4_nl *) (src0 + offset0);
+    device const float        * y = (device const float        *) (src1 + offset1);
 
     const int ix = tiisg/2;  // 0...15
     const int it = tiisg%2;  // 0 or 1
 
-    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
+    shmem_f32[tiisg] = kvalues_iq4nl_f[tiisg%16];
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     float4 yl[4];
@@ -5937,7 +5121,7 @@ void kernel_mul_mv_iq4_nl_f32_impl(
         device const float4 * y4 = (device const float4 *)yb;
         yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
 
-        for (int row = 0; row < 2 && first_row + row < ne01; ++row) {
+        for (int row = 0; row < 2 && first_row + row < args.ne01; ++row) {
 
             device const block_iq4_nl & xb = x[row*nb + ib];
             device const uint16_t * q4 = (device const uint16_t *)(xb.qs + 8*it);
@@ -5947,16 +5131,16 @@ void kernel_mul_mv_iq4_nl_f32_impl(
             aux32[0] = q4[0] | (q4[1] << 16);
             aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
             aux32[0] &= 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            qf1 = {shmem_f32[q8[0]], shmem_f32[q8[1]], shmem_f32[q8[2]], shmem_f32[q8[3]]};
+            qf2 = {shmem_f32[q8[4]], shmem_f32[q8[5]], shmem_f32[q8[6]], shmem_f32[q8[7]]};
             acc1 += yl[0] * qf1;
             acc2 += yl[1] * qf2;
 
             aux32[0] = q4[2] | (q4[3] << 16);
             aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
             aux32[0] &= 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            qf1 = {shmem_f32[q8[0]], shmem_f32[q8[1]], shmem_f32[q8[2]], shmem_f32[q8[3]]};
+            qf2 = {shmem_f32[q8[4]], shmem_f32[q8[5]], shmem_f32[q8[6]], shmem_f32[q8[7]]};
             acc1 += yl[2] * qf1;
             acc2 += yl[3] * qf2;
 
@@ -5969,60 +5153,49 @@ void kernel_mul_mv_iq4_nl_f32_impl(
         yb += 16 * QK4_NL;
     }
 
-    for (int row = 0; row < 2 && first_row + row < ne01; ++row) {
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
+    for (int row = 0; row < 2 && first_row + row < args.ne01; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
+template<typename args_t>
 void kernel_mul_mv_iq4_xs_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values_i8,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    threadgroup float * shared_values = (threadgroup float *)shared_values_i8;
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
     const int first_row = (r0 * 2 + sgitg) * 2;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq4_xs * x = (device const block_iq4_xs *) ((device char *) src0 + offset0);
-    device const float        * y = (device const float        *) ((device char *) src1 + offset1);
+    device const block_iq4_xs * x = (device const block_iq4_xs *) (src0 + offset0);
+    device const float        * y = (device const float        *) (src1 + offset1);
 
     const int ix = tiisg/16;  // 0 or 1
     const int it = tiisg%16;  // 0...15
     const int ib = it/2;
     const int il = it%2;
 
-    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
+    shmem_f32[tiisg] = kvalues_iq4nl_f[tiisg%16];
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     float4 yl[4];
@@ -6036,28 +5209,26 @@ void kernel_mul_mv_iq4_xs_f32_impl(
     float4 qf1, qf2;
 
     for (int ibl = ix; ibl < nb; ibl += 2) {
-
         device const float4 * y4 = (device const float4 *)yb;
         yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
 
         for (int row = 0; row < 2; ++row) {
-
             device const block_iq4_xs & xb = x[row*nb + ibl];
             device const uint32_t * q4 = (device const uint32_t *)(xb.qs + 16*ib + 8*il);
 
             float4 acc1 = {0.f}, acc2 = {0.f};
 
-            aux32[0] = q4[0] & 0x0f0f0f0f;
+            aux32[0] = (q4[0]     ) & 0x0f0f0f0f;
             aux32[1] = (q4[0] >> 4) & 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            qf1 = {shmem_f32[q8[0]], shmem_f32[q8[1]], shmem_f32[q8[2]], shmem_f32[q8[3]]};
+            qf2 = {shmem_f32[q8[4]], shmem_f32[q8[5]], shmem_f32[q8[6]], shmem_f32[q8[7]]};
             acc1 += yl[0] * qf1;
             acc2 += yl[1] * qf2;
 
-            aux32[0] = q4[1] & 0x0f0f0f0f;
+            aux32[0] = (q4[1]     ) & 0x0f0f0f0f;
             aux32[1] = (q4[1] >> 4) & 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            qf1 = {shmem_f32[q8[0]], shmem_f32[q8[1]], shmem_f32[q8[2]], shmem_f32[q8[3]]};
+            qf2 = {shmem_f32[q8[4]], shmem_f32[q8[5]], shmem_f32[q8[6]], shmem_f32[q8[7]]};
             acc1 += yl[2] * qf1;
             acc2 += yl[3] * qf2;
 
@@ -6071,134 +5242,68 @@ void kernel_mul_mv_iq4_xs_f32_impl(
         yb += 2 * QK_K;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < 2; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_iq1_s_f32")]]
 kernel void kernel_mul_mv_iq1_s_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq1_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq1_s_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 [[host_name("kernel_mul_mv_iq1_m_f32")]]
 kernel void kernel_mul_mv_iq1_m_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq1_m_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq1_m_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 [[host_name("kernel_mul_mv_iq4_nl_f32")]]
 kernel void kernel_mul_mv_iq4_nl_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq4_nl_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
 [[host_name("kernel_mul_mv_iq4_xs_f32")]]
 kernel void kernel_mul_mv_iq4_xs_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq4_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq4_xs_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
@@ -6302,38 +5407,26 @@ kernel void kernel_get_rows_i32(
 
 // each block_q contains 16*nl weights
 template<typename T, typename T4x4, typename simdgroup_T8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
-kernel void kernel_mul_mm(device const  uchar * src0,
-                          device const  uchar * src1,
-                          device        float * dst,
-                          constant    int64_t & ne00,
-                          constant    int64_t & ne02,
-                          constant   uint64_t & nb01,
-                          constant   uint64_t & nb02,
-                          constant   uint64_t & nb03,
-                          constant    int64_t & ne12,
-                          constant   uint64_t & nb10,
-                          constant   uint64_t & nb11,
-                          constant   uint64_t & nb12,
-                          constant   uint64_t & nb13,
-                          constant    int64_t & ne0,
-                          constant    int64_t & ne1,
-                          constant       uint & r2,
-                          constant       uint & r3,
-                          threadgroup   uchar * shared_memory [[threadgroup(0)]],
-                          uint3                 tgpig[[threadgroup_position_in_grid]],
-                          uint                  tiitg[[thread_index_in_threadgroup]],
-                          uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    threadgroup T     * sa = (threadgroup T     *)(shared_memory);
-    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
-
-    const uint r0 = tgpig.y;
-    const uint r1 = tgpig.x;
-    const uint im = tgpig.z;
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup T     * sa = (threadgroup T     *)(shmem);
+    threadgroup float * sb = (threadgroup float *)(shmem + 4096);
+
+    const int r0 = tgpig.y;
+    const int r1 = tgpig.x;
+    const int im = tgpig.z;
 
     // if this block is of 64x32 shape or smaller
-    short n_rows = (ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    short n_cols = (ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
+    short n_cols = (args.ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (args.ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
 
     // a thread shouldn't load data outside of the matrix
     short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
@@ -6349,20 +5442,20 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 
     short il = (tiitg % THREAD_PER_ROW);
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const int i12 = im%args.ne12;
+    const int i13 = im/args.ne12;
 
-    uint   offset0 = (i12/r2)*nb02 + (i13/r3)*nb03;
-    ushort offset1 = il/nl;
+    uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    short    offset1 = il/nl;
 
-    device const block_q * x = (device const block_q *)(src0 + (r0*BLOCK_SIZE_M + thread_row)*nb01 + offset0) + offset1;
+    device const block_q * x = (device const block_q *)(src0 + (r0*BLOCK_SIZE_M + thread_row)*args.nb01 + offset0) + offset1;
     device const float   * y = (device const float   *)(src1
-        + nb13 * i13
-        + nb12 * i12
-        + nb11 * (r1 * BLOCK_SIZE_N + thread_col)
-        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*(r1 * BLOCK_SIZE_N + thread_col)
+        + args.nb10*(BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
 
-    for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
         // load data and store to threadgroup memory
         T4x4 temp_a;
         dequantize_func(x, il, temp_a);
@@ -6409,16 +5502,18 @@ kernel void kernel_mul_mm(device const  uchar * src0,
         }
     }
 
-    if ((r0 + 1) * BLOCK_SIZE_M <= ne0 && (r1 + 1) * BLOCK_SIZE_N <= ne1) {
-        device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg &  1)) \
-                               + (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
+    if ((r0 + 1) * BLOCK_SIZE_M <= args.ne0 && (r1 + 1) * BLOCK_SIZE_N <= args.ne1) {
+        device float * C = (device float *) dst +
+            (BLOCK_SIZE_M * r0 + 32 * (sgitg &  1)) + \
+            (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+
         for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
+            simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.ne0 * (i/4), args.ne0);
         }
     } else {
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *) shared_memory) \
+        threadgroup float * temp_str = ((threadgroup float *) shmem) \
                                       + 32 * (sgitg&1) + (16 * (sgitg>>1))*BLOCK_SIZE_M;
         for (short i = 0; i < 8; i++) {
             simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
@@ -6428,7 +5523,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 
         if (sgitg == 0) {
             for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
-                device float  * D  = dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*ne0 + im*ne1*ne0;
+                device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*args.ne0 + im*args.ne1*args.ne0;
                 device float4 * D4 = (device float4 *) D;
 
                 threadgroup float  * C  = temp_str + (j*BLOCK_SIZE_M);
@@ -6449,36 +5544,37 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 }
 
 // same as kernel_mul_mm_impl, but src1 and dst are accessed via indices stored in rowids
+// TODO: this kernel needs to be reimplemented from scratch for better performance
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
 void kernel_mul_mm_id_impl(
-        device const  uchar * src0,
-        device const  uchar * src1,
+        int32_t  ne00,
+        int32_t  ne02,
+        uint64_t nb01,
+        uint64_t nb02,
+        int32_t  ne11,
+        int32_t  ne12,
+        uint64_t nb10,
+        uint64_t nb11,
+        uint64_t nb12,
+        int32_t  ne0,
+        int32_t  ne1,
+        int64_t  ne0ne1,
+        device   const char * src0,
+        device   const char * src1,
         threadgroup ushort2 * rowids,
-        device        float * dst,
-        constant    int64_t & ne00,
-        constant    int64_t & ne02,
-        constant   uint64_t & nb01,
-        constant   uint64_t & nb02,
-        constant    int64_t & ne11,
-        constant    int64_t & ne12,
-        constant   uint64_t & nb10,
-        constant   uint64_t & nb11,
-        constant   uint64_t & nb12,
-        constant    int64_t & ne0,
-                    int64_t   ne1,
-                    int64_t   ne0ne1,
-        threadgroup   uchar * shared_memory,
-        uint3                 tgpig[[threadgroup_position_in_grid]],
-        uint                  tiitg[[thread_index_in_threadgroup]],
-        uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    threadgroup half  * sa = (threadgroup half  *)(shared_memory);
-    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
-
-    const uint r0 = tgpig.y;
-    const uint r1 = tgpig.x;
-
-    if (r1 * BLOCK_SIZE_N >= ne1) return;
+        device         char * dst,
+        threadgroup    char * shmem,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup half  * sa = (threadgroup half  *)(shmem);
+    threadgroup float * sb = (threadgroup float *)(shmem + 4096);
+
+    const int r0 = tgpig.y;
+    const int r1 = tgpig.x;
+
+    if (r1*BLOCK_SIZE_N >= ne1) return;
 
     // if this block is of 64x32 shape or smaller
     short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
@@ -6490,9 +5586,9 @@ void kernel_mul_mm_id_impl(
 
     simdgroup_half8x8  ma[4];
     simdgroup_float8x8 mb[2];
-    simdgroup_float8x8 c_res[8];
+    simdgroup_float8x8 mc[8];
     for (int i = 0; i < 8; i++){
-        c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
     }
     short il = (tiitg % THREAD_PER_ROW);
 
@@ -6530,11 +5626,14 @@ void kernel_mul_mm_id_impl(
         threadgroup half  * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
         threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
 
+        #pragma unroll(BLOCK_SIZE_K/8)
         for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
+            #pragma unroll(4)
             for (int i = 0; i < 4; i++) {
                 simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
             }
             simdgroup_barrier(mem_flags::mem_none);
+            #pragma unroll(2)
             for (int i = 0; i < 2; i++) {
                 simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
             }
@@ -6542,29 +5641,42 @@ void kernel_mul_mm_id_impl(
             lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
             lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
 
+            #pragma unroll(8)
             for (int i = 0; i < 8; i++){
-                simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
             }
         }
     }
 
     {
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
+        threadgroup float * temp_str = ((threadgroup float *) shmem) \
                                       + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
         for (int i = 0; i < 8; i++) {
-            simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
+            simdgroup_store(mc[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
         }
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        device float * C = dst + (BLOCK_SIZE_M * r0);
         if (sgitg == 0) {
             for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
                 threadgroup const auto & jid = rowids[r1 * BLOCK_SIZE_N + j];
-                int joff =  jid[0] * ne0 + jid[1] * ne0ne1;
-                for (int i = 0; i < n_rows; i++) {
-                    *(C + i + joff) = *(temp_str + i + j * BLOCK_SIZE_M);
+                int64_t joff = jid[0]*ne0 + jid[1]*ne0ne1;
+
+                device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + joff;
+                device float4 * D4 = (device float4 *) D;
+
+                threadgroup float  * C  = temp_str + (j*BLOCK_SIZE_M);
+                threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+                int i = 0;
+                for (; i < n_rows/4; i++) {
+                    *(D4 + i) = *(C4 + i);
+                }
+
+                i *= 4;
+                for (; i < n_rows; i++) {
+                    *(D + i) = *(C + i);
                 }
             }
         }
@@ -6573,48 +5685,34 @@ void kernel_mul_mm_id_impl(
 
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
 kernel void kernel_mul_mm_id(
-        device const   uchar * src0s,
-        device const   uchar * src1,
-        device         float * dst,
-        device const   uchar * ids,
-        constant     int64_t & nei0,
-        constant     int64_t & nei1,
-        constant    uint64_t & nbi1,
-        constant     int64_t & ne00,
-        constant     int64_t & ne02,
-        constant    uint64_t & nb01,
-        constant    uint64_t & nb02,
-        constant     int64_t & ne11,
-        constant     int64_t & ne12,
-        constant     int64_t & ne13,
-        constant    uint64_t & nb10,
-        constant    uint64_t & nb11,
-        constant    uint64_t & nb12,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant    uint64_t & nb1,
-        threadgroup    uchar * shared_memory [[threadgroup(0)]],
-        uint3                  tgpig[[threadgroup_position_in_grid]],
-        uint                   tiitg[[thread_index_in_threadgroup]],
-        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mm_id & args,
+        device const char * src0s,
+        device const char * src1,
+        device       char * dst,
+        device const char * ids,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const int32_t i02 = tgpig.z;
+
     tgpig.z = 0;
 
-    device const uchar * src0 = src0s + i02*nb02;
+    device const char * src0 = src0s + i02*args.nb02;
 
     // row indices
-    threadgroup ushort2 * rowids = (threadgroup ushort2 *)(shared_memory + 8192);
+    threadgroup ushort2 * rowids = (threadgroup ushort2 *)(shmem + 8192);
 
     // TODO: parallelize this loop
-    int64_t _ne1 = 0;
-    for (ushort ii1 = 0; ii1 < nei1; ii1++) {
-        for (ushort ii0 = 0; ii0 < nei0; ii0++) {
-            int32_t id = ((device int32_t *) (ids + ii1*nbi1))[ii0];
+    int32_t _ne1 = 0;
+    for (ushort ii1 = 0; ii1 < args.nei1; ii1++) {
+        for (ushort ii0 = 0; ii0 < args.nei0; ii0++) {
+            int32_t id = ((device int32_t *) (ids + ii1*args.nbi1))[ii0];
             if (id == i02) {
-                //if (tiitg == 0) {
+                if (tiitg == 0) {
                     rowids[_ne1] = ushort2(ii0, ii1);
-                //}
+                }
                 _ne1++;
             }
         }
@@ -6623,23 +5721,23 @@ kernel void kernel_mul_mm_id(
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     kernel_mul_mm_id_impl<block_q, nl, dequantize_func>(
+        args.ne00,
+        args.ne02,
+        args.nb01,
+        args.nb02,
+        args.ne11,
+        args.ne12,
+        args.nb10,
+        args.nb11,
+        args.nb12,
+        args.ne0,
+        _ne1,
+        (int64_t)args.ne0*args.ne1,
         src0,
         src1,
         rowids,
         dst,
-        ne00,
-        ne02,
-        nb01,
-        nb02,
-        ne11,
-        ne12,
-        nb10,
-        nb11,
-        nb12,
-        ne0,
-        _ne1,
-        ne0*ne1,
-        shared_memory,
+        shmem,
         tgpig,
         tiitg,
         sgitg);
@@ -6748,194 +5846,110 @@ template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mat_mm_id_t kernel
 //
 
 typedef void (kernel_mul_mv_impl_t)(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb00,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne11,
-                   int64_t   ne12,
-                  uint64_t   nb10,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-                   uint3     tgpig,
-                   uint      tiisg);
+        ggml_metal_kargs_mul_mv args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig,
+        ushort tiisg);
 
 typedef void (kernel_mul_mv2_impl_t)(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg);
+        ggml_metal_kargs_mul_mv args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg);
 
 template<kernel_mul_mv_impl_t impl_fn>
 void mmv_fn(
-        device const    char * src0,
-        device const    char * src1,
-        device         float * dst,
-                     int64_t   ne00,
-                     int64_t   ne01,
-                     int64_t   ne02,
-                    uint64_t   nb00,
-                    uint64_t   nb01,
-                    uint64_t   nb02,
-                    uint64_t   nb03,
-                     int64_t   ne10,
-                     int64_t   ne11,
-                     int64_t   ne12,
-                     int64_t   ne13,
-                    uint64_t   nb10,
-                    uint64_t   nb11,
-                    uint64_t   nb12,
-                    uint64_t   nb13,
-                     int64_t   ne0,
-                     int64_t   ne1,
-                    uint64_t   nb1,
-                        uint   r2,
-                        uint   r3,
-        threadgroup int8_t   * shared_values,
-        uint3                  tgpig,
-        uint                   tiitg,
-        uint                   tiisg,
-        uint                   sgitg) {
-    impl_fn(src0,src1,dst,ne00,ne01,ne02,nb00,nb01,nb02,nb03,ne10,ne11,ne12,nb10,nb11,nb12,nb13,ne0,ne1,r2,r3,tgpig,tiisg);
+        ggml_metal_kargs_mul_mv args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiitg,
+        ushort tiisg,
+        ushort sgitg) {
+    impl_fn(args, src0, src1, dst, tgpig, tiisg);
 }
 
 template<kernel_mul_mv2_impl_t impl_fn>
 void mmv_fn(
-        device const    char * src0,
-        device const    char * src1,
-        device         float * dst,
-                     int64_t   ne00,
-                     int64_t   ne01,
-                     int64_t   ne02,
-                    uint64_t   nb00,
-                    uint64_t   nb01,
-                    uint64_t   nb02,
-                    uint64_t   nb03,
-                     int64_t   ne10,
-                     int64_t   ne11,
-                     int64_t   ne12,
-                     int64_t   ne13,
-                    uint64_t   nb10,
-                    uint64_t   nb11,
-                    uint64_t   nb12,
-                    uint64_t   nb13,
-                     int64_t   ne0,
-                     int64_t   ne1,
-                    uint64_t   nb1,
-                        uint   r2,
-                        uint   r3,
-        threadgroup int8_t   * shared_values,
-        uint3                  tgpig,
-        uint                   tiitg,
-        uint                   tiisg,
-        uint                   sgitg) {
-    impl_fn(src0,(const device float *)src1,dst,ne00,ne01,ne02,nb01,nb02,nb03,ne10,ne12,nb11,nb12,nb13,ne0,ne1,r2,r3,shared_values,tgpig,tiisg,sgitg);
-}
-
-typedef decltype(mmv_fn<kernel_mul_mv_impl<half, half4, half, half4>>) mul_mv_impl_fn_t;
+        ggml_metal_kargs_mul_mv args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiitg,
+        ushort tiisg,
+        ushort sgitg) {
+    impl_fn(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
+}
+
+typedef decltype(mmv_fn<kernel_mul_mv_impl<half, half4, half, half4, ggml_metal_kargs_mul_mv>>) mul_mv_impl_fn_t;
 
 template<mul_mv_impl_fn_t impl_fn>
 kernel void kernel_mul_mv_id(
-        device const    char * src0s,
-        device const    char * src1,
-        device         float * dst,
-        device const    char * ids,
-        constant     int64_t & nei0,
-        constant     int64_t & nei1,
-        constant    uint64_t & nbi1,
-        constant     int64_t & ne00,
-        constant     int64_t & ne01,
-        constant     int64_t & ne02,
-        constant    uint64_t & nb00,
-        constant    uint64_t & nb01,
-        constant    uint64_t & nb02,
-        constant     int64_t & ne10,
-        constant     int64_t & ne11,
-        constant     int64_t & ne12,
-        constant     int64_t & ne13,
-        constant    uint64_t & nb10,
-        constant    uint64_t & nb11,
-        constant    uint64_t & nb12,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant    uint64_t & nb1,
-        threadgroup int8_t   * shared_values [[threadgroup(0)]],
-        uint3                  tgpig[[threadgroup_position_in_grid]],
-        uint                   tiitg[[thread_index_in_threadgroup]],
-        uint                   tiisg[[thread_index_in_simdgroup]],
-        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
-    const int iid1 = tgpig.z/nei0;
-    const int idx = tgpig.z%nei0;
+        constant ggml_metal_kargs_mul_mv_id & args,
+        device const char * src0s,
+        device const char * src1,
+        device       char * dst,
+        device const char * ids,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    const int iid1 = tgpig.z/args.nei0;
+    const int idx  = tgpig.z%args.nei0;
 
     tgpig.z = 0;
 
-    const int32_t i02 = ((device const int32_t *) (ids + iid1*nbi1))[idx];
+    const int32_t i02 = ((device const int32_t *) (ids + iid1*args.nbi1))[idx];
 
-    const int64_t i11 = idx % ne11;
+    const int64_t i11 = idx % args.ne11;
     const int64_t i12 = iid1;
 
     const int64_t i1 = idx;
     const int64_t i2 = i12;
 
-    device const char * src0_cur = src0s + i02*nb02;
-    device const char * src1_cur = src1  + i11*nb11 + i12*nb12;
-    device      float *  dst_cur = dst   + i1*ne0   + i2*ne1*ne0;
+    device const char * src0_cur = src0s + i02*args.nb02;
+    device const char * src1_cur = src1  + i11*args.nb11 + i12*args.nb12;
+
+    device char * dst_cur = dst + (i1*args.ne0 + i2*args.ne1*args.ne0)*sizeof(float);
+
+    ggml_metal_kargs_mul_mv args0 = {
+        /*.ne00 =*/ args.ne00,
+        /*.ne01 =*/ args.ne01,
+        /*.ne02 =*/ 1, // args.ne02,
+        /*.nb00 =*/ args.nb00,
+        /*.nb01 =*/ args.nb01,
+        /*.nb02 =*/ args.nb02,
+        /*.nb03 =*/ args.nb02, // args.ne02 == 1
+        /*.ne10 =*/ args.ne10,
+        /*.ne11 =*/ 1, // args.ne11,
+        /*.ne12 =*/ 1, // args.ne12,
+        /*.nb10 =*/ args.nb10,
+        /*.nb11 =*/ args.nb11,
+        /*.nb12 =*/ args.nb12,
+        /*.nb13 =*/ args.nb12, // ne12 == 1
+        /*.ne0  =*/ args.ne0,
+        /*.ne1  =*/ 1, // args.ne1,
+        /*.r2   =*/ 1,
+        /*.r3   =*/ 1,
+    };
 
     impl_fn(
+        args0,
         /* src0 */ src0_cur,
         /* src1 */ src1_cur,
         /* dst  */ dst_cur,
-        /* ne00 */ ne00,
-        /* ne01 */ ne01,
-        /* ne02 */ 1, // ne02,
-        /* nb00 */ nb00,
-        /* nb01 */ nb01,
-        /* nb02 */ nb02,
-        /* nb03 */ nb02, // ne02 == 1
-        /* ne10 */ ne10,
-        /* ne11 */ 1, // ne11,
-        /* ne12 */ 1, // ne12,
-        /* ne13 */ 1, // ne13,
-        /* nb10 */ nb10,
-        /* nb11 */ nb11,
-        /* nb12 */ nb12,
-        /* ne13 */ nb12, // ne12 == 1
-        /* ne0  */ ne0,
-        /* ne1  */ 1, // ne1,
-        /* nb1  */ nb1,
-        /* r2   */ 1,
-        /* r3   */ 1,
-        shared_values,
+        shmem,
         tgpig,
         tiitg,
         tiisg,

From 20a780c7b64c38071fe70ad23e16e80c23c0147b Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sun, 17 Nov 2024 13:12:22 +0200
Subject: [PATCH 019/102] gitignore : ignore local run scripts [no ci]

---
 .gitignore | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/.gitignore b/.gitignore
index 0b56bcc7ab682..307c065f793e6 100644
--- a/.gitignore
+++ b/.gitignore
@@ -134,3 +134,7 @@ poetry.toml
 
 # Test models for lora adapters
 /lora-tests
+
+# Local scripts
+/run-vim.sh
+/run-chat.sh

From be5caccef945546ee9fd25a151330a88d785faf9 Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Sun, 17 Nov 2024 12:25:45 +0100
Subject: [PATCH 020/102] llama : only use default buffer types for the KV
 cache (#10358)

---
 ggml/src/ggml-backend.cpp |  9 +++++----
 src/llama.cpp             | 16 ++++------------
 2 files changed, 9 insertions(+), 16 deletions(-)

diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index 9a6010d361224..9dcde8d11952a 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -689,7 +689,7 @@ static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backen
 }
 
 static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor, const struct ggml_tensor * op) {
-    ggml_backend_buffer_t buffer = tensor->buffer;
+    ggml_backend_buffer_t buffer = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
     if (buffer == NULL) {
         return -1;
     }
@@ -722,8 +722,6 @@ static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML
 
 // returns the backend that should be used for the node based on the current locations
 static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) {
-    // TODO: use supports_op to check if the backend supports the op
-
     // assign pre-allocated nodes to their backend
     int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor, tensor);
     if (cur_backend_id != -1) {
@@ -742,7 +740,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 
     if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) {
         // since the tensor is pre-allocated, it cannot be moved to another backend
-        GGML_ABORT("pre-allocated tensor in a backend that cannot run the operation");
+        GGML_ABORT("pre-allocated tensor (%s) in a backend that cannot run the operation", tensor->name);
     }
 
     // graph input
@@ -886,6 +884,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         int * node_backend_id = &tensor_backend_id(node);
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
         // do not overwrite user assignments
         if (*node_backend_id == -1) {
             *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
diff --git a/src/llama.cpp b/src/llama.cpp
index 1703104fb3680..de96959f26632 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -3460,21 +3460,13 @@ static bool llama_kv_cache_init(
         const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
         const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
 
-        const llama_model::buft_list_t * buft_list;
+        ggml_backend_buffer_type_t buft;
         if (offload) {
-            buft_list = model.dev_layer.at(i).buft_list;
+            auto * dev = model.dev_layer.at(i).dev;
+            buft = ggml_backend_dev_buffer_type(dev);
         } else {
-            buft_list = &model.cpu_buft_list;
+            buft = ggml_backend_cpu_buffer_type();
         }
-        ggml_backend_buffer_type_t buft = select_buft(*buft_list,
-            [&](ggml_context * ctx) {
-                ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
-                if (hparams.rope_type == LLAMA_ROPE_TYPE_NONE) {
-                    return k;
-                }
-                ggml_tensor * p = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 1);
-                return ggml_rope(ctx, k, p, hparams.n_rot, hparams.rope_type);
-            });
         ggml_context * ctx = ctx_for_buft(buft);
 
         if (!ctx) {

From ce2e59ba107cf71ed566040ff20a15d1c58e09c2 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Sun, 17 Nov 2024 12:59:38 +0100
Subject: [PATCH 021/102] CMake: fix typo in comment [no ci] (#10360)

---
 ggml/src/ggml-cuda/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt
index 3dde0f3666187..e592f79896db2 100644
--- a/ggml/src/ggml-cuda/CMakeLists.txt
+++ b/ggml/src/ggml-cuda/CMakeLists.txt
@@ -11,7 +11,7 @@ if (CUDAToolkit_FOUND)
         # 60     == P100, FP16 CUDA intrinsics
         # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
         # 70     == V100, FP16 tensor cores
-        # 75     == Turing, int6 tensor cores
+        # 75     == Turing, int8 tensor cores
         if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6")
             set(CMAKE_CUDA_ARCHITECTURES "native")
         elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)

From 76e9e58b7847112848aa1f40b65d1cbcd6d5f5a3 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Sun, 17 Nov 2024 23:20:42 +0100
Subject: [PATCH 022/102] CUDA: fix MMV kernel being used for FP16 src1
 (#10357)

---
 ggml/src/ggml-cuda/ggml-cuda.cu | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index ef56e944d01ab..dd94ab03d5b6c 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -1760,11 +1760,13 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && src0->type == GGML_TYPE_F16 && src1->ne[1] == 1 && dst->ne[3] == 1 && (src0->ne[1] < MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+    if (!split && use_mul_mat_vec && dst->ne[3] == 1 && (src0->ne[1] < MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+        // the custom F16 vector kernel can be used over batched cuBLAS GEMM
+        // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
         ggml_cuda_mul_mat_vec(ctx, src0, src1, dst);
     } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
                && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
-        // KQ + KQV multi-batch without FlashAttention
+        // general KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_mul_mat_vec) {
         ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec, nullptr);

From 75207b3a887f91f813de1eb6e9fd135d3cb2b8c6 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Mon, 18 Nov 2024 00:21:53 +0100
Subject: [PATCH 023/102] docker: use GGML_NATIVE=OFF (#10368)

---
 .devops/full-cuda.Dockerfile           | 2 +-
 .devops/full-musa.Dockerfile           | 2 +-
 .devops/llama-cli-cann.Dockerfile      | 2 +-
 .devops/llama-cli-cuda.Dockerfile      | 2 +-
 .devops/llama-cli-intel.Dockerfile     | 2 +-
 .devops/llama-cli-musa.Dockerfile      | 2 +-
 .devops/llama-cli-vulkan.Dockerfile    | 2 +-
 .devops/llama-server-cuda.Dockerfile   | 2 +-
 .devops/llama-server-intel.Dockerfile  | 2 +-
 .devops/llama-server-musa.Dockerfile   | 2 +-
 .devops/llama-server-vulkan.Dockerfile | 2 +-
 11 files changed, 11 insertions(+), 11 deletions(-)

diff --git a/.devops/full-cuda.Dockerfile b/.devops/full-cuda.Dockerfile
index d5acd35e204d3..05bff1bdf6a22 100644
--- a/.devops/full-cuda.Dockerfile
+++ b/.devops/full-cuda.Dockerfile
@@ -26,7 +26,7 @@ COPY . .
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc) && \
     cp build/bin/* .
 
diff --git a/.devops/full-musa.Dockerfile b/.devops/full-musa.Dockerfile
index 34ba856d3d1ca..575e81b486d13 100644
--- a/.devops/full-musa.Dockerfile
+++ b/.devops/full-musa.Dockerfile
@@ -19,7 +19,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc) && \
     cp build/bin/* .
 
diff --git a/.devops/llama-cli-cann.Dockerfile b/.devops/llama-cli-cann.Dockerfile
index 45c0585b0b3d9..02dce501ce286 100644
--- a/.devops/llama-cli-cann.Dockerfile
+++ b/.devops/llama-cli-cann.Dockerfile
@@ -22,7 +22,7 @@ ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
 
 RUN echo "Building with static libs" && \
     source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \
-    cmake -B build -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF  && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF  && \
     cmake --build build --config Release --target llama-cli
 
 # TODO: use image with NNRT
diff --git a/.devops/llama-cli-cuda.Dockerfile b/.devops/llama-cli-cuda.Dockerfile
index 3279c8da41c8a..7796891d5b53c 100644
--- a/.devops/llama-cli-cuda.Dockerfile
+++ b/.devops/llama-cli-cuda.Dockerfile
@@ -22,7 +22,7 @@ COPY . .
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release --target llama-cli -j$(nproc) && \
     mkdir -p /app/lib && \
     find build -name "*.so" -exec cp {} /app/lib \;
diff --git a/.devops/llama-cli-intel.Dockerfile b/.devops/llama-cli-intel.Dockerfile
index 1f4ce0730ec7d..0706f732a98f0 100644
--- a/.devops/llama-cli-intel.Dockerfile
+++ b/.devops/llama-cli-intel.Dockerfile
@@ -15,7 +15,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
         export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
     fi && \
     echo "Building with static libs" && \
-    cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \
     ${OPT_SYCL_F16} -DBUILD_SHARED_LIBS=OFF && \
     cmake --build build --config Release --target llama-cli
 
diff --git a/.devops/llama-cli-musa.Dockerfile b/.devops/llama-cli-musa.Dockerfile
index 1edf75cf2a1b2..3372749bee70e 100644
--- a/.devops/llama-cli-musa.Dockerfile
+++ b/.devops/llama-cli-musa.Dockerfile
@@ -15,7 +15,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release --target llama-cli -j$(nproc) && \
     mkdir -p /app/lib && \
     find build -name "*.so" -exec cp {} /app/lib \;
diff --git a/.devops/llama-cli-vulkan.Dockerfile b/.devops/llama-cli-vulkan.Dockerfile
index 9b0dad8bf7a13..92a6e04793491 100644
--- a/.devops/llama-cli-vulkan.Dockerfile
+++ b/.devops/llama-cli-vulkan.Dockerfile
@@ -14,7 +14,7 @@ RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key
 # Build it
 WORKDIR /app
 COPY . .
-RUN cmake -B build -DGGML_VULKAN=1 && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 && \
     cmake --build build --config Release --target llama-cli
 
 # Clean up
diff --git a/.devops/llama-server-cuda.Dockerfile b/.devops/llama-server-cuda.Dockerfile
index ea07a4d52474e..bf8a198f99f73 100644
--- a/.devops/llama-server-cuda.Dockerfile
+++ b/.devops/llama-server-cuda.Dockerfile
@@ -22,7 +22,7 @@ COPY . .
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release --target llama-server -j$(nproc) && \
     mkdir -p /app/lib && \
     find build -name "*.so" -exec cp {} /app/lib \;
diff --git a/.devops/llama-server-intel.Dockerfile b/.devops/llama-server-intel.Dockerfile
index 773f030a7b031..b503b8cfe1084 100644
--- a/.devops/llama-server-intel.Dockerfile
+++ b/.devops/llama-server-intel.Dockerfile
@@ -15,7 +15,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
         export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
     fi && \
     echo "Building with dynamic libs" && \
-    cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
     cmake --build build --config Release --target llama-server
 
 FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime
diff --git a/.devops/llama-server-musa.Dockerfile b/.devops/llama-server-musa.Dockerfile
index 2598774684e7b..eb67201c18775 100644
--- a/.devops/llama-server-musa.Dockerfile
+++ b/.devops/llama-server-musa.Dockerfile
@@ -15,7 +15,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release --target llama-server -j$(nproc) && \
     mkdir -p /app/lib && \
     find build -name "*.so" -exec cp {} /app/lib \;
diff --git a/.devops/llama-server-vulkan.Dockerfile b/.devops/llama-server-vulkan.Dockerfile
index 93c5e0c26e691..6aa7867791a38 100644
--- a/.devops/llama-server-vulkan.Dockerfile
+++ b/.devops/llama-server-vulkan.Dockerfile
@@ -14,7 +14,7 @@ RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key
 # Build it
 WORKDIR /app
 COPY . .
-RUN cmake -B build -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \
     cmake --build build --config Release --target llama-server
 
 # Clean up

From 9b75f03cd2ec9cc482084049d87a0f08f9f01517 Mon Sep 17 00:00:00 2001
From: 0cc4m <picard12@live.de>
Date: Mon, 18 Nov 2024 11:02:43 +0100
Subject: [PATCH 024/102] Vulkan: Fix device info output format specifiers
 (#10366)

* Vulkan: Fix device info output format specifiers

* Vulkan: Use zu printf specifier for size_t instead of ld
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp | 9 ++++-----
 1 file changed, 4 insertions(+), 5 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 19951d9b6d60b..ef067356ec794 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -1764,11 +1764,11 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     fp16 = fp16 && vk12_features.shaderFloat16;
 
     std::string device_name = props2.properties.deviceName.data();
-    GGML_LOG_DEBUG("ggml_vulkan: %d = %s (%s) | uma: %d | fp16: %d | warp size: %d\n",
-              idx, device_name.c_str(), driver_props.driverName, uma, fp16, subgroup_size);
+    GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu\n",
+              idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size);
 
     if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) {
-        std::cerr << "ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want." << std::endl;
+        GGML_LOG_DEBUG("ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want.\n");
     }
 }
 
@@ -1937,8 +1937,7 @@ void ggml_vk_instance_init() {
             vk_instance.device_indices.push_back(0);
         }
     }
-    GGML_LOG_DEBUG("ggml_vulkan: Found %d Vulkan devices:\n", vk_instance.device_indices.size());
-
+    GGML_LOG_DEBUG("ggml_vulkan: Found %zu Vulkan devices:\n", vk_instance.device_indices.size());
 
     for (size_t i = 0; i < vk_instance.device_indices.size(); i++) {
         ggml_vk_print_gpu_info(i);

From 2eb76b2a5e4ea395b971a419c95b473ab6f253e4 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 18 Nov 2024 16:08:20 +0200
Subject: [PATCH 025/102] flake.lock: Update (#10346)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Flake lock file updates:

• Updated input 'nixpkgs':
    'github:NixOS/nixpkgs/4aa36568d413aca0ea84a1684d2d46f55dbabad7?narHash=sha256-Zwl8YgTVJTEum%2BL%2B0zVAWvXAGbWAuXHax3KzuejaDyo%3D' (2024-11-05)
  → 'github:NixOS/nixpkgs/5e4fbfb6b3de1aa2872b76d49fafc942626e2add?narHash=sha256-OZiZ3m8SCMfh3B6bfGC/Bm4x3qc1m2SVEAlkV6iY7Yg%3D' (2024-11-15)

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
---
 flake.lock | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/flake.lock b/flake.lock
index 470363a2f8fad..ee8cf07e37220 100644
--- a/flake.lock
+++ b/flake.lock
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1730785428,
-        "narHash": "sha256-Zwl8YgTVJTEum+L+0zVAWvXAGbWAuXHax3KzuejaDyo=",
+        "lastModified": 1731676054,
+        "narHash": "sha256-OZiZ3m8SCMfh3B6bfGC/Bm4x3qc1m2SVEAlkV6iY7Yg=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "4aa36568d413aca0ea84a1684d2d46f55dbabad7",
+        "rev": "5e4fbfb6b3de1aa2872b76d49fafc942626e2add",
         "type": "github"
       },
       "original": {

From f139d2ea611c5604395c95160d3c53f7c4eaf220 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Mon, 18 Nov 2024 08:28:42 -0600
Subject: [PATCH 026/102] vulkan: remove use of null initializer (#10372)

Seems like this isn't working for vulkan-over-metal when the array is sized
by a spec constant. Maybe a spirv-cross limitation?
---
 ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
index ddbac5d2c9334..970aac6ef72cc 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
@@ -5,8 +5,6 @@
 #endif
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
-#extension GL_EXT_null_initializer : enable
-
 #include "mul_mat_vec_base.comp"
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
@@ -57,7 +55,11 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 
     y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
 
-    FLOAT_TYPE temp[NUM_ROWS] = {};
+    FLOAT_TYPE temp[NUM_ROWS];
+
+    for (uint i = 0; i < NUM_ROWS; ++i) {
+        temp[i] = FLOAT_TYPE(0);
+    }
 
     const int unroll_count = 8;
 

From 531cb1c233800e6acb021dc56d69595e314db072 Mon Sep 17 00:00:00 2001
From: bandoti <141645996+bandoti@users.noreply.github.com>
Date: Mon, 18 Nov 2024 11:23:58 -0400
Subject: [PATCH 027/102] Skip searching root path for cross-compile builds
 (#10383)

---
 cmake/llama-config.cmake.in | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/cmake/llama-config.cmake.in b/cmake/llama-config.cmake.in
index ef68417b412cf..a7f1efb88f2aa 100644
--- a/cmake/llama-config.cmake.in
+++ b/cmake/llama-config.cmake.in
@@ -66,11 +66,15 @@ endif()
 
 find_library(ggml_LIBRARY ggml
     REQUIRED
-    HINTS ${LLAMA_LIB_DIR})
+    HINTS ${LLAMA_LIB_DIR}
+    NO_CMAKE_FIND_ROOT_PATH
+)
 
 find_library(llama_LIBRARY llama
     REQUIRED
-    HINTS ${LLAMA_LIB_DIR})
+    HINTS ${LLAMA_LIB_DIR}
+    NO_CMAKE_FIND_ROOT_PATH
+)
 
 set(_llama_link_deps "${ggml_LIBRARY}" "@GGML_LINK_LIBRARIES@")
 set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@")

From d3481e631661b5e9517f78908cdd58cee63c4903 Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Mon, 18 Nov 2024 18:43:40 +0100
Subject: [PATCH 028/102] cuda : only use native when supported by cmake
 (#10389)

---
 ggml/src/ggml-cuda/CMakeLists.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt
index e592f79896db2..25858ece8ddd7 100644
--- a/ggml/src/ggml-cuda/CMakeLists.txt
+++ b/ggml/src/ggml-cuda/CMakeLists.txt
@@ -12,7 +12,7 @@ if (CUDAToolkit_FOUND)
         # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
         # 70     == V100, FP16 tensor cores
         # 75     == Turing, int8 tensor cores
-        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6")
+        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
             set(CMAKE_CUDA_ARCHITECTURES "native")
         elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
             set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")

From 557924f22237c76387a39c4db5abae154d57e754 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Alberto=20Cabrera=20P=C3=A9rez?=
 <alberto.cabrera@codeplay.com>
Date: Tue, 19 Nov 2024 00:50:04 +0000
Subject: [PATCH 029/102] sycl: Revert MUL_MAT_OP support changes (#10385)

---
 ggml/src/ggml-sycl/ggml-sycl.cpp | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index 2dba15d237e92..255bc64c6badd 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -4350,10 +4350,6 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 if (op->op == GGML_OP_MUL_MAT) {
                     a = op->src[0];
                     b = op->src[1];
-                    if (ggml_is_permuted(a) || ggml_is_permuted(b)) {
-                        // TODO: fix like https://github.com/ggerganov/llama.cpp/pull/10021
-                        return false;
-                    }
                 } else {
                     a = op->src[2];
                     b = op->src[1];

From b3e585988fc65d3a8083c6d94dfc0629f9ce226d Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Tue, 19 Nov 2024 01:25:17 -0600
Subject: [PATCH 030/102] vulkan: Optimize soft_max (#10301)

* vulkan: Optimize soft_max

Large soft_max could already saturate memory, but small/medium sizes were
pretty slow. The bulk of the gains for them comes from using a smaller
workgroup size, and making the workgroup size match the subgroup size also
makes the barriers much cheaper.

Cache some values in locals to avoid refetching/recomputing. And stamp
out a few "template instantiations" so smaller cases will fully unroll.

Add a missing early return for OOB rows. This happens when there are more
than 512 rows and the dispatch is 512 x H.

* vulkan: Further soft_max optimizations

Restore the workgroup size of 512 case, use it for >1024.

Use unrollable loops for more iteration counts.
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          |  13 +-
 .../ggml-vulkan/vulkan-shaders/soft_max.comp  | 112 ++++++++++++++----
 tests/test-backend-ops.cpp                    |   8 ++
 3 files changed, 106 insertions(+), 27 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index ef067356ec794..21fee2f3d7533 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -218,6 +218,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_tanh_f32;
     vk_pipeline pipeline_diag_mask_inf_f32;
     vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
+    vk_pipeline pipeline_soft_max_f32_wg512, pipeline_soft_max_f32_f16_wg512;
     vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16;
     vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
     vk_pipeline pipeline_argsort_f32;
@@ -388,6 +389,7 @@ struct vk_op_soft_max_push_constants {
     float m0;
     float m1;
     uint32_t n_head_log2;
+    uint32_t nrows_x;
 };
 
 struct vk_op_argsort_push_constants {
@@ -1497,8 +1499,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
@@ -3932,10 +3936,10 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);
 
         if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_soft_max_f32;
+            return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_wg512 : ctx->device->pipeline_soft_max_f32;
         }
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_soft_max_f32_f16;
+            return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_f16_wg512 : ctx->device->pipeline_soft_max_f32_f16;
         }
         return nullptr;
     case GGML_OP_ROPE:
@@ -4581,6 +4585,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
         scale, max_bias,
         m0, m1,
         n_head_log2,
+        nrows_x,
     }, dryrun);
 }
 
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
index 0bd51ecab5870..f9727679ec5f2 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
@@ -1,6 +1,7 @@
 #version 450
 
-#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_EXT_control_flow_attributes : enable
 
 layout (push_constant) uniform parameter
 {
@@ -11,14 +12,13 @@ layout (push_constant) uniform parameter
     float m0;
     float m1;
     uint n_head_log2;
+    uint nrows_x;
 } p;
 
 #include "types.comp"
 
-#extension GL_EXT_control_flow_attributes : enable
-#define BLOCK_SIZE 512
-
-layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+layout(constant_id = 0) const uint BLOCK_SIZE = 32;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
@@ -26,11 +26,18 @@ layout (binding = 2) buffer D {D_TYPE data_d[];};
 
 shared FLOAT_TYPE vals[BLOCK_SIZE];
 
-void main() {
+// num_iters is the number of BLOCK_SIZE loop iterations we need to iterate
+// over all the columns. The main function tries to pass a constant here,
+// as if it were a template function, to allow unrolling.
+void soft_max(uint num_iters) {
     const uint tid = gl_LocalInvocationID.x;
     const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint rowy = rowx % p.KY;
 
+    if (rowx >= p.nrows_x) {
+        return;
+    }
+
     float slope = 1.0f;
 
     // ALiBi
@@ -46,19 +53,37 @@ void main() {
     // Find max
     FLOAT_TYPE max_val = uintBitsToFloat(0xFF800000);
 
-    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
+    // Cache values while we compute the max, so we don't need to read them
+    // again when we're ready to compute exp(x-max).
+    const uint DATA_CACHE_SIZE = 16;
+    FLOAT_TYPE data_cache[DATA_CACHE_SIZE];
+
+    [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
         const uint col = col0 + tid;
 
-        if (col >= p.KX) {
-            break;
+        FLOAT_TYPE a = FLOAT_TYPE(0);
+        if (col < p.KX) {
+            a = data_a[rowx * p.KX + col];
         }
 
-        max_val = max(max_val, FLOAT_TYPE(data_a[rowx * p.KX + col]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)));
+        FLOAT_TYPE b = FLOAT_TYPE(0);
+        if (p.KY > 0 && col < p.KX) {
+            b = data_b[rowy * p.KX + col];
+        }
+
+        FLOAT_TYPE v = a * p.scale + slope * b;
+
+        max_val = max(max_val, v);
+
+        if (idx < DATA_CACHE_SIZE) {
+            data_cache[idx] = v;
+        }
     }
-    vals[tid] = max_val;
 
+    // reduce across the workgroup
+    vals[tid] = max_val;
     barrier();
-    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
         if (tid < s) {
             vals[tid] = max(vals[tid], vals[tid + s]);
         }
@@ -68,39 +93,80 @@ void main() {
     max_val = vals[0];
     barrier();
 
-    // Sum up values
-    vals[tid] = FLOAT_TYPE(0.0f);
+    FLOAT_TYPE sum = FLOAT_TYPE(0.0f);
 
-    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
+    // Compute sum{exp(x - max)}
+    [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
         const uint col = col0 + tid;
 
         if (col >= p.KX) {
             break;
         }
 
+        // compute exp(a*scale+b*slope), add it to sum, and cache the new value
+        // in data_cache if possible.
         const uint i = rowx * p.KX + col;
-        const FLOAT_TYPE val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
-        vals[tid] += val;
-        data_d[i] = D_TYPE(val);
+        FLOAT_TYPE val;
+        if (idx < DATA_CACHE_SIZE) {
+            val = exp(data_cache[idx] - max_val);
+        } else {
+            val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
+        }
+        sum += val;
+        if (idx < DATA_CACHE_SIZE) {
+            data_cache[idx] = val;
+        } else {
+            data_d[i] = D_TYPE(val);
+        }
     }
 
+    // reduce across the workgroup
+    vals[tid] = sum;
     barrier();
-    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
         if (tid < s) {
             vals[tid] += vals[tid + s];
         }
         barrier();
     }
+    sum = vals[0];
 
-    const D_TYPE divisor = D_TYPE(vals[0]);
+    FLOAT_TYPE rcpdivisor = 1.0/sum;
 
-    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
+    [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
         const uint col = col0 + tid;
 
         if (col >= p.KX) {
-            break;
+            continue;
+        }
+
+        if (idx < DATA_CACHE_SIZE) {
+            data_d[rowx*p.KX + col] = D_TYPE(data_cache[idx] * rcpdivisor);
+        } else {
+            data_d[rowx*p.KX + col] *= D_TYPE(rcpdivisor);
         }
+    }
+}
 
-        data_d[rowx*p.KX + col] /= divisor;
+void main() {
+    // instantiate the soft_max function for several different
+    // dimensions, to allow loop unrolling
+    uint num_blocks = (p.KX + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    if (num_blocks > 32) {
+        soft_max(num_blocks);
+    } else if (num_blocks > 16) {
+        soft_max(32);
+    } else if (num_blocks > 8) {
+        soft_max(16);
+    } else if (num_blocks > 4) {
+        soft_max(8);
+    } else if (num_blocks == 4) {
+        soft_max(4);
+    } else if (num_blocks == 3) {
+        soft_max(3);
+    } else if (num_blocks == 2) {
+        soft_max(2);
+    } else if (num_blocks == 1) {
+        soft_max(1);
     }
 }
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index f8a59b6df175f..01ac7166e0e22 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -3823,6 +3823,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
 
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F16, {512, 3072, 1, 1}));
 
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {4096, 4096, 5, 1}, false, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 4096, 5, 1}, false, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {1024, 1024, 10, 1}, false, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 1024, 10, 1}, false, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {256, 256, 20, 1}, false, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {64, 64, 20, 1}, false, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 64, 20, 1}, false, 1.0f, 0.0f));
+
     for (int bs : {1, 512}) {
         for (ggml_type type_a : all_types) {
             for (ggml_type type_b : {GGML_TYPE_F32}) {

From 2a1507c1629975d9d20a503d6a14f44eff292c25 Mon Sep 17 00:00:00 2001
From: Romain Biessy <romain.biessy@codeplay.com>
Date: Tue, 19 Nov 2024 09:02:23 +0100
Subject: [PATCH 031/102] sycl : Add option to set the SYCL architecture for
 all targets (#10266)

* Add option to set the SYCL architecture for all targets
* Convert GGML_SYCL_HIP_TARGET to the more generic GGML_SYCL_ARCH option
* Document that setting GGML_SYCL_ARCH can improve the performance
---
 docs/backend/SYCL.md              | 12 ++++++++----
 ggml/CMakeLists.txt               |  2 ++
 ggml/src/ggml-sycl/CMakeLists.txt | 10 +++++++---
 3 files changed, 17 insertions(+), 7 deletions(-)

diff --git a/docs/backend/SYCL.md b/docs/backend/SYCL.md
index 38185f73897ee..e431f51f184cc 100644
--- a/docs/backend/SYCL.md
+++ b/docs/backend/SYCL.md
@@ -312,12 +312,14 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_
 export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
 
 # Build LLAMA with Nvidia BLAS acceleration through SYCL
+# Setting GGML_SYCL_DEVICE_ARCH is optional but can improve performance
+GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture
 
 # Option 1: Use FP32 (recommended for better performance in most cases)
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
 # Option 2: Use FP16
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
 
 # build all binary
 cmake --build build --config Release -j -v
@@ -335,8 +337,9 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithrocBLAS/include:$CPLUS_INCLUDE
 
 ## AMD
 # Use FP32, FP16 is not supported
-# Find your GGML_SYCL_HIP_TARGET with rocminfo, under the key 'Name:'
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=AMD -DGGML_SYCL_HIP_TARGET=${GGML_SYCL_HIP_TARGET} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+# Find your GGML_SYCL_DEVICE_ARCH with rocminfo, under the key 'Name:'
+GGML_SYCL_DEVICE_ARCH=gfx90a # Example architecture
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=AMD -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
 # build all binary
 cmake --build build --config Release -j -v
@@ -646,6 +649,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 |--------------------|---------------------------------------|---------------------------------------------|
 | GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
 | GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA \| AMD    | Set the SYCL target device type.            |
+| GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)          | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path.      |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index a82818d607f5e..b16a0e9ada7ed 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -164,6 +164,8 @@ option(GGML_SYCL                            "ggml: use SYCL"
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
                                             "ggml: sycl target device")
+set   (GGML_SYCL_DEVICE_ARCH "" CACHE STRING
+                                            "ggml: sycl device architecture")
 
 # extra artifacts
 option(GGML_BUILD_TESTS    "ggml: build tests"    ${GGML_STANDALONE})
diff --git a/ggml/src/ggml-sycl/CMakeLists.txt b/ggml/src/ggml-sycl/CMakeLists.txt
index 03bf3cb929b32..d1d0ff83d636c 100644
--- a/ggml/src/ggml-sycl/CMakeLists.txt
+++ b/ggml/src/ggml-sycl/CMakeLists.txt
@@ -72,10 +72,14 @@ else()
         set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
         target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl)
     elseif (GGML_SYCL_TARGET STREQUAL "AMD")
-        if (GGML_SYCL_HIP_TARGET STREQUAL "")
-            message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_HIP_TARGET has not been set.")
+        if (NOT GGML_SYCL_DEVICE_ARCH)
+            message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
         endif()
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=amdgcn-amd-amdhsa -Xsycl-target-backend --offload-arch=${GGML_SYCL_HIP_TARGET}")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=amdgcn-amd-amdhsa")
         target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl)
     endif()
+
+    if (GGML_SYCL_DEVICE_ARCH)
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH}")
+  endif()
 endif()

From a88ad007de3eb5d92cb538fd269ff94c4bf0c8d2 Mon Sep 17 00:00:00 2001
From: Shane A <shanea@allenai.org>
Date: Tue, 19 Nov 2024 01:04:08 -0800
Subject: [PATCH 032/102] llama : add OLMo November 2024 support (#10394)

* Add OLMo November 2024 constants

* Add OLMo November 2024 converter

* Add loading of OLMo November 2024 tensors and hyper parameters

* Add building of OLMo November 2024 model
---
 convert_hf_to_gguf.py          |   5 +
 gguf-py/gguf/constants.py      |  18 ++++
 gguf-py/gguf/tensor_mapping.py |  28 ++---
 src/llama.cpp                  | 186 +++++++++++++++++++++++++++++++++
 4 files changed, 223 insertions(+), 14 deletions(-)

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 39afa5ef4f272..9f4b8154b88a8 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -3040,6 +3040,11 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("Olmo1124ForCausalLM")
+class Olmo1124Model(Model):
+    model_arch = gguf.MODEL_ARCH.OLMO_1124
+
+
 @Model.register("OlmoeForCausalLM")
 class OlmoeModel(Model):
     model_arch = gguf.MODEL_ARCH.OLMOE
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index bc2b649d1a83d..d83b72f761951 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -243,6 +243,7 @@ class MODEL_ARCH(IntEnum):
     COMMAND_R    = auto()
     DBRX         = auto()
     OLMO         = auto()
+    OLMO_1124    = auto()
     OLMOE        = auto()
     OPENELM      = auto()
     ARCTIC       = auto()
@@ -404,6 +405,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.COMMAND_R:      "command-r",
     MODEL_ARCH.DBRX:           "dbrx",
     MODEL_ARCH.OLMO:           "olmo",
+    MODEL_ARCH.OLMO_1124:      "olmo_1124",
     MODEL_ARCH.OLMOE:          "olmoe",
     MODEL_ARCH.OPENELM:        "openelm",
     MODEL_ARCH.ARCTIC:         "arctic",
@@ -1069,6 +1071,22 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.OLMO_1124: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.FFN_POST_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     MODEL_ARCH.OLMOE: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py
index f4a787c56993a..4cbd39e03e942 100644
--- a/gguf-py/gguf/tensor_mapping.py
+++ b/gguf-py/gguf/tensor_mapping.py
@@ -13,7 +13,7 @@ class TensorNameMap:
             "transformer.wte",                           # gpt2 gpt-j mpt refact qwen dbrx jais exaone
             "transformer.word_embeddings",               # falcon
             "word_embeddings",                           # bloom
-            "model.embed_tokens",                        # llama-hf nemotron olmoe
+            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo_1124
             "tok_embeddings",                            # llama-pth
             "embeddings.word_embeddings",                # bert nomic-bert
             "language_model.embedding.word_embeddings",  # persimmon
@@ -54,7 +54,7 @@ class TensorNameMap:
         # Output
         MODEL_TENSOR.OUTPUT: (
             "embed_out",                 # gptneox
-            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe
+            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo_1124
             "output",                    # llama-pth bloom internlm2
             "word_embeddings_for_head",  # persimmon
             "lm_head.linear",            # phi2
@@ -66,7 +66,7 @@ class TensorNameMap:
         MODEL_TENSOR.OUTPUT_NORM: (
             "gpt_neox.final_layer_norm",               # gptneox
             "transformer.ln_f",                        # gpt2 gpt-j falcon jais exaone
-            "model.norm",                              # llama-hf baichuan internlm2 olmoe
+            "model.norm",                              # llama-hf baichuan internlm2 olmoe olmo_1124
             "norm",                                    # llama-pth
             "transformer.norm_f",                      # mpt dbrx
             "ln_f",                                    # refact bloom qwen gpt2
@@ -145,7 +145,7 @@ class TensorNameMap:
 
         # Attention query
         MODEL_TENSOR.ATTN_Q: (
-            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron olmoe
+            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron olmoe olmo_1124
             "layers.{bid}.attention.wq",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.query",                  # bert
             "transformer.h.{bid}.attn.q_proj",                           # gpt-j
@@ -157,7 +157,7 @@ class TensorNameMap:
 
         # Attention key
         MODEL_TENSOR.ATTN_K: (
-            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron olmoe
+            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron olmoe olmo_1124
             "layers.{bid}.attention.wk",                               # llama-pth
             "encoder.layer.{bid}.attention.self.key",                  # bert
             "transformer.h.{bid}.attn.k_proj",                         # gpt-j
@@ -170,7 +170,7 @@ class TensorNameMap:
 
         # Attention value
         MODEL_TENSOR.ATTN_V: (
-            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron olmoe
+            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron olmoe olmo_1124
             "layers.{bid}.attention.wv",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.value",                  # bert
             "transformer.h.{bid}.attn.v_proj",                           # gpt-j
@@ -188,7 +188,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.attn.out_proj",                       # mpt
             "transformer.h.{bid}.self_attention.dense",                     # falcon
             "h.{bid}.self_attention.dense",                                 # bloom
-            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron olmoe
+            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron olmoe olmo_1124
             "layers.{bid}.attention.wo",                                    # llama-pth
             "encoder.layer.{bid}.attention.output.dense",                   # bert
             "transformer.h.{bid}.attn.out_proj",                            # gpt-j
@@ -215,7 +215,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.ATTN_POST_NORM: (
-            "model.layers.{bid}.post_attention_layernorm",     # gemma2
+            "model.layers.{bid}.post_attention_layernorm",     # gemma2 olmo_1124
         ),
 
         # Rotary embeddings
@@ -250,7 +250,7 @@ class TensorNameMap:
 
         # Post feed-forward norm
         MODEL_TENSOR.FFN_POST_NORM: (
-            "model.layers.{bid}.post_feedforward_layernorm", # gemma2
+            "model.layers.{bid}.post_feedforward_layernorm", # gemma2 olmo_1124
         ),
 
         MODEL_TENSOR.FFN_GATE_INP: (
@@ -273,7 +273,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.ffn.up_proj",                   # mpt
             "transformer.h.{bid}.mlp.dense_h_to_4h",                  # falcon
             "h.{bid}.mlp.dense_h_to_4h",                              # bloom
-            "model.layers.{bid}.mlp.up_proj",                         # llama-hf refact nemotron
+            "model.layers.{bid}.mlp.up_proj",                         # llama-hf refact nemotron olmo_1124
             "layers.{bid}.feed_forward.w3",                           # llama-pth
             "encoder.layer.{bid}.intermediate.dense",                 # bert
             "transformer.h.{bid}.mlp.fc_in",                          # gpt-j
@@ -314,7 +314,7 @@ class TensorNameMap:
 
         # Feed-forward gate
         MODEL_TENSOR.FFN_GATE: (
-            "model.layers.{bid}.mlp.gate_proj",           # llama-hf refact
+            "model.layers.{bid}.mlp.gate_proj",           # llama-hf refact olmo_1124
             "layers.{bid}.feed_forward.w1",               # llama-pth
             "transformer.h.{bid}.mlp.w2",                 # qwen
             "transformer.h.{bid}.mlp.c_fc2",              # jais
@@ -346,7 +346,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.ffn.down_proj",                 # mpt
             "transformer.h.{bid}.mlp.dense_4h_to_h",                  # falcon
             "h.{bid}.mlp.dense_4h_to_h",                              # bloom
-            "model.layers.{bid}.mlp.down_proj",                       # llama-hf nemotron
+            "model.layers.{bid}.mlp.down_proj",                       # llama-hf nemotron olmo_1124
             "layers.{bid}.feed_forward.w2",                           # llama-pth
             "encoder.layer.{bid}.output.dense",                       # bert
             "transformer.h.{bid}.mlp.fc_out",                         # gpt-j
@@ -383,7 +383,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_Q_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
             "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon
+            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon olmo_1124
             "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_q",                # jina-bert-v2
             "transformer.layers.{bid}.attn.q_norm",                           # openelm
@@ -392,7 +392,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_K_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
             "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon
+            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon olmo_1124
             "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_k",                # jina-bert-v2
             "transformer.layers.{bid}.attn.k_norm",                           # openelm
diff --git a/src/llama.cpp b/src/llama.cpp
index de96959f26632..4f31f25b1d355 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -179,6 +179,7 @@ enum llm_arch {
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_DBRX,
     LLM_ARCH_OLMO,
+    LLM_ARCH_OLMO_1124,
     LLM_ARCH_OLMOE,
     LLM_ARCH_OPENELM,
     LLM_ARCH_ARCTIC,
@@ -232,6 +233,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COMMAND_R,       "command-r"    },
     { LLM_ARCH_DBRX,            "dbrx"         },
     { LLM_ARCH_OLMO,            "olmo"         },
+    { LLM_ARCH_OLMO_1124,       "olmo_1124"    },
     { LLM_ARCH_OLMOE,           "olmoe"        },
     { LLM_ARCH_OPENELM,         "openelm"      },
     { LLM_ARCH_ARCTIC,          "arctic"       },
@@ -1207,6 +1209,25 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_OLMO_1124,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_OLMOE,
         {
@@ -5877,6 +5898,17 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_OLMO_1124:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    case 16: model.type = e_model::MODEL_1B; break;
+                    case 32: model.type = e_model::MODEL_7B; break;
+                    case 40: model.type = e_model::MODEL_13B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_OLMOE:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -8559,6 +8591,31 @@ static bool llm_load_tensors(
                         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
                     }
                 } break;
+            case LLM_ARCH_OLMO_1124:
+                {
+                    model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    model.output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    model.output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = model.layers[i];
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+                        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+                    }
+                } break;
             case LLM_ARCH_OLMOE:
                 {
                     model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -14424,6 +14481,130 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_olmo_1124() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            cur = inpL;
+
+            // self_attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].attn_q_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(Qcur, "Qcur_normed", il);
+
+                Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].attn_k_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(Kcur, "Kcur_normed", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur_rope", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur_rope", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                    model.layers[il].attn_post_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_post_norm", il);
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_ffn(ctx0, lctx, ffn_inp,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+            cb(cur, "ffn_post_norm", -1);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
     // based on the build_qwen2moe() function, changes:
     //   * removed shared experts
     //   * removed bias
@@ -16616,6 +16797,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_olmo();
             } break;
+        case LLM_ARCH_OLMO_1124:
+            {
+                result = llm.build_olmo_1124();
+            } break;
         case LLM_ARCH_OLMOE:
             {
                 result = llm.build_olmoe();
@@ -19885,6 +20070,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_QWEN:
         case LLM_ARCH_QWEN2:
         case LLM_ARCH_QWEN2MOE:
+        case LLM_ARCH_OLMO_1124:
         case LLM_ARCH_OLMOE:
         case LLM_ARCH_PHI2:
         case LLM_ARCH_PHI3:

From 8e752a777b272606f22cb741b03e062de4ddb8fe Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Tue, 19 Nov 2024 13:29:26 +0200
Subject: [PATCH 033/102] llama : add check for KV cache shifts (#10401)

ggml-ci
---
 common/common.cpp | 6 ++++++
 include/llama.h   | 3 +++
 src/llama.cpp     | 6 +++++-
 3 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/common/common.cpp b/common/common.cpp
index 93037462127d5..d314523db4c62 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -875,6 +875,12 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
+    if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) {
+        LOG_ERR("%s: KV cache shifting is not supported for this model (--no-context-shift to disable)'\n", __func__);
+        llama_free_model(model);
+        return iparams;
+    }
+
     if (!params.control_vectors.empty()) {
         if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
         if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_n_layer(model);
diff --git a/include/llama.h b/include/llama.h
index bc268e7996baa..90791d5f5ea12 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -667,6 +667,9 @@ extern "C" {
     // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
     LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
 
+    // Check if the context supports KV cache shifting
+    LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx);
+
     //
     // State / sessions
     //
diff --git a/src/llama.cpp b/src/llama.cpp
index 4f31f25b1d355..c51b36e66042e 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -18213,7 +18213,7 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) {
 
     // apply K-shift if needed
     if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE && lctx.kv_self.has_shift) {
-        if (lctx.model.arch == LLM_ARCH_DEEPSEEK2) { // not supported due to MLA
+        if (!llama_kv_cache_can_shift(&lctx)) {
             GGML_ABORT("Deepseek2 does not support K-shift");
         }
 
@@ -20462,6 +20462,10 @@ void llama_kv_cache_update(struct llama_context * ctx) {
     llama_kv_cache_update_internal(*ctx);
 }
 
+bool llama_kv_cache_can_shift(struct llama_context * ctx) {
+    return ctx->model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA
+}
+
 // deprecated
 size_t llama_get_state_size(struct llama_context * ctx) {
     return llama_state_get_size(ctx);

From 3ee6382d48b07b31e64983969c16019490e19740 Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Tue, 19 Nov 2024 14:29:38 +0100
Subject: [PATCH 034/102] cuda : fix CUDA_FLAGS not being applied (#10403)

---
 ggml/src/ggml-blas/CMakeLists.txt | 1 -
 ggml/src/ggml-cuda/CMakeLists.txt | 2 +-
 2 files changed, 1 insertion(+), 2 deletions(-)

diff --git a/ggml/src/ggml-blas/CMakeLists.txt b/ggml/src/ggml-blas/CMakeLists.txt
index ffe39e8bd1c97..e2cbabf0dae74 100644
--- a/ggml/src/ggml-blas/CMakeLists.txt
+++ b/ggml/src/ggml-blas/CMakeLists.txt
@@ -75,7 +75,6 @@ if (BLAS_FOUND)
 
     message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
 
-    #add_compile_options(${BLAS_LINKER_FLAGS})
     target_compile_options(ggml-blas PRIVATE ${BLAS_LINKER_FLAGS})
 
     if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt
index 25858ece8ddd7..e1482a269d698 100644
--- a/ggml/src/ggml-cuda/CMakeLists.txt
+++ b/ggml/src/ggml-cuda/CMakeLists.txt
@@ -149,7 +149,7 @@ if (CUDAToolkit_FOUND)
         list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
     endif()
 
-    add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+    target_compile_options(ggml-cuda PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
 else()
     message(FATAL_ERROR "CUDA Toolkit not found")
 endif()

From 2a11b6b0946c1abab2ab150725610e5ee736b3af Mon Sep 17 00:00:00 2001
From: bandoti <141645996+bandoti@users.noreply.github.com>
Date: Tue, 19 Nov 2024 12:10:30 -0400
Subject: [PATCH 035/102] Add required ggml-base and backend libs to cmake pkg
 (#10407)

---
 cmake/llama-config.cmake.in | 52 +++++++++++++++++++++++--------------
 ggml/CMakeLists.txt         |  8 ++----
 ggml/src/CMakeLists.txt     |  2 +-
 3 files changed, 35 insertions(+), 27 deletions(-)

diff --git a/cmake/llama-config.cmake.in b/cmake/llama-config.cmake.in
index a7f1efb88f2aa..28a8c18b65176 100644
--- a/cmake/llama-config.cmake.in
+++ b/cmake/llama-config.cmake.in
@@ -3,17 +3,11 @@ set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
 set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
 set(LLAMA_SHARED_LIB   @BUILD_SHARED_LIBS@)
 
-set(GGML_BLAS       @GGML_BLAS@)
-set(GGML_CUDA       @GGML_CUDA@)
-set(GGML_METAL      @GGML_METAL@)
-set(GGML_HIP        @GGML_HIP@)
 set(GGML_ACCELERATE @GGML_ACCELERATE@)
-set(GGML_VULKAN @GGML_VULKAN@)
 set(GGML_VULKAN_CHECK_RESULTS @GGML_VULKAN_CHECK_RESULTS@)
 set(GGML_VULKAN_DEBUG @GGML_VULKAN_DEBUG@)
 set(GGML_VULKAN_MEMORY_DEBUG @GGML_VULKAN_MEMORY_DEBUG@)
 set(GGML_VULKAN_VALIDATE @GGML_VULKAN_VALIDATE@)
-set(GGML_SYCL @GGML_SYCL@)
 set(GGML_OPENMP @GGML_OPENMP@)
 
 @PACKAGE_INIT@
@@ -22,10 +16,39 @@ set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@")
 set_and_check(LLAMA_LIB_DIR     "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
 set_and_check(LLAMA_BIN_DIR     "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
 
-# Ensure transient dependencies satisfied
-
 find_package(Threads REQUIRED)
 
+set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@")
+set(_llama_link_deps "")
+foreach(_ggml_lib ggml ggml-base)
+    string(REPLACE "-" "_" _ggml_lib_var "${_ggml_lib}_LIBRARY")
+    find_library(${_ggml_lib_var} ${_ggml_lib}
+        REQUIRED
+        HINTS ${LLAMA_LIB_DIR}
+        NO_CMAKE_FIND_ROOT_PATH
+    )
+    list(APPEND _llama_link_deps "${${_ggml_lib_var}}")
+    message(STATUS "Found ${${_ggml_lib_var}}")
+endforeach()
+
+foreach(backend amx blas cann cpu cuda hip kompute metal musa rpc sycl vulkan)
+    string(TOUPPER "GGML_${backend}" backend_id)
+    set(_ggml_lib "ggml-${backend}")
+    string(REPLACE "-" "_" _ggml_lib_var "${_ggml_lib}_LIBRARY")
+
+    find_library(${_ggml_lib_var} ${_ggml_lib}
+        HINTS ${LLAMA_LIB_DIR}
+        NO_CMAKE_FIND_ROOT_PATH
+    )
+    if(${_ggml_lib_var})
+        list(APPEND _llama_link_deps "${${_ggml_lib_var}}")
+        set(${backend_id} ON)
+        message(STATUS "Found backend ${${_ggml_lib_var}}")
+    else()
+        set(${backend_id} OFF)
+    endif()
+endforeach()
+
 if (APPLE AND GGML_ACCELERATE)
     find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
 endif()
@@ -48,7 +71,7 @@ if (GGML_VULKAN)
     find_package(Vulkan REQUIRED)
 endif()
 
-if (GGML_HIPBLAS)
+if (GGML_HIP)
     find_package(hip REQUIRED)
     find_package(hipblas REQUIRED)
     find_package(rocblas REQUIRED)
@@ -63,24 +86,13 @@ if (GGML_OPENMP)
     find_package(OpenMP REQUIRED)
 endif()
 
-
-find_library(ggml_LIBRARY ggml
-    REQUIRED
-    HINTS ${LLAMA_LIB_DIR}
-    NO_CMAKE_FIND_ROOT_PATH
-)
-
 find_library(llama_LIBRARY llama
     REQUIRED
     HINTS ${LLAMA_LIB_DIR}
     NO_CMAKE_FIND_ROOT_PATH
 )
 
-set(_llama_link_deps "${ggml_LIBRARY}" "@GGML_LINK_LIBRARIES@")
-set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@")
-
 add_library(llama UNKNOWN IMPORTED)
-
 set_target_properties(llama
     PROPERTIES
         INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index b16a0e9ada7ed..9ab91421a7d25 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -235,12 +235,8 @@ set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
 #if (GGML_METAL)
 #    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
 #endif()
-install(TARGETS ggml PUBLIC_HEADER)
-
-if (BUILD_SHARED_LIBS)
-    install(TARGETS ggml      LIBRARY)
-    install(TARGETS ggml-base LIBRARY)
-endif()
+install(TARGETS ggml LIBRARY PUBLIC_HEADER)
+install(TARGETS ggml-base LIBRARY)
 
 # FIXME: this should be done in the backend cmake files
 if (GGML_METAL)
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index ae7d3abc8de32..8df0e85c0d092 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -239,8 +239,8 @@ function(ggml_add_backend backend)
             if (${BUILD_SHARED_LIBS})
                 target_compile_definitions(${backend_target} PRIVATE GGML_BACKEND_BUILD)
                 target_compile_definitions(${backend_target} PUBLIC  GGML_BACKEND_SHARED)
-                install(TARGETS ${backend_target} LIBRARY)
             endif()
+            install(TARGETS ${backend_target} LIBRARY)
             target_link_libraries(ggml PUBLIC ${backend_target})
             string(TOUPPER "GGML_USE_${backend}" backend_use)
             target_compile_definitions(ggml PUBLIC ${backend_use})

From 342397dc7edb311e0373205134d0d3a928b891b3 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=95=AD=E6=BE=A7=E9=82=A6?=
 <45505768+shou692199@users.noreply.github.com>
Date: Wed, 20 Nov 2024 01:42:00 +0800
Subject: [PATCH 036/102] cmake: force MSVC compiler charset to utf-8 (#9989)

---
 CMakeLists.txt | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 93c60ef431d66..994e61e45fedd 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -46,6 +46,13 @@ if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
 endif()
 
+if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:/source-charset:utf-8>")
+    add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/source-charset:utf-8>")
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:/execution-charset:utf-8>")
+    add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/execution-charset:utf-8>")
+endif()
+
 #
 # option list
 #

From 12b0ad953a59563ea8d973708760d747321d8432 Mon Sep 17 00:00:00 2001
From: PAB <pierreantoine.bannier@gmail.com>
Date: Mon, 18 Nov 2024 10:02:49 +0100
Subject: [PATCH 037/102] metal : add `GGML_UNARY_OP_ELU` kernel (ggml/1018)

---
 ggml/src/ggml-metal/ggml-metal.m     | 15 +++++++++++++++
 ggml/src/ggml-metal/ggml-metal.metal |  8 ++++++++
 2 files changed, 23 insertions(+)

diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index 58fee4bfd1296..d1abb3cef0ec4 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -126,6 +126,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,
     GGML_METAL_KERNEL_TYPE_SILU,
     GGML_METAL_KERNEL_TYPE_SILU_4,
+    GGML_METAL_KERNEL_TYPE_ELU,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,
@@ -649,6 +650,7 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,                  gelu_quick_4,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                          silu,                           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4,                        silu_4,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ELU,                           elu,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,                  soft_max_f16,                   has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,                soft_max_f16_4,                 has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,                  soft_max_f32,                   has_simdgroup_reduction);
@@ -968,6 +970,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
+                case GGML_UNARY_OP_ELU:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -1589,6 +1592,18 @@ static void ggml_metal_encode_node(
 
                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                 } break;
+                case GGML_UNARY_OP_ELU:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ELU].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
                 default:
                 {
                     GGML_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));
diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
index 86fdf1c18cfb6..819b20ba89bec 100644
--- a/ggml/src/ggml-metal/ggml-metal.metal
+++ b/ggml/src/ggml-metal/ggml-metal.metal
@@ -782,6 +782,14 @@ kernel void kernel_silu_4(
     dst[tpig] = x / (1.0f + exp(-x));
 }
 
+kernel void kernel_elu(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+    dst[tpig] = (x > 0.0f) ? x : (exp(x) - 1.0f);
+}
+
 kernel void kernel_sqr(
         device const float * src0,
         device       float * dst,

From 611fabd7922050e1e99bd276d3544527cd46047b Mon Sep 17 00:00:00 2001
From: Plamen Minev <pacominev@gmail.com>
Date: Mon, 18 Nov 2024 15:02:27 +0200
Subject: [PATCH 038/102] metal : fox offset integer overflows in im2col
 (ggml/1015)

-- While running StableDiffusion.cpp locally with Metal some offsets overflow and results in incorrect calculations
---
 ggml/src/ggml-metal/ggml-metal.metal | 52 ++++++++++++++++++----------
 1 file changed, 33 insertions(+), 19 deletions(-)

diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
index 819b20ba89bec..971f5054bce2a 100644
--- a/ggml/src/ggml-metal/ggml-metal.metal
+++ b/ggml/src/ggml-metal/ggml-metal.metal
@@ -2145,20 +2145,34 @@ kernel void kernel_im2col(
         uint3  tgpg[[threadgroups_per_grid]],
         uint3 tpitg[[thread_position_in_threadgroup]],
         uint3   ntg[[threads_per_threadgroup]]) {
-    const int32_t iiw = tgpig[2] * s0 + tpitg[2] * d0 - p0;
-    const int32_t iih = tgpig[1] * s1 + tpitg[1] * d1 - p1;
+//    const int64_t IC = tgpg[0];
+    const int64_t OH = tgpg[1];
+    const int64_t OW = tgpg[2];
 
-    const int32_t offset_dst =
-        (tpitg[0] * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW +
-        (tgpig[0] * (ntg[1] * ntg[2]) + tpitg[1] * ntg[2] + tpitg[2]);
+//    const int64_t N  = ntg[0];
+    const int64_t KH = ntg[1];
+    const int64_t KW = ntg[2];
+
+    const int64_t in  = tpitg[0];
+    const int64_t ikh = tpitg[1];
+    const int64_t ikw = tpitg[2];
+
+    const int64_t iic = tgpig[0];
+    const int64_t ioh = tgpig[1];
+    const int64_t iow = tgpig[2];
+
+    const int64_t iiw = iow*s0 + ikw*d0 - p0;
+    const int64_t iih = ioh*s1 + ikh*d1 - p1;
+
+    const int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*CHW + (iic*(KH*KW) + ikh*KW + ikw);
 
     device T * pdst = (device T *) (dst);
 
     if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
         pdst[offset_dst] = 0.0f;
     } else {
-        const int32_t offset_src = tpitg[0] * ofs0 + tgpig[0] * ofs1;
-        pdst[offset_dst] = x[offset_src + iih * IW + iiw];
+        const int64_t offset_src = in*ofs0 + iic*ofs1 + iih*IW + iiw;
+        pdst[offset_dst] = x[offset_src];
     }
 }
 
@@ -2209,25 +2223,25 @@ kernel void kernel_im2col_ext(
         uint3  tgpg[[threadgroups_per_grid]],      // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW
         uint3 tpitg[[thread_position_in_threadgroup]],
         uint3   ntg[[threads_per_threadgroup]]) {  // [M, 1, 1]
-    const int32_t KHW = KH * KW;             // KHW == ntg[1] * ntg[2], KW == ntg[2]
+    const int64_t KHW = KH * KW;             // KHW == ntg[1] * ntg[2], KW == ntg[2]
 
-    const int32_t d = tgpig[0] / CHW;
-    const int32_t chw = tgpig[0] % CHW;
-    const int32_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
-    const int32_t HW = tgpig[0] % KHW;
+    const int64_t d = tgpig[0] / CHW;
+    const int64_t chw = tgpig[0] % CHW;
+    const int64_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
+    const int64_t HW = tgpig[0] % KHW;
 
-    const int32_t tpitg_0 = (d * ntg[0]) + tpitg[0];
+    const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0];
     if (tpitg_0 >= N) {
         return;
     }
 
-    const int32_t tpitg_1 = HW / KW;
-    const int32_t tpitg_2 = HW % KW;
+    const int64_t tpitg_1 = HW / KW;
+    const int64_t tpitg_2 = HW % KW;
 
-    const int32_t iiw = tgpig[2] * s0 + tpitg_2 * d0 - p0;
-    const int32_t iih = tgpig[1] * s1 + tpitg_1 * d1 - p1;
+    const int64_t iiw = tgpig[2] * s0 + tpitg_2 * d0 - p0;
+    const int64_t iih = tgpig[1] * s1 + tpitg_1 * d1 - p1;
 
-    const int32_t offset_dst =
+    const int64_t offset_dst =
         (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW +
         (tgpig_0 * KHW + tpitg_1 * KW + tpitg_2);
 
@@ -2236,7 +2250,7 @@ kernel void kernel_im2col_ext(
     if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
         pdst[offset_dst] = 0.0f;
     } else {
-        const int32_t offset_src = tpitg_0 * ofs0 + tgpig_0 * ofs1;
+        const int64_t offset_src = tpitg_0 * ofs0 + tgpig_0 * ofs1;
         pdst[offset_dst] = x[offset_src + iih * IW + iiw];
     }
 }

From 9fe0fb062630728e3c21b5839e3bce87bff2440a Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Tue, 19 Nov 2024 19:15:50 +0200
Subject: [PATCH 039/102] sync : ggml

---
 scripts/sync-ggml.last | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last
index 6ddb71ab133bf..e9bd2dbb05e2c 100644
--- a/scripts/sync-ggml.last
+++ b/scripts/sync-ggml.last
@@ -1 +1 @@
-9d0708e863f3aa2fc1eb0b75d433303c30bd0dbc
+2884dd72fea8922910fe53387c3d17ab928d3a8e

From 42ae10bbcd7b56f29a302c86796542a6dadf46c9 Mon Sep 17 00:00:00 2001
From: haopeng <657407891@qq.com>
Date: Wed, 20 Nov 2024 04:10:31 +0800
Subject: [PATCH 040/102] add cmake rvv support (#10411)

---
 ggml/CMakeLists.txt              | 1 +
 ggml/src/ggml-cpu/CMakeLists.txt | 5 +++++
 2 files changed, 6 insertions(+)

diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 9ab91421a7d25..2d32da1b6d879 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -109,6 +109,7 @@ if (NOT MSVC)
 endif()
 option(GGML_LASX        "ggml: enable lasx"             ON)
 option(GGML_LSX         "ggml: enable lsx"              ON)
+option(GGML_RVV         "ggml: enable rvv"              ON)
 option(GGML_SVE         "ggml: enable SVE"              OFF)
 
 if (WIN32)
diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt
index cef41a0743cef..2880523331dbd 100644
--- a/ggml/src/ggml-cpu/CMakeLists.txt
+++ b/ggml/src/ggml-cpu/CMakeLists.txt
@@ -244,6 +244,11 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
     if (GGML_LSX)
         list(APPEND ARCH_FLAGS -mlsx)
     endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
+    message(STATUS "RISC-V detected")
+    if (GGML_RVV)
+        list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
+    endif()
 else()
     message(STATUS "Unknown architecture")
 endif()

From 3952a221af54b8a6549bc2bd4a7363ef7ad3081e Mon Sep 17 00:00:00 2001
From: Anthony Van de Gejuchte <anthonyvdgent@gmail.com>
Date: Tue, 19 Nov 2024 23:18:17 +0100
Subject: [PATCH 041/102] Fix missing file renames in Makefile due to changes
 in commit ae8de6d50a (#10413)

---
 Makefile | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/Makefile b/Makefile
index 95110d4eb7d8b..5c899438515e1 100644
--- a/Makefile
+++ b/Makefile
@@ -730,10 +730,10 @@ GLSLC_CMD  = glslc
 _ggml_vk_genshaders_cmd = $(shell pwd)/vulkan-shaders-gen
 _ggml_vk_header = ggml/src/ggml-vulkan-shaders.hpp
 _ggml_vk_source = ggml/src/ggml-vulkan-shaders.cpp
-_ggml_vk_input_dir = ggml/src/vulkan-shaders
+_ggml_vk_input_dir = ggml/src/ggml-vulkan/vulkan-shaders
 _ggml_vk_shader_deps = $(echo $(_ggml_vk_input_dir)/*.comp)
 
-ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source)
+ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source)
 	$(CXX) $(CXXFLAGS) $(shell pkg-config --cflags vulkan) -c $< -o $@
 
 $(_ggml_vk_header): $(_ggml_vk_source)
@@ -745,8 +745,8 @@ $(_ggml_vk_source): $(_ggml_vk_shader_deps) vulkan-shaders-gen
 		--target-hpp $(_ggml_vk_header) \
 		--target-cpp $(_ggml_vk_source)
 
-vulkan-shaders-gen: ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
-	$(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
+vulkan-shaders-gen: ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+	$(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
 
 endif # GGML_VULKAN
 

From ad21c9e1f14d82b8c15ae369a8839019e3d498b4 Mon Sep 17 00:00:00 2001
From: Neo Zhang Jianyu <jianyu.zhang@intel.com>
Date: Wed, 20 Nov 2024 13:54:25 +0800
Subject: [PATCH 042/102] update rel to 4040 (#10395)

Co-authored-by: arthw <14088817+arthw@users.noreply.github.com>
---
 docs/backend/SYCL.md | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/docs/backend/SYCL.md b/docs/backend/SYCL.md
index e431f51f184cc..8d8312e9158f1 100644
--- a/docs/backend/SYCL.md
+++ b/docs/backend/SYCL.md
@@ -34,9 +34,10 @@ The SYCL backend would be broken by some PRs due to no online CI.
 
 The following release is verified with good quality:
 
-|Commit ID|Tag|Release|Verified  Platform|
-|-|-|-|-|
-|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1|
+|Commit ID|Tag|Release|Verified  Platform| Update date|
+|-|-|-|-|-|
+|3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
+|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||
 
 
 ## News

From 1bacb9f62514b520bdf74ed6feb46c80508dad38 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Wed, 20 Nov 2024 01:11:00 -0600
Subject: [PATCH 043/102] vulkan: further optimize mul_mat_vec using larger
 loads (#10387)

* vulkan: Use pipeline_robustness to disable robustness in mul_mat_vec.

Add some early returns for nonexistent rows in mul_mat_vec shaders. These
can only be hit when dispatching a 2D grid of workgroups. Fix the logic
for the 2D grid of workgroups to round up.

Enable the pipeline robustness extension if it's available, and use it to
disable robustness for these pipelines. The instructions to do the bounds
checking contend for the same ALU resources as the bit twiddling dequant
instructions.

* vulkan: Add GLSL structure aliases for quant types to allow larger loads

In Vulkan it's not possible to cast pointer types, so instead you have to
declare an aliased binding for the memory with a different type. This
commit adds aliases for the quant formats using 16b ints, and in a few
places where the struct size is a multiple of 4 also using 32b ints.
Currently only q4_k's aliases are used, but others will be used in
subsequent commits.

* vulkan: use larger loads in q5_k and q6_k shaders.

Similar to the optimization I did in q4_k recently, this vectorizes some loads
and reduces the number of bit twiddling instructions.

* vulkan: use larger K step per iteration in mul_mat_vec.

Add vec4 dequantization functions, and use them to do K=8 per iteration in
mul_mat_vec. This uses 16b loads for the quant values and 128b loads for B
which helps reduce the load on the memory system.

The K_PER_ITER==2 logic is still there, just for F16/F32, and really only
because they support unaligned sizes.

Tweak the num_iters/unrolling logic to be simpler and catch a couple missed
unrolling opportunities.
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp          | 101 +++++++-----
 .../vulkan-shaders/dequant_funcs.comp         |  48 ++++++
 .../vulkan-shaders/mul_mat_vec.comp           |  82 ++++++++--
 .../vulkan-shaders/mul_mat_vec_base.comp      |   3 +
 .../vulkan-shaders/mul_mat_vec_q2_k.comp      |   4 +
 .../vulkan-shaders/mul_mat_vec_q3_k.comp      |   4 +
 .../vulkan-shaders/mul_mat_vec_q4_k.comp      |  34 ++---
 .../vulkan-shaders/mul_mat_vec_q5_k.comp      | 144 +++++++++++-------
 .../vulkan-shaders/mul_mat_vec_q6_k.comp      |  69 ++++++---
 .../src/ggml-vulkan/vulkan-shaders/types.comp | 113 +++++++++++++-
 .../vulkan-shaders/vulkan-shaders-gen.cpp     |   6 +-
 11 files changed, 459 insertions(+), 149 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 21fee2f3d7533..ca71da2f7b7f5 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -158,6 +158,7 @@ struct vk_device_struct {
     std::string name;
     uint64_t max_memory_allocation_size;
     bool fp16;
+    bool pipeline_robustness;
     vk::Device device;
     uint32_t vendor_id;
     vk_queue compute_queue;
@@ -654,7 +655,7 @@ static uint32_t compile_count = 0;
 static std::mutex compile_count_mutex;
 static std::condition_variable compile_count_cond;
 
-static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants, uint32_t align) {
+static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants, uint32_t align, bool disable_robustness) {
     VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size << ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align << ")");
     GGML_ASSERT(parameter_count > 0);
     GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT
@@ -724,6 +725,15 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
         vk::PipelineCreateFlags(),
         pipeline_shader_create_info,
         pipeline->layout);
+
+    vk::PipelineRobustnessCreateInfoEXT rci;
+
+    if (device->pipeline_robustness && disable_robustness) {
+        rci.storageBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled;
+        rci.uniformBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled;
+        compute_pipeline_create_info.setPNext(&rci);
+    }
+
     pipeline->pipeline = device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value;
 
     {
@@ -1261,7 +1271,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
 
     std::vector<std::future<void>> compiles;
-    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants, uint32_t align) {
+    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants, uint32_t align, bool disable_robustness = false) {
         {
             // wait until fewer than N compiles are in progress
             uint32_t N = std::max(1u, std::thread::hardware_concurrency());
@@ -1271,7 +1281,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             }
             compile_count++;
         }
-        compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align));
+        compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align, disable_robustness));
     };
 
     if (device->fp16) {
@@ -1370,45 +1380,45 @@ static void ggml_vk_load_shaders(vk_device& device) {
     // computing two rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0.
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size}, 1, true);
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size, 1}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1, true);
 
     // dequant shaders
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@@ -1591,12 +1601,15 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         bool fp16_storage = false;
         bool fp16_compute = false;
+        bool pipeline_robustness = false;
 
         for (const auto& properties : ext_props) {
             if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
                 fp16_storage = true;
             } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
                 fp16_compute = true;
+            } else if (strcmp("VK_EXT_pipeline_robustness", properties.extensionName) == 0) {
+                pipeline_robustness = true;
             }
         }
 
@@ -1642,10 +1655,22 @@ static vk_device ggml_vk_get_device(size_t idx) {
         vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES;
         vk11_features.pNext = &vk12_features;
 
+        VkPhysicalDevicePipelineRobustnessFeaturesEXT pl_robustness_features;
+        pl_robustness_features.pNext = nullptr;
+        pl_robustness_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_ROBUSTNESS_FEATURES_EXT;
+        pl_robustness_features.pipelineRobustness = VK_FALSE;
+
+        if (pipeline_robustness) {
+            vk12_features.pNext = &pl_robustness_features;
+            device_extensions.push_back("VK_EXT_pipeline_robustness");
+        }
+
         vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2);
 
         device->fp16 = device->fp16 && vk12_features.shaderFloat16;
 
+        device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
+
         if (!vk11_features.storageBuffer16BitAccess) {
             std::cerr << "ggml_vulkan: device " << GGML_VK_NAME << idx << " does not support 16-bit storage." << std::endl;
             throw std::runtime_error("Unsupported device");
@@ -3190,7 +3215,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
     if (ne01 > max_groups_x) {
         groups_z = 64;
-        groups_x /= groups_z;
+        groups_x = CEIL_DIV(groups_x, groups_z);
     }
 
     // compute
@@ -3767,7 +3792,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
 
     if (ne01 > max_groups_x) {
         groups_z = 64;
-        groups_x /= groups_z;
+        groups_x = CEIL_DIV(groups_x, groups_z);
     }
 
     // compute
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
index d5b989735bc0b..5fc1ba4ad364b 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
@@ -2,6 +2,15 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
 #endif
 
+#include "types.comp"
+
+#if defined(A_TYPE_PACKED16)
+layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
+#endif
+#if defined(A_TYPE_PACKED32)
+layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
+#endif
+
 #if defined(DATA_A_F32)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
@@ -20,6 +29,11 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
     return (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a_packed16[a_offset + ib].d);
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return (vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, (vui >> 12) & 0xF) - 8.0f) * d;
+}
 #endif
 
 #if defined(DATA_A_Q4_1)
@@ -29,6 +43,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
     return vec2(vui & 0xF, vui >> 4) * d + m;
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a_packed16[a_offset + ib].d);
+    const float m = float(data_a_packed16[a_offset + ib].m);
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, (vui >> 12) & 0xF) * d + m;
+}
 #endif
 
 #if defined(DATA_A_Q5_0)
@@ -39,6 +59,14 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
     return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a_packed16[a_offset + ib].d);
+    const uint uint_qh = uint(data_a_packed16[a_offset + ib].qh[1]) << 16 | data_a_packed16[a_offset + ib].qh[0];
+    const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+    const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10);
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return (vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, ((vui >> 12) & 0xF) | qh1.y) - 16.0f) * d;
+}
 #endif
 
 #if defined(DATA_A_Q5_1)
@@ -50,6 +78,15 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
     return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a_packed16[a_offset + ib].d);
+    const float m = float(data_a_packed16[a_offset + ib].m);
+    const uint uint_qh = data_a_packed16[a_offset + ib].qh;
+    const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+    const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10);
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return vec4(((vui >> 0) & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, ((vui >> 12) & 0xF) | qh1.y) * d + m;
+}
 #endif
 
 #if defined(DATA_A_Q8_0)
@@ -57,6 +94,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const float d = float(data_a[a_offset + ib].d);
     return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d;
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a_packed16[a_offset + ib].d);
+    uint32_t v0 = data_a_packed16[a_offset + ib].qs[iqs/2];
+    uint32_t v1 = data_a_packed16[a_offset + ib].qs[iqs/2 + 1];
+    return vec4(int8_t(v0 & 0xFF), int8_t((v0 >> 8) & 0xFF), int8_t(v1 & 0xFF), int8_t((v1 >> 8) & 0xFF)) * d;
+}
 #endif
 
 #if defined(DATA_A_IQ4_NL)
@@ -65,4 +108,9 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
     return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d;
 }
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a_packed16[a_offset + ib].d);
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return vec4(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[(vui >> 4) & 0xF], kvalues_iq4nl[(vui >> 8) & 0xF], kvalues_iq4nl[(vui >> 12) & 0xF]) * d;
+}
 #endif
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
index 970aac6ef72cc..00807a060c46b 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
@@ -3,7 +3,7 @@
 #ifdef FLOAT16
 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
 #endif
-#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+#extension GL_EXT_shader_explicit_arithmetic_types : require
 
 #include "mul_mat_vec_base.comp"
 
@@ -12,16 +12,48 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 layout (constant_id = 0) const uint BLOCK_SIZE = 32;
 layout (constant_id = 1) const uint NUM_ROWS = 1;
 
+#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
+#define K_PER_ITER 8
+#else
+#define K_PER_ITER 2
+#endif
+
+
 uint a_offset, b_offset, d_offset, y_offset;
 
 shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
 
 void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter)
 {
-    const uint col = i*BLOCK_SIZE + 2*tid;
+    const uint col = i*BLOCK_SIZE + K_PER_ITER*tid;
     const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
     const uint iybs = col - col%QUANT_K; // y block start index
 
+#if K_PER_ITER == 8
+#if QUANT_R == 2
+    B_TYPE_VEC4 bv02 = data_b_v4[(b_offset + iybs + iqs) / 4];
+    B_TYPE_VEC4 bv13 = data_b_v4[(b_offset + iybs + iqs + y_offset) / 4];
+    FLOAT_TYPE b0 = FLOAT_TYPE(bv02.x);
+    FLOAT_TYPE b1 = FLOAT_TYPE(bv13.x);
+    FLOAT_TYPE b2 = FLOAT_TYPE(bv02.y);
+    FLOAT_TYPE b3 = FLOAT_TYPE(bv13.y);
+    FLOAT_TYPE b4 = FLOAT_TYPE(bv02.z);
+    FLOAT_TYPE b5 = FLOAT_TYPE(bv13.z);
+    FLOAT_TYPE b6 = FLOAT_TYPE(bv02.w);
+    FLOAT_TYPE b7 = FLOAT_TYPE(bv13.w);
+#else
+    B_TYPE_VEC4 bv0 = data_b_v4[(b_offset + iybs + iqs) / 4];
+    B_TYPE_VEC4 bv1 = data_b_v4[(b_offset + iybs + iqs) / 4 + 1];
+    FLOAT_TYPE b0 = FLOAT_TYPE(bv0.x);
+    FLOAT_TYPE b1 = FLOAT_TYPE(bv0.y);
+    FLOAT_TYPE b2 = FLOAT_TYPE(bv0.z);
+    FLOAT_TYPE b3 = FLOAT_TYPE(bv0.w);
+    FLOAT_TYPE b4 = FLOAT_TYPE(bv1.x);
+    FLOAT_TYPE b5 = FLOAT_TYPE(bv1.y);
+    FLOAT_TYPE b6 = FLOAT_TYPE(bv1.z);
+    FLOAT_TYPE b7 = FLOAT_TYPE(bv1.w);
+#endif
+#else
     // Check if the second of the pair of elements is OOB, and don't fetch B or
     // accumulate it. We still fetch a pair of elements for A, which is fine for
     // quantized formats since they'll be within the same block. We should
@@ -34,9 +66,24 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_
     if (!OOB) {
         b1 = FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
     }
+#endif
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const uint ib = ((first_row + n)*p.ncols + col)/QUANT_K; // block index
 
+#if K_PER_ITER == 8
+        const vec4 v = dequantize4(ib, iqs, a_offset);
+        const vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset);
+
+        // matrix multiplication
+        temp[n] = fma(FLOAT_TYPE(v.x), b0, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v.z), b2, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v.w), b3, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v2.x), b4, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v2.y), b5, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v2.z), b6, temp[n]);
+        temp[n] = fma(FLOAT_TYPE(v2.w), b7, temp[n]);
+#else
         const vec2 v = dequantize(ib, iqs, a_offset);
 
         // matrix multiplication
@@ -44,6 +91,7 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_
         if (!OOB) {
             temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]);
         }
+#endif
     }
 }
 
@@ -61,22 +109,33 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
         temp[i] = FLOAT_TYPE(0);
     }
 
-    const int unroll_count = 8;
-
-    const uint num_iters = (p.ncols >= 2*tid) ? ((p.ncols - 2*tid + BLOCK_SIZE - 1) / BLOCK_SIZE) : 0;
-    const uint unrolled_iters = num_iters & ~(2*unroll_count - 1);
+    uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE);
+    if (num_iters * K_PER_ITER * BLOCK_SIZE + K_PER_ITER*tid < p.ncols) {
+        num_iters++;
+    }
+    int unroll_count = 4;
+    uint unrolled_iters = num_iters & ~(unroll_count - 1);
 
     uint i = 0;
     while (i < unrolled_iters) {
         // Manually partially unroll the loop
         [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
-            iter(temp, first_row, num_rows, tid, i, false);
-            i += 2;
+            iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false);
+            i++;
+        }
+    }
+    unroll_count = 2;
+    unrolled_iters = num_iters & ~(unroll_count - 1);
+    while (i < unrolled_iters) {
+        // Manually partially unroll the loop
+        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
+            iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false);
+            i++;
         }
     }
     while (i < num_iters) {
-        iter(temp, first_row, num_rows, tid, i, true);
-        i += 2;
+        iter(temp, first_row, num_rows, tid, i*K_PER_ITER, true);
+        i++;
     }
 
     // sum up partial sums and write back result
@@ -106,6 +165,9 @@ void main() {
     if (first_row + NUM_ROWS <= p.stride_d) {
         compute_outputs(first_row, NUM_ROWS);
     } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
         compute_outputs(first_row, p.stride_d - first_row);
     }
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
index 5920bc93641c8..8d0a5791374d3 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
@@ -12,6 +12,9 @@
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];};
+layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
index ec8eadcd5828a..e2625d32b902d 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
@@ -9,6 +9,10 @@ shared FLOAT_TYPE tmp[32];
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
 
+    if (row >= p.stride_d) {
+        return;
+    }
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
index 3ca4ad85a5ca0..a28804533c320 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
@@ -9,6 +9,10 @@ shared FLOAT_TYPE tmp[32];
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
 
+    if (row >= p.stride_d) {
+        return;
+    }
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
index b7c9b722d126b..5846f2e86f3e3 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
@@ -8,30 +8,14 @@ layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
 
 shared FLOAT_TYPE tmp[32];
 
-// Declare aliased versions of A and B bindings that can use 16b/32b loads for
-// the quantized values, and vec4 loads for B.
-struct block_q4_K_u32
-{
-    f16vec2 d;
-    uint32_t scales[3*QUANT_K/64/4];
-    uint32_t qs[QUANT_K/2/4];
-};
-
-struct block_q4_K_u16
-{
-    f16vec2 d;
-    uint16_t scales[3*QUANT_K/64/2];
-    uint16_t qs[QUANT_K/2/2];
-};
-
-layout (binding = 0) readonly buffer A_u32 {block_q4_K_u32 data_a_u32[];};
-layout (binding = 0) readonly buffer A_u16 {block_q4_K_u16 data_a_u16[];};
-layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
-
 // This shader assumes K_QUANTS_PER_ITERATION == 2 for alignment of loads
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
 
+    if (row >= p.stride_d) {
+        return;
+    }
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
@@ -64,9 +48,9 @@ void main() {
         const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
         const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
 
-        uint32_t scale0_u32 = data_a_u16[ib0 + i].scales[v_im    ];
-        uint32_t scale4_u32 = data_a_u16[ib0 + i].scales[v_im + 2];
-        uint32_t scale8_u32 = data_a_u16[ib0 + i].scales[v_im + 4];
+        uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
+        uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
+        uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
         uvec4 scale0 = uvec4(unpack8(scale0_u32));
         uvec4 scale4 = uvec4(unpack8(scale4_u32));
         uvec4 scale8 = uvec4(unpack8(scale8_u32));
@@ -80,8 +64,8 @@ void main() {
         const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
         const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
 
-        uint32_t qs0_u32 = data_a_u32[ib0 + i].qs[q_offset / 4];
-        uint32_t qs64_u32 = data_a_u32[ib0 + i].qs[q_offset / 4 + 16];
+        uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
+        uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
 
         uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
         uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
index 2306785af4226..22a6bfae4fb61 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
@@ -1,5 +1,7 @@
 #version 450
 
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
 #include "mul_mat_vec_base.comp"
 
 layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
@@ -9,6 +11,10 @@ shared FLOAT_TYPE tmp[32];
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
 
+    if (row >= p.stride_d) {
+        return;
+    }
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
@@ -31,70 +37,106 @@ void main() {
     const uint8_t hm1 = uint8_t(1 << (2*v_im));
     const uint8_t hm2 = uint8_t(hm1 << 4);
 
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += 2) {
         const uint y1_idx = i * QUANT_K + y_offset;
         const uint y2_idx = y1_idx + 128;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
-
-        const uint8_t sc0 = uint8_t(  data_a[ib0 + i].scales[v_im * 2    ]       & 0x3f);
-        const uint8_t sc1 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 1]       & 0x3f);
-        const uint8_t sc2 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 4]       & 0x3f);
-        const uint8_t sc3 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 5]       & 0x3f);
-        const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2    ] & 0xc0) >> 2));
-        const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2));
-        const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2));
-        const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2));
-
-        const uint8_t q4_0  = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
-        const uint8_t q4_1  = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
-        const uint8_t q4_2  = uint8_t(data_a[ib0 + i].qs[q_offset + 16] & 0xf);
-        const uint8_t q4_3  = uint8_t(data_a[ib0 + i].qs[q_offset + 17] & 0xf);
-        const uint8_t q4_4  = uint8_t(data_a[ib0 + i].qs[q_offset     ]  >> 4);
-        const uint8_t q4_5  = uint8_t(data_a[ib0 + i].qs[q_offset +  1]  >> 4);
-        const uint8_t q4_6  = uint8_t(data_a[ib0 + i].qs[q_offset + 16]  >> 4);
-        const uint8_t q4_7  = uint8_t(data_a[ib0 + i].qs[q_offset + 17]  >> 4);
-        const uint8_t q4_8  = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
-        const uint8_t q4_9  = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
-        const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] & 0xf);
-        const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] & 0xf);
-        const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
-        const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
-        const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 80]  >> 4);
-        const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 81]  >> 4);
+        f16vec2 d = data_a[ib0 + i].d;
+        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+
+        uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
+        uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
+        uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
+        uvec4 scale0 = uvec4(unpack8(scale0_u32));
+        uvec4 scale4 = uvec4(unpack8(scale4_u32));
+        uvec4 scale8 = uvec4(unpack8(scale8_u32));
+
+        const uint32_t sc0 = (  scale0.x       & 0x3f);
+        const uint32_t sc1 = (  scale0.y       & 0x3f);
+        const uint32_t sc2 = (  scale4.x       & 0x3f);
+        const uint32_t sc3 = (  scale4.y       & 0x3f);
+        const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
+        const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
+        const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
+        const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
+
+        uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
+        uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
+
+        uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
+        uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
+        uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
+        uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
+
+        uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
+        uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
+        uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
+        uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4));
+
+        const uint32_t q4_0  = qs0_16_lo4.x;
+        const uint32_t q4_1  = qs0_16_lo4.y;
+        const uint32_t q4_2  = qs0_16_lo4.z;
+        const uint32_t q4_3  = qs0_16_lo4.w;
+        const uint32_t q4_4  = qs0_16_hi4.x;
+        const uint32_t q4_5  = qs0_16_hi4.y;
+        const uint32_t q4_6  = qs0_16_hi4.z;
+        const uint32_t q4_7  = qs0_16_hi4.w;
+        const uint32_t q4_8  = qs64_80_lo4.x;
+        const uint32_t q4_9  = qs64_80_lo4.y;
+        const uint32_t q4_10 = qs64_80_lo4.z;
+        const uint32_t q4_11 = qs64_80_lo4.w;
+        const uint32_t q4_12 = qs64_80_hi4.x;
+        const uint32_t q4_13 = qs64_80_hi4.y;
+        const uint32_t q4_14 = qs64_80_hi4.z;
+        const uint32_t q4_15 = qs64_80_hi4.w;
+
+        B_TYPE_VEC2 by10 =  data_b_v2[(b_offset + y1_idx) / 2];
+        B_TYPE_VEC2 by116 = data_b_v2[(b_offset + y1_idx) / 2 + 8];
+        B_TYPE_VEC2 by132 = data_b_v2[(b_offset + y1_idx) / 2 + 16];
+        B_TYPE_VEC2 by148 = data_b_v2[(b_offset + y1_idx) / 2 + 24];
+        B_TYPE_VEC2 by20 =  data_b_v2[(b_offset + y2_idx) / 2];
+        B_TYPE_VEC2 by216 = data_b_v2[(b_offset + y2_idx) / 2 + 8];
+        B_TYPE_VEC2 by232 = data_b_v2[(b_offset + y2_idx) / 2 + 16];
+        B_TYPE_VEC2 by248 = data_b_v2[(b_offset + y2_idx) / 2 + 24];
+
+        uint32_t qh0 = data_a_packed16[ib0 + i].qh[l0 / 2];
+        uint32_t qh1 = qh0 >> 8;
+        uint32_t qh16 = data_a_packed16[ib0 + i].qh[l0 / 2 + 8];
+        uint32_t qh17 = qh16 >> 8;
 
         const FLOAT_TYPE sx =
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]), (q4_0 + (((data_a[ib0 + i].qh[l0     ] & hm1) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx +  1]), (q4_1 + (((data_a[ib0 + i].qh[l0 +  1] & hm1) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 16]), (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0)),
-             FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0)))));
+          fma(FLOAT_TYPE(by10.x), (q4_0 + (((qh0 & hm1) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by10.y), (q4_1 + (((qh1 & hm1) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by116.x), (q4_2 + (((qh16 & hm1) != 0) ? 16 : 0)),
+             FLOAT_TYPE(by116.y) * (q4_3 + (((qh17 & hm1) != 0) ? 16 : 0)))));
         const FLOAT_TYPE sy =
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), (q4_4 + (((data_a[ib0 + i].qh[l0     ] & (hm1 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), (q4_5 + (((data_a[ib0 + i].qh[l0 +  1] & (hm1 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 48]), (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0)),
-             FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0)))));
+          fma(FLOAT_TYPE(by132.x), (q4_4 + (((qh0 & (hm1 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by132.y), (q4_5 + (((qh1 & (hm1 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by148.x), (q4_6 + (((qh16 & (hm1 << 1)) != 0) ? 16 : 0)),
+             FLOAT_TYPE(by148.y) * (q4_7 + (((qh17 & (hm1 << 1)) != 0) ? 16 : 0)))));
         const FLOAT_TYPE sz =
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]), (q4_8  + (((data_a[ib0 + i].qh[l0     ] & hm2) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx +  1]), (q4_9  + (((data_a[ib0 + i].qh[l0 +  1] & hm2) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 16]), (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0)),
-             FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0)))));
+          fma(FLOAT_TYPE(by20.x), (q4_8  + (((qh0 & hm2) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by20.y), (q4_9  + (((qh1 & hm2) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by216.x), (q4_10 + (((qh16 & hm2) != 0) ? 16 : 0)),
+             FLOAT_TYPE(by216.y) * (q4_11 + (((qh17 & hm2) != 0) ? 16 : 0)))));
         const FLOAT_TYPE sw =
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), (q4_12 + (((data_a[ib0 + i].qh[l0     ] & (hm2 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), (q4_13 + (((data_a[ib0 + i].qh[l0 +  1] & (hm2 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 48]), (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0)),
-             FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0)))));
+          fma(FLOAT_TYPE(by232.x), (q4_12 + (((qh0 & (hm2 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by232.y), (q4_13 + (((qh1 & (hm2 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(by248.x), (q4_14 + (((qh16 & (hm2 << 1)) != 0) ? 16 : 0)),
+             FLOAT_TYPE(by248.y) * (q4_15 + (((qh17 & (hm2 << 1)) != 0) ? 16 : 0)))));
         const FLOAT_TYPE smin =
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]), sc2,
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]), sc3,
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]), sc6,
-              (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7)));
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx]));
+          fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
+          fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
+          fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
+              (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
+        temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
     }
 
+    tmp[gl_LocalInvocationID.x] = temp;
+
     // sum up partial sums and write back result
     barrier();
     [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp
index 95c286eeb17e1..0b392d68d04d0 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp
@@ -1,5 +1,7 @@
 #version 450
 
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
 #include "mul_mat_vec_base.comp"
 
 layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
@@ -9,6 +11,10 @@ shared FLOAT_TYPE tmp[32];
 void main() {
     const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
 
+    if (row >= p.stride_d) {
+        return;
+    }
+
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
@@ -36,41 +42,66 @@ void main() {
     const uint s_offset  =  8*v_im + is;
     const uint y_offset = 128*v_im + l0;
 
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
         const uint y_idx   = i * QUANT_K + y_offset;
 
         const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
 
-#if K_QUANTS_PER_ITERATION == 1
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(FLOAT_TYPE(data_b[b_offset + y_idx +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x03) << 4)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x0c) << 2)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x30) >> 0)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0xc0) >> 2)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32), tmp[tmp_idx]))))))));
-#else
+        FLOAT_TYPE scales[4];
+        scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]);
+        scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]);
+        scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]);
+        scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]);
+
+        uint32_t ql0_u32 =  uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
+        uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
+
+        uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
+        uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
+        uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
+        uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
+
+        uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
+        uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
+        uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
+        uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0;
+        uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
+
+        uint32_t q0_u32 = ql0_u32_lo4  | qh0_u32;
+        uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
+        uint32_t q2_u32 = ql0_u32_hi4  | qh4_u32;
+        uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
+
+        uvec4 q0 = uvec4(unpack8(q0_u32));
+        uvec4 q1 = uvec4(unpack8(q1_u32));
+        uvec4 q2 = uvec4(unpack8(q2_u32));
+        uvec4 q3 = uvec4(unpack8(q3_u32));
+
+        B_TYPE_VEC4 by0  = data_b_v4[(b_offset + y_idx) / 4];
+        B_TYPE_VEC4 by32 = data_b_v4[(b_offset + y_idx) / 4 + 8];
+        B_TYPE_VEC4 by64 = data_b_v4[(b_offset + y_idx) / 4 + 16];
+        B_TYPE_VEC4 by96 = data_b_v4[(b_offset + y_idx) / 4 + 24];
+
         FLOAT_TYPE sum = FLOAT_TYPE(0.0);
         [[unroll]] for (int l = 0; l < 4; ++l) {
-            sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32), sum))));
+            sum = fma(FLOAT_TYPE(by0[l])  * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32),
+                  fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32),
+                  fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32),
+                  fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum))));
         }
-        tmp[16 * ix + tid] += sum;
-#endif
+        temp += sum * d;
     }
 
+    tmp[gl_LocalInvocationID.x] = temp;
+
     // sum up partial sums and write back result
     barrier();
     [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
         if (tid < s) {
             tmp[tid] += tmp[tid + s];
-       }
+        }
         barrier();
     }
     if (tid == 0) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
index 21dce72fc7dfb..7a34820bc2789 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
@@ -1,6 +1,8 @@
-#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
-#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
-#endif
+
+#if !defined(GGML_TYPES_COMP)
+#define GGML_TYPES_COMP
+
+#extension GL_EXT_shader_explicit_arithmetic_types : require
 
 #if defined(DATA_A_F32)
 #define QUANT_K 1
@@ -38,8 +40,14 @@ struct block_q4_0
     float16_t d;
     uint8_t qs[16];
 };
+struct block_q4_0_packed16
+{
+    float16_t d;
+    uint16_t qs[16/2];
+};
 
 #define A_TYPE block_q4_0
+#define A_TYPE_PACKED16 block_q4_0_packed16
 #endif
 
 #if defined(DATA_A_Q4_1)
@@ -54,7 +62,15 @@ struct block_q4_1
     uint8_t qs[16];
 };
 
+struct block_q4_1_packed16
+{
+    float16_t d;
+    float16_t m;
+    uint16_t qs[16/2];
+};
+
 #define A_TYPE block_q4_1
+#define A_TYPE_PACKED16 block_q4_1_packed16
 #endif
 
 #if defined(DATA_A_Q5_0)
@@ -70,7 +86,15 @@ struct block_q5_0
     uint8_t qs[16];
 };
 
+struct block_q5_0_packed16
+{
+    float16_t d;
+    uint16_t qh[2];
+    uint16_t qs[16/2];
+};
+
 #define A_TYPE block_q5_0
+#define A_TYPE_PACKED16 block_q5_0_packed16
 #endif
 
 #if defined(DATA_A_Q5_1)
@@ -87,7 +111,16 @@ struct block_q5_1
     uint8_t qs[16];
 };
 
+struct block_q5_1_packed16
+{
+    float16_t d;
+    float16_t m;
+    uint qh;
+    uint16_t qs[16/2];
+};
+
 #define A_TYPE block_q5_1
+#define A_TYPE_PACKED16 block_q5_1_packed16
 #endif
 
 #if defined(DATA_A_Q8_0)
@@ -100,8 +133,14 @@ struct block_q8_0
     float16_t d;
     int8_t qs[32];
 };
+struct block_q8_0_packed16
+{
+    float16_t d;
+    uint16_t qs[32/2];
+};
 
 #define A_TYPE block_q8_0
+#define A_TYPE_PACKED16 block_q8_0_packed16
 #endif
 
 // K-quants
@@ -116,7 +155,23 @@ struct block_q2_K
     f16vec2 d;
 };
 
+struct block_q2_K_packed16
+{
+    uint16_t scales[QUANT_K/16/2];
+    uint16_t qs[QUANT_K/4/2];
+    f16vec2 d;
+};
+
+struct block_q2_K_packed32
+{
+    uint32_t scales[QUANT_K/16/4];
+    uint32_t qs[QUANT_K/4/4];
+    f16vec2 d;
+};
+
 #define A_TYPE block_q2_K
+#define A_TYPE_PACKED16 block_q2_K_packed16
+#define A_TYPE_PACKED32 block_q2_K_packed32
 #endif
 
 #if defined(DATA_A_Q3_K)
@@ -131,7 +186,16 @@ struct block_q3_K
     float16_t d;
 };
 
+struct block_q3_K_packed16
+{
+    uint16_t hmask[QUANT_K/8/2];
+    uint16_t qs[QUANT_K/4/2];
+    uint16_t scales[12/2];
+    float16_t d;
+};
+
 #define A_TYPE block_q3_K
+#define A_TYPE_PACKED16 block_q3_K_packed16
 #endif
 
 #if defined(DATA_A_Q4_K)
@@ -145,7 +209,23 @@ struct block_q4_K
     uint8_t qs[QUANT_K/2];
 };
 
+struct block_q4_K_packed16
+{
+    f16vec2 d;
+    uint16_t scales[3*QUANT_K/64/2];
+    uint16_t qs[QUANT_K/2/2];
+};
+
+struct block_q4_K_packed32
+{
+    f16vec2 d;
+    uint32_t scales[3*QUANT_K/64/4];
+    uint32_t qs[QUANT_K/2/4];
+};
+
 #define A_TYPE block_q4_K
+#define A_TYPE_PACKED16 block_q4_K_packed16
+#define A_TYPE_PACKED32 block_q4_K_packed32
 #endif
 
 #if defined(DATA_A_Q5_K)
@@ -160,7 +240,16 @@ struct block_q5_K
     uint8_t qs[QUANT_K/2];
 };
 
+struct block_q5_K_packed16
+{
+    f16vec2 d;
+    uint16_t scales[12/2];
+    uint16_t qh[QUANT_K/8/2];
+    uint16_t qs[QUANT_K/2/2];
+};
+
 #define A_TYPE block_q5_K
+#define A_TYPE_PACKED16 block_q5_K_packed16
 #endif
 
 #if defined(DATA_A_Q6_K)
@@ -175,7 +264,16 @@ struct block_q6_K
     float16_t d;
 };
 
+struct block_q6_K_packed16
+{
+    uint16_t ql[QUANT_K/2/2];
+    uint16_t qh[QUANT_K/4/2];
+    int8_t scales[QUANT_K/16];
+    float16_t d;
+};
+
 #define A_TYPE block_q6_K
+#define A_TYPE_PACKED16 block_q6_K_packed16
 #endif
 
 // IQuants
@@ -191,10 +289,19 @@ struct block_iq4_nl
     uint8_t qs[QUANT_K/2];
 };
 
+struct block_iq4_nl_packed16
+{
+    float16_t d;
+    uint16_t qs[QUANT_K/2/2];
+};
+
 #define A_TYPE block_iq4_nl
+#define A_TYPE_PACKED16 block_iq4_nl_packed16
 
 const int8_t kvalues_iq4nl[16] = {
     int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10),
     int8_t(1), int8_t(13), int8_t(25), int8_t(38), int8_t(53), int8_t(69), int8_t(89), int8_t(113)
 };
 #endif
+
+#endif // !defined(GGML_TYPES_COMP)
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index fe3e4cb39ff47..f753109556f93 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -317,10 +317,10 @@ void process_shaders() {
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
         std::string shader = (string_ends_with(tname, "_k")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
 
-        string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
-        string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}}));
 
-        string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
 
         // Dequant shaders
         if (tname != "f16") {

From 8fd4b7fa29c3061b2e02e897d818dfcbc593430a Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Wed, 20 Nov 2024 01:40:18 -0600
Subject: [PATCH 044/102] vulkan: copy iq4_nl LUT into shared memory (#10409)

---
 .../ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp  |  2 ++
 .../ggml-vulkan/vulkan-shaders/get_rows_quant.comp  |  4 ++++
 .../src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp |  4 ++++
 ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp     |  4 ++++
 ggml/src/ggml-vulkan/vulkan-shaders/types.comp      | 13 ++++++++++++-
 .../vulkan-shaders/vulkan-shaders-gen.cpp           |  6 +++---
 6 files changed, 29 insertions(+), 4 deletions(-)

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp
index 34ef3da30b82c..8de14fc03f102 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp
@@ -10,6 +10,8 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
 
+    init_iq4nl_shmem();
+
     const uint tid = gl_LocalInvocationID.x % 64;
     const uint il  = tid/32;
     const uint ir  = tid%32;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp
index 8d30b63c165b2..7f608315b6844 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp
@@ -12,6 +12,10 @@ void main() {
     const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
     const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
 
+#if defined(DATA_A_IQ4_NL)
+    init_iq4nl_shmem();
+#endif
+
     if (i00 >= p.ne00) {
         return;
     }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
index 00807a060c46b..2d5b8e4661312 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
@@ -161,6 +161,10 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 void main() {
     const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
 
+#if defined(DATA_A_IQ4_NL)
+    init_iq4nl_shmem();
+#endif
+
     // do NUM_ROWS at a time, unless there aren't enough remaining rows
     if (first_row + NUM_ROWS <= p.stride_d) {
         compute_outputs(first_row, NUM_ROWS);
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
index fffdd18189d55..2ff5c430519bc 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
@@ -75,6 +75,10 @@ shared u16vec2 row_ids[3072];
 #endif
 
 void main() {
+#if defined(DATA_A_IQ4_NL)
+    init_iq4nl_shmem();
+#endif
+
 #ifdef MUL_MAT_ID
     const uint expert_idx = gl_GlobalInvocationID.z;
 #else
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
index 7a34820bc2789..bc28e0ab857aa 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
@@ -298,10 +298,21 @@ struct block_iq4_nl_packed16
 #define A_TYPE block_iq4_nl
 #define A_TYPE_PACKED16 block_iq4_nl_packed16
 
-const int8_t kvalues_iq4nl[16] = {
+const int8_t kvalues_iq4nl_const[16] = {
     int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10),
     int8_t(1), int8_t(13), int8_t(25), int8_t(38), int8_t(53), int8_t(69), int8_t(89), int8_t(113)
 };
+
+shared FLOAT_TYPE kvalues_iq4nl[16];
+
+void init_iq4nl_shmem()
+{
+    // copy the table into shared memory and sync
+    if (gl_LocalInvocationIndex.x < 16) {
+        kvalues_iq4nl[gl_LocalInvocationIndex.x] = FLOAT_TYPE(kvalues_iq4nl_const[gl_LocalInvocationIndex.x]);
+    }
+    barrier();
+}
 #endif
 
 #endif // !defined(GGML_TYPES_COMP)
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index f753109556f93..6bbe8e96edd9f 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -331,11 +331,11 @@ void process_shaders() {
             shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp";
 
             if (tname == "f16") {
-                string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+                string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}));
             } else {
-                string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}});
+                string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}}));
             }
-            string_to_spv("get_rows_" + tname + "_f32", shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}});
+            string_to_spv("get_rows_" + tname + "_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}}));
         }
     }
 

From fab5d30ff6729ff6ff615c41e8c0215d6bc30393 Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Wed, 20 Nov 2024 12:57:53 +0100
Subject: [PATCH 045/102] llama : add .clang-format file (#10415)

---
 .clang-format                        | 161 ++++++
 examples/llama-bench/llama-bench.cpp | 702 +++++++++++++++------------
 2 files changed, 551 insertions(+), 312 deletions(-)
 create mode 100644 .clang-format

diff --git a/.clang-format b/.clang-format
new file mode 100644
index 0000000000000..45232b80ed8cd
--- /dev/null
+++ b/.clang-format
@@ -0,0 +1,161 @@
+---
+Language:        Cpp
+AlignAfterOpenBracket: Align
+AlignArrayOfStructures: Left
+AlignConsecutiveAssignments: AcrossComments
+AlignConsecutiveBitFields: AcrossComments
+AlignConsecutiveDeclarations: AcrossComments
+AlignConsecutiveMacros: AcrossComments
+# AlignConsecutiveShortCaseStatements: AcrossComments
+AlignEscapedNewlines: Left # LeftWithLastLine
+AlignOperands:   Align
+AlignTrailingComments:
+  Kind: Always
+  OverEmptyLines: 1
+AllowAllArgumentsOnNextLine: true
+AllowAllParametersOfDeclarationOnNextLine: false
+# AllowBreakBeforeNoexceptSpecifier: OnlyWithParen
+AllowShortBlocksOnASingleLine: Never
+AllowShortCaseLabelsOnASingleLine: false
+AllowShortFunctionsOnASingleLine: Inline
+AllowShortIfStatementsOnASingleLine: Never
+AllowShortLambdasOnASingleLine: Inline
+AllowShortLoopsOnASingleLine: false
+AlwaysBreakBeforeMultilineStrings: true
+BinPackArguments: true
+BinPackParameters: true # OnePerLine
+BitFieldColonSpacing: Both
+BreakBeforeBraces: Custom # Attach
+BraceWrapping:
+  AfterCaseLabel:  true
+  AfterClass:      false
+  AfterControlStatement: false
+  AfterEnum:       false
+  AfterFunction:   false
+  AfterNamespace:  false
+  AfterObjCDeclaration: false
+  AfterStruct:     false
+  AfterUnion:      false
+  AfterExternBlock: false
+  BeforeCatch:     false
+  BeforeElse:      false
+  BeforeLambdaBody: false
+  BeforeWhile: false
+  IndentBraces:    false
+  SplitEmptyFunction: false
+  SplitEmptyRecord: false
+  SplitEmptyNamespace: false
+# BreakAdjacentStringLiterals: true
+BreakAfterAttributes: Never
+BreakBeforeBinaryOperators: None
+BreakBeforeInlineASMColon: OnlyMultiline
+BreakBeforeTernaryOperators: false
+# BreakBinaryOperations: Never
+BreakConstructorInitializers: AfterColon
+# BreakFunctionDefinitionParameters: false
+BreakInheritanceList: AfterComma
+BreakStringLiterals: true
+# BreakTemplateDeclarations: Yes
+ColumnLimit:     120
+CommentPragmas:  '^ IWYU pragma:'
+CompactNamespaces: false
+ConstructorInitializerIndentWidth: 4
+ContinuationIndentWidth: 4
+Cpp11BracedListStyle: false
+DerivePointerAlignment: false
+DisableFormat:   false
+EmptyLineBeforeAccessModifier: Leave
+EmptyLineAfterAccessModifier: Never
+ExperimentalAutoDetectBinPacking: false
+FixNamespaceComments: true
+IncludeBlocks:   Regroup
+IncludeCategories:
+  - Regex:           '^<.*\.h>'
+    Priority:        1
+    SortPriority:    0
+  - Regex:           '^<.*'
+    Priority:        2
+    SortPriority:    0
+  - Regex:           '.*'
+    Priority:        3
+    SortPriority:    0
+IncludeIsMainRegex: '([-_](test|unittest))?$'
+IncludeIsMainSourceRegex: ''
+IndentAccessModifiers: false
+IndentCaseBlocks: true
+IndentCaseLabels: true
+IndentExternBlock: NoIndent
+IndentGotoLabels: false
+IndentPPDirectives: AfterHash
+IndentWidth:     4
+IndentWrappedFunctionNames: false
+InsertBraces:    true # NOTE: may lead to incorrect formatting
+InsertNewlineAtEOF: true
+JavaScriptQuotes: Leave
+JavaScriptWrapImports: true
+KeepEmptyLinesAtTheStartOfBlocks: false
+LambdaBodyIndentation: Signature
+LineEnding: LF
+MacroBlockBegin: ''
+MacroBlockEnd:   ''
+MaxEmptyLinesToKeep: 1
+NamespaceIndentation: None
+ObjCBinPackProtocolList: Auto
+ObjCBlockIndentWidth: 4
+ObjCSpaceAfterProperty: true
+ObjCSpaceBeforeProtocolList: true
+PPIndentWidth: -1
+PackConstructorInitializers: CurrentLine
+PenaltyBreakAssignment: 2
+PenaltyBreakBeforeFirstCallParameter: 1
+PenaltyBreakComment: 300
+PenaltyBreakFirstLessLess: 120
+PenaltyBreakString: 1000
+PenaltyBreakTemplateDeclaration: 10
+PenaltyExcessCharacter: 1000000
+PenaltyReturnTypeOnItsOwnLine: 200
+PointerAlignment: Middle
+QualifierAlignment: Left
+#QualifierOrder: ['static', 'inline', 'friend', 'constexpr', 'const', 'volatile', 'type', 'restrict']
+RawStringFormats:
+  - Language:        Cpp
+    Delimiters:
+      - cc
+      - CC
+      - cpp
+      - Cpp
+      - CPP
+      - 'c++'
+      - 'C++'
+    CanonicalDelimiter: ''
+ReferenceAlignment: Middle
+ReflowComments:  false # IndentOnly
+SeparateDefinitionBlocks: Always
+SortIncludes:    CaseInsensitive
+SortUsingDeclarations: LexicographicNumeric
+SpaceAfterCStyleCast: true
+SpaceAfterLogicalNot: false
+SpaceAfterTemplateKeyword: true
+SpaceBeforeAssignmentOperators: true
+SpaceBeforeCpp11BracedList: false
+SpaceBeforeCtorInitializerColon: true
+SpaceBeforeInheritanceColon: true
+SpaceBeforeParens: ControlStatements
+SpaceBeforeRangeBasedForLoopColon: true
+SpaceInEmptyBlock: false
+SpaceInEmptyParentheses: false
+SpacesBeforeTrailingComments: 2
+SpacesInAngles:  Never
+SpacesInContainerLiterals: true
+SpacesInLineCommentPrefix:
+  Minimum: 1
+  Maximum: -1
+SpacesInParentheses: false
+SpacesInSquareBrackets: false
+SpaceBeforeSquareBrackets: false
+Standard:        c++17
+TabWidth:        4
+UseTab:          Never
+WhitespaceSensitiveMacros: ['STRINGIZE']
+...
+
diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp
index 8f4e0e206a27e..3dc84a75cbec7 100644
--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
@@ -6,28 +6,28 @@
 #include <clocale>
 #include <cmath>
 #include <cstdio>
+#include <cstdlib>
 #include <cstring>
 #include <ctime>
-#include <cstdlib>
 #include <iterator>
 #include <map>
 #include <numeric>
 #include <regex>
 #include <sstream>
 #include <string>
-#include <vector>
 #include <thread>
+#include <vector>
 
+#include "common.h"
 #include "ggml.h"
 #include "llama.h"
-#include "common.h"
 
 #ifdef _WIN32
-#define WIN32_LEAN_AND_MEAN
-#ifndef NOMINMAX
-#   define NOMINMAX
-#endif
-#include <windows.h>
+#    define WIN32_LEAN_AND_MEAN
+#    ifndef NOMINMAX
+#        define NOMINMAX
+#    endif
+#    include <windows.h>
 #endif
 
 // utils
@@ -36,8 +36,7 @@ static uint64_t get_time_ns() {
     return std::chrono::nanoseconds(clock::now().time_since_epoch()).count();
 }
 
-template<class T>
-static std::string join(const std::vector<T> & values, const std::string & delim) {
+template <class T> static std::string join(const std::vector<T> & values, const std::string & delim) {
     std::ostringstream str;
     for (size_t i = 0; i < values.size(); i++) {
         str << values[i];
@@ -48,38 +47,35 @@ static std::string join(const std::vector<T> & values, const std::string & delim
     return str.str();
 }
 
-template<typename T, typename F>
-static std::vector<std::string> transform_to_str(const std::vector<T> & values, F f) {
+template <typename T, typename F> static std::vector<std::string> transform_to_str(const std::vector<T> & values, F f) {
     std::vector<std::string> str_values;
     std::transform(values.begin(), values.end(), std::back_inserter(str_values), f);
     return str_values;
 }
 
-template<typename T>
-static T avg(const std::vector<T> & v) {
+template <typename T> static T avg(const std::vector<T> & v) {
     if (v.empty()) {
         return 0;
     }
     T sum = std::accumulate(v.begin(), v.end(), T(0));
-    return sum / (T)v.size();
+    return sum / (T) v.size();
 }
 
-template<typename T>
-static T stdev(const std::vector<T> & v) {
+template <typename T> static T stdev(const std::vector<T> & v) {
     if (v.size() <= 1) {
         return 0;
     }
-    T mean = avg(v);
+    T mean   = avg(v);
     T sq_sum = std::inner_product(v.begin(), v.end(), v.begin(), T(0));
-    T stdev = std::sqrt(sq_sum / (T)(v.size() - 1) - mean * mean * (T)v.size() / (T)(v.size() - 1));
+    T stdev  = std::sqrt(sq_sum / (T) (v.size() - 1) - mean * mean * (T) v.size() / (T) (v.size() - 1));
     return stdev;
 }
 
 static std::string get_cpu_info() {
     std::vector<std::string> cpu_list;
     for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
-        auto * dev = ggml_backend_dev_get(i);
-        auto dev_type = ggml_backend_dev_type(dev);
+        auto * dev      = ggml_backend_dev_get(i);
+        auto   dev_type = ggml_backend_dev_type(dev);
         if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU || dev_type == GGML_BACKEND_DEVICE_TYPE_ACCEL) {
             cpu_list.push_back(ggml_backend_dev_description(dev));
         }
@@ -90,8 +86,8 @@ static std::string get_cpu_info() {
 static std::string get_gpu_info() {
     std::vector<std::string> gpu_list;
     for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
-        auto * dev = ggml_backend_dev_get(i);
-        auto dev_type = ggml_backend_dev_type(dev);
+        auto * dev      = ggml_backend_dev_get(i);
+        auto   dev_type = ggml_backend_dev_type(dev);
         if (dev_type == GGML_BACKEND_DEVICE_TYPE_GPU) {
             gpu_list.push_back(ggml_backend_dev_description(dev));
         }
@@ -100,17 +96,24 @@ static std::string get_gpu_info() {
 }
 
 // command line params
-enum output_formats {NONE, CSV, JSON, JSONL, MARKDOWN, SQL};
+enum output_formats { NONE, CSV, JSON, JSONL, MARKDOWN, SQL };
 
 static const char * output_format_str(output_formats format) {
     switch (format) {
-        case NONE:     return "none";
-        case CSV:      return "csv";
-        case JSON:     return "json";
-        case JSONL:    return "jsonl";
-        case MARKDOWN: return "md";
-        case SQL:      return "sql";
-        default: GGML_ABORT("invalid output format");
+        case NONE:
+            return "none";
+        case CSV:
+            return "csv";
+        case JSON:
+            return "json";
+        case JSONL:
+            return "jsonl";
+        case MARKDOWN:
+            return "md";
+        case SQL:
+            return "sql";
+        default:
+            GGML_ABORT("invalid output format");
     }
 }
 
@@ -135,10 +138,14 @@ static bool output_format_from_str(const std::string & s, output_formats & forma
 
 static const char * split_mode_str(llama_split_mode mode) {
     switch (mode) {
-        case LLAMA_SPLIT_MODE_NONE:  return "none";
-        case LLAMA_SPLIT_MODE_LAYER: return "layer";
-        case LLAMA_SPLIT_MODE_ROW:   return "row";
-        default: GGML_ABORT("invalid split mode");
+        case LLAMA_SPLIT_MODE_NONE:
+            return "none";
+        case LLAMA_SPLIT_MODE_LAYER:
+            return "layer";
+        case LLAMA_SPLIT_MODE_ROW:
+            return "row";
+        default:
+            GGML_ABORT("invalid split mode");
     }
 }
 
@@ -149,59 +156,59 @@ static std::string pair_str(const std::pair<int, int> & p) {
 }
 
 struct cmd_params {
-    std::vector<std::string> model;
-    std::vector<int> n_prompt;
-    std::vector<int> n_gen;
+    std::vector<std::string>         model;
+    std::vector<int>                 n_prompt;
+    std::vector<int>                 n_gen;
     std::vector<std::pair<int, int>> n_pg;
-    std::vector<int> n_batch;
-    std::vector<int> n_ubatch;
-    std::vector<ggml_type> type_k;
-    std::vector<ggml_type> type_v;
-    std::vector<int> n_threads;
-    std::vector<std::string> cpu_mask;
-    std::vector<bool> cpu_strict;
-    std::vector<int> poll;
-    std::vector<int> n_gpu_layers;
-    std::vector<std::string> rpc_servers;
-    std::vector<llama_split_mode> split_mode;
-    std::vector<int> main_gpu;
-    std::vector<bool> no_kv_offload;
-    std::vector<bool> flash_attn;
-    std::vector<std::vector<float>> tensor_split;
-    std::vector<bool> use_mmap;
-    std::vector<bool> embeddings;
-    ggml_numa_strategy numa;
-    int reps;
-    ggml_sched_priority prio;
-    int delay;
-    bool verbose;
-    bool progress;
-    output_formats output_format;
-    output_formats output_format_stderr;
+    std::vector<int>                 n_batch;
+    std::vector<int>                 n_ubatch;
+    std::vector<ggml_type>           type_k;
+    std::vector<ggml_type>           type_v;
+    std::vector<int>                 n_threads;
+    std::vector<std::string>         cpu_mask;
+    std::vector<bool>                cpu_strict;
+    std::vector<int>                 poll;
+    std::vector<int>                 n_gpu_layers;
+    std::vector<std::string>         rpc_servers;
+    std::vector<llama_split_mode>    split_mode;
+    std::vector<int>                 main_gpu;
+    std::vector<bool>                no_kv_offload;
+    std::vector<bool>                flash_attn;
+    std::vector<std::vector<float>>  tensor_split;
+    std::vector<bool>                use_mmap;
+    std::vector<bool>                embeddings;
+    ggml_numa_strategy               numa;
+    int                              reps;
+    ggml_sched_priority              prio;
+    int                              delay;
+    bool                             verbose;
+    bool                             progress;
+    output_formats                   output_format;
+    output_formats                   output_format_stderr;
 };
 
 static const cmd_params cmd_params_defaults = {
-    /* model                */ {"models/7B/ggml-model-q4_0.gguf"},
-    /* n_prompt             */ {512},
-    /* n_gen                */ {128},
+    /* model                */ { "models/7B/ggml-model-q4_0.gguf" },
+    /* n_prompt             */ { 512 },
+    /* n_gen                */ { 128 },
     /* n_pg                 */ {},
-    /* n_batch              */ {2048},
-    /* n_ubatch             */ {512},
-    /* type_k               */ {GGML_TYPE_F16},
-    /* type_v               */ {GGML_TYPE_F16},
-    /* n_threads            */ {cpu_get_num_math()},
-    /* cpu_mask             */ {"0x0"},
-    /* cpu_strict           */ {false},
-    /* poll                 */ {50},
-    /* n_gpu_layers         */ {99},
-    /* rpc_servers          */ {""},
-    /* split_mode           */ {LLAMA_SPLIT_MODE_LAYER},
-    /* main_gpu             */ {0},
-    /* no_kv_offload        */ {false},
-    /* flash_attn           */ {false},
-    /* tensor_split         */ {std::vector<float>(llama_max_devices(), 0.0f)},
-    /* use_mmap             */ {true},
-    /* embeddings           */ {false},
+    /* n_batch              */ { 2048 },
+    /* n_ubatch             */ { 512 },
+    /* type_k               */ { GGML_TYPE_F16 },
+    /* type_v               */ { GGML_TYPE_F16 },
+    /* n_threads            */ { cpu_get_num_math() },
+    /* cpu_mask             */ { "0x0" },
+    /* cpu_strict           */ { false },
+    /* poll                 */ { 50 },
+    /* n_gpu_layers         */ { 99 },
+    /* rpc_servers          */ { "" },
+    /* split_mode           */ { LLAMA_SPLIT_MODE_LAYER },
+    /* main_gpu             */ { 0 },
+    /* no_kv_offload        */ { false },
+    /* flash_attn           */ { false },
+    /* tensor_split         */ { std::vector<float>(llama_max_devices(), 0.0f) },
+    /* use_mmap             */ { true },
+    /* embeddings           */ { false },
     /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
     /* reps                 */ 5,
     /* prio                 */ GGML_SCHED_PRIO_NORMAL,
@@ -218,38 +225,59 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("options:\n");
     printf("  -h, --help\n");
     printf("  -m, --model <filename>                    (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
-    printf("  -p, --n-prompt <n>                        (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
+    printf("  -p, --n-prompt <n>                        (default: %s)\n",
+           join(cmd_params_defaults.n_prompt, ",").c_str());
     printf("  -n, --n-gen <n>                           (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
-    printf("  -pg <pp,tg>                               (default: %s)\n", join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str());
-    printf("  -b, --batch-size <n>                      (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    printf("  -ub, --ubatch-size <n>                    (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str());
-    printf("  -ctk, --cache-type-k <t>                  (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
-    printf("  -ctv, --cache-type-v <t>                  (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
-    printf("  -t, --threads <n>                         (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
-    printf("  -C, --cpu-mask <hex,hex>                  (default: %s)\n", join(cmd_params_defaults.cpu_mask, ",").c_str());
-    printf("  --cpu-strict <0|1>                        (default: %s)\n", join(cmd_params_defaults.cpu_strict, ",").c_str());
+    printf("  -pg <pp,tg>                               (default: %s)\n",
+           join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str());
+    printf("  -b, --batch-size <n>                      (default: %s)\n",
+           join(cmd_params_defaults.n_batch, ",").c_str());
+    printf("  -ub, --ubatch-size <n>                    (default: %s)\n",
+           join(cmd_params_defaults.n_ubatch, ",").c_str());
+    printf("  -ctk, --cache-type-k <t>                  (default: %s)\n",
+           join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
+    printf("  -ctv, --cache-type-v <t>                  (default: %s)\n",
+           join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
+    printf("  -t, --threads <n>                         (default: %s)\n",
+           join(cmd_params_defaults.n_threads, ",").c_str());
+    printf("  -C, --cpu-mask <hex,hex>                  (default: %s)\n",
+           join(cmd_params_defaults.cpu_mask, ",").c_str());
+    printf("  --cpu-strict <0|1>                        (default: %s)\n",
+           join(cmd_params_defaults.cpu_strict, ",").c_str());
     printf("  --poll <0...100>                          (default: %s)\n", join(cmd_params_defaults.poll, ",").c_str());
-    printf("  -ngl, --n-gpu-layers <n>                  (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+    printf("  -ngl, --n-gpu-layers <n>                  (default: %s)\n",
+           join(cmd_params_defaults.n_gpu_layers, ",").c_str());
     if (llama_supports_rpc()) {
-        printf("  -rpc, --rpc <rpc_servers>                 (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str());
+        printf("  -rpc, --rpc <rpc_servers>                 (default: %s)\n",
+               join(cmd_params_defaults.rpc_servers, ",").c_str());
     }
-    printf("  -sm, --split-mode <none|layer|row>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
-    printf("  -mg, --main-gpu <i>                       (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
-    printf("  -nkvo, --no-kv-offload <0|1>              (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
-    printf("  -fa, --flash-attn <0|1>                   (default: %s)\n", join(cmd_params_defaults.flash_attn, ",").c_str());
-    printf("  -mmp, --mmap <0|1>                        (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
+    printf("  -sm, --split-mode <none|layer|row>        (default: %s)\n",
+           join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
+    printf("  -mg, --main-gpu <i>                       (default: %s)\n",
+           join(cmd_params_defaults.main_gpu, ",").c_str());
+    printf("  -nkvo, --no-kv-offload <0|1>              (default: %s)\n",
+           join(cmd_params_defaults.no_kv_offload, ",").c_str());
+    printf("  -fa, --flash-attn <0|1>                   (default: %s)\n",
+           join(cmd_params_defaults.flash_attn, ",").c_str());
+    printf("  -mmp, --mmap <0|1>                        (default: %s)\n",
+           join(cmd_params_defaults.use_mmap, ",").c_str());
     printf("  --numa <distribute|isolate|numactl>       (default: disabled)\n");
-    printf("  -embd, --embeddings <0|1>                 (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
+    printf("  -embd, --embeddings <0|1>                 (default: %s)\n",
+           join(cmd_params_defaults.embeddings, ",").c_str());
     printf("  -ts, --tensor-split <ts0/ts1/..>          (default: 0)\n");
     printf("  -r, --repetitions <n>                     (default: %d)\n", cmd_params_defaults.reps);
     printf("  --prio <0|1|2|3>                          (default: %d)\n", cmd_params_defaults.prio);
     printf("  --delay <0...N> (seconds)                 (default: %d)\n", cmd_params_defaults.delay);
-    printf("  -o, --output <csv|json|jsonl|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
-    printf("  -oe, --output-err <csv|json|jsonl|md|sql> (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr));
+    printf("  -o, --output <csv|json|jsonl|md|sql>      (default: %s)\n",
+           output_format_str(cmd_params_defaults.output_format));
+    printf("  -oe, --output-err <csv|json|jsonl|md|sql> (default: %s)\n",
+           output_format_str(cmd_params_defaults.output_format_stderr));
     printf("  -v, --verbose                             (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
     printf("  --progress                                (default: %s)\n", cmd_params_defaults.progress ? "1" : "0");
     printf("\n");
-    printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
+    printf(
+        "Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter "
+        "multiple times.\n");
 }
 
 static ggml_type ggml_type_from_name(const std::string & s) {
@@ -281,22 +309,21 @@ static ggml_type ggml_type_from_name(const std::string & s) {
     return GGML_TYPE_COUNT;
 }
 
-
 static cmd_params parse_cmd_params(int argc, char ** argv) {
-    cmd_params params;
-    std::string arg;
-    bool invalid_param = false;
-    const std::string arg_prefix = "--";
-    const char split_delim = ',';
-
-    params.verbose = cmd_params_defaults.verbose;
-    params.output_format = cmd_params_defaults.output_format;
+    cmd_params        params;
+    std::string       arg;
+    bool              invalid_param = false;
+    const std::string arg_prefix    = "--";
+    const char        split_delim   = ',';
+
+    params.verbose              = cmd_params_defaults.verbose;
+    params.output_format        = cmd_params_defaults.output_format;
     params.output_format_stderr = cmd_params_defaults.output_format_stderr;
-    params.reps = cmd_params_defaults.reps;
-    params.numa = cmd_params_defaults.numa;
-    params.prio = cmd_params_defaults.prio;
-    params.delay = cmd_params_defaults.delay;
-    params.progress = cmd_params_defaults.progress;
+    params.reps                 = cmd_params_defaults.reps;
+    params.numa                 = cmd_params_defaults.numa;
+    params.prio                 = cmd_params_defaults.prio;
+    params.delay                = cmd_params_defaults.delay;
+    params.progress             = cmd_params_defaults.progress;
 
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
@@ -338,7 +365,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            params.n_pg.push_back({std::stoi(p[0]), std::stoi(p[1])});
+            params.n_pg.push_back({ std::stoi(p[0]), std::stoi(p[1]) });
         } else if (arg == "-b" || arg == "--batch-size") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -358,7 +385,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            auto p = string_split<std::string>(argv[i], split_delim);
+            auto                   p = string_split<std::string>(argv[i], split_delim);
             std::vector<ggml_type> types;
             for (const auto & t : p) {
                 ggml_type gt = ggml_type_from_name(t);
@@ -377,7 +404,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            auto p = string_split<std::string>(argv[i], split_delim);
+            auto                   p = string_split<std::string>(argv[i], split_delim);
             std::vector<ggml_type> types;
             for (const auto & t : p) {
                 ggml_type gt = ggml_type_from_name(t);
@@ -437,7 +464,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 invalid_param = true;
                 break;
             }
-            auto p = string_split<std::string>(argv[i], split_delim);
+            auto                          p = string_split<std::string>(argv[i], split_delim);
             std::vector<llama_split_mode> modes;
             for (const auto & m : p) {
                 llama_split_mode mode;
@@ -476,10 +503,16 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 break;
             } else {
                 std::string value(argv[i]);
-                /**/ if (value == "distribute" || value == "" ) { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
-                else if (value == "isolate")                    { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
-                else if (value == "numactl")                    { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
-                else { invalid_param = true; break; }
+                /**/ if (value == "distribute" || value == "") {
+                    params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE;
+                } else if (value == "isolate") {
+                    params.numa = GGML_NUMA_STRATEGY_ISOLATE;
+                } else if (value == "numactl") {
+                    params.numa = GGML_NUMA_STRATEGY_NUMACTL;
+                } else {
+                    invalid_param = true;
+                    break;
+                }
             }
         } else if (arg == "-fa" || arg == "--flash-attn") {
             if (++i >= argc) {
@@ -509,9 +542,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             for (auto ts : string_split<std::string>(argv[i], split_delim)) {
                 // split string by ; and /
-                const std::regex regex{R"([;/]+)"};
-                std::sregex_token_iterator it{ts.begin(), ts.end(), regex, -1};
-                std::vector<std::string> split_arg{it, {}};
+                const std::regex           regex{ R"([;/]+)" };
+                std::sregex_token_iterator it{ ts.begin(), ts.end(), regex, -1 };
+                std::vector<std::string>   split_arg{ it, {} };
                 GGML_ASSERT(split_arg.size() <= llama_max_devices());
 
                 std::vector<float> tensor_split(llama_max_devices());
@@ -570,52 +603,94 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     }
 
     // set defaults
-    if (params.model.empty())        { params.model = cmd_params_defaults.model; }
-    if (params.n_prompt.empty())     { params.n_prompt = cmd_params_defaults.n_prompt; }
-    if (params.n_gen.empty())        { params.n_gen = cmd_params_defaults.n_gen; }
-    if (params.n_pg.empty())         { params.n_pg = cmd_params_defaults.n_pg; }
-    if (params.n_batch.empty())      { params.n_batch = cmd_params_defaults.n_batch; }
-    if (params.n_ubatch.empty())     { params.n_ubatch = cmd_params_defaults.n_ubatch; }
-    if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
-    if (params.type_v.empty())       { params.type_v = cmd_params_defaults.type_v; }
-    if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
-    if (params.rpc_servers.empty())  { params.rpc_servers = cmd_params_defaults.rpc_servers; }
-    if (params.split_mode.empty())   { params.split_mode = cmd_params_defaults.split_mode; }
-    if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
-    if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; }
-    if (params.flash_attn.empty())   { params.flash_attn = cmd_params_defaults.flash_attn; }
-    if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; }
-    if (params.use_mmap.empty())     { params.use_mmap = cmd_params_defaults.use_mmap; }
-    if (params.embeddings.empty())   { params.embeddings = cmd_params_defaults.embeddings; }
-    if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
-    if (params.cpu_mask.empty())     { params.cpu_mask  = cmd_params_defaults.cpu_mask;  }
-    if (params.cpu_strict.empty())   { params.cpu_strict = cmd_params_defaults.cpu_strict; }
-    if (params.poll.empty())         { params.poll = cmd_params_defaults.poll; }
+    if (params.model.empty()) {
+        params.model = cmd_params_defaults.model;
+    }
+    if (params.n_prompt.empty()) {
+        params.n_prompt = cmd_params_defaults.n_prompt;
+    }
+    if (params.n_gen.empty()) {
+        params.n_gen = cmd_params_defaults.n_gen;
+    }
+    if (params.n_pg.empty()) {
+        params.n_pg = cmd_params_defaults.n_pg;
+    }
+    if (params.n_batch.empty()) {
+        params.n_batch = cmd_params_defaults.n_batch;
+    }
+    if (params.n_ubatch.empty()) {
+        params.n_ubatch = cmd_params_defaults.n_ubatch;
+    }
+    if (params.type_k.empty()) {
+        params.type_k = cmd_params_defaults.type_k;
+    }
+    if (params.type_v.empty()) {
+        params.type_v = cmd_params_defaults.type_v;
+    }
+    if (params.n_gpu_layers.empty()) {
+        params.n_gpu_layers = cmd_params_defaults.n_gpu_layers;
+    }
+    if (params.rpc_servers.empty()) {
+        params.rpc_servers = cmd_params_defaults.rpc_servers;
+    }
+    if (params.split_mode.empty()) {
+        params.split_mode = cmd_params_defaults.split_mode;
+    }
+    if (params.main_gpu.empty()) {
+        params.main_gpu = cmd_params_defaults.main_gpu;
+    }
+    if (params.no_kv_offload.empty()) {
+        params.no_kv_offload = cmd_params_defaults.no_kv_offload;
+    }
+    if (params.flash_attn.empty()) {
+        params.flash_attn = cmd_params_defaults.flash_attn;
+    }
+    if (params.tensor_split.empty()) {
+        params.tensor_split = cmd_params_defaults.tensor_split;
+    }
+    if (params.use_mmap.empty()) {
+        params.use_mmap = cmd_params_defaults.use_mmap;
+    }
+    if (params.embeddings.empty()) {
+        params.embeddings = cmd_params_defaults.embeddings;
+    }
+    if (params.n_threads.empty()) {
+        params.n_threads = cmd_params_defaults.n_threads;
+    }
+    if (params.cpu_mask.empty()) {
+        params.cpu_mask = cmd_params_defaults.cpu_mask;
+    }
+    if (params.cpu_strict.empty()) {
+        params.cpu_strict = cmd_params_defaults.cpu_strict;
+    }
+    if (params.poll.empty()) {
+        params.poll = cmd_params_defaults.poll;
+    }
 
     return params;
 }
 
 struct cmd_params_instance {
-    std::string model;
-    int n_prompt;
-    int n_gen;
-    int n_batch;
-    int n_ubatch;
-    ggml_type type_k;
-    ggml_type type_v;
-    int n_threads;
-    std::string cpu_mask;
-    bool cpu_strict;
-    int poll;
-    int n_gpu_layers;
-    std::string rpc_servers;
-    llama_split_mode split_mode;
-    int main_gpu;
-    bool no_kv_offload;
-    bool flash_attn;
+    std::string        model;
+    int                n_prompt;
+    int                n_gen;
+    int                n_batch;
+    int                n_ubatch;
+    ggml_type          type_k;
+    ggml_type          type_v;
+    int                n_threads;
+    std::string        cpu_mask;
+    bool               cpu_strict;
+    int                poll;
+    int                n_gpu_layers;
+    std::string        rpc_servers;
+    llama_split_mode   split_mode;
+    int                main_gpu;
+    bool               no_kv_offload;
+    bool               flash_attn;
     std::vector<float> tensor_split;
-    bool use_mmap;
-    bool embeddings;
+    bool               use_mmap;
+    bool               embeddings;
 
     llama_model_params to_llama_mparams() const {
         llama_model_params mparams = llama_model_default_params();
@@ -624,35 +699,31 @@ struct cmd_params_instance {
         if (!rpc_servers.empty()) {
             mparams.rpc_servers = rpc_servers.c_str();
         }
-        mparams.split_mode = split_mode;
-        mparams.main_gpu = main_gpu;
+        mparams.split_mode   = split_mode;
+        mparams.main_gpu     = main_gpu;
         mparams.tensor_split = tensor_split.data();
-        mparams.use_mmap = use_mmap;
+        mparams.use_mmap     = use_mmap;
 
         return mparams;
     }
 
     bool equal_mparams(const cmd_params_instance & other) const {
-        return model == other.model &&
-               n_gpu_layers == other.n_gpu_layers &&
-               rpc_servers == other.rpc_servers &&
-               split_mode == other.split_mode &&
-               main_gpu == other.main_gpu &&
-               use_mmap == other.use_mmap &&
+        return model == other.model && n_gpu_layers == other.n_gpu_layers && rpc_servers == other.rpc_servers &&
+               split_mode == other.split_mode && main_gpu == other.main_gpu && use_mmap == other.use_mmap &&
                tensor_split == other.tensor_split;
     }
 
     llama_context_params to_llama_cparams() const {
         llama_context_params cparams = llama_context_default_params();
 
-        cparams.n_ctx = n_prompt + n_gen;
-        cparams.n_batch = n_batch;
-        cparams.n_ubatch = n_ubatch;
-        cparams.type_k = type_k;
-        cparams.type_v = type_v;
+        cparams.n_ctx       = n_prompt + n_gen;
+        cparams.n_batch     = n_batch;
+        cparams.n_ubatch    = n_ubatch;
+        cparams.type_k      = type_k;
+        cparams.type_v      = type_v;
         cparams.offload_kqv = !no_kv_offload;
-        cparams.flash_attn = flash_attn;
-        cparams.embeddings = embeddings;
+        cparams.flash_attn  = flash_attn;
+        cparams.embeddings  = embeddings;
 
         return cparams;
     }
@@ -662,6 +733,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     std::vector<cmd_params_instance> instances;
 
     // this ordering minimizes the number of times that each model needs to be reloaded
+    // clang-format off
     for (const auto & m : params.model)
     for (const auto & nl : params.n_gpu_layers)
     for (const auto & rpc : params.rpc_servers)
@@ -767,100 +839,94 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
             instances.push_back(instance);
         }
     }
+    // clang-format on
 
     return instances;
 }
 
 struct test {
     static const std::string build_commit;
-    static const int build_number;
+    static const int         build_number;
     static const std::string cpu_info;
     static const std::string gpu_info;
-    std::string model_filename;
-    std::string model_type;
-    uint64_t model_size;
-    uint64_t model_n_params;
-    int n_batch;
-    int n_ubatch;
-    int n_threads;
-    std::string cpu_mask;
-    bool cpu_strict;
-    int poll;
-    ggml_type type_k;
-    ggml_type type_v;
-    int n_gpu_layers;
-    llama_split_mode split_mode;
-    int main_gpu;
-    bool no_kv_offload;
-    bool flash_attn;
-    std::vector<float> tensor_split;
-    bool use_mmap;
-    bool embeddings;
-    int n_prompt;
-    int n_gen;
-    std::string test_time;
-    std::vector<uint64_t> samples_ns;
+    std::string              model_filename;
+    std::string              model_type;
+    uint64_t                 model_size;
+    uint64_t                 model_n_params;
+    int                      n_batch;
+    int                      n_ubatch;
+    int                      n_threads;
+    std::string              cpu_mask;
+    bool                     cpu_strict;
+    int                      poll;
+    ggml_type                type_k;
+    ggml_type                type_v;
+    int                      n_gpu_layers;
+    llama_split_mode         split_mode;
+    int                      main_gpu;
+    bool                     no_kv_offload;
+    bool                     flash_attn;
+    std::vector<float>       tensor_split;
+    bool                     use_mmap;
+    bool                     embeddings;
+    int                      n_prompt;
+    int                      n_gen;
+    std::string              test_time;
+    std::vector<uint64_t>    samples_ns;
 
     test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) {
         model_filename = inst.model;
         char buf[128];
         llama_model_desc(lmodel, buf, sizeof(buf));
-        model_type = buf;
-        model_size = llama_model_size(lmodel);
+        model_type     = buf;
+        model_size     = llama_model_size(lmodel);
         model_n_params = llama_model_n_params(lmodel);
-        n_batch = inst.n_batch;
-        n_ubatch = inst.n_ubatch;
-        n_threads = inst.n_threads;
-        cpu_mask = inst.cpu_mask;
-        cpu_strict = inst.cpu_strict;
-        poll = inst.poll;
-        type_k = inst.type_k;
-        type_v = inst.type_v;
-        n_gpu_layers = inst.n_gpu_layers;
-        split_mode = inst.split_mode;
-        main_gpu = inst.main_gpu;
-        no_kv_offload = inst.no_kv_offload;
-        flash_attn = inst.flash_attn;
-        tensor_split = inst.tensor_split;
-        use_mmap = inst.use_mmap;
-        embeddings = inst.embeddings;
-        n_prompt = inst.n_prompt;
-        n_gen = inst.n_gen;
+        n_batch        = inst.n_batch;
+        n_ubatch       = inst.n_ubatch;
+        n_threads      = inst.n_threads;
+        cpu_mask       = inst.cpu_mask;
+        cpu_strict     = inst.cpu_strict;
+        poll           = inst.poll;
+        type_k         = inst.type_k;
+        type_v         = inst.type_v;
+        n_gpu_layers   = inst.n_gpu_layers;
+        split_mode     = inst.split_mode;
+        main_gpu       = inst.main_gpu;
+        no_kv_offload  = inst.no_kv_offload;
+        flash_attn     = inst.flash_attn;
+        tensor_split   = inst.tensor_split;
+        use_mmap       = inst.use_mmap;
+        embeddings     = inst.embeddings;
+        n_prompt       = inst.n_prompt;
+        n_gen          = inst.n_gen;
         // RFC 3339 date-time format
-        time_t t = time(NULL);
+        time_t t       = time(NULL);
         std::strftime(buf, sizeof(buf), "%FT%TZ", gmtime(&t));
         test_time = buf;
 
         (void) ctx;
     }
 
-    uint64_t avg_ns() const {
-        return ::avg(samples_ns);
-    }
+    uint64_t avg_ns() const { return ::avg(samples_ns); }
 
-    uint64_t stdev_ns() const {
-        return ::stdev(samples_ns);
-    }
+    uint64_t stdev_ns() const { return ::stdev(samples_ns); }
 
     std::vector<double> get_ts() const {
-        int n_tokens = n_prompt + n_gen;
+        int                 n_tokens = n_prompt + n_gen;
         std::vector<double> ts;
-        std::transform(samples_ns.begin(), samples_ns.end(), std::back_inserter(ts), [n_tokens](uint64_t t) { return 1e9 * n_tokens / t; });
+        std::transform(samples_ns.begin(), samples_ns.end(), std::back_inserter(ts),
+                       [n_tokens](uint64_t t) { return 1e9 * n_tokens / t; });
         return ts;
     }
 
-    double avg_ts() const {
-        return ::avg(get_ts());
-    }
+    double avg_ts() const { return ::avg(get_ts()); }
 
-    double stdev_ts() const {
-        return ::stdev(get_ts());
-    }
+    double stdev_ts() const { return ::stdev(get_ts()); }
 
     static std::string get_backend() {
         std::vector<std::string> backends;
         for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
-            auto * reg = ggml_backend_reg_get(i);
+            auto *      reg  = ggml_backend_reg_get(i);
             std::string name = ggml_backend_reg_name(reg);
             if (name != "CPU") {
                 backends.push_back(ggml_backend_reg_name(reg));
@@ -871,36 +937,27 @@ struct test {
 
     static const std::vector<std::string> & get_fields() {
         static const std::vector<std::string> fields = {
-            "build_commit", "build_number",
-            "cpu_info", "gpu_info", "backends",
-            "model_filename", "model_type", "model_size", "model_n_params",
-            "n_batch", "n_ubatch",
-            "n_threads", "cpu_mask", "cpu_strict", "poll",
-            "type_k", "type_v",
-            "n_gpu_layers", "split_mode",
-            "main_gpu", "no_kv_offload", "flash_attn",
-            "tensor_split", "use_mmap", "embeddings",
-            "n_prompt", "n_gen", "test_time",
-            "avg_ns", "stddev_ns",
-            "avg_ts", "stddev_ts",
+            "build_commit", "build_number", "cpu_info",       "gpu_info",   "backends",     "model_filename",
+            "model_type",   "model_size",   "model_n_params", "n_batch",    "n_ubatch",     "n_threads",
+            "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
+            "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "use_mmap",
+            "embeddings",   "n_prompt",     "n_gen",          "test_time",  "avg_ns",       "stddev_ns",
+            "avg_ts",       "stddev_ts",
         };
         return fields;
     }
 
-    enum field_type {STRING, BOOL, INT, FLOAT};
+    enum field_type { STRING, BOOL, INT, FLOAT };
 
     static field_type get_field_type(const std::string & field) {
-        if (field == "build_number" || field == "n_batch" || field == "n_ubatch" ||
-            field == "n_threads" || field == "poll" ||
-            field == "model_size" || field == "model_n_params" ||
-            field == "n_gpu_layers" || field == "main_gpu" ||
-            field == "n_prompt" || field == "n_gen" ||
-            field == "avg_ns" || field == "stddev_ns") {
+        if (field == "build_number" || field == "n_batch" || field == "n_ubatch" || field == "n_threads" ||
+            field == "poll" || field == "model_size" || field == "model_n_params" || field == "n_gpu_layers" ||
+            field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "avg_ns" ||
+            field == "stddev_ns") {
             return INT;
         }
-        if (field == "f16_kv" || field == "no_kv_offload" ||
-            field == "cpu_strict" ||
-            field == "flash_attn" || field == "use_mmap" || field == "embeddings") {
+        if (field == "f16_kv" || field == "no_kv_offload" || field == "cpu_strict" || field == "flash_attn" ||
+            field == "use_mmap" || field == "embeddings") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -911,7 +968,7 @@ struct test {
 
     std::vector<std::string> get_values() const {
         std::string tensor_split_str;
-        int max_nonzero = 0;
+        int         max_nonzero = 0;
         for (size_t i = 0; i < llama_max_devices(); i++) {
             if (tensor_split[i] > 0) {
                 max_nonzero = i;
@@ -925,29 +982,47 @@ struct test {
                 tensor_split_str += "/";
             }
         }
-        std::vector<std::string> values = {
-            build_commit, std::to_string(build_number),
-            cpu_info, gpu_info, get_backend(),
-            model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
-            std::to_string(n_batch), std::to_string(n_ubatch),
-            std::to_string(n_threads), cpu_mask, std::to_string(cpu_strict), std::to_string(poll),
-            ggml_type_name(type_k), ggml_type_name(type_v),
-            std::to_string(n_gpu_layers), split_mode_str(split_mode),
-            std::to_string(main_gpu), std::to_string(no_kv_offload), std::to_string(flash_attn),
-            tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
-            std::to_string(n_prompt), std::to_string(n_gen), test_time,
-            std::to_string(avg_ns()), std::to_string(stdev_ns()),
-            std::to_string(avg_ts()), std::to_string(stdev_ts())
-        };
+        std::vector<std::string> values = { build_commit,
+                                            std::to_string(build_number),
+                                            cpu_info,
+                                            gpu_info,
+                                            get_backend(),
+                                            model_filename,
+                                            model_type,
+                                            std::to_string(model_size),
+                                            std::to_string(model_n_params),
+                                            std::to_string(n_batch),
+                                            std::to_string(n_ubatch),
+                                            std::to_string(n_threads),
+                                            cpu_mask,
+                                            std::to_string(cpu_strict),
+                                            std::to_string(poll),
+                                            ggml_type_name(type_k),
+                                            ggml_type_name(type_v),
+                                            std::to_string(n_gpu_layers),
+                                            split_mode_str(split_mode),
+                                            std::to_string(main_gpu),
+                                            std::to_string(no_kv_offload),
+                                            std::to_string(flash_attn),
+                                            tensor_split_str,
+                                            std::to_string(use_mmap),
+                                            std::to_string(embeddings),
+                                            std::to_string(n_prompt),
+                                            std::to_string(n_gen),
+                                            test_time,
+                                            std::to_string(avg_ns()),
+                                            std::to_string(stdev_ns()),
+                                            std::to_string(avg_ts()),
+                                            std::to_string(stdev_ts()) };
         return values;
     }
 
     std::map<std::string, std::string> get_map() const {
         std::map<std::string, std::string> map;
-        auto fields = get_fields();
-        auto values = get_values();
-        std::transform(fields.begin(), fields.end(), values.begin(),
-                std::inserter(map, map.end()), std::make_pair<const std::string &, const std::string &>);
+        auto                               fields = get_fields();
+        auto                               values = get_values();
+        std::transform(fields.begin(), fields.end(), values.begin(), std::inserter(map, map.end()),
+                       std::make_pair<const std::string &, const std::string &>);
         return map;
     }
 };
@@ -961,9 +1036,12 @@ struct printer {
     virtual ~printer() {}
 
     FILE * fout;
+
     virtual void print_header(const cmd_params & params) { (void) params; }
+
     virtual void print_test(const test & t) = 0;
-    virtual void print_footer() { }
+
+    virtual void print_footer() {}
 };
 
 struct csv_printer : public printer {
@@ -979,7 +1057,7 @@ struct csv_printer : public printer {
         return escaped;
     }
 
-    void print_header(const cmd_params & params) override  {
+    void print_header(const cmd_params & params) override {
         std::vector<std::string> fields = test::get_fields();
         fprintf(fout, "%s\n", join(fields, ",").c_str());
         (void) params;
@@ -992,7 +1070,6 @@ struct csv_printer : public printer {
     }
 };
 
-
 static std::string escape_json(const std::string & value) {
     std::string escaped;
     for (auto c : value) {
@@ -1000,7 +1077,7 @@ static std::string escape_json(const std::string & value) {
             escaped += "\\\"";
         } else if (c == '\\') {
             escaped += "\\\\";
-        } else  if (c <= 0x1f) {
+        } else if (c <= 0x1f) {
             char buf[8];
             snprintf(buf, sizeof(buf), "\\u%04x", c);
             escaped += buf;
@@ -1033,7 +1110,8 @@ struct json_printer : public printer {
     void print_fields(const std::vector<std::string> & fields, const std::vector<std::string> & values) {
         assert(fields.size() == values.size());
         for (size_t i = 0; i < fields.size(); i++) {
-            fprintf(fout, "    \"%s\": %s,\n", fields.at(i).c_str(), format_json_value(fields.at(i), values.at(i)).c_str());
+            fprintf(fout, "    \"%s\": %s,\n", fields.at(i).c_str(),
+                    format_json_value(fields.at(i), values.at(i)).c_str());
         }
     }
 
@@ -1051,12 +1129,9 @@ struct json_printer : public printer {
         fflush(fout);
     }
 
-    void print_footer() override {
-        fprintf(fout, "\n]\n");
-    }
+    void print_footer() override { fprintf(fout, "\n]\n"); }
 };
 
-
 struct jsonl_printer : public printer {
     void print_fields(const std::vector<std::string> & fields, const std::vector<std::string> & values) {
         assert(fields.size() == values.size());
@@ -1116,7 +1191,7 @@ struct markdown_printer : public printer {
             return 13;
         }
 
-        int width = std::max((int)field.length(), 10);
+        int width = std::max((int) field.length(), 10);
 
         if (test::get_field_type(field) == test::STRING) {
             return -width;
@@ -1230,18 +1305,18 @@ struct markdown_printer : public printer {
         fprintf(fout, "|");
         for (const auto & field : fields) {
             std::string value;
-            char buf[128];
+            char        buf[128];
             if (field == "model") {
                 value = t.model_type;
             } else if (field == "size") {
-                if (t.model_size < 1024*1024*1024) {
+                if (t.model_size < 1024 * 1024 * 1024) {
                     snprintf(buf, sizeof(buf), "%.2f MiB", t.model_size / 1024.0 / 1024.0);
                 } else {
                     snprintf(buf, sizeof(buf), "%.2f GiB", t.model_size / 1024.0 / 1024.0 / 1024.0);
                 }
                 value = buf;
             } else if (field == "params") {
-                if (t.model_n_params < 1000*1000*1000) {
+                if (t.model_n_params < 1000 * 1000 * 1000) {
                     snprintf(buf, sizeof(buf), "%.2f M", t.model_n_params / 1e6);
                 } else {
                     snprintf(buf, sizeof(buf), "%.2f B", t.model_n_params / 1e9);
@@ -1303,7 +1378,8 @@ struct sql_printer : public printer {
         std::vector<std::string> fields = test::get_fields();
         fprintf(fout, "CREATE TABLE IF NOT EXISTS test (\n");
         for (size_t i = 0; i < fields.size(); i++) {
-            fprintf(fout, "  %s %s%s\n", fields.at(i).c_str(), get_sql_field_type(fields.at(i)).c_str(),  i < fields.size() - 1 ? "," : "");
+            fprintf(fout, "  %s %s%s\n", fields.at(i).c_str(), get_sql_field_type(fields.at(i)).c_str(),
+                    i < fields.size() - 1 ? "," : "");
         }
         fprintf(fout, ");\n");
         fprintf(fout, "\n");
@@ -1324,8 +1400,8 @@ struct sql_printer : public printer {
 static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_threads) {
     llama_set_n_threads(ctx, n_threads, n_threads);
 
-    const llama_model * model = llama_get_model(ctx);
-    const int32_t n_vocab = llama_n_vocab(model);
+    const llama_model * model   = llama_get_model(ctx);
+    const int32_t       n_vocab = llama_n_vocab(model);
 
     std::vector<llama_token> tokens(n_batch);
 
@@ -1333,7 +1409,7 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_th
 
     while (n_processed < n_prompt) {
         int n_tokens = std::min(n_prompt - n_processed, n_batch);
-        tokens[0] = n_processed == 0 && llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
+        tokens[0]    = n_processed == 0 && llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
         for (int i = 1; i < n_tokens; i++) {
             tokens[i] = std::rand() % n_vocab;
         }
@@ -1347,8 +1423,8 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_th
 static void test_gen(llama_context * ctx, int n_gen, int n_threads) {
     llama_set_n_threads(ctx, n_threads, n_threads);
 
-    const llama_model * model = llama_get_model(ctx);
-    const int32_t n_vocab = llama_n_vocab(model);
+    const llama_model * model   = llama_get_model(ctx);
+    const int32_t       n_vocab = llama_n_vocab(model);
 
     llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
 
@@ -1411,7 +1487,7 @@ int main(int argc, char ** argv) {
     set_process_priority(params.prio);
 
     // initialize printer
-    std::unique_ptr<printer> p = create_printer(params.output_format);
+    std::unique_ptr<printer> p     = create_printer(params.output_format);
     std::unique_ptr<printer> p_err = create_printer(params.output_format_stderr);
 
     if (p) {
@@ -1426,13 +1502,13 @@ int main(int argc, char ** argv) {
 
     std::vector<cmd_params_instance> params_instances = get_cmd_params_instances(params);
 
-    llama_model * lmodel = nullptr;
+    llama_model *               lmodel    = nullptr;
     const cmd_params_instance * prev_inst = nullptr;
 
-    int params_idx = 0;
+    int  params_idx   = 0;
     auto params_count = params_instances.size();
     for (const auto & inst : params_instances) {
-        params_idx ++;
+        params_idx++;
         if (params.progress) {
             fprintf(stderr, "llama-bench: benchmark %d/%ld: starting\n", params_idx, params_count);
         }
@@ -1475,7 +1551,7 @@ int main(int argc, char ** argv) {
         tpp.poll       = t.poll;
         tpp.prio       = params.prio;
 
-        struct ggml_threadpool* threadpool = ggml_threadpool_new(&tpp);
+        struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp);
         if (!threadpool) {
             fprintf(stderr, "%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
             exit(1);
@@ -1505,13 +1581,15 @@ int main(int argc, char ** argv) {
 
             if (t.n_prompt > 0) {
                 if (params.progress) {
-                    fprintf(stderr, "llama-bench: benchmark %d/%ld: prompt run %d/%d\n", params_idx, params_count, i + 1, params.reps);
+                    fprintf(stderr, "llama-bench: benchmark %d/%ld: prompt run %d/%d\n", params_idx, params_count,
+                            i + 1, params.reps);
                 }
                 test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
             }
             if (t.n_gen > 0) {
                 if (params.progress) {
-                    fprintf(stderr, "llama-bench: benchmark %d/%ld: generation run %d/%d\n", params_idx, params_count, i + 1, params.reps);
+                    fprintf(stderr, "llama-bench: benchmark %d/%ld: generation run %d/%d\n", params_idx, params_count,
+                            i + 1, params.reps);
                 }
                 test_gen(ctx, t.n_gen, t.n_threads);
             }

From f95caa79546271722ada703da20ffb1cfcd21fed Mon Sep 17 00:00:00 2001
From: bandoti <141645996+bandoti@users.noreply.github.com>
Date: Wed, 20 Nov 2024 12:22:19 -0400
Subject: [PATCH 046/102] cmake: add link dependencies to cmake find pkg
 (#10433)

* cmake pkg: find accelerate, openmp, memkind libs

* cmake pkg: find BLAS libs

* try BLAS_LIBRARIES instead

* Add BLAS link opts

* Add more link deps. and set GGML_ vars
---
 cmake/llama-config.cmake.in | 136 +++++++++++++++++++++++++++---------
 1 file changed, 104 insertions(+), 32 deletions(-)

diff --git a/cmake/llama-config.cmake.in b/cmake/llama-config.cmake.in
index 28a8c18b65176..5c55bc6b822a6 100644
--- a/cmake/llama-config.cmake.in
+++ b/cmake/llama-config.cmake.in
@@ -3,12 +3,60 @@ set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
 set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
 set(LLAMA_SHARED_LIB   @BUILD_SHARED_LIBS@)
 
+set(GGML_STATIC @GGML_STATIC@)
+set(GGML_NATIVE @GGML_NATIVE@)
+set(GGML_LTO    @GGML_LTO@)
+set(GGML_CCACHE @GGML_CCACHE@)
+set(GGML_AVX    @GGML_AVX@)
+set(GGML_AVX2   @GGML_AVX2@)
+set(GGML_AVX512 @GGML_AVX512@)
+set(GGML_AVX512_VBMI @GGML_AVX512_VBMI@)
+set(GGML_AVX512_VNNI @GGML_AVX512_VNNI@)
+set(GGML_AVX512_BF16 @GGML_AVX512_BF16@)
+set(GGML_AMX_TILE @GGML_AMX_TILE@)
+set(GGML_AMX_INT8 @GGML_AMX_INT8@)
+set(GGML_AMX_BF16 @GGML_AMX_BF16@)
+set(GGML_FMA  @GGML_FMA@)
+set(GGML_LASX @GGML_LASX@)
+set(GGML_LSX  @GGML_LSX@)
+set(GGML_RVV  @GGML_RVV@)
+set(GGML_SVE  @GGML_SVE@)
+
 set(GGML_ACCELERATE @GGML_ACCELERATE@)
+set(GGML_OPENMP  @GGML_OPENMP@)
+set(GGML_CPU_HBM @GGML_CPU_HBM@)
+set(GGML_BLAS_VENDOR @GGML_BLAS_VENDOR@)
+
+set(GGML_CUDA_FORCE_MMQ    @GGML_CUDA_FORCE_MMQ@)
+set(GGML_CUDA_FORCE_CUBLAS @GGML_CUDA_FORCE_CUBLAS@)
+set(GGML_CUDA_F16          @GGML_CUDA_F16@)
+set(GGML_CUDA_PEER_MAX_BATCH_SIZE @GGML_CUDA_PEER_MAX_BATCH_SIZE@)
+set(GGML_CUDA_NO_PEER_COPY  @GGML_CUDA_NO_PEER_COPY@)
+set(GGML_CUDA_NO_VMM        @GGML_CUDA_NO_VMM@)
+set(GGML_CUDA_FA_ALL_QUANTS @GGML_CUDA_FA_ALL_QUANTS@)
+set(GGML_CUDA_GRAPHS        @GGML_CUDA_GRAPHS@)
+
+set(GGML_HIP_UMA @GGML_HIP_UMA@)
+
 set(GGML_VULKAN_CHECK_RESULTS @GGML_VULKAN_CHECK_RESULTS@)
-set(GGML_VULKAN_DEBUG @GGML_VULKAN_DEBUG@)
-set(GGML_VULKAN_MEMORY_DEBUG @GGML_VULKAN_MEMORY_DEBUG@)
-set(GGML_VULKAN_VALIDATE @GGML_VULKAN_VALIDATE@)
-set(GGML_OPENMP @GGML_OPENMP@)
+set(GGML_VULKAN_DEBUG         @GGML_VULKAN_DEBUG@)
+set(GGML_VULKAN_MEMORY_DEBUG  @GGML_VULKAN_MEMORY_DEBUG@)
+set(GGML_VULKAN_SHADER_DEBUG_INFO @GGML_VULKAN_SHADER_DEBUG_INFO@)
+set(GGML_VULKAN_PERF      @GGML_VULKAN_PERF@)
+set(GGML_VULKAN_VALIDATE  @GGML_VULKAN_VALIDATE@)
+set(GGML_VULKAN_RUN_TESTS @GGML_VULKAN_RUN_TESTS@)
+
+set(GGML_METAL_USE_BF16 @GGML_METAL_USE_BF16@)
+set(GGML_METAL_NDEBUG   @GGML_METAL_NDEBUG@)
+set(GGML_METAL_SHADER_DEBUG  @GGML_METAL_SHADER_DEBUG@)
+set(GGML_METAL_EMBED_LIBRARY @GGML_METAL_EMBED_LIBRARY@)
+set(GGML_METAL_MACOSX_VERSION_MIN @GGML_METAL_MACOSX_VERSION_MIN@)
+set(GGML_METAL_STD @GGML_METAL_STD@)
+
+set(GGML_SYCL_F16    @GGML_SYCL_F16@)
+set(GGML_SYCL_TARGET @GGML_SYCL_TARGET@)
+set(GGML_SYCL_DEVICE_ARCH @GGML_SYCL_DEVICE_ARCH@)
+
 
 @PACKAGE_INIT@
 
@@ -20,6 +68,7 @@ find_package(Threads REQUIRED)
 
 set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@")
 set(_llama_link_deps "")
+set(_llama_link_opts "")
 foreach(_ggml_lib ggml ggml-base)
     string(REPLACE "-" "_" _ggml_lib_var "${_ggml_lib}_LIBRARY")
     find_library(${_ggml_lib_var} ${_ggml_lib}
@@ -49,41 +98,63 @@ foreach(backend amx blas cann cpu cuda hip kompute metal musa rpc sycl vulkan)
     endif()
 endforeach()
 
-if (APPLE AND GGML_ACCELERATE)
-    find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
-endif()
+if (NOT LLAMA_SHARED_LIB)
+    if (APPLE AND GGML_ACCELERATE)
+        find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
+        list(APPEND _llama_link_deps ${ACCELERATE_FRAMEWORK})
+    endif()
 
-if (GGML_BLAS)
-    find_package(BLAS REQUIRED)
-endif()
+    if (GGML_OPENMP)
+        find_package(OpenMP REQUIRED)
+        list(APPEND _llama_link_deps OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+    endif()
 
-if (GGML_CUDA)
-    find_package(CUDAToolkit REQUIRED)
-endif()
+    if (GGML_CPU_HBM)
+        find_library(memkind memkind REQUIRED)
+        list(APPEND _llama_link_deps memkind)
+    endif()
 
-if (GGML_METAL)
-    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
-    find_library(METAL_FRAMEWORK Metal REQUIRED)
-    find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
-endif()
+    if (GGML_BLAS)
+        find_package(BLAS REQUIRED)
+        list(APPEND _llama_link_deps ${BLAS_LIBRARIES})
+        list(APPEND _llama_link_opts ${BLAS_LINKER_FLAGS})
+    endif()
 
-if (GGML_VULKAN)
-    find_package(Vulkan REQUIRED)
-endif()
+    if (GGML_CUDA)
+        find_package(CUDAToolkit REQUIRED)
+    endif()
 
-if (GGML_HIP)
-    find_package(hip REQUIRED)
-    find_package(hipblas REQUIRED)
-    find_package(rocblas REQUIRED)
-endif()
+    if (GGML_METAL)
+        find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+        find_library(METAL_FRAMEWORK    Metal REQUIRED)
+        find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
+        list(APPEND _llama_link_deps ${FOUNDATION_LIBRARY}
+                                     ${METAL_FRAMEWORK} ${METALKIT_FRAMEWORK})
+    endif()
 
-if (GGML_SYCL)
-    find_package(IntelSYCL REQUIRED)
-    find_package(MKL REQUIRED)
-endif()
+    if (GGML_VULKAN)
+        find_package(Vulkan REQUIRED)
+        list(APPEND _llama_link_deps Vulkan::Vulkan)
+    endif()
 
-if (GGML_OPENMP)
-    find_package(OpenMP REQUIRED)
+    if (GGML_HIP)
+        find_package(hip     REQUIRED)
+        find_package(hipblas REQUIRED)
+        find_package(rocblas REQUIRED)
+        list(APPEND _llama_link_deps hip::host roc::rocblas roc::hipblas)
+    endif()
+
+    if (GGML_SYCL)
+        find_package(DNNL)
+        if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
+            list(APPEND _llama_link_deps DNNL::dnnl)
+        endif()
+        if (WIN32)
+            find_package(IntelSYCL REQUIRED)
+            find_package(MKL       REQUIRED)
+            list(APPEND _llama_link_deps IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
+        endif()
+    endif()
 endif()
 
 find_library(llama_LIBRARY llama
@@ -97,6 +168,7 @@ set_target_properties(llama
     PROPERTIES
         INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
         INTERFACE_LINK_LIBRARIES "${_llama_link_deps}"
+        INTERFACE_LINK_OPTIONS   "${_llama_link_opts}"
         INTERFACE_COMPILE_DEFINITIONS "${_llama_transient_defines}"
         IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
         IMPORTED_LOCATION "${llama_LIBRARY}"

From 9abe9eeae98b11fa93b82632b264126a010225ff Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Wed, 20 Nov 2024 13:47:36 -0600
Subject: [PATCH 047/102] vulkan: predicate max operation in soft_max
 shaders/soft_max (#10437)

Fixes #10434
---
 ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
index f9727679ec5f2..6e20b6411cadc 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
@@ -73,7 +73,9 @@ void soft_max(uint num_iters) {
 
         FLOAT_TYPE v = a * p.scale + slope * b;
 
-        max_val = max(max_val, v);
+        if (col < p.KX) {
+            max_val = max(max_val, v);
+        }
 
         if (idx < DATA_CACHE_SIZE) {
             data_cache[idx] = v;

From 02e4eaf22f229a114054b053a9eff61483653670 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Wed, 20 Nov 2024 14:56:04 +0100
Subject: [PATCH 048/102] ggml-opt: fix data corruption (ggml/1022)

---
 ggml/src/ggml-backend.cpp  |   2 +
 ggml/src/ggml-impl.h       |   3 +
 ggml/src/ggml-opt.cpp      | 147 +++++++++++++++++--------------------
 ggml/src/ggml.c            |  94 ++++++++++++++----------
 tests/test-backend-ops.cpp |   1 -
 5 files changed, 129 insertions(+), 118 deletions(-)

diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index 9dcde8d11952a..3433d082e775b 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -252,6 +252,7 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
 }
 
 void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor);
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
@@ -266,6 +267,7 @@ void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, siz
 }
 
 void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor);
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
index 92a64fe5a5930..3965be78755a1 100644
--- a/ggml/src/ggml-impl.h
+++ b/ggml/src/ggml-impl.h
@@ -295,6 +295,9 @@ struct ggml_cgraph {
     enum ggml_cgraph_eval_order order;
 };
 
+// returns a slice of cgraph with nodes [i0, i1)
+// the slice does not have leafs or gradients
+// if you need the gradients, get them from the original graph
 struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
 
 // Memory allocation
diff --git a/ggml/src/ggml-opt.cpp b/ggml/src/ggml-opt.cpp
index 040205a31bab0..7c3e24103a250 100644
--- a/ggml/src/ggml-opt.cpp
+++ b/ggml/src/ggml-opt.cpp
@@ -14,51 +14,51 @@
 #include <vector>
 
 struct ggml_opt_dataset {
-    struct ggml_context   * ctx;
-    ggml_backend_buffer_t   buf;
-    struct ggml_tensor    * data;
-    struct ggml_tensor    * labels;
+    struct ggml_context   * ctx    = nullptr;
+    ggml_backend_buffer_t   buf    = nullptr;
+    struct ggml_tensor    * data   = nullptr;
+    struct ggml_tensor    * labels = nullptr;
 
-    int64_t ndata;
-    int64_t ndata_shard;
-    size_t  nbs_data;
-    size_t  nbs_labels;
+    int64_t ndata       = -1;
+    int64_t ndata_shard = -1;
+    size_t  nbs_data    = -1;
+    size_t  nbs_labels  = -1;
 
     std::vector<int64_t> permutation;
 };
 
 struct ggml_opt_context {
-    ggml_backend_sched_t    backend_sched;
-    ggml_cgraph           * allocated_graph;
-    ggml_cgraph           * allocated_graph_copy;
-    struct ggml_context   * ctx_static;
-    struct ggml_context   * ctx_static_cpu;
-    struct ggml_context   * ctx_compute;
-    struct ggml_context   * ctx_copy;
-    ggml_backend_buffer_t   buf_static;
-    ggml_backend_buffer_t   buf_static_cpu;
+    ggml_backend_sched_t    backend_sched        = nullptr;
+    ggml_cgraph           * allocated_graph      = nullptr;
+    ggml_cgraph           * allocated_graph_copy = nullptr;
+    struct ggml_context   * ctx_static           = nullptr;
+    struct ggml_context   * ctx_static_cpu       = nullptr;
+    struct ggml_context   * ctx_compute          = nullptr;
+    struct ggml_context   * ctx_copy             = nullptr;
+    ggml_backend_buffer_t   buf_static           = nullptr;
+    ggml_backend_buffer_t   buf_static_cpu       = nullptr;
     std::mt19937            rng;
 
-    struct ggml_tensor * inputs;
-    struct ggml_tensor * outputs;
-    struct ggml_tensor * labels;
+    struct ggml_tensor * inputs  = nullptr;
+    struct ggml_tensor * outputs = nullptr;
+    struct ggml_tensor * labels  = nullptr;
 
-    struct ggml_tensor * loss;
-    struct ggml_tensor * pred;
-    struct ggml_tensor * ncorrect;
+    struct ggml_tensor * loss     = nullptr;
+    struct ggml_tensor * pred     = nullptr;
+    struct ggml_tensor * ncorrect = nullptr;
 
-    struct ggml_cgraph * gf;
-    struct ggml_cgraph * gb_grad;
-    struct ggml_cgraph * gb_opt;
+    struct ggml_cgraph * gf      = nullptr;
+    struct ggml_cgraph * gb_grad = nullptr;
+    struct ggml_cgraph * gb_opt  = nullptr;
 
-    int64_t iter;
-    int32_t opt_period;
-    int32_t opt_i;
-    bool    loss_per_datapoint;
+    int64_t iter               = 1;
+    int32_t opt_period         = 1;
+    int32_t opt_i              = 0;
+    bool    loss_per_datapoint = false;
 
-    ggml_opt_get_optimizer_params get_opt_pars;
-    void * get_opt_pars_ud;
-    struct ggml_tensor * adamw_params;
+    ggml_opt_get_optimizer_params get_opt_pars = nullptr;
+    void * get_opt_pars_ud                     = nullptr;
+    struct ggml_tensor * adamw_params          = nullptr;
 };
 
 struct ggml_opt_result {
@@ -67,8 +67,8 @@ struct ggml_opt_result {
     std::vector<int32_t> pred;
     int64_t              ncorrect = 0;
 
-    bool loss_per_datapoint = false;
-    int64_t opt_period = -1;
+    int64_t opt_period         = -1;
+    bool    loss_per_datapoint = false;
 };
 
 // ====== Dataset ======
@@ -188,11 +188,11 @@ struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * us
 }
 
 struct ggml_opt_params ggml_opt_default_params(
-        ggml_backend_sched_t backend_sched,
-        struct ggml_context * ctx_compute,
-        struct ggml_tensor * inputs,
-        struct ggml_tensor * outputs,
-        enum ggml_opt_loss_type loss_type) {
+        ggml_backend_sched_t      backend_sched,
+        struct ggml_context     * ctx_compute,
+        struct ggml_tensor      * inputs,
+        struct ggml_tensor      * outputs,
+        enum ggml_opt_loss_type   loss_type) {
     return {
         /*backend_sched   =*/ backend_sched,
         /*ctx_compute     =*/ ctx_compute,
@@ -237,25 +237,33 @@ static ggml_tensor * map_tensor(std::map<ggml_tensor *, ggml_tensor *> & tensor_
     return new_tensor;
 }
 
-static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * graph) {
+static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * src) {
     std::map<ggml_tensor *, ggml_tensor *> tensor_map;
 
-    ggml_cgraph * new_graph = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true);
+    ggml_cgraph * dst = ggml_new_graph_custom(ctx, src->size, /*grads =*/ true);
 
-    for (int i = 0; i < graph->n_leafs; i++) {
-        ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->leafs[i]));
+    for (int i = 0; i < src->n_leafs; i++) {
+        ggml_build_forward_expand(dst, map_tensor(tensor_map, ctx, src->leafs[i]));
     }
-    for (int i = 0; i < graph->n_nodes; i++) {
-        ggml_build_forward_expand(new_graph, map_tensor(tensor_map, ctx, graph->nodes[i]));
+    GGML_ASSERT(dst->n_leafs == src->n_leafs);
+    for (int i = 0; i < src->n_nodes; i++) {
+        ggml_build_forward_expand(dst, map_tensor(tensor_map, ctx, src->nodes[i]));
     }
-    for (int i = 0; i < graph->n_nodes; ++i) {
-        const size_t igrad_src = ggml_hash_find(&graph->visited_hash_set, graph->nodes[i]);
-        const size_t igrad_dst = ggml_hash_find(&new_graph->visited_hash_set, new_graph->nodes[i]);
-        graph->grads[igrad_dst]     = new_graph->grads[igrad_src];
-        graph->grad_accs[igrad_dst] = new_graph->grad_accs[igrad_src];
+    GGML_ASSERT(dst->n_nodes == src->n_nodes);
+    for (int i = 0; i < src->n_nodes; ++i) {
+        const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]);
+        const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]);
+
+        GGML_ASSERT(igrad_src != GGML_HASHSET_FULL);
+        GGML_ASSERT(ggml_bitset_get(src->visited_hash_set.used, igrad_src));
+        GGML_ASSERT(igrad_dst != GGML_HASHSET_FULL);
+        GGML_ASSERT(ggml_bitset_get(dst->visited_hash_set.used, igrad_dst));
+
+        dst->grads[igrad_dst]     = src->grads[igrad_src];
+        dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src];
     }
 
-    return new_graph;
+    return dst;
 }
 
 static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) {
@@ -284,18 +292,13 @@ static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph
 
 ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
     ggml_opt_context_t result = new struct ggml_opt_context;
-    result->backend_sched        = params.backend_sched;
-    result->allocated_graph      = nullptr;
-    result->allocated_graph_copy = nullptr;
-    result->ctx_compute          = params.ctx_compute;
-    result->ctx_copy             = nullptr;
-    result->inputs               = params.inputs;
-    result->outputs              = params.outputs;
-    result->iter                 = 1;
-    result->opt_period           = params.opt_period;
-    result->opt_i                = 0;
-    result->get_opt_pars         = params.get_opt_pars;
-    result->get_opt_pars_ud      = params.get_opt_pars_ud;
+    result->backend_sched   = params.backend_sched;
+    result->ctx_compute     = params.ctx_compute;
+    result->inputs          = params.inputs;
+    result->outputs         = params.outputs;
+    result->opt_period      = params.opt_period;
+    result->get_opt_pars    = params.get_opt_pars;
+    result->get_opt_pars_ud = params.get_opt_pars_ud;
 
     GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically");
     GGML_ASSERT(result->opt_period >= 1);
@@ -348,7 +351,6 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
 
     switch (params.loss_type) {
         case GGML_OPT_LOSS_TYPE_MEAN: {
-            result->labels = nullptr;
             result->loss = ggml_sum(result->ctx_static, result->outputs);
             ggml_set_name(result->loss, "loss_sum");
             const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
@@ -358,7 +360,6 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
             break;
         }
         case GGML_OPT_LOSS_TYPE_SUM: {
-            result->labels = nullptr;
             result->loss = ggml_sum(result->ctx_static, result->outputs);
             ggml_set_name(result->loss, "loss_sum");
             result->loss_per_datapoint = false;
@@ -413,14 +414,7 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
     }
 
     if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
-        result->gb_grad = nullptr;
-        result->gb_opt  = nullptr;
-
         result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-        result->buf_static_cpu = nullptr;
-
-        ggml_opt_alloc_graph(result, result->gf);
-
         return result;
     }
 
@@ -429,14 +423,8 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
     ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate);
 
     if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) {
-        result->gb_opt  = nullptr;
-
         result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-        result->buf_static_cpu = nullptr;
-
-        ggml_opt_alloc_graph(result, result->gb_grad);
         ggml_graph_reset(result->gb_grad);
-
         return result;
     }
 
@@ -466,7 +454,6 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
 
     result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type());
 
-    ggml_opt_alloc_graph(result, result->gb_opt);
     ggml_graph_reset(result->gb_opt);
 
     return result;
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index ee72a173e07fd..719d75c70c75d 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -5019,8 +5019,10 @@ static void ggml_hash_map_free(struct hash_map * map) {
 }
 
 // utility functions to change gradients
-// if a is in acc_table, modify gradients in-place and mark result as gradient accumulator
-// else if a is in zero_table, replace a
+// isrc is the index of tensor in cgraph->visited_has_set.keys
+// the corresponding gradient (accumulators) are also at position isrc
+// if tensor has a gradient accumulator, modify that accumulator in-place
+// else if there is no gradient for tensor, set the corresponding value
 // else, just add/subtract/etc. the gradients
 
 static void ggml_add_or_set(
@@ -5028,11 +5030,14 @@ static void ggml_add_or_set(
         struct ggml_cgraph  * cgraph,
         size_t                isrc,
         struct ggml_tensor  * tensor) {
+    struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
+    GGML_ASSERT(src);
     if (cgraph->grads[isrc]) {
-        cgraph->grads[isrc] = ggml_add_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]);
+        cgraph->grads[isrc] = ggml_add_impl(ctx, cgraph->grads[isrc], tensor, /*inplace =*/ cgraph->grad_accs[isrc]);
     } else {
         cgraph->grads[isrc] = tensor;
     }
+    ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name);
     ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
 }
 
@@ -5040,18 +5045,20 @@ static void ggml_acc_or_set(
         struct ggml_context * ctx,
         struct ggml_cgraph  * cgraph,
         size_t                isrc,
-        struct ggml_tensor  * src,
         struct ggml_tensor  * tensor,
         const  size_t         nb1,
         const  size_t         nb2,
         const  size_t         nb3,
         const  size_t         offset) {
+    struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
+    GGML_ASSERT(src);
     if (cgraph->grads[isrc]) {
         cgraph->grads[isrc] = ggml_acc_impl(ctx, cgraph->grads[isrc], tensor, nb1, nb2, nb3, offset, cgraph->grad_accs[isrc]);
     } else {
         struct ggml_tensor * a_zero = ggml_scale(ctx, src, 0.0f); // FIXME this is going to produce NaN if a contains inf/NaN
         cgraph->grads[isrc] = ggml_acc_impl(ctx, a_zero, tensor, nb1, nb2, nb3, offset, false);
     }
+    ggml_format_name(cgraph->grads[isrc], "grad for %s", cgraph->visited_hash_set.keys[isrc]->name);
     ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
 }
 
@@ -5059,13 +5066,15 @@ static void ggml_add1_or_set(
         struct ggml_context * ctx,
         struct ggml_cgraph  * cgraph,
         size_t                isrc,
-        struct ggml_tensor  * src,
         struct ggml_tensor  * tensor) {
+    struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
+    GGML_ASSERT(src);
     if (cgraph->grads[isrc]) {
         cgraph->grads[isrc] = ggml_add1_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]);
     } else {
         cgraph->grads[isrc] = ggml_repeat(ctx, tensor, src);
     }
+    ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name);
     ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
 }
 
@@ -5074,11 +5083,14 @@ static void ggml_sub_or_set(
         struct ggml_cgraph  * cgraph,
         size_t                isrc,
         struct ggml_tensor  * tensor) {
+    struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
+    GGML_ASSERT(src);
     if (cgraph->grads[isrc]) {
         cgraph->grads[isrc] = ggml_sub_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]);
     } else {
         cgraph->grads[isrc] = ggml_neg(ctx, tensor);
     }
+    ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name);
     ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
 }
 
@@ -5095,12 +5107,12 @@ static void ggml_compute_backward(
     struct ggml_tensor * src1 = tensor->src[1];
     struct ggml_tensor * src2 = tensor->src[2];
     struct ggml_hash_set * hash_set = &cgraph->visited_hash_set;
-    const size_t isrc0 = ggml_hash_find(hash_set, src0);
-    const size_t isrc1 = ggml_hash_find(hash_set, src1);
-    const size_t isrc2 = ggml_hash_find(hash_set, src2);
-    const bool src0_needs_grads = isrc0 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc0) && grads_needed[isrc0];
-    const bool src1_needs_grads = isrc1 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc1) && grads_needed[isrc1];
-    const bool src2_needs_grads = isrc2 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc2) && grads_needed[isrc2];
+    const size_t isrc0 = src0 ? ggml_hash_find(hash_set, src0) : (size_t) -1;
+    const size_t isrc1 = src1 ? ggml_hash_find(hash_set, src1) : (size_t) -1;
+    const size_t isrc2 = src2 ? ggml_hash_find(hash_set, src2) : (size_t) -1;
+    const bool src0_needs_grads = src0 && isrc0 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc0) && grads_needed[isrc0];
+    const bool src1_needs_grads = src1 && isrc1 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc1) && grads_needed[isrc1];
+    const bool src2_needs_grads = src2 && isrc2 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc2) && grads_needed[isrc2];
 
     switch (tensor->op) {
         case GGML_OP_DUP: {
@@ -5200,7 +5212,7 @@ static void ggml_compute_backward(
         } break;
         case GGML_OP_SUM: {
             if (src0_needs_grads) {
-                ggml_add1_or_set(ctx, cgraph, isrc0, src0, grad);
+                ggml_add1_or_set(ctx, cgraph, isrc0, grad);
             }
         } break;
         case GGML_OP_SUM_ROWS: {
@@ -5210,7 +5222,7 @@ static void ggml_compute_backward(
         } break;
         case GGML_OP_MEAN: {
             if (src0_needs_grads) {
-                ggml_add1_or_set(ctx, cgraph, isrc0, src0, ggml_scale_impl(ctx, grad, 1.0f/src0->ne[0], false));
+                ggml_add1_or_set(ctx, cgraph, isrc0, ggml_scale_impl(ctx, grad, 1.0f/src0->ne[0], false));
             }
         } break;
         case GGML_OP_REPEAT: {
@@ -5363,7 +5375,7 @@ static void ggml_compute_backward(
                     nb3 = (nb3 / n0) * ng;
                 }
 
-                ggml_acc_or_set(ctx, cgraph, isrc0, src0, grad, nb1, nb2, nb3, offset);
+                ggml_acc_or_set(ctx, cgraph, isrc0, grad, nb1, nb2, nb3, offset);
             }
         } break;
         case GGML_OP_PERMUTE: {
@@ -5597,10 +5609,9 @@ void ggml_build_backward_expand(
 
     const int n_nodes_f = cgraph->n_nodes;
 
-    const size_t hash_size = ggml_hash_size(2*cgraph->size);
-    memset(cgraph->grads,     0, hash_size*sizeof(struct ggml_tensor *));
-    memset(cgraph->grad_accs, 0, hash_size*sizeof(struct ggml_tensor *));
-    bool * grads_needed = calloc(hash_size, sizeof(bool));
+    memset(cgraph->grads,     0, cgraph->visited_hash_set.size*sizeof(struct ggml_tensor *));
+    memset(cgraph->grad_accs, 0, cgraph->visited_hash_set.size*sizeof(struct ggml_tensor *));
+    bool * grads_needed = calloc(cgraph->visited_hash_set.size, sizeof(bool));
 
     {
         bool any_params = false;
@@ -5621,7 +5632,7 @@ void ggml_build_backward_expand(
             continue;
         }
 
-        bool node_needs_grad = node->flags & GGML_TENSOR_FLAG_PARAM;
+        bool node_needs_grad = (node->flags & GGML_TENSOR_FLAG_PARAM) || (node->flags & GGML_TENSOR_FLAG_LOSS);
         bool ignore_src[GGML_MAX_SRC] = {false};
         switch (node->op) {
             // gradients in node->src[0] for one reason or another have no effect on output gradients
@@ -5638,7 +5649,7 @@ void ggml_build_backward_expand(
             } break;
 
             // gradients in node->src[1] for one reason or another have no effect on output gradients
-            case GGML_OP_CPY:           // gradients in CPY target  are irrelevant
+            case GGML_OP_CPY:           // gradients in CPY target are irrelevant
             case GGML_OP_GET_ROWS:      // row indices not differentiable
             case GGML_OP_GET_ROWS_BACK: // same as for GET_ROWS
             case GGML_OP_ROPE:          // positions not differentiable
@@ -5665,9 +5676,12 @@ void ggml_build_backward_expand(
             node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE);
 
         const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
+        GGML_ASSERT(igrad != GGML_HASHSET_FULL);
+        GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, igrad));
         if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) {
-            cgraph->grads[igrad]     = ggml_dup_tensor(ctx_static, node);
-            cgraph->grad_accs[igrad] = cgraph->grads[igrad];
+            cgraph->grad_accs[igrad] = ggml_dup_tensor(ctx_static, node);
+            cgraph->grads[igrad]     = cgraph->grad_accs[igrad];
+            ggml_format_name(cgraph->grad_accs[igrad], "grad acc for %s", node->name);
         }
         grads_needed[igrad] = true;
     }
@@ -5761,15 +5775,15 @@ struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) {
 
 struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1) {
     struct ggml_cgraph cgraph = {
-        /*.size         =*/ 0,
-        /*.n_nodes      =*/ i1 - i0,
-        /*.n_leafs      =*/ 0,
-        /*.nodes        =*/ cgraph0->nodes + i0,
-        /*.grads        =*/ cgraph0->grads ? cgraph0->grads + i0 : NULL,
-        /*.grad_accs    =*/ cgraph0->grad_accs ? cgraph0->grad_accs + i0 : NULL,
-        /*.leafs        =*/ NULL,
-        /*.hash_table   =*/ { 0, NULL, NULL },
-        /*.order        =*/ cgraph0->order,
+        /*.size             =*/ 0,
+        /*.n_nodes          =*/ i1 - i0,
+        /*.n_leafs          =*/ 0,
+        /*.nodes            =*/ cgraph0->nodes + i0,
+        /*.grads            =*/ NULL, // gradients would need visited_hash_set
+        /*.grad_accs        =*/ NULL,
+        /*.leafs            =*/ NULL,
+        /*.visited_hash_set =*/ { 0, NULL, NULL },
+        /*.order            =*/ cgraph0->order,
     };
 
     return cgraph;
@@ -5799,12 +5813,22 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
         }
     }
 
+    if (dst->grads) {
+        memset(dst->grads,     0, dst->visited_hash_set.size*sizeof(struct ggml_tensor *));
+        memset(dst->grad_accs, 0, dst->visited_hash_set.size*sizeof(struct ggml_tensor *));
+    }
     if (src->grads) {
         GGML_ASSERT(dst->grads     != NULL);
         GGML_ASSERT(dst->grad_accs != NULL);
         for (int i = 0; i < src->n_nodes; ++i) {
             const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]);
             const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]);
+
+            GGML_ASSERT(igrad_src != GGML_HASHSET_FULL);
+            GGML_ASSERT(ggml_bitset_get(src->visited_hash_set.used, igrad_src));
+            GGML_ASSERT(igrad_dst != GGML_HASHSET_FULL);
+            GGML_ASSERT(ggml_bitset_get(dst->visited_hash_set.used, igrad_dst));
+
             dst->grads[igrad_dst]     = src->grads[igrad_src];
             dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src];
         }
@@ -5839,12 +5863,8 @@ void ggml_graph_reset(struct ggml_cgraph * cgraph) {
 
         if (node->op == GGML_OP_OPT_STEP_ADAMW) {
             // clear momenta
-            if (node->src[2]->data) {
-                ggml_set_zero(node->src[2]);
-            }
-            if (node->src[3]->data) {
-                ggml_set_zero(node->src[3]);
-            }
+            ggml_set_zero(node->src[2]);
+            ggml_set_zero(node->src[3]);
         }
 
         // initial gradients of loss should be 1, 0 otherwise
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 01ac7166e0e22..37342c156d7c1 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -819,7 +819,6 @@ struct test_case {
             }
         }
 
-        // TODO: refactor so that this check is only needed once
         for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
             if (!ggml_backend_supports_op(backend, t)) {
                 printf("not supported [%s] ", ggml_backend_name(backend));

From 59b917282236eadfb82bf1f46a31eb119941da08 Mon Sep 17 00:00:00 2001
From: slaren <slarengh@gmail.com>
Date: Wed, 20 Nov 2024 13:25:08 +0100
Subject: [PATCH 049/102] ggml/sched : do not skip views in pre-assignments

---
 ggml/src/ggml-backend.cpp | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index 3433d082e775b..45da0c27daffa 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -886,9 +886,6 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         int * node_backend_id = &tensor_backend_id(node);
-        if (ggml_is_view_op(node->op)) {
-            continue;
-        }
         // do not overwrite user assignments
         if (*node_backend_id == -1) {
             *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);

From 87a533be57e602f8ca469d14ad15ee851265b655 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Thu, 21 Nov 2024 09:22:11 +0200
Subject: [PATCH 050/102] sync : ggml

---
 scripts/sync-ggml.last | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last
index e9bd2dbb05e2c..d101d2b57f8fa 100644
--- a/scripts/sync-ggml.last
+++ b/scripts/sync-ggml.last
@@ -1 +1 @@
-2884dd72fea8922910fe53387c3d17ab928d3a8e
+6fcbd60bc72ac3f7ad43f78c87e535f2e6206f58

From 1bb30bf28cb5a7adf111bc41c935bdaf128397e7 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Thu, 21 Nov 2024 10:22:47 +0200
Subject: [PATCH 051/102] llama : handle KV shift for recurrent models (#10402)

ggml-ci
---
 src/llama.cpp | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/src/llama.cpp b/src/llama.cpp
index c51b36e66042e..001711037d5d1 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -18211,13 +18211,13 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
 static void llama_kv_cache_update_internal(struct llama_context & lctx) {
     bool need_reserve = false;
 
-    // apply K-shift if needed
-    if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE && lctx.kv_self.has_shift) {
+    if (lctx.kv_self.has_shift) {
         if (!llama_kv_cache_can_shift(&lctx)) {
-            GGML_ABORT("Deepseek2 does not support K-shift");
+            GGML_ABORT("The current context does not support K-shift");
         }
 
-        {
+        // apply K-shift if needed
+        if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE) {
             ggml_backend_sched_reset(lctx.sched.get());
 
             ggml_cgraph * gf = llama_build_graph_k_shift(lctx);
@@ -20463,7 +20463,7 @@ void llama_kv_cache_update(struct llama_context * ctx) {
 }
 
 bool llama_kv_cache_can_shift(struct llama_context * ctx) {
-    return ctx->model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA
+    return !ctx->kv_self.recurrent && ctx->model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA
 }
 
 // deprecated

From a5e47592b6171ae21f3eaa1aba6fb2b707875063 Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Thu, 21 Nov 2024 18:18:50 +0100
Subject: [PATCH 052/102] cuda : optimize argmax (#10441)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

* cuda : optimize argmax

* remove unused parameter

ggml-ci

* fixup : use full warps

ggml-ci

* Apply suggestions from code review

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

* fix ub

* ggml : check ne00 <= INT32_MAX in argmax and argsort

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---
 ggml/src/ggml-cuda/argmax.cu   | 96 +++++++++++++++++++---------------
 ggml/src/ggml-cuda/common.cuh  | 30 +++++------
 ggml/src/ggml-cuda/quantize.cu |  8 +--
 ggml/src/ggml.c                |  2 +
 tests/test-backend-ops.cpp     | 29 ++++++++++
 5 files changed, 104 insertions(+), 61 deletions(-)

diff --git a/ggml/src/ggml-cuda/argmax.cu b/ggml/src/ggml-cuda/argmax.cu
index aab04eca7a385..5340eedc08916 100644
--- a/ggml/src/ggml-cuda/argmax.cu
+++ b/ggml/src/ggml-cuda/argmax.cu
@@ -1,57 +1,69 @@
-#include "common.cuh"
+#include <algorithm>
+#include <cstdint>
+
 #include "argmax.cuh"
+#include "common.cuh"
 #include "sum.cuh"
 
-#include <cstdint>
+static __global__ void argmax_f32(const float * __restrict__ x, int32_t * __restrict__ dst, const int64_t ncols) {
+    const int64_t row = blockIdx.x;
 
-static __global__ void argmax_f32(
-    const float * x, int32_t * dst, const int64_t ncols, const int64_t nrows) {
+    float maxval = -FLT_MAX;
+    int   argmax = -1;
+    const float * rowx = x + row * ncols;
 
-    int argmax_thread = 0;
-    const int64_t row0 = (int64_t)blockIdx.x*WARP_SIZE;
+    for (int32_t col = threadIdx.x; col < ncols; col += blockDim.x) {
+        const float val = rowx[col];
+        if (val > maxval) {
+            maxval = val;
+            argmax = col;
+        }
+    }
 
 #pragma unroll
-    for (int64_t row1 = 0; row1 < WARP_SIZE; ++row1) {
-        const int64_t row = row0 + row1;
-
-        if (row >= nrows) {
-            break;
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, offset, WARP_SIZE);
+        const int   col = __shfl_xor_sync(0xFFFFFFFF, argmax, offset, WARP_SIZE);
+        if (val > maxval) {
+            maxval = val;
+            argmax = col;
         }
+    }
 
-        float maxval = -FLT_MAX;
-        int   argmax = -1;
-
-        for (int32_t col = threadIdx.x; col < ncols; col += WARP_SIZE) {
-            const float val        = x[row*ncols + col];
-            const int   bigger     = val > maxval;
-            const int   not_bigger = bigger ^ 0x00000001;
-
-            maxval = maxval*not_bigger + val*bigger;
-            argmax = argmax*not_bigger + col*bigger;
+    const int n_warps = blockDim.x / WARP_SIZE;
+    const int lane_id = threadIdx.x % WARP_SIZE;
+    const int warp_id = threadIdx.x / WARP_SIZE;
+    if (n_warps > 1) {
+        constexpr int    max_warps = 1024 / WARP_SIZE;
+        __shared__ float shared_maxval[max_warps];
+        __shared__ int   shared_argmax[max_warps];
+        if (lane_id == 0) {
+            shared_maxval[warp_id] = maxval;
+            shared_argmax[warp_id] = argmax;
         }
 
+        __syncthreads();
+
+        if (warp_id == 0) {
+            if (lane_id < n_warps) {
+                maxval = shared_maxval[lane_id];
+                argmax = shared_argmax[lane_id];
+            }
 #pragma unroll
-        for (int mask = 16; mask > 0; mask >>= 1) {
-            const float val        = __shfl_xor_sync(0xFFFFFFFF, maxval, mask, WARP_SIZE);
-            const int   col        = __shfl_xor_sync(0xFFFFFFFF, argmax, mask, WARP_SIZE);
-            const int   bigger     = val > maxval;
-            const int   not_bigger = bigger ^ 0x00000001;
-
-            maxval = maxval*not_bigger + val*bigger;
-            argmax = argmax*not_bigger + col*bigger;
+            for (int offset = 16; offset > 0; offset >>= 1) {
+                const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, offset, WARP_SIZE);
+                const int   col = __shfl_xor_sync(0xFFFFFFFF, argmax, offset, WARP_SIZE);
+                if (val > maxval) {
+                    maxval = val;
+                    argmax = col;
+                }
+            }
         }
-
-        const int store = row1 == threadIdx.x;
-        argmax_thread += store*argmax;
     }
 
-    const int row = row0 + threadIdx.x;
-
-    if (row >= nrows) {
-        return;
+    if (warp_id == 0 && lane_id == 0) {
+        dst[row] = argmax;
     }
-
-    dst[row] = argmax_thread;
 }
 
 void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -70,10 +82,10 @@ void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     cudaStream_t stream = ctx.stream();
 
-    const int64_t num_blocks = (nrows + WARP_SIZE - 1) / WARP_SIZE;
-
-    const dim3 blocks_dim(WARP_SIZE, 1, 1);
+    const int64_t num_blocks = nrows;
+    const int64_t num_threads = std::min<int64_t>(1024, (ne00 + WARP_SIZE - 1) / WARP_SIZE * WARP_SIZE);
+    const dim3 blocks_dim(num_threads, 1, 1);
     const dim3 blocks_num(num_blocks, 1, 1);
 
-    argmax_f32<<<blocks_num, blocks_dim, 0, stream>>>(src0_d, dst_d, ne00, nrows);
+    argmax_f32<<<blocks_num, blocks_dim, 0, stream>>>(src0_d, dst_d, ne00);
 }
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index e146c691c6f87..b0dd16066b4ba 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -180,8 +180,8 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
     return __reduce_add_sync(0xffffffff, x);
 #else
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, offset, 32);
     }
     return x;
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
@@ -189,17 +189,17 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
 
 static __device__ __forceinline__ float warp_reduce_sum(float x) {
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, offset, 32);
     }
     return x;
 }
 
 static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
-        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        a.x += __shfl_xor_sync(0xffffffff, a.x, offset, 32);
+        a.y += __shfl_xor_sync(0xffffffff, a.y, offset, 32);
     }
     return a;
 }
@@ -209,16 +209,16 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        const half2 a_other = __shfl_xor_sync(0xffffffff, a, offset, 32);
         reinterpret_cast<half&>(a.x) +=  __low2half(a_other);
         reinterpret_cast<half&>(a.y) += __high2half(a_other);
     }
     return a;
 #else
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, offset, 32));
     }
     return a;
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
@@ -231,8 +231,8 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, offset, 32));
     }
     return x;
 }
@@ -275,8 +275,8 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
 #pragma unroll
-   for (int mask = 16; mask > 0; mask >>= 1) {
-       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+   for (int offset = 16; offset > 0; offset >>= 1) {
+       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, offset, 32));
    }
    return x;
 #else
diff --git a/ggml/src/ggml-cuda/quantize.cu b/ggml/src/ggml-cuda/quantize.cu
index 45408ce8684e4..1702e4ce2feba 100644
--- a/ggml/src/ggml-cuda/quantize.cu
+++ b/ggml/src/ggml-cuda/quantize.cu
@@ -69,8 +69,8 @@ static __global__ void quantize_mmq_q8_1(
 
     // Exchange max. abs. value between vals_per_scale/4 threads.
 #pragma unroll
-    for (int mask = vals_per_scale/8; mask > 0; mask >>= 1) {
-        amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, mask, WARP_SIZE));
+    for (int offset = vals_per_scale/8; offset > 0; offset >>= 1) {
+        amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, offset, WARP_SIZE));
     }
 
     float sum;
@@ -79,8 +79,8 @@ static __global__ void quantize_mmq_q8_1(
 
         // Exchange calculate sum across vals_per_sum/4 threads.
 #pragma unroll
-        for (int mask = vals_per_sum/8; mask > 0; mask >>= 1) {
-            sum += __shfl_xor_sync(0xFFFFFFFF, sum, mask, WARP_SIZE);
+        for (int offset = vals_per_sum/8; offset > 0; offset >>= 1) {
+            sum += __shfl_xor_sync(0xFFFFFFFF, sum, offset, WARP_SIZE);
         }
     }
 
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 719d75c70c75d..78e7874dee04d 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -2255,6 +2255,7 @@ struct ggml_tensor * ggml_argmax(
         struct ggml_context * ctx,
         struct ggml_tensor  * a) {
     GGML_ASSERT(ggml_is_matrix(a));
+    GGML_ASSERT(a->ne[0] <= INT32_MAX);
 
     struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, a->ne[1]);
 
@@ -4138,6 +4139,7 @@ struct ggml_tensor * ggml_argsort(
         struct ggml_context  * ctx,
         struct ggml_tensor   * a,
         enum ggml_sort_order   order) {
+    GGML_ASSERT(a->ne[0] <= INT32_MAX);
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
 
     ggml_set_op_params_i32(result, 0, (int32_t) order);
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 37342c156d7c1..b2b5705243ea2 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -1154,6 +1154,26 @@ struct test_argmax : public test_case {
         return out;
     }
 
+    void initialize_tensors(ggml_context * ctx) override {
+        std::random_device rd;
+        std::default_random_engine rng(rd());
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            if (t->type == GGML_TYPE_F32) {
+                // initialize with unique values to avoid ties
+                for (int64_t r = 0; r < ggml_nrows(t); r++) {
+                    std::vector<float> data(t->ne[0]);
+                    for (int i = 0; i < t->ne[0]; i++) {
+                        data[i] = i;
+                    }
+                    std::shuffle(data.begin(), data.end(), rng);
+                    ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(float));
+                }
+            } else {
+                init_tensor_uniform(t);
+            }
+        }
+    }
+
     double max_nmse_err() override {
         return 0.0;
     }
@@ -3440,6 +3460,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_conv_transpose_1d({2,1,1,1}, {3,1,1,1}, 1, 0, 1));
 
     test_cases.emplace_back(new test_argmax());
+    test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32, 1, 1, 1}));
+    test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {100, 10, 1, 1}));
+    test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 10, 1, 1}));
+    test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {2000, 10, 1, 1}));
+
     test_cases.emplace_back(new test_count_equal());
 
     for (int ne3 : {1, 3}) { // CUDA backward pass only supports ne3 == 1
@@ -3830,6 +3855,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {64, 64, 20, 1}, false, 1.0f, 0.0f));
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 64, 20, 1}, false, 1.0f, 0.0f));
 
+    test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32, 10, 1, 1}));
+    test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 10, 1, 1}));
+    test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32000, 512, 1, 1}));
+
     for (int bs : {1, 512}) {
         for (ggml_type type_a : all_types) {
             for (ggml_type type_b : {GGML_TYPE_F32}) {

From c18610b4ee29ca056bb4f2d375a4ad1b16f44ef7 Mon Sep 17 00:00:00 2001
From: leo-pony <nengjunma@outlook.com>
Date: Fri, 22 Nov 2024 14:07:20 +0800
Subject: [PATCH 053/102] CANN: Support Ascend310P to accelerate F32 and F16
 Model (#10216)

* CANN Support Ascend310P to accelerate F32 and F16 Model

* Add compile option soc type macro ASCEND_310P to ggml-cann lib

* Remove unused code

* Remove the ascend soc_type hard code compile option in CMakelist.txt
---
 ggml/src/ggml-cann/CMakeLists.txt           | 29 ++++++++++++++++
 ggml/src/ggml-cann/aclnn_ops.cpp            | 18 ++++++++++
 ggml/src/ggml-cann/kernels/CMakeLists.txt   |  7 ++--
 ggml/src/ggml-cann/kernels/dup.cpp          | 32 +++++++++++++-----
 ggml/src/ggml-cann/kernels/get_row_f16.cpp  | 37 +++++++++++++--------
 ggml/src/ggml-cann/kernels/get_row_f32.cpp  | 36 ++++++++++++--------
 ggml/src/ggml-cann/kernels/get_row_q4_0.cpp |  5 ++-
 7 files changed, 123 insertions(+), 41 deletions(-)

diff --git a/ggml/src/ggml-cann/CMakeLists.txt b/ggml/src/ggml-cann/CMakeLists.txt
index c8e15c6d40ce9..756200b893d02 100644
--- a/ggml/src/ggml-cann/CMakeLists.txt
+++ b/ggml/src/ggml-cann/CMakeLists.txt
@@ -3,6 +3,33 @@ if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOM
     message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$ENV{ASCEND_TOOLKIT_HOME}")
 endif()
 
+# Auto-detech Soc type and Soc version, if detect failed, will abort build
+set(SOC_VERSION "")
+function(detect_ascend_soc_type SOC_VERSION)
+    execute_process(
+        COMMAND bash -c "npu-smi info|awk -F' ' 'NF > 0 && NR==7 {print $3}'"
+        OUTPUT_VARIABLE npu_info
+        RESULT_VARIABLE npu_result
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    if("${npu_info}" STREQUAL "" OR ${npu_result})
+        message(FATAL_ERROR "Auto-detech ascend soc type failed, please specify manually or check ascend device working normally.")
+    endif()
+    set(${SOC_VERSION} "Ascend${npu_info}" PARENT_SCOPE)
+endfunction()
+
+if(NOT SOC_TYPE)
+    detect_ascend_soc_type(SOC_VERSION)
+    set(SOC_TYPE "${SOC_VERSION}")
+    message(STATUS "CANN: SOC_VERSION auto-detected is:${SOC_VERSION}")
+else()
+    string(TOLOWER ${SOC_TYPE} SOC_VERSION)
+endif()
+
+# Construct Soc specify compile option: ASCEND_#Soc_Major_SN. Such as ASCEND_910B, ASCEND310P.
+string(REGEX MATCH "[0-9]+[a-zA-Z]" SOC_TYPE_MAJOR_SN "${SOC_VERSION}")
+set(SOC_TYPE_COMPILE_OPTION "ASCEND_${SOC_TYPE_MAJOR_SN}")
+
 if (CANN_INSTALL_DIR)
     # Only Support Linux.
     if (NOT UNIX)
@@ -39,6 +66,8 @@ if (CANN_INSTALL_DIR)
     target_include_directories(ggml-cann PRIVATE . .. ${CANN_INCLUDE_DIRS})
     target_link_directories(ggml-cann PRIVATE ${CANN_INSTALL_DIR}/lib64)
 
+    target_compile_definitions(ggml-cann PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
+
     message(STATUS "CANN: CANN_INCLUDE_DIRS =  ${CANN_INCLUDE_DIRS}")
     message(STATUS "CANN: CANN_LIBRARIES =  ${CANN_LIBRARIES}")
 else()
diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index a4ec8418e2ab3..1f4ee986ceb89 100644
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -2312,6 +2312,14 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
     switch (src0->type) {
         case GGML_TYPE_F32:
+        {
+#ifdef ASCEND_310P
+             // Special operation for get_row_f32 kernel of 310P: clear the content of dest data buffer when row is not aligned to 32 bytes
+            if ((src0->ne[0] % 8) != 0) {
+                size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * ggml_type_size(GGML_TYPE_F32);
+                ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
+            }
+#endif
             aclrtlaunch_ascendc_get_row_f32(
                 24, ctx.stream(), src0->data, src1->data, dst->data,
                 ((ggml_tensor*)src0->extra)->ne,
@@ -2320,7 +2328,16 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
                 ((ggml_tensor*)dst->extra)->nb);
             break;
+        }
         case GGML_TYPE_F16:
+        {
+#ifdef ASCEND_310P
+             // Special operation for get_row_f16 kernel of 310P: clear the content of dest data buffer when row is not aligned to 32 bytes
+            if ((src0->ne[0] % 16) != 0) {
+                size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * ggml_type_size(GGML_TYPE_F32); // out is also f32, even input is f16
+                ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
+            }
+#endif
             aclrtlaunch_ascendc_get_row_f16(
                 24, ctx.stream(), src0->data, src1->data, dst->data,
                 ((ggml_tensor*)src0->extra)->ne,
@@ -2329,6 +2346,7 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
                 ((ggml_tensor*)dst->extra)->nb);
             break;
+        }
         case GGML_TYPE_Q4_0:
             aclrtlaunch_ascendc_get_row_q4_0(
                 24, ctx.stream(), src0->data, src1->data, dst->data,
diff --git a/ggml/src/ggml-cann/kernels/CMakeLists.txt b/ggml/src/ggml-cann/kernels/CMakeLists.txt
index 5b4fef91b5877..6a4e17cce54c9 100644
--- a/ggml/src/ggml-cann/kernels/CMakeLists.txt
+++ b/ggml/src/ggml-cann/kernels/CMakeLists.txt
@@ -1,7 +1,3 @@
-if (NOT SOC_TYPE)
-    set (SOC_TYPE "Ascend910B3")
-endif()
-
 file(GLOB SRC_FILES
     get_row_f32.cpp
     get_row_f16.cpp
@@ -13,7 +9,6 @@ file(GLOB SRC_FILES
     dup.cpp
 )
 
-string(TOLOWER ${SOC_TYPE} SOC_VERSION)
 set(ASCEND_CANN_PACKAGE_PATH ${CANN_INSTALL_DIR})
 set(RUN_MODE "npu" CACHE STRING "run mode: npu/sim")
 
@@ -30,4 +25,6 @@ ascendc_library(ascendc_kernels STATIC
     ${SRC_FILES}
 )
 
+message(STATUS "CANN: compile ascend kernels witch SOC_VERSION:${SOC_VERSION}.")
+ascendc_compile_definitions(ascendc_kernels PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
 # ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)
diff --git a/ggml/src/ggml-cann/kernels/dup.cpp b/ggml/src/ggml-cann/kernels/dup.cpp
index e2c651152f486..99f03e05883aa 100644
--- a/ggml/src/ggml-cann/kernels/dup.cpp
+++ b/ggml/src/ggml-cann/kernels/dup.cpp
@@ -5,6 +5,7 @@
 using namespace AscendC;
 
 #define BUFFER_NUM 2
+const int64_t SUPPORTED_MAX_DIM = 65535;  // currently the limit of max block dim supportted by dup kernel is 65535template <typename SRC_T, typename DST_T>
 
 template <typename SRC_T, typename DST_T>
 class DupByRows {
@@ -19,6 +20,7 @@ class DupByRows {
         // Input has four dims.
         int64_t op_block_num = GetBlockNum();
         int64_t op_block_idx = GetBlockIdx();
+        assert(op_block_idx < SUPPORTED_MAX_DIM && op_block_idx >= 0, "Invalid block index:%d, max is:%d\n", op_block_idx, SUPPORTED_MAX_DIM);
 
         // param
         num_rows = input_ne_ub[1] * input_ne_ub[2] * input_ne_ub[3];
@@ -51,24 +53,36 @@ class DupByRows {
 
     __aicore__ inline void copy_in() {
         LocalTensor<SRC_T> src_local = src_queue.AllocTensor<SRC_T>();
-
-        DataCopyExtParams dataCopyParams;
-        dataCopyParams.blockCount = 1;
-        dataCopyParams.blockLen = num_elem * sizeof(SRC_T);
-        DataCopyPadExtParams<SRC_T> padParams;
-        DataCopyPad(src_local, src_gm, dataCopyParams, padParams);
-
+        const size_t elem_per_block = 32 / sizeof(SRC_T);
+        size_t tail = num_elem % elem_per_block;
+        size_t cpy_elements_len = tail > 0 ? num_elem + 1 : num_elem;
+        DataCopy(src_local, src_gm, cpy_elements_len);
         src_queue.EnQue(src_local);
     }
 
     __aicore__ inline void copy_out() {
         LocalTensor<DST_T> dst_local = dst_queue.DeQue<DST_T>();
-
+#ifdef ASCEND_310P
+        const size_t elem_per_block = 32 / sizeof(DST_T);
+        size_t tail = num_elem % elem_per_block;
+        size_t len = num_elem & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(dst_gm, dst_local, len);
+        }
+        if(tail != 0) {
+            for (size_t i = tail; i < elem_per_block; i++) {
+                dst_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(dst_gm[len], dst_local[len], elem_per_block);
+            SetAtomicNone();
+        }
+#else
         DataCopyExtParams dataCopyParams;
         dataCopyParams.blockCount = 1;
         dataCopyParams.blockLen = num_elem * sizeof(DST_T);
         DataCopyPad(dst_gm, dst_local, dataCopyParams);
-
+#endif
         dst_queue.FreeTensor(dst_local);
     }
 
diff --git a/ggml/src/ggml-cann/kernels/get_row_f16.cpp b/ggml/src/ggml-cann/kernels/get_row_f16.cpp
index c704b5b2ec0f3..416b45104de5b 100644
--- a/ggml/src/ggml-cann/kernels/get_row_f16.cpp
+++ b/ggml/src/ggml-cann/kernels/get_row_f16.cpp
@@ -14,7 +14,7 @@ class GET_ROW_F16 {
                                 int64_t *output_ne_ub, size_t *output_nb_ub) {
         // TODO, use template for F16/f32
         int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
+        op_block_idx = GetBlockIdx();
 
         for (int i = 0; i < 4; i++) {
             input_ne[i] = input_ne_ub[i];
@@ -59,32 +59,42 @@ class GET_ROW_F16 {
     }
 
     __aicore__ inline void copy_in(uint32_t offset, size_t len) {
+        size_t origin_len = len;
         LocalTensor<half> input_local = input_queue.AllocTensor<half>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(input_local, input_gm[offset], len);
+        const size_t elem_per_block = 32 / sizeof(half);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
         if(tail != 0) {
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = tail * sizeof(half);
-            DataCopyPadExtParams<half> padParams;
-            DataCopyPad(input_local[len], input_gm[offset + len],
-                        dataCopyParams, padParams);
+            len += elem_per_block;
         }
+        DataCopy(input_local, input_gm[offset], len);
         input_queue.EnQue(input_local);
     }
 
     __aicore__ inline void copy_out(uint32_t offset, size_t len) {
         LocalTensor<float> output_local = output_queue.DeQue<float>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(output_gm[offset], output_local, len);
+        const size_t elem_per_block = 32 / sizeof(float);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(output_gm[offset], output_local, len);
+        }
+
         if(tail != 0) {
+#ifdef ASCEND_310P
+            for (size_t i = tail; i < elem_per_block; i++) {
+                output_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
+            SetAtomicNone();
+#else
             DataCopyExtParams dataCopyParams;
             dataCopyParams.blockCount = 1;
             dataCopyParams.blockLen = tail * sizeof(float);
             DataCopyPad(output_gm[offset + len], output_local[len],
                         dataCopyParams);
+#endif
         }
         output_queue.FreeTensor(output_local);
     }
@@ -150,6 +160,7 @@ class GET_ROW_F16 {
     GlobalTensor<float> output_gm;
     TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
     TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    int64_t op_block_idx;
 };
 
 template <typename T>
diff --git a/ggml/src/ggml-cann/kernels/get_row_f32.cpp b/ggml/src/ggml-cann/kernels/get_row_f32.cpp
index 9db080af36998..02116905b18e4 100644
--- a/ggml/src/ggml-cann/kernels/get_row_f32.cpp
+++ b/ggml/src/ggml-cann/kernels/get_row_f32.cpp
@@ -13,7 +13,7 @@ class GET_ROW_F32 {
                                 int64_t *indices_ne_ub, size_t *indices_nb_ub,
                                 int64_t *output_ne_ub, size_t *output_nb_ub) {
         int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
+        op_block_idx = GetBlockIdx();
 
         for (int i = 0; i < 4; i++) {
             input_ne[i] = input_ne_ub[i];
@@ -55,31 +55,40 @@ class GET_ROW_F32 {
 
     __aicore__ inline void copy_in(uint32_t offset, size_t len) {
         LocalTensor<float> input_local = input_queue.AllocTensor<float>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(input_local, input_gm[offset], len);
+        const size_t elem_per_block = 32 / sizeof(float);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
         if(tail != 0) {
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = tail * sizeof(float);
-            DataCopyPadExtParams<float> padParams;
-            DataCopyPad(input_local[len], input_gm[offset + len],
-                        dataCopyParams, padParams);
+            len += elem_per_block;
         }
+        DataCopy(input_local, input_gm[offset], len);
         input_queue.EnQue(input_local);
     }
 
     __aicore__ inline void copy_out(uint32_t offset, size_t len) {
         LocalTensor<float> output_local = output_queue.DeQue<float>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(output_gm[offset], output_local, len);
+        const size_t elem_per_block = 32 / sizeof(float);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(output_gm[offset], output_local, len);
+        }
+
         if(tail != 0) {
+#ifdef ASCEND_310P
+            for (size_t i = tail; i < elem_per_block; i++) {
+                output_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
+            SetAtomicNone();
+#else
             DataCopyExtParams dataCopyParams;
             dataCopyParams.blockCount = 1;
             dataCopyParams.blockLen = tail * sizeof(float);
             DataCopyPad(output_gm[offset + len], output_local[len],
                         dataCopyParams);
+#endif
         }
         output_queue.FreeTensor(output_local);
     }
@@ -144,6 +153,7 @@ class GET_ROW_F32 {
     GlobalTensor<float> output_gm;
     TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
     TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    int64_t op_block_idx;
 };
 
 template <typename T>
diff --git a/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp b/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp
index a80bfeec2417d..377211096e1f5 100644
--- a/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp
+++ b/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp
@@ -110,9 +110,12 @@ class GET_ROW_Q4_0 {
         LocalTensor<float> output_local = output_queue.AllocTensor<float>();
 
         // TODO: cast more data to speed up.
+#ifdef ASCEND_310P
+        // TODO: 310P support quantification
+#else
         Cast(cast_local, input_local, RoundMode::CAST_NONE, QK4_0);
         Cast(output_local, cast_local, RoundMode::CAST_NONE, QK4_0);
-
+#endif
         // Only mul need compile by group.
         half scale = scale_gm.GetValue(scale_offset);
 

From 599b3e0cd40432cd1975a8906f3db70bbe53b627 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Fri, 22 Nov 2024 08:32:40 +0100
Subject: [PATCH 054/102] GitHub: ask for more info in issue templates (#10426)

* GitHub: ask for more info in issues [no ci]

* refactor issue templates to be component-specific

* more understandable issue description

* add dropdown for llama.cpp module
---
 .github/ISSUE_TEMPLATE/01-bug-low.yml         | 50 ----------
 .../ISSUE_TEMPLATE/010-bug-compilation.yml    | 73 ++++++++++++++
 .github/ISSUE_TEMPLATE/011-bug-results.yml    | 98 +++++++++++++++++++
 .github/ISSUE_TEMPLATE/019-bug-misc.yml       | 78 +++++++++++++++
 .github/ISSUE_TEMPLATE/02-bug-medium.yml      | 50 ----------
 ...05-enhancement.yml => 020-enhancement.yml} |  2 +-
 .github/ISSUE_TEMPLATE/03-bug-high.yml        | 50 ----------
 .../{06-research.yml => 030-research.yml}     |  2 +-
 .github/ISSUE_TEMPLATE/04-bug-critical.yml    | 50 ----------
 .../{07-refactor.yml => 040-refactor.yml}     |  2 +-
 10 files changed, 252 insertions(+), 203 deletions(-)
 delete mode 100644 .github/ISSUE_TEMPLATE/01-bug-low.yml
 create mode 100644 .github/ISSUE_TEMPLATE/010-bug-compilation.yml
 create mode 100644 .github/ISSUE_TEMPLATE/011-bug-results.yml
 create mode 100644 .github/ISSUE_TEMPLATE/019-bug-misc.yml
 delete mode 100644 .github/ISSUE_TEMPLATE/02-bug-medium.yml
 rename .github/ISSUE_TEMPLATE/{05-enhancement.yml => 020-enhancement.yml} (97%)
 delete mode 100644 .github/ISSUE_TEMPLATE/03-bug-high.yml
 rename .github/ISSUE_TEMPLATE/{06-research.yml => 030-research.yml} (97%)
 delete mode 100644 .github/ISSUE_TEMPLATE/04-bug-critical.yml
 rename .github/ISSUE_TEMPLATE/{07-refactor.yml => 040-refactor.yml} (95%)

diff --git a/.github/ISSUE_TEMPLATE/01-bug-low.yml b/.github/ISSUE_TEMPLATE/01-bug-low.yml
deleted file mode 100644
index 54785854f776e..0000000000000
--- a/.github/ISSUE_TEMPLATE/01-bug-low.yml
+++ /dev/null
@@ -1,50 +0,0 @@
-name: Low Severity Bugs
-description: Used to report low severity bugs in llama.cpp (e.g. cosmetic issues, non critical UI glitches)
-title: "Bug: "
-labels: ["bug-unconfirmed", "low severity"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for taking the time to fill out this bug report!
-        Please include information about your system, the steps to reproduce the bug,
-        and the version of llama.cpp that you are using.
-        If possible, please provide a minimal code example that reproduces the bug.
-  - type: textarea
-    id: what-happened
-    attributes:
-      label: What happened?
-      description: Also tell us, what did you expect to happen?
-      placeholder: Tell us what you see!
-    validations:
-      required: true
-  - type: textarea
-    id: version
-    attributes:
-      label: Name and Version
-      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
-      placeholder: |
-        $./llama-cli --version
-        version: 2999 (42b4109e)
-        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
-    validations:
-      required: true
-  - type: dropdown
-    id: operating-system
-    attributes:
-      label: What operating system are you seeing the problem on?
-      multiple: true
-      options:
-        - Linux
-        - Mac
-        - Windows
-        - BSD
-        - Other? (Please let us know in description)
-    validations:
-      required: false
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant log output
-      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
-      render: shell
diff --git a/.github/ISSUE_TEMPLATE/010-bug-compilation.yml b/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
new file mode 100644
index 0000000000000..550ee1b498ab9
--- /dev/null
+++ b/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
@@ -0,0 +1,73 @@
+name: Bug (compilation)
+description: Something goes wrong when trying to compile llama.cpp.
+title: "Compile bug: "
+labels: ["bug-unconfirmed", "compilation"]
+body:
+  - type: markdown
+    attributes:
+      value: >
+        Thanks for taking the time to fill out this bug report!
+        This issue template is intended for bug reports where the compilation of llama.cpp fails.
+        Before opening an issue, please confirm that the compilation still fails with `-DGGML_CCACHE=OFF`.
+        If the compilation succeeds with ccache disabled you should be able to permanently fix the issue
+        by clearing `~/.cache/ccache` (on Linux).
+  - type: textarea
+    id: commit
+    attributes:
+      label: Git commit
+      description: Which commit are you trying to compile?
+      placeholder: |
+        $git rev-parse HEAD
+        84a07a17b1b08cf2b9747c633a2372782848a27f
+    validations:
+      required: true
+  - type: dropdown
+    id: operating-system
+    attributes:
+      label: Which operating systems do you know to be affected?
+      multiple: true
+      options:
+        - Linux
+        - Mac
+        - Windows
+        - BSD
+        - Other? (Please let us know in description)
+    validations:
+      required: true
+  - type: dropdown
+    id: backends
+    attributes:
+        label: GGML backends
+        description: Which GGML backends do you know to be affected?
+        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
+        multiple: true
+  - type: textarea
+    id: steps_to_reproduce
+    attributes:
+      label: Steps to Reproduce
+      description: >
+        Please tell us how to reproduce the bug and any additional information that you think could be useful for fixing it.
+        If you can narrow down the bug to specific compile flags, that information would be very much appreciated by us.
+      placeholder: >
+        Here are the exact commands that I used: ...
+    validations:
+      required: true
+  - type: textarea
+    id: first_bad_commit
+    attributes:
+      label: First Bad Commit
+      description: >
+        If the bug was not present on an earlier version: when did it start appearing?
+        If possible, please do a git bisect and identify the exact commit that introduced the bug.
+    validations:
+      required: false
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant log output
+      description: >
+          Please copy and paste any relevant log output, including the command that you entered and any generated text.
+          This will be automatically formatted into code, so no need for backticks.
+      render: shell
+    validations:
+      required: true
diff --git a/.github/ISSUE_TEMPLATE/011-bug-results.yml b/.github/ISSUE_TEMPLATE/011-bug-results.yml
new file mode 100644
index 0000000000000..1adb162b79a77
--- /dev/null
+++ b/.github/ISSUE_TEMPLATE/011-bug-results.yml
@@ -0,0 +1,98 @@
+name: Bug (model use)
+description: Something goes wrong when using a model (in general, not specific to a single llama.cpp module).
+title: "Eval bug: "
+labels: ["bug-unconfirmed", "model evaluation"]
+body:
+  - type: markdown
+    attributes:
+      value: >
+        Thanks for taking the time to fill out this bug report!
+        This issue template is intended for bug reports where the model evaluation results
+        (i.e. the generated text) are incorrect or llama.cpp crashes during model evaluation.
+        If you encountered the issue while using an external UI (e.g. ollama),
+        please reproduce your issue using one of the examples/binaries in this repository.
+        The `llama-cli` binary can be used for simple and reproducible model inference.
+  - type: textarea
+    id: version
+    attributes:
+      label: Name and Version
+      description: Which version of our software are you running? (use `--version` to get a version string)
+      placeholder: |
+        $./llama-cli --version
+        version: 2999 (42b4109e)
+        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
+    validations:
+      required: true
+  - type: dropdown
+    id: operating-system
+    attributes:
+      label: Which operating systems do you know to be affected?
+      multiple: true
+      options:
+        - Linux
+        - Mac
+        - Windows
+        - BSD
+        - Other? (Please let us know in description)
+    validations:
+      required: true
+  - type: dropdown
+    id: backends
+    attributes:
+        label: GGML backends
+        description: Which GGML backends do you know to be affected?
+        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
+        multiple: true
+  - type: textarea
+    id: hardware
+    attributes:
+      label: Hardware
+      description: Which CPUs/GPUs are you using?
+      placeholder: >
+        e.g. Ryzen 5950X + 2x RTX 4090
+    validations:
+      required: true
+  - type: textarea
+    id: model
+    attributes:
+      label: Model
+      description: >
+        Which model at which quantization were you using when encountering the bug?
+        If you downloaded a GGUF file off of Huggingface, please provide a link.
+      placeholder: >
+        e.g. Meta LLaMA 3.1 Instruct 8b q4_K_M
+    validations:
+      required: false
+  - type: textarea
+    id: steps_to_reproduce
+    attributes:
+      label: Steps to Reproduce
+      description: >
+        Please tell us how to reproduce the bug and any additional information that you think could be useful for fixing it.
+        If you can narrow down the bug to specific hardware, compile flags, or command line arguments,
+        that information would be very much appreciated by us.
+      placeholder: >
+        e.g. when I run llama-cli with -ngl 99 I get garbled outputs.
+        When I use -ngl 0 it works correctly.
+        Here are the exact commands that I used: ...
+    validations:
+      required: true
+  - type: textarea
+    id: first_bad_commit
+    attributes:
+      label: First Bad Commit
+      description: >
+        If the bug was not present on an earlier version: when did it start appearing?
+        If possible, please do a git bisect and identify the exact commit that introduced the bug.
+    validations:
+      required: false
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant log output
+      description: >
+          Please copy and paste any relevant log output, including the command that you entered and any generated text.
+          This will be automatically formatted into code, so no need for backticks.
+      render: shell
+    validations:
+      required: true
diff --git a/.github/ISSUE_TEMPLATE/019-bug-misc.yml b/.github/ISSUE_TEMPLATE/019-bug-misc.yml
new file mode 100644
index 0000000000000..124cdee919089
--- /dev/null
+++ b/.github/ISSUE_TEMPLATE/019-bug-misc.yml
@@ -0,0 +1,78 @@
+name: Bug (misc.)
+description: Something is not working the way it should (and it's not covered by any of the above cases).
+title: "Misc. bug: "
+labels: ["bug-unconfirmed"]
+body:
+  - type: markdown
+    attributes:
+      value: >
+        Thanks for taking the time to fill out this bug report!
+        This issue template is intended for miscellaneous bugs that don't fit into any other category.
+        If you encountered the issue while using an external UI (e.g. ollama),
+        please reproduce your issue using one of the examples/binaries in this repository.
+  - type: textarea
+    id: version
+    attributes:
+      label: Name and Version
+      description: Which version of our software are you running? (use `--version` to get a version string)
+      placeholder: |
+        $./llama-cli --version
+        version: 2999 (42b4109e)
+        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
+    validations:
+      required: true
+  - type: dropdown
+    id: operating-system
+    attributes:
+      label: Which operating systems do you know to be affected?
+      multiple: true
+      options:
+        - Linux
+        - Mac
+        - Windows
+        - BSD
+        - Other? (Please let us know in description)
+    validations:
+      required: true
+  - type: dropdown
+    id: module
+    attributes:
+      label: Which llama.cpp modules do you know to be affected?
+      multiple: true
+      options:
+        - libllama (core library)
+        - llama-cli
+        - llama-server
+        - llama-bench
+        - llama-quantize
+        - Python/Bash scripts
+        - Other (Please specify in the next section)
+    validations:
+      required: true
+  - type: textarea
+    id: steps_to_reproduce
+    attributes:
+      label: Steps to Reproduce
+      description: >
+        Please tell us how to reproduce the bug and any additional information that you think could be useful for fixing it.
+    validations:
+      required: true
+  - type: textarea
+    id: first_bad_commit
+    attributes:
+      label: First Bad Commit
+      description: >
+        If the bug was not present on an earlier version: when did it start appearing?
+        If possible, please do a git bisect and identify the exact commit that introduced the bug.
+    validations:
+      required: false
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant log output
+      description: >
+          Please copy and paste any relevant log output, including the command that you entered and any generated text.
+          This will be automatically formatted into code, so no need for backticks.
+      render: shell
+    validations:
+      required: true
diff --git a/.github/ISSUE_TEMPLATE/02-bug-medium.yml b/.github/ISSUE_TEMPLATE/02-bug-medium.yml
deleted file mode 100644
index a6285c6f05bac..0000000000000
--- a/.github/ISSUE_TEMPLATE/02-bug-medium.yml
+++ /dev/null
@@ -1,50 +0,0 @@
-name: Medium Severity Bug
-description: Used to report medium severity bugs in llama.cpp (e.g. Malfunctioning Features but generally still useable)
-title: "Bug: "
-labels: ["bug-unconfirmed", "medium severity"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for taking the time to fill out this bug report!
-        Please include information about your system, the steps to reproduce the bug,
-        and the version of llama.cpp that you are using.
-        If possible, please provide a minimal code example that reproduces the bug.
-  - type: textarea
-    id: what-happened
-    attributes:
-      label: What happened?
-      description: Also tell us, what did you expect to happen?
-      placeholder: Tell us what you see!
-    validations:
-      required: true
-  - type: textarea
-    id: version
-    attributes:
-      label: Name and Version
-      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
-      placeholder: |
-        $./llama-cli --version
-        version: 2999 (42b4109e)
-        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
-    validations:
-      required: true
-  - type: dropdown
-    id: operating-system
-    attributes:
-      label: What operating system are you seeing the problem on?
-      multiple: true
-      options:
-        - Linux
-        - Mac
-        - Windows
-        - BSD
-        - Other? (Please let us know in description)
-    validations:
-      required: false
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant log output
-      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
-      render: shell
diff --git a/.github/ISSUE_TEMPLATE/05-enhancement.yml b/.github/ISSUE_TEMPLATE/020-enhancement.yml
similarity index 97%
rename from .github/ISSUE_TEMPLATE/05-enhancement.yml
rename to .github/ISSUE_TEMPLATE/020-enhancement.yml
index 58fca73183d41..02dd4f575a686 100644
--- a/.github/ISSUE_TEMPLATE/05-enhancement.yml
+++ b/.github/ISSUE_TEMPLATE/020-enhancement.yml
@@ -1,5 +1,5 @@
 name: Enhancement
-description: Used to request enhancements for llama.cpp
+description: Used to request enhancements for llama.cpp.
 title: "Feature Request: "
 labels: ["enhancement"]
 body:
diff --git a/.github/ISSUE_TEMPLATE/03-bug-high.yml b/.github/ISSUE_TEMPLATE/03-bug-high.yml
deleted file mode 100644
index ff816b93769c3..0000000000000
--- a/.github/ISSUE_TEMPLATE/03-bug-high.yml
+++ /dev/null
@@ -1,50 +0,0 @@
-name: High Severity Bug
-description: Used to report high severity bugs in llama.cpp (e.g. Malfunctioning features hindering important common workflow)
-title: "Bug: "
-labels: ["bug-unconfirmed", "high severity"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for taking the time to fill out this bug report!
-        Please include information about your system, the steps to reproduce the bug,
-        and the version of llama.cpp that you are using.
-        If possible, please provide a minimal code example that reproduces the bug.
-  - type: textarea
-    id: what-happened
-    attributes:
-      label: What happened?
-      description: Also tell us, what did you expect to happen?
-      placeholder: Tell us what you see!
-    validations:
-      required: true
-  - type: textarea
-    id: version
-    attributes:
-      label: Name and Version
-      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
-      placeholder: |
-        $./llama-cli --version
-        version: 2999 (42b4109e)
-        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
-    validations:
-      required: true
-  - type: dropdown
-    id: operating-system
-    attributes:
-      label: What operating system are you seeing the problem on?
-      multiple: true
-      options:
-        - Linux
-        - Mac
-        - Windows
-        - BSD
-        - Other? (Please let us know in description)
-    validations:
-      required: false
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant log output
-      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
-      render: shell
diff --git a/.github/ISSUE_TEMPLATE/06-research.yml b/.github/ISSUE_TEMPLATE/030-research.yml
similarity index 97%
rename from .github/ISSUE_TEMPLATE/06-research.yml
rename to .github/ISSUE_TEMPLATE/030-research.yml
index 3ae4e9f8caaa4..18975dbbfd0fe 100644
--- a/.github/ISSUE_TEMPLATE/06-research.yml
+++ b/.github/ISSUE_TEMPLATE/030-research.yml
@@ -1,5 +1,5 @@
 name: Research
-description: Track new technical research area
+description: Track new technical research area.
 title: "Research: "
 labels: ["research 🔬"]
 body:
diff --git a/.github/ISSUE_TEMPLATE/04-bug-critical.yml b/.github/ISSUE_TEMPLATE/04-bug-critical.yml
deleted file mode 100644
index 7af42a80b3b93..0000000000000
--- a/.github/ISSUE_TEMPLATE/04-bug-critical.yml
+++ /dev/null
@@ -1,50 +0,0 @@
-name: Critical Severity Bug
-description: Used to report critical severity bugs in llama.cpp (e.g. Crashing, Corrupted, Dataloss)
-title: "Bug: "
-labels: ["bug-unconfirmed", "critical severity"]
-body:
-  - type: markdown
-    attributes:
-      value: |
-        Thanks for taking the time to fill out this bug report!
-        Please include information about your system, the steps to reproduce the bug,
-        and the version of llama.cpp that you are using.
-        If possible, please provide a minimal code example that reproduces the bug.
-  - type: textarea
-    id: what-happened
-    attributes:
-      label: What happened?
-      description: Also tell us, what did you expect to happen?
-      placeholder: Tell us what you see!
-    validations:
-      required: true
-  - type: textarea
-    id: version
-    attributes:
-      label: Name and Version
-      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
-      placeholder: |
-        $./llama-cli --version
-        version: 2999 (42b4109e)
-        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
-    validations:
-      required: true
-  - type: dropdown
-    id: operating-system
-    attributes:
-      label: What operating system are you seeing the problem on?
-      multiple: true
-      options:
-        - Linux
-        - Mac
-        - Windows
-        - BSD
-        - Other? (Please let us know in description)
-    validations:
-      required: false
-  - type: textarea
-    id: logs
-    attributes:
-      label: Relevant log output
-      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
-      render: shell
diff --git a/.github/ISSUE_TEMPLATE/07-refactor.yml b/.github/ISSUE_TEMPLATE/040-refactor.yml
similarity index 95%
rename from .github/ISSUE_TEMPLATE/07-refactor.yml
rename to .github/ISSUE_TEMPLATE/040-refactor.yml
index 3a68d3d5355d6..b6e6ab36defd6 100644
--- a/.github/ISSUE_TEMPLATE/07-refactor.yml
+++ b/.github/ISSUE_TEMPLATE/040-refactor.yml
@@ -1,5 +1,5 @@
 name: Refactor (Maintainers)
-description: Used to track refactoring opportunities
+description: Used to track refactoring opportunities.
 title: "Refactor: "
 labels: ["refactor"]
 body:

From 6dfcfef0787e9902df29f510b63621f60a09a50b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E8=95=AD=E6=BE=A7=E9=82=A6?=
 <45505768+shou692199@users.noreply.github.com>
Date: Fri, 22 Nov 2024 17:44:08 +0800
Subject: [PATCH 055/102] ci: Update oneAPI runtime dll packaging (#10428)

This is the minimum runtime dll dependencies for oneAPI 2025.0
---
 .github/workflows/build.yml | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 6ef0770f3fa2f..572f91643fbba 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -986,13 +986,14 @@ jobs:
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.4.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
 
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_win_proxy_loader.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_level_zero.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin

From 55ed008b2de01592659b9eba068ea01bb2f72160 Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Sat, 23 Nov 2024 14:41:12 +0100
Subject: [PATCH 056/102] ggml : do not use ARM features not included in the
 build (#10457)

---
 ggml/src/ggml-cpu/ggml-cpu.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
index 0d23669c25bea..4b58254e7d108 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -13896,7 +13896,7 @@ int ggml_cpu_has_vsx(void) {
 }
 
 int ggml_cpu_has_neon(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_NEON)
     return ggml_arm_arch_features.has_neon;
 #else
     return 0;
@@ -13904,7 +13904,7 @@ int ggml_cpu_has_neon(void) {
 }
 
 int ggml_cpu_has_sve(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
     return ggml_arm_arch_features.has_sve;
 #else
     return 0;
@@ -13912,7 +13912,7 @@ int ggml_cpu_has_sve(void) {
 }
 
 int ggml_cpu_has_matmul_int8(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8)
     return ggml_arm_arch_features.has_i8mm;
 #else
     return 0;
@@ -13920,7 +13920,7 @@ int ggml_cpu_has_matmul_int8(void) {
 }
 
 int ggml_cpu_get_sve_cnt(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
     return ggml_arm_arch_features.sve_cnt;
 #else
     return 0;

From 96fa2c5e2d6cb5319f34c3a7fb0cec05694b22f1 Mon Sep 17 00:00:00 2001
From: momonga <146910567+mmngays@users.noreply.github.com>
Date: Sun, 24 Nov 2024 09:09:22 +0900
Subject: [PATCH 057/102] fix gguf-py:  Conversion error when multiple licenses
 are configured (#9807)

* fix general.license list to str

* fix join license list

---------

Co-authored-by: momonga <115213907+mmnga@users.noreply.github.com>
---
 gguf-py/gguf/metadata.py | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/gguf-py/gguf/metadata.py b/gguf-py/gguf/metadata.py
index 321cbcd4c5507..962c27b204464 100644
--- a/gguf-py/gguf/metadata.py
+++ b/gguf-py/gguf/metadata.py
@@ -545,7 +545,10 @@ def set_gguf_meta_model(self, gguf_writer: gguf.GGUFWriter):
             gguf_writer.add_size_label(self.size_label)
 
         if self.license is not None:
-            gguf_writer.add_license(self.license)
+            if isinstance(self.license, list):
+                gguf_writer.add_license(",".join(self.license))
+            else:
+                gguf_writer.add_license(self.license)
         if self.license_name is not None:
             gguf_writer.add_license_name(self.license_name)
         if self.license_link is not None:

From 9336db462c0c34bbe2055413fe4e16442626c38b Mon Sep 17 00:00:00 2001
From: Gabe Goodhart <ghart@us.ibm.com>
Date: Sun, 24 Nov 2024 02:02:34 -0700
Subject: [PATCH 058/102] convert : XLMRoberta Type Vocab Size (#10458)

This matches the key in common bert-based embedding models and may have a
value other than 1 in it.

Branch: XLMRobertaTypeVocabSize

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
---
 convert_hf_to_gguf.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 9f4b8154b88a8..80a179b86af7e 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -2707,7 +2707,7 @@ def set_vocab(self):
         self.gguf_writer.add_token_scores(scores)
         self.gguf_writer.add_token_types(toktypes)
         self.gguf_writer.add_add_space_prefix(add_prefix)
-        self.gguf_writer.add_token_type_count(1)
+        self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
         self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
         if precompiled_charsmap:
             self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)

From dc39012cbaf8752fabecaeb60af78ccdd1dfb73b Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Sun, 24 Nov 2024 16:10:26 +0100
Subject: [PATCH 059/102] llama : fix op mul check with command-r-plus (#10476)

---
 src/llama.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/llama.cpp b/src/llama.cpp
index 001711037d5d1..20df09b133bfb 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -7181,12 +7181,12 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
             } break;
         case GGML_OP_ADD:
             {
-                ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, w->ne[0], 512);
+                ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]);
                 op_tensor = ggml_add(ctx, a, w);
             } break;
         case GGML_OP_MUL:
             {
-                ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, w->ne[0], 512);
+                ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]);
                 op_tensor = ggml_mul(ctx, a, w);
             } break;
         case GGML_OP_DIV:

From cce5a9007572c6e9fa522296b77571d2e5071357 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sun, 24 Nov 2024 18:03:25 +0200
Subject: [PATCH 060/102] flake.lock: Update (#10470)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Flake lock file updates:

• Updated input 'nixpkgs':
    'github:NixOS/nixpkgs/5e4fbfb6b3de1aa2872b76d49fafc942626e2add?narHash=sha256-OZiZ3m8SCMfh3B6bfGC/Bm4x3qc1m2SVEAlkV6iY7Yg%3D' (2024-11-15)
  → 'github:NixOS/nixpkgs/23e89b7da85c3640bbc2173fe04f4bd114342367?narHash=sha256-y/MEyuJ5oBWrWAic/14LaIr/u5E0wRVzyYsouYY3W6w%3D' (2024-11-19)

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
---
 flake.lock | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/flake.lock b/flake.lock
index ee8cf07e37220..d114f4422a36a 100644
--- a/flake.lock
+++ b/flake.lock
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1731676054,
-        "narHash": "sha256-OZiZ3m8SCMfh3B6bfGC/Bm4x3qc1m2SVEAlkV6iY7Yg=",
+        "lastModified": 1732014248,
+        "narHash": "sha256-y/MEyuJ5oBWrWAic/14LaIr/u5E0wRVzyYsouYY3W6w=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "5e4fbfb6b3de1aa2872b76d49fafc942626e2add",
+        "rev": "23e89b7da85c3640bbc2173fe04f4bd114342367",
         "type": "github"
       },
       "original": {

From d9d54e498d38ec99bbc0031022f9c92711e97bbc Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 25 Nov 2024 09:58:41 +0200
Subject: [PATCH 061/102] speculative : refactor and add a simpler example
 (#10362)

* speculative : refactor and add a simpler example

ggml-ci

* speculative : clean-up and add comments and TODOs [no ci]

* speculative : manage context in common_speculative

ggml-ci

* speculative : simplify

ggml-ci

* speculative : simplify (cont)

ggml-ci

* speculative : add --draft-min CLI arg

* speculative : minor fixup

* make : build fixes

* speculative : do not redraft previous drafts

ggml-ci

* speculative : fix the draft sampling

ggml-ci

* speculative : fix compile warning

* common : refactor args

ggml-ci

* common : change defaults [no ci]

* common : final touches

ggml-ci
---
 Makefile                                      |   1 +
 common/CMakeLists.txt                         |   2 +
 common/arg.cpp                                | 438 +++++++++---------
 common/common.cpp                             |  76 ++-
 common/common.h                               |  41 +-
 common/sampling.cpp                           |  45 +-
 common/sampling.h                             |  23 +-
 common/speculative.cpp                        | 269 +++++++++++
 common/speculative.h                          |  28 ++
 examples/CMakeLists.txt                       |   1 +
 examples/batched/batched.cpp                  |   8 +-
 examples/infill/infill.cpp                    |   2 +-
 examples/llava/llava-cli.cpp                  |   2 +-
 examples/llava/minicpmv-cli.cpp               |   2 +-
 examples/lookahead/lookahead.cpp              |   2 +-
 examples/lookup/lookup-stats.cpp              |   3 +-
 examples/lookup/lookup.cpp                    |   4 +-
 examples/main/main.cpp                        |   2 +-
 examples/parallel/parallel.cpp                |   2 +-
 examples/retrieval/retrieval.cpp              |   4 +-
 examples/save-load-state/save-load-state.cpp  |   8 +-
 examples/server/server.cpp                    |  10 +-
 examples/server/utils.hpp                     |  57 ---
 examples/speculative-simple/CMakeLists.txt    |   5 +
 examples/speculative-simple/README.md         |  12 +
 .../speculative-simple/speculative-simple.cpp | 273 +++++++++++
 examples/speculative/speculative.cpp          |  30 +-
 tests/test-arg-parser.cpp                     |   4 +-
 28 files changed, 1028 insertions(+), 326 deletions(-)
 create mode 100644 common/speculative.cpp
 create mode 100644 common/speculative.h
 create mode 100644 examples/speculative-simple/CMakeLists.txt
 create mode 100644 examples/speculative-simple/README.md
 create mode 100644 examples/speculative-simple/speculative-simple.cpp

diff --git a/Makefile b/Makefile
index 5c899438515e1..dd6d864ad513a 100644
--- a/Makefile
+++ b/Makefile
@@ -966,6 +966,7 @@ OBJ_COMMON = \
 	$(DIR_COMMON)/console.o \
 	$(DIR_COMMON)/ngram-cache.o \
 	$(DIR_COMMON)/sampling.o \
+	$(DIR_COMMON)/speculative.o \
 	$(DIR_COMMON)/build-info.o \
 	$(DIR_COMMON)/json-schema-to-grammar.o
 
diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt
index 5ab1ffa1922aa..62a8a7db5652f 100644
--- a/common/CMakeLists.txt
+++ b/common/CMakeLists.txt
@@ -66,6 +66,8 @@ add_library(${TARGET} STATIC
     ngram-cache.h
     sampling.cpp
     sampling.h
+    speculative.cpp
+    speculative.h
     )
 
 if (BUILD_SHARED_LIBS)
diff --git a/common/arg.cpp b/common/arg.cpp
index 4115b2f7511d3..32240f21f2469 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -233,10 +233,11 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         }
     }
 
-    postprocess_cpu_params(params.cpuparams, nullptr);
+    postprocess_cpu_params(params.cpuparams,       nullptr);
     postprocess_cpu_params(params.cpuparams_batch, &params.cpuparams);
-    postprocess_cpu_params(params.draft_cpuparams, &params.cpuparams);
-    postprocess_cpu_params(params.draft_cpuparams_batch, &params.cpuparams_batch);
+
+    postprocess_cpu_params(params.speculative.cpuparams,       &params.cpuparams);
+    postprocess_cpu_params(params.speculative.cpuparams_batch, &params.cpuparams_batch);
 
     if (params.prompt_cache_all && (params.interactive || params.interactive_first)) {
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
@@ -251,7 +252,7 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         for (auto & antiprompt : params.antiprompt) {
             string_process_escapes(antiprompt);
         }
-        for (auto & seq_breaker : params.sparams.dry_sequence_breakers) {
+        for (auto & seq_breaker : params.sampling.dry_sequence_breakers) {
             string_process_escapes(seq_breaker);
         }
     }
@@ -329,7 +330,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
 
     std::string sampler_type_chars;
     std::string sampler_type_names;
-    for (const auto & sampler : params.sparams.samplers) {
+    for (const auto & sampler : params.sampling.samplers) {
         sampler_type_chars += common_sampler_type_to_chr(sampler);
         sampler_type_names += common_sampler_type_to_str(sampler) + ";";
     }
@@ -407,26 +408,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ));
-    add_opt(common_arg(
-        {"-td", "--threads-draft"}, "N",
-        "number of threads to use during generation (default: same as --threads)",
-        [](common_params & params, int value) {
-            params.draft_cpuparams.n_threads = value;
-            if (params.draft_cpuparams.n_threads <= 0) {
-                params.draft_cpuparams.n_threads = std::thread::hardware_concurrency();
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"-tbd", "--threads-batch-draft"}, "N",
-        "number of threads to use during batch and prompt processing (default: same as --threads-draft)",
-        [](common_params & params, int value) {
-            params.draft_cpuparams_batch.n_threads = value;
-            if (params.draft_cpuparams_batch.n_threads <= 0) {
-                params.draft_cpuparams_batch.n_threads = std::thread::hardware_concurrency();
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
     add_opt(common_arg(
         {"-C", "--cpu-mask"}, "M",
         "CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: \"\")",
@@ -515,108 +496,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.cpuparams_batch.poll = value;
         }
     ));
-    add_opt(common_arg(
-        {"-Cd", "--cpu-mask-draft"}, "M",
-        "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)",
-        [](common_params & params, const std::string & mask) {
-            params.draft_cpuparams.mask_valid = true;
-            if (!parse_cpu_mask(mask, params.draft_cpuparams.cpumask)) {
-                throw std::invalid_argument("invalid cpumask");
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"-Crd", "--cpu-range-draft"}, "lo-hi",
-        "Ranges of CPUs for affinity. Complements --cpu-mask-draft",
-        [](common_params & params, const std::string & range) {
-            params.draft_cpuparams.mask_valid = true;
-            if (!parse_cpu_range(range, params.draft_cpuparams.cpumask)) {
-                throw std::invalid_argument("invalid range");
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--cpu-strict-draft"}, "<0|1>",
-        "Use strict CPU placement for draft model (default: same as --cpu-strict)",
-        [](common_params & params, int value) {
-            params.draft_cpuparams.strict_cpu = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--prio-draft"}, "N",
-        string_format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.draft_cpuparams.priority),
-        [](common_params & params, int prio) {
-            if (prio < 0 || prio > 3) {
-                throw std::invalid_argument("invalid value");
-            }
-            params.draft_cpuparams.priority = (enum ggml_sched_priority) prio;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--poll-draft"}, "<0|1>",
-        "Use polling to wait for draft model work (default: same as --poll])",
-        [](common_params & params, int value) {
-            params.draft_cpuparams.poll = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"-Cbd", "--cpu-mask-batch-draft"}, "M",
-        "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)",
-        [](common_params & params, const std::string & mask) {
-            params.draft_cpuparams_batch.mask_valid = true;
-            if (!parse_cpu_mask(mask, params.draft_cpuparams_batch.cpumask)) {
-                throw std::invalid_argument("invalid cpumask");
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"-Crbd", "--cpu-range-batch-draft"}, "lo-hi",
-        "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)",
-        [](common_params & params, const std::string & range) {
-            params.draft_cpuparams_batch.mask_valid = true;
-            if (!parse_cpu_range(range, params.draft_cpuparams_batch.cpumask)) {
-                throw std::invalid_argument("invalid cpumask");
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--cpu-strict-batch-draft"}, "<0|1>",
-        "Use strict CPU placement for draft model (default: --cpu-strict-draft)",
-        [](common_params & params, int value) {
-            params.draft_cpuparams_batch.strict_cpu = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--prio-batch-draft"}, "N",
-        string_format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.draft_cpuparams_batch.priority),
-        [](common_params & params, int prio) {
-            if (prio < 0 || prio > 3) {
-                throw std::invalid_argument("invalid value");
-            }
-            params.draft_cpuparams_batch.priority = (enum ggml_sched_priority) prio;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--poll-batch-draft"}, "<0|1>",
-        "Use polling to wait for draft model work (default: --poll-draft)",
-        [](common_params & params, int value) {
-            params.draft_cpuparams_batch.poll = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
-    add_opt(common_arg(
-        {"--draft"}, "N",
-        string_format("number of tokens to draft for speculative decoding (default: %d)", params.n_draft),
-        [](common_params & params, int value) {
-            params.n_draft = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP}));
-    add_opt(common_arg(
-        {"-ps", "--p-split"}, "N",
-        string_format("speculative decoding split probability (default: %.1f)", (double)params.p_split),
-        [](common_params & params, const std::string & value) {
-            params.p_split = std::stof(value);
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
     add_opt(common_arg(
         {"-lcs", "--lookup-cache-static"}, "FNAME",
         "path to static lookup cache to use for lookup decoding (not updated by generation)",
@@ -701,7 +580,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"),
         [](common_params & params) {
             params.no_perf = true;
-            params.sparams.no_perf = true;
+            params.sampling.no_perf = true;
         }
     ).set_env("LLAMA_ARG_NO_PERF"));
     add_opt(common_arg(
@@ -883,155 +762,155 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         string_format("samplers that will be used for generation in the order, separated by \';\'\n(default: %s)", sampler_type_names.c_str()),
         [](common_params & params, const std::string & value) {
             const auto sampler_names = string_split<std::string>(value, ';');
-            params.sparams.samplers = common_sampler_types_from_names(sampler_names, true);
+            params.sampling.samplers = common_sampler_types_from_names(sampler_names, true);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"-s", "--seed"}, "SEED",
-        string_format("RNG seed (default: %d, use random seed for %d)", params.sparams.seed, LLAMA_DEFAULT_SEED),
+        string_format("RNG seed (default: %d, use random seed for %d)", params.sampling.seed, LLAMA_DEFAULT_SEED),
         [](common_params & params, const std::string & value) {
-            params.sparams.seed = std::stoul(value);
+            params.sampling.seed = std::stoul(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--sampling-seq"}, "SEQUENCE",
         string_format("simplified sequence for samplers that will be used (default: %s)", sampler_type_chars.c_str()),
         [](common_params & params, const std::string & value) {
-            params.sparams.samplers = common_sampler_types_from_chars(value);
+            params.sampling.samplers = common_sampler_types_from_chars(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--ignore-eos"},
         "ignore end of stream token and continue generating (implies --logit-bias EOS-inf)",
         [](common_params & params) {
-            params.sparams.ignore_eos = true;
+            params.sampling.ignore_eos = true;
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--penalize-nl"},
-        string_format("penalize newline tokens (default: %s)", params.sparams.penalize_nl ? "true" : "false"),
+        string_format("penalize newline tokens (default: %s)", params.sampling.penalize_nl ? "true" : "false"),
         [](common_params & params) {
-            params.sparams.penalize_nl = true;
+            params.sampling.penalize_nl = true;
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--temp"}, "N",
-        string_format("temperature (default: %.1f)", (double)params.sparams.temp),
+        string_format("temperature (default: %.1f)", (double)params.sampling.temp),
         [](common_params & params, const std::string & value) {
-            params.sparams.temp = std::stof(value);
-            params.sparams.temp = std::max(params.sparams.temp, 0.0f);
+            params.sampling.temp = std::stof(value);
+            params.sampling.temp = std::max(params.sampling.temp, 0.0f);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--top-k"}, "N",
-        string_format("top-k sampling (default: %d, 0 = disabled)", params.sparams.top_k),
+        string_format("top-k sampling (default: %d, 0 = disabled)", params.sampling.top_k),
         [](common_params & params, int value) {
-            params.sparams.top_k = value;
+            params.sampling.top_k = value;
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--top-p"}, "N",
-        string_format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sparams.top_p),
+        string_format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sampling.top_p),
         [](common_params & params, const std::string & value) {
-            params.sparams.top_p = std::stof(value);
+            params.sampling.top_p = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--min-p"}, "N",
-        string_format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sparams.min_p),
+        string_format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sampling.min_p),
         [](common_params & params, const std::string & value) {
-            params.sparams.min_p = std::stof(value);
+            params.sampling.min_p = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--xtc-probability"}, "N",
-        string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sparams.xtc_probability),
+        string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
         [](common_params & params, const std::string & value) {
-            params.sparams.xtc_probability = std::stof(value);
+            params.sampling.xtc_probability = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--xtc-threshold"}, "N",
-        string_format("xtc threshold (default: %.1f, 1.0 = disabled)", (double)params.sparams.xtc_threshold),
+        string_format("xtc threshold (default: %.1f, 1.0 = disabled)", (double)params.sampling.xtc_threshold),
         [](common_params & params, const std::string & value) {
-            params.sparams.xtc_threshold = std::stof(value);
+            params.sampling.xtc_threshold = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--typical"}, "N",
-        string_format("locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)", (double)params.sparams.typ_p),
+        string_format("locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)", (double)params.sampling.typ_p),
         [](common_params & params, const std::string & value) {
-            params.sparams.typ_p = std::stof(value);
+            params.sampling.typ_p = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--repeat-last-n"}, "N",
-        string_format("last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", params.sparams.penalty_last_n),
+        string_format("last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", params.sampling.penalty_last_n),
         [](common_params & params, int value) {
-            params.sparams.penalty_last_n = value;
-            params.sparams.n_prev = std::max(params.sparams.n_prev, params.sparams.penalty_last_n);
+            params.sampling.penalty_last_n = value;
+            params.sampling.n_prev = std::max(params.sampling.n_prev, params.sampling.penalty_last_n);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--repeat-penalty"}, "N",
-        string_format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sparams.penalty_repeat),
+        string_format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
         [](common_params & params, const std::string & value) {
-            params.sparams.penalty_repeat = std::stof(value);
+            params.sampling.penalty_repeat = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--presence-penalty"}, "N",
-        string_format("repeat alpha presence penalty (default: %.1f, 0.0 = disabled)", (double)params.sparams.penalty_present),
+        string_format("repeat alpha presence penalty (default: %.1f, 0.0 = disabled)", (double)params.sampling.penalty_present),
         [](common_params & params, const std::string & value) {
-            params.sparams.penalty_present = std::stof(value);
+            params.sampling.penalty_present = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--frequency-penalty"}, "N",
-        string_format("repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)", (double)params.sparams.penalty_freq),
+        string_format("repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)", (double)params.sampling.penalty_freq),
         [](common_params & params, const std::string & value) {
-            params.sparams.penalty_freq = std::stof(value);
+            params.sampling.penalty_freq = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dry-multiplier"}, "N",
-        string_format("set DRY sampling multiplier (default: %.1f, 0.0 = disabled)", (double)params.sparams.dry_multiplier),
+        string_format("set DRY sampling multiplier (default: %.1f, 0.0 = disabled)", (double)params.sampling.dry_multiplier),
         [](common_params & params, const std::string & value) {
-            params.sparams.dry_multiplier = std::stof(value);
+            params.sampling.dry_multiplier = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dry-base"}, "N",
-        string_format("set DRY sampling base value (default: %.2f)", (double)params.sparams.dry_base),
+        string_format("set DRY sampling base value (default: %.2f)", (double)params.sampling.dry_base),
         [](common_params & params, const std::string & value) {
             float potential_base = std::stof(value);
             if (potential_base >= 1.0f)
             {
-                params.sparams.dry_base = potential_base;
+                params.sampling.dry_base = potential_base;
             }
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dry-allowed-length"}, "N",
-        string_format("set allowed length for DRY sampling (default: %d)", params.sparams.dry_allowed_length),
+        string_format("set allowed length for DRY sampling (default: %d)", params.sampling.dry_allowed_length),
         [](common_params & params, int value) {
-            params.sparams.dry_allowed_length = value;
+            params.sampling.dry_allowed_length = value;
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dry-penalty-last-n"}, "N",
-        string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable, -1 = context size)", params.sparams.dry_penalty_last_n),
+        string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable, -1 = context size)", params.sampling.dry_penalty_last_n),
         [](common_params & params, int value) {
-            params.sparams.dry_penalty_last_n = value;
+            params.sampling.dry_penalty_last_n = value;
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dry-sequence-breaker"}, "STRING",
         string_format("add sequence breaker for DRY sampling, clearing out default breakers (%s) in the process; use \"none\" to not use any sequence breakers\n",
-            params.sparams.dry_sequence_breakers.empty() ? "none" :
-            std::accumulate(std::next(params.sparams.dry_sequence_breakers.begin()),
-                params.sparams.dry_sequence_breakers.end(),
-                std::string("'") + (params.sparams.dry_sequence_breakers[0] == "\n" ? "\\n" : params.sparams.dry_sequence_breakers[0]) + "'",
+            params.sampling.dry_sequence_breakers.empty() ? "none" :
+            std::accumulate(std::next(params.sampling.dry_sequence_breakers.begin()),
+                params.sampling.dry_sequence_breakers.end(),
+                std::string("'") + (params.sampling.dry_sequence_breakers[0] == "\n" ? "\\n" : params.sampling.dry_sequence_breakers[0]) + "'",
                 [](const std::string& a, const std::string& b) {
                     std::string formatted_b = (b == "\n") ? "\\n" : b;
                     return a + ", '" + formatted_b + "'";
@@ -1040,51 +919,51 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             static bool defaults_cleared = false;
 
             if (!defaults_cleared) {
-                params.sparams.dry_sequence_breakers.clear();
+                params.sampling.dry_sequence_breakers.clear();
                 defaults_cleared = true;
             }
 
             if (value == "none") {
-                params.sparams.dry_sequence_breakers.clear();
+                params.sampling.dry_sequence_breakers.clear();
             } else {
-                params.sparams.dry_sequence_breakers.emplace_back(value);
+                params.sampling.dry_sequence_breakers.emplace_back(value);
             }
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dynatemp-range"}, "N",
-        string_format("dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)params.sparams.dynatemp_range),
+        string_format("dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)params.sampling.dynatemp_range),
         [](common_params & params, const std::string & value) {
-            params.sparams.dynatemp_range = std::stof(value);
+            params.sampling.dynatemp_range = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dynatemp-exp"}, "N",
-        string_format("dynamic temperature exponent (default: %.1f)", (double)params.sparams.dynatemp_exponent),
+        string_format("dynamic temperature exponent (default: %.1f)", (double)params.sampling.dynatemp_exponent),
         [](common_params & params, const std::string & value) {
-            params.sparams.dynatemp_exponent = std::stof(value);
+            params.sampling.dynatemp_exponent = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--mirostat"}, "N",
         string_format("use Mirostat sampling.\nTop K, Nucleus and Locally Typical samplers are ignored if used.\n"
-        "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sparams.mirostat),
+        "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sampling.mirostat),
         [](common_params & params, int value) {
-            params.sparams.mirostat = value;
+            params.sampling.mirostat = value;
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--mirostat-lr"}, "N",
-        string_format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sparams.mirostat_eta),
+        string_format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sampling.mirostat_eta),
         [](common_params & params, const std::string & value) {
-            params.sparams.mirostat_eta = std::stof(value);
+            params.sampling.mirostat_eta = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--mirostat-ent"}, "N",
-        string_format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sparams.mirostat_tau),
+        string_format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sampling.mirostat_tau),
         [](common_params & params, const std::string & value) {
-            params.sparams.mirostat_tau = std::stof(value);
+            params.sampling.mirostat_tau = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1100,7 +979,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             try {
                 if (ss >> key && ss >> sign && std::getline(ss, value_str) && (sign == '+' || sign == '-')) {
                     const float bias = std::stof(value_str) * ((sign == '-') ? -1.0f : 1.0f);
-                    params.sparams.logit_bias.push_back({key, bias});
+                    params.sampling.logit_bias.push_back({key, bias});
                 } else {
                     throw std::invalid_argument("invalid input format");
                 }
@@ -1111,9 +990,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--grammar"}, "GRAMMAR",
-        string_format("BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '%s')", params.sparams.grammar.c_str()),
+        string_format("BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '%s')", params.sampling.grammar.c_str()),
         [](common_params & params, const std::string & value) {
-            params.sparams.grammar = value;
+            params.sampling.grammar = value;
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1127,7 +1006,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             std::copy(
                 std::istreambuf_iterator<char>(file),
                 std::istreambuf_iterator<char>(),
-                std::back_inserter(params.sparams.grammar)
+                std::back_inserter(params.sampling.grammar)
             );
         }
     ).set_sparam());
@@ -1135,7 +1014,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-j", "--json-schema"}, "SCHEMA",
         "JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\nFor schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead",
         [](common_params & params, const std::string & value) {
-            params.sparams.grammar = json_schema_to_grammar(json::parse(value));
+            params.sampling.grammar = json_schema_to_grammar(json::parse(value));
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1444,17 +1323,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_env("LLAMA_ARG_N_GPU_LAYERS"));
-    add_opt(common_arg(
-        {"-ngld", "--gpu-layers-draft", "--n-gpu-layers-draft"}, "N",
-        "number of layers to store in VRAM for the draft model",
-        [](common_params & params, int value) {
-            params.n_gpu_layers_draft = value;
-            if (!llama_supports_gpu_offload()) {
-                fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers-draft option will be ignored\n");
-                fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
     add_opt(common_arg(
         {"-sm", "--split-mode"}, "{none,layer,row}",
         "how to split the model across multiple GPUs, one of:\n"
@@ -1593,13 +1461,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.model = value;
         }
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}).set_env("LLAMA_ARG_MODEL"));
-    add_opt(common_arg(
-        {"-md", "--model-draft"}, "FNAME",
-        "draft model for speculative decoding (default: unused)",
-        [](common_params & params, const std::string & value) {
-            params.model_draft = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
     add_opt(common_arg(
         {"-mu", "--model-url"}, "MODEL_URL",
         "model download url (default: unused)",
@@ -2037,5 +1898,168 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_LOG_TIMESTAMPS"));
 
+    // speculative parameters
+    add_opt(common_arg(
+        {"-td", "--threads-draft"}, "N",
+        "number of threads to use during generation (default: same as --threads)",
+        [](common_params & params, int value) {
+            params.speculative.cpuparams.n_threads = value;
+            if (params.speculative.cpuparams.n_threads <= 0) {
+                params.speculative.cpuparams.n_threads = std::thread::hardware_concurrency();
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"-tbd", "--threads-batch-draft"}, "N",
+        "number of threads to use during batch and prompt processing (default: same as --threads-draft)",
+        [](common_params & params, int value) {
+            params.speculative.cpuparams_batch.n_threads = value;
+            if (params.speculative.cpuparams_batch.n_threads <= 0) {
+                params.speculative.cpuparams_batch.n_threads = std::thread::hardware_concurrency();
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"-Cd", "--cpu-mask-draft"}, "M",
+        "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)",
+        [](common_params & params, const std::string & mask) {
+            params.speculative.cpuparams.mask_valid = true;
+            if (!parse_cpu_mask(mask, params.speculative.cpuparams.cpumask)) {
+                throw std::invalid_argument("invalid cpumask");
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"-Crd", "--cpu-range-draft"}, "lo-hi",
+        "Ranges of CPUs for affinity. Complements --cpu-mask-draft",
+        [](common_params & params, const std::string & range) {
+            params.speculative.cpuparams.mask_valid = true;
+            if (!parse_cpu_range(range, params.speculative.cpuparams.cpumask)) {
+                throw std::invalid_argument("invalid range");
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--cpu-strict-draft"}, "<0|1>",
+        "Use strict CPU placement for draft model (default: same as --cpu-strict)",
+        [](common_params & params, int value) {
+            params.speculative.cpuparams.strict_cpu = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--prio-draft"}, "N",
+        string_format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.speculative.cpuparams.priority),
+        [](common_params & params, int prio) {
+            if (prio < 0 || prio > 3) {
+                throw std::invalid_argument("invalid value");
+            }
+            params.speculative.cpuparams.priority = (enum ggml_sched_priority) prio;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--poll-draft"}, "<0|1>",
+        "Use polling to wait for draft model work (default: same as --poll])",
+        [](common_params & params, int value) {
+            params.speculative.cpuparams.poll = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"-Cbd", "--cpu-mask-batch-draft"}, "M",
+        "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)",
+        [](common_params & params, const std::string & mask) {
+            params.speculative.cpuparams_batch.mask_valid = true;
+            if (!parse_cpu_mask(mask, params.speculative.cpuparams_batch.cpumask)) {
+                throw std::invalid_argument("invalid cpumask");
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"-Crbd", "--cpu-range-batch-draft"}, "lo-hi",
+        "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)",
+        [](common_params & params, const std::string & range) {
+            params.speculative.cpuparams_batch.mask_valid = true;
+            if (!parse_cpu_range(range, params.speculative.cpuparams_batch.cpumask)) {
+                throw std::invalid_argument("invalid cpumask");
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--cpu-strict-batch-draft"}, "<0|1>",
+        "Use strict CPU placement for draft model (default: --cpu-strict-draft)",
+        [](common_params & params, int value) {
+            params.speculative.cpuparams_batch.strict_cpu = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--prio-batch-draft"}, "N",
+        string_format("set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.speculative.cpuparams_batch.priority),
+        [](common_params & params, int prio) {
+            if (prio < 0 || prio > 3) {
+                throw std::invalid_argument("invalid value");
+            }
+            params.speculative.cpuparams_batch.priority = (enum ggml_sched_priority) prio;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--poll-batch-draft"}, "<0|1>",
+        "Use polling to wait for draft model work (default: --poll-draft)",
+        [](common_params & params, int value) {
+            params.speculative.cpuparams_batch.poll = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--draft-max", "--draft", "--draft-n"}, "N",
+        string_format("number of tokens to draft for speculative decoding (default: %d)", params.speculative.n_max),
+        [](common_params & params, int value) {
+            params.speculative.n_max = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--draft-min", "--draft-n-min"}, "N",
+        string_format("minimum number of draft tokens to use for speculative decoding (default: %d)", params.speculative.n_min),
+        [](common_params & params, int value) {
+            params.speculative.n_min = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--draft-p-split"}, "P",
+        string_format("speculative decoding split probability (default: %.1f)", (double)params.speculative.p_split),
+        [](common_params & params, const std::string & value) {
+            params.speculative.p_split = std::stof(value);
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    add_opt(common_arg(
+        {"--draft-p-min"}, "P",
+        string_format("minimum speculative decoding probability (greedy) (default: %.1f)", (double)params.speculative.p_min),
+        [](common_params & params, const std::string & value) {
+            params.speculative.p_min = std::stof(value);
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-cd", "--ctx-size-draft"}, "N",
+        string_format("size of the prompt context for the draft model (default: %d, 0 = loaded from model)", params.speculative.n_ctx),
+        [](common_params & params, int value) {
+            params.speculative.n_ctx = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-ngld", "--gpu-layers-draft", "--n-gpu-layers-draft"}, "N",
+        "number of layers to store in VRAM for the draft model",
+        [](common_params & params, int value) {
+            params.speculative.n_gpu_layers = value;
+            if (!llama_supports_gpu_offload()) {
+                fprintf(stderr, "warning: not compiled with GPU offload support, --gpu-layers-draft option will be ignored\n");
+                fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-md", "--model-draft"}, "FNAME",
+        "draft model for speculative decoding (default: unused)",
+        [](common_params & params, const std::string & value) {
+            params.speculative.model = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+
     return ctx_arg;
 }
diff --git a/common/common.cpp b/common/common.cpp
index d314523db4c62..c398329d05bf5 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -536,12 +536,12 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
                     [](const unsigned char c) { return !std::isprint(c); }),
                 detokenized.end());
 
-        buf << "\n" << std::to_string(i)
-            << ":token '" << detokenized << "'"
-            << ":pos " << std::to_string(batch.pos[i])
-            << ":n_seq_id  " << std::to_string(batch.n_seq_id[i])
-            << ":seq_id " << std::to_string(batch.seq_id[i][0])
-            << ":logits " << std::to_string(batch.logits[i]);
+        buf << "\n"          << std::to_string(i)
+            << ", token '"   << detokenized << "'"
+            << ", pos "      << std::to_string(batch.pos[i])
+            << ", n_seq_id " << std::to_string(batch.n_seq_id[i])
+            << ", seq_id "   << std::to_string(batch.seq_id[i][0])
+            << ", logits "   << std::to_string(batch.logits[i]);
     }
 
     buf << " ]";
@@ -925,9 +925,9 @@ struct common_init_result common_init_from_params(common_params & params) {
         common_lora_adapters_apply(lctx, iparams.lora_adapters);
     }
 
-    if (params.sparams.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
+    if (params.sampling.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
         LOG_WRN("%s: warning: model does not have an EOS token, ignoring --ignore-eos\n", __func__);
-        params.sparams.ignore_eos = false;
+        params.sampling.ignore_eos = false;
     }
 
     if (params.warmup) {
@@ -1490,6 +1490,66 @@ void common_batch_add(
     batch.n_tokens++;
 }
 
+//
+// Token utils
+//
+
+size_t common_lcp(const llama_tokens & a, const llama_tokens & b) {
+    size_t i;
+    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
+
+    return i;
+}
+
+size_t common_lcs(const llama_tokens & a, const llama_tokens & b) {
+    // check for empty sequences
+    if (a.empty() || b.empty()) {
+        return 0;
+    }
+
+    // get the lengths of the input sequences
+    size_t a_len = a.size();
+    size_t b_len = b.size();
+
+    // initialize the maximum length of the longest common subsequence (LCS)
+    size_t max_length = 0;
+
+    // use two rows instead of a 2D matrix to optimize space
+    std::vector<size_t> prev_row(b_len + 1, 0);
+    std::vector<size_t> curr_row(b_len + 1, 0);
+
+    // iterate through the elements of a
+    for (size_t i = 1; i <= a_len; i++) {
+        // iterate through the elements of b
+        for (size_t j = 1; j <= b_len; j++) {
+            // if elements at the current positions match
+            if (a[i - 1] == b[j - 1]) {
+                // if it's the first element of either sequences, set LCS length to 1
+                if (i == 1 || j == 1) {
+                    curr_row[j] = 1;
+                } else {
+                    // increment LCS length by 1 compared to the previous element
+                    curr_row[j] = prev_row[j - 1] + 1;
+                }
+
+                // update max_length if necessary
+                if (curr_row[j] > max_length) {
+                    max_length = curr_row[j];
+                }
+            } else {
+                // reset LCS length if elements don't match
+                curr_row[j] = 0;
+            }
+        }
+
+        // update the previous row for the next iteration
+        prev_row = curr_row;
+    }
+
+    // return the maximum length of the LCS
+    return max_length;
+}
+
 //
 // Vocab utils
 //
diff --git a/common/common.h b/common/common.h
index 7977cc7a99a78..5c579b5abfe03 100644
--- a/common/common.h
+++ b/common/common.h
@@ -33,6 +33,8 @@ struct common_lora_adapter_container : common_lora_adapter_info {
     struct llama_lora_adapter * adapter;
 };
 
+using llama_tokens = std::vector<llama_token>;
+
 // build info
 extern int LLAMA_BUILD_NUMBER;
 extern char const * LLAMA_COMMIT;
@@ -101,8 +103,8 @@ enum dimre_method {
     DIMRE_METHOD_MEAN,
 };
 
-// sampler parameters
-struct common_sampler_params {
+// sampling parameters
+struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
 
     int32_t n_prev             = 64;    // number of previous tokens to remember
@@ -153,19 +155,30 @@ struct common_sampler_params {
     std::string print() const;
 };
 
+struct common_params_speculative {
+    int32_t n_ctx        =     0; // draft context size
+    int32_t n_max        =    16; // maximum number of tokens to draft during speculative decoding
+    int32_t n_min        =     5; // minimum number of draft tokens to use for speculative decoding
+    int32_t n_gpu_layers =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
+    float   p_split      =  0.1f; // speculative decoding split probability
+    float   p_min        =  0.9f; // minimum speculative decoding probability (greedy)
+
+    struct cpu_params cpuparams;
+    struct cpu_params cpuparams_batch;
+
+    std::string model = ""; // draft model for speculative decoding                          // NOLINT
+};
+
 struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =  4096; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt
-    int32_t n_draft               =     5; // number of tokens to draft during speculative decoding
     int32_t n_chunks              =    -1; // max number of chunks to process (-1 = unlimited)
     int32_t n_parallel            =     1; // number of parallel sequences to decode
     int32_t n_sequences           =     1; // number of sequences to decode
-    float   p_split               =  0.1f; // speculative decoding split probability
     int32_t n_gpu_layers          =    -1; // number of layers to store in VRAM (-1 - use default)
-    int32_t n_gpu_layers_draft    =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
     int32_t main_gpu              =     0; // the GPU that is used for scratch and small tensors
     float   tensor_split[128]     =   {0}; // how split tensors should be distributed across GPUs
     int32_t grp_attn_n            =     1; // group-attention factor
@@ -182,8 +195,6 @@ struct common_params {
 
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
-    struct cpu_params draft_cpuparams;
-    struct cpu_params draft_cpuparams_batch;
 
     ggml_backend_sched_eval_callback cb_eval = nullptr;
     void * cb_eval_user_data                 = nullptr;
@@ -195,10 +206,10 @@ struct common_params {
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
     enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
 
-    struct common_sampler_params sparams;
+    struct common_params_sampling sampling;
+    struct common_params_speculative speculative;
 
     std::string model                = ""; // model path                                                    // NOLINT
-    std::string model_draft          = ""; // draft model for speculative decoding                          // NOLINT
     std::string model_alias          = "unknown"; // model alias                                            // NOLINT
     std::string model_url            = ""; // model url to download                                         // NOLINT
     std::string hf_token             = ""; // HF token                                                      // NOLINT
@@ -461,7 +472,9 @@ struct llama_model * common_load_model_from_hf(const char * repo, const char * f
 // clear LoRA adapters from context, then apply new list of adapters
 void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters);
 
+//
 // Batch utils
+//
 
 void common_batch_clear(struct llama_batch & batch);
 
@@ -472,6 +485,16 @@ void common_batch_add(
     const std::vector<llama_seq_id> & seq_ids,
                                bool   logits);
 
+//
+// Token utils
+//
+
+// longest common prefix
+size_t common_lcp(const llama_tokens & a, const llama_tokens & b);
+
+// longet common subsequence
+size_t common_lcs(const llama_tokens & a, const llama_tokens & b);
+
 //
 // Vocab utils
 //
diff --git a/common/sampling.cpp b/common/sampling.cpp
index 7922fde47d369..0c4699a89c8b2 100644
--- a/common/sampling.cpp
+++ b/common/sampling.cpp
@@ -99,7 +99,7 @@ struct ring_buffer {
 };
 
 struct common_sampler {
-    common_sampler_params params;
+    common_params_sampling params;
 
     struct llama_sampler * grmr;
     struct llama_sampler * chain;
@@ -125,7 +125,7 @@ struct common_sampler {
     }
 };
 
-std::string common_sampler_params::print() const {
+std::string common_params_sampling::print() const {
     char result[1024];
 
     snprintf(result, sizeof(result),
@@ -141,7 +141,7 @@ std::string common_sampler_params::print() const {
     return std::string(result);
 }
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_sampler_params & params) {
+struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
     llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
 
     lparams.no_perf = params.no_perf;
@@ -320,6 +320,45 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
     return cur_p.data[cur_p.selected].id;
 }
 
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first) {
+    GGML_ASSERT(idxs.size() == draft.size() + 1 && "idxs.size() must be draft.size() + 1");
+
+    std::vector<llama_token> result;
+    result.reserve(idxs.size());
+
+    size_t i = 0;
+    for (; i < draft.size(); i++) {
+        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
+
+        common_sampler_accept(gsmpl, id, true);
+
+        result.push_back(id);
+
+        if (draft[i] != id) {
+            break;
+        }
+    }
+
+    if (i == draft.size()) {
+        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
+
+        common_sampler_accept(gsmpl, id, true);
+
+        result.push_back(id);
+    }
+
+    return result;
+}
+
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first) {
+    std::vector<int> idxs(draft.size() + 1);
+    for (size_t i = 0; i < idxs.size(); ++i) {
+        idxs[i] = i;
+    }
+
+    return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft, grammar_first);
+}
+
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
     return llama_sampler_get_seed(gsmpl->chain);
 }
diff --git a/common/sampling.h b/common/sampling.h
index d37f25ad37c4a..348911b18888b 100644
--- a/common/sampling.h
+++ b/common/sampling.h
@@ -36,7 +36,7 @@ struct common_sampler;
 
 // llama_sampler API overloads
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_sampler_params & params);
+struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params);
 
 void common_sampler_free(struct common_sampler * gsmpl);
 
@@ -60,6 +60,27 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
 //
 llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
 
+// generalized version of common_sampler_sample
+//
+// will cross-reference the sampled tokens with a batch of draft tokens and accept those that match
+// if the sampler disagrees at some point, we stop and return the accepted tokens up to now
+//
+//      common_sampler_sample_n(gsmpl, ctx, { idx }, {});
+//
+// is equivalent to
+//
+//      common_sampler_sample(gsmpl, ctx, idx);
+//      common_sampler_accept(gsmpl, token, true);
+//
+// requires: idxs.size() == draft.size() + 1
+//
+// returns at least 1 token, up to idxs.size()
+//
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first = false);
+
+// assume idxs == [ 0, 1, 2, ..., draft.size() ]
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first = false);
+
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
 
 // helpers
diff --git a/common/speculative.cpp b/common/speculative.cpp
new file mode 100644
index 0000000000000..fe315a2703e9c
--- /dev/null
+++ b/common/speculative.cpp
@@ -0,0 +1,269 @@
+#include "speculative.h"
+
+#include "log.h"
+#include "common.h"
+#include "sampling.h"
+
+#include <cstring>
+
+#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
+#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
+
+struct common_speculative {
+    struct llama_context * ctx;
+    struct common_sampler * smpl;
+
+    llama_batch batch;
+    llama_tokens prompt;
+};
+
+struct common_speculative * common_speculative_init(
+        struct llama_context * ctx_dft) {
+    auto * result = new common_speculative {
+        /* .ctx    = */ ctx_dft,
+        /* .smpl   = */ nullptr,
+        /* .batch  = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
+        /* .prompt = */ {},
+    };
+
+    // TODO: optimize or pass from outside?
+#if 0
+    {
+        common_params_sampling params;
+        params.no_perf = false;
+
+        params.top_k = 40;
+        params.top_p = 0.9;
+
+        params.samplers = {
+            COMMON_SAMPLER_TYPE_TOP_K,
+            COMMON_SAMPLER_TYPE_TOP_P,
+            COMMON_SAMPLER_TYPE_INFILL,
+        };
+
+        result->smpl = common_sampler_init(llama_get_model(ctx_dft), params);
+    }
+#else
+    {
+        common_params_sampling params;
+        params.no_perf = false;
+
+        params.top_k = 10;
+
+        params.samplers = {
+            COMMON_SAMPLER_TYPE_TOP_K,
+        };
+
+        result->smpl = common_sampler_init(llama_get_model(ctx_dft), params);
+    }
+#endif
+
+    return result;
+}
+
+void common_speculative_free(struct common_speculative * spec) {
+    common_sampler_free(spec->smpl);
+
+    llama_batch_free(spec->batch);
+
+    delete spec;
+}
+
+bool common_speculative_are_compatible(
+        const struct llama_context * ctx_tgt,
+        const struct llama_context * ctx_dft) {
+    const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
+    const struct llama_model * model_dft = llama_get_model(ctx_dft);
+
+    const bool vocab_type_tgt = llama_vocab_type(model_tgt);
+    LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
+
+    const bool vocab_type_dft = llama_vocab_type(model_dft);
+    LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
+
+    if (vocab_type_tgt != vocab_type_dft) {
+        LOG_ERR("%s: draft model vocab type must match target model to use speculation but "
+                     "vocab_type_dft = %d while vocab_type_tgt = %d\n", __func__, vocab_type_dft, vocab_type_tgt);
+        return false;
+    }
+
+    if (llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
+        llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
+        llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
+        llama_token_eos(model_tgt) != llama_token_eos(model_dft)
+    ) {
+        LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
+        return false;
+    }
+
+    {
+        const int n_vocab_tgt = llama_n_vocab(model_tgt);
+        const int n_vocab_dft = llama_n_vocab(model_dft);
+
+        const int vocab_diff = std::abs(n_vocab_tgt - n_vocab_dft);
+
+        if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
+            LOG_ERR("%s: draft model vocab must closely match target model to use speculation but "
+                         "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
+                    __func__, n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
+            return false;
+        }
+
+        for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
+            const char * token_text_tgt = llama_token_get_text(model_tgt, i);
+            const char * token_text_dft = llama_token_get_text(model_dft, i);
+            if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
+                LOG_ERR("%s: draft model vocab must match target model to use speculation but "
+                             "token %d content differs - target '%s', draft '%s'\n", __func__, i,
+                        common_token_to_piece(ctx_tgt, i).c_str(),
+                        common_token_to_piece(ctx_dft, i).c_str());
+                return false;
+            }
+        }
+    }
+
+    return true;
+}
+
+llama_tokens common_speculative_gen_draft(
+        struct common_speculative * spec,
+        struct common_speculative_params params,
+        const llama_tokens & prompt_tgt,
+        llama_token id_last) {
+    auto & batch  = spec->batch;
+    auto & ctx    = spec->ctx;
+    auto & smpl   = spec->smpl;
+    auto & prompt = spec->prompt;
+
+    int reuse_i = 0;
+    int reuse_n = 0;
+
+    const int n_ctx = llama_n_ctx(ctx) - params.n_draft;
+
+    const int i_start = std::max<int>(0, (int) prompt_tgt.size() - n_ctx);
+
+    // reuse as much as possible from the old draft context
+    // ideally, the draft context should be as big as the target context and we will always reuse the entire prompt
+    for (int i = 0; i < (int) prompt.size(); ++i) {
+        int cur = 0;
+        while (i_start + cur < (int) prompt_tgt.size() &&
+               i       + cur < (int) prompt.size() &&
+               prompt_tgt[i_start + cur] == prompt[i + cur]) {
+            cur++;
+        }
+
+        if ((cur >= params.n_reuse || n_ctx >= (int) prompt_tgt.size()) && cur > reuse_n) {
+            reuse_i = i;
+            reuse_n = cur;
+        }
+    }
+
+    LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt.size());
+
+    llama_tokens result;
+    result.reserve(params.n_draft);
+
+    if (reuse_n == 0) {
+        llama_kv_cache_clear(ctx);
+
+        prompt.clear();
+    } else {
+        // this happens when a previous draft has been discarded (for example, due to being too small), but the
+        // target model agreed with it. in this case, we simply pass back the previous results to save compute
+        if (reuse_i + reuse_n < (int) prompt.size() && prompt[reuse_i + reuse_n] == id_last) {
+            for (int i = reuse_i + reuse_n + 1; i < (int) prompt.size(); ++i) {
+                result.push_back(prompt[i]);
+
+                if (params.n_draft <= (int) result.size()) {
+                    break;
+                }
+            }
+
+            return result;
+        }
+
+        if (reuse_i > 0) {
+            llama_kv_cache_seq_rm (ctx, 0, 0, reuse_i);
+            llama_kv_cache_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
+
+            prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
+        }
+
+        if (reuse_n < (int) prompt.size()) {
+            llama_kv_cache_seq_rm (ctx, 0, reuse_n, -1);
+
+            prompt.erase(prompt.begin() + reuse_n, prompt.end());
+        }
+    }
+
+    // prepare a batch to evaluate any new tokens in the prompt
+    common_batch_clear(batch);
+
+    for (size_t i = i_start + reuse_n; i < prompt_tgt.size(); ++i) {
+        //LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_tgt[i]);
+        common_batch_add(batch, prompt_tgt[i], i - i_start, { 0 }, false);
+
+        prompt.push_back(prompt_tgt[i]);
+    }
+
+    // we should rarely end-up here during normal decoding
+    if (batch.n_tokens > 0) {
+        //LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str());
+
+        llama_decode(ctx, batch);
+    }
+
+    const llama_pos n_past = prompt.size();
+
+    LOG_DBG("%s: n_past = %d\n", __func__, n_past);
+
+    common_batch_clear(batch);
+    common_batch_add  (batch, id_last, n_past, { 0 }, true);
+
+    prompt.push_back(id_last);
+
+    //LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx, prompt).c_str());
+
+    llama_decode(ctx, batch);
+
+    common_sampler_reset(smpl);
+
+    // sample n_draft tokens from the draft model
+    for (int i = 0; i < params.n_draft; ++i) {
+        common_batch_clear(batch);
+
+        common_sampler_sample(smpl, ctx, 0, true);
+
+        const auto * cur_p = common_sampler_get_candidates(smpl);
+
+        for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
+            LOG_DBG(" - draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
+                    k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx, cur_p->data[k].id).c_str());
+        }
+
+        // add drafted token for each sequence
+        const llama_token id = cur_p->data[0].id;
+
+        // only collect very high-confidence draft tokens
+        if (cur_p->data[0].p < params.p_min) {
+            break;
+        }
+
+        common_sampler_accept(smpl, id, true);
+
+        result.push_back(id);
+
+        if (params.n_draft <= (int) result.size()) {
+            break;
+        }
+
+        common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
+
+        // evaluate the drafted tokens on the draft model
+        llama_decode(ctx, batch);
+
+        prompt.push_back(id);
+    }
+
+    return result;
+}
diff --git a/common/speculative.h b/common/speculative.h
new file mode 100644
index 0000000000000..50ec0344618aa
--- /dev/null
+++ b/common/speculative.h
@@ -0,0 +1,28 @@
+#pragma once
+
+#include "llama.h"
+#include "common.h"
+
+struct common_speculative;
+
+struct common_speculative_params {
+    int n_draft = 16;  // max drafted tokens
+    int n_reuse = 256;
+
+    float p_min = 0.9f; // min probabiliy required to accept a token in the draft
+};
+
+struct common_speculative * common_speculative_init(struct llama_context * ctx_dft);
+
+void common_speculative_free(struct common_speculative * spec);
+
+bool common_speculative_are_compatible(
+        const struct llama_context * ctx_tgt,
+        const struct llama_context * ctx_dft);
+
+// sample up to n_draft tokens and add them to the batch using the draft model
+llama_tokens common_speculative_gen_draft(
+               struct common_speculative * spec,
+        struct common_speculative_params   params,
+                      const llama_tokens & prompt,
+                             llama_token   id_last);
diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index d63a96c1c2547..9bd099d4ef8a5 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -50,5 +50,6 @@ else()
     add_subdirectory(simple)
     add_subdirectory(simple-chat)
     add_subdirectory(speculative)
+    add_subdirectory(speculative-simple)
     add_subdirectory(tokenize)
 endif()
diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp
index 3b554033e7ee4..ba219cd4b32ae 100644
--- a/examples/batched/batched.cpp
+++ b/examples/batched/batched.cpp
@@ -68,10 +68,10 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl, llama_sampler_init_top_k(params.sparams.top_k));
-    llama_sampler_chain_add(smpl, llama_sampler_init_top_p(params.sparams.top_p, params.sparams.min_keep));
-    llama_sampler_chain_add(smpl, llama_sampler_init_temp (params.sparams.temp));
-    llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sparams.seed));
+    llama_sampler_chain_add(smpl, llama_sampler_init_top_k(params.sampling.top_k));
+    llama_sampler_chain_add(smpl, llama_sampler_init_top_p(params.sampling.top_p, params.sampling.min_keep));
+    llama_sampler_chain_add(smpl, llama_sampler_init_temp (params.sampling.temp));
+    llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sampling.seed));
 
     if (ctx == NULL) {
         LOG_ERR("%s: error: failed to create the llama_context\n" , __func__);
diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp
index 15b358dc4e854..ef700895720ff 100644
--- a/examples/infill/infill.cpp
+++ b/examples/infill/infill.cpp
@@ -73,7 +73,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    auto & sparams = params.sparams;
+    auto & sparams = params.sampling;
 
     console::init(params.simple_io, params.use_color);
     atexit([]() { console::cleanup(); });
diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp
index 1610985858fc9..2691c6e6b2dd2 100644
--- a/examples/llava/llava-cli.cpp
+++ b/examples/llava/llava-cli.cpp
@@ -191,7 +191,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
 
     LOG("\n");
 
-    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sparams);
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
     if (!smpl) {
         LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
         exit(1);
diff --git a/examples/llava/minicpmv-cli.cpp b/examples/llava/minicpmv-cli.cpp
index cbecec343c640..e9cbb51ed90ab 100644
--- a/examples/llava/minicpmv-cli.cpp
+++ b/examples/llava/minicpmv-cli.cpp
@@ -237,7 +237,7 @@ static struct common_sampler * llama_init(struct llava_context * ctx_llava, comm
 
     LOG_INF("\n");
 
-    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sparams);
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
     return smpl;
 }
 
diff --git a/examples/lookahead/lookahead.cpp b/examples/lookahead/lookahead.cpp
index 3c0ccfea2ccd7..8d0ef8b3d75e6 100644
--- a/examples/lookahead/lookahead.cpp
+++ b/examples/lookahead/lookahead.cpp
@@ -115,7 +115,7 @@ int main(int argc, char ** argv) {
     llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1);
 
     // target model sampling context
-    struct common_sampler * smpl = common_sampler_init(model, params.sparams);
+    struct common_sampler * smpl = common_sampler_init(model, params.sampling);
 
     // verification n-grams
     std::vector<ngram_data> ngrams_cur(G);
diff --git a/examples/lookup/lookup-stats.cpp b/examples/lookup/lookup-stats.cpp
index 7faebe7ba11fc..dff07c075c47f 100644
--- a/examples/lookup/lookup-stats.cpp
+++ b/examples/lookup/lookup-stats.cpp
@@ -21,7 +21,7 @@ int main(int argc, char ** argv){
 
     common_init();
 
-    const int n_draft = params.n_draft;
+    const int n_draft = params.speculative.n_max;
 
     // init llama.cpp
     llama_backend_init();
@@ -40,6 +40,7 @@ int main(int argc, char ** argv){
     common_ngram_cache ngram_cache_context;
     common_ngram_cache ngram_cache_dynamic;
     common_ngram_cache ngram_cache_static;
+
     int64_t t_draft_flat_us = 0;
     int64_t t_draft_us = 0;
 
diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp
index a04728b1834cc..4d92bb2385358 100644
--- a/examples/lookup/lookup.cpp
+++ b/examples/lookup/lookup.cpp
@@ -22,7 +22,7 @@ int main(int argc, char ** argv){
     common_init();
 
     // max. number of additional tokens to draft if match is found
-    const int n_draft = params.n_draft;
+    const int n_draft = params.speculative.n_max;
 
     const bool dump_kv_cache = params.dump_kv_cache;
 
@@ -102,7 +102,7 @@ int main(int argc, char ** argv){
 
     bool has_eos = false;
 
-    struct common_sampler * smpl = common_sampler_init(model, params.sparams);
+    struct common_sampler * smpl = common_sampler_init(model, params.sampling);
 
     std::vector<llama_token> draft;
 
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index 7c4ce4be2abae..957451af7ce0a 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -100,7 +100,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    auto & sparams = params.sparams;
+    auto & sparams = params.sampling;
 
     // save choice to use color for later
     // (note for later: this is a slightly awkward choice)
diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp
index 43c8f3ed56ba9..fd2b1c0112838 100644
--- a/examples/parallel/parallel.cpp
+++ b/examples/parallel/parallel.cpp
@@ -160,7 +160,7 @@ int main(int argc, char ** argv) {
     for (size_t i = 0; i < clients.size(); ++i) {
         auto & client = clients[i];
         client.id = i;
-        client.smpl = common_sampler_init(model, params.sparams);
+        client.smpl = common_sampler_init(model, params.sampling);
     }
 
     std::vector<llama_token> tokens_system;
diff --git a/examples/retrieval/retrieval.cpp b/examples/retrieval/retrieval.cpp
index 1768aae510067..e78a8596d8cfe 100644
--- a/examples/retrieval/retrieval.cpp
+++ b/examples/retrieval/retrieval.cpp
@@ -282,8 +282,8 @@ int main(int argc, char ** argv) {
                 return a.second > b.second;
             });
 
-            LOG("Top %d similar chunks:\n", params.sparams.top_k);
-            for (int i = 0; i < std::min(params.sparams.top_k, (int) chunks.size()); i++) {
+            LOG("Top %d similar chunks:\n", params.sampling.top_k);
+            for (int i = 0; i < std::min(params.sampling.top_k, (int) chunks.size()); i++) {
                 LOG("filename: %s\n", chunks[similarities[i].first].filename.c_str());
                 LOG("filepos: %lld\n", (long long int) chunks[similarities[i].first].filepos);
                 LOG("similarity: %f\n", similarities[i].second);
diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp
index 8c49a52a66124..2f0cf9baa32b7 100644
--- a/examples/save-load-state/save-load-state.cpp
+++ b/examples/save-load-state/save-load-state.cpp
@@ -9,7 +9,7 @@ int main(int argc, char ** argv) {
     common_params params;
 
     params.prompt = "The quick brown fox";
-    params.sparams.seed = 1234;
+    params.sampling.seed = 1234;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
@@ -42,7 +42,7 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sparams.seed));
+    llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sampling.seed));
 
     // tokenize prompt
     auto tokens = common_tokenize(ctx, params.prompt, true);
@@ -106,7 +106,7 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl2, llama_sampler_init_dist(params.sparams.seed));
+    llama_sampler_chain_add(smpl2, llama_sampler_init_dist(params.sampling.seed));
 
     printf("\nsecond run: %s", params.prompt.c_str());
 
@@ -169,7 +169,7 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl3, llama_sampler_init_dist(params.sparams.seed));
+    llama_sampler_chain_add(smpl3, llama_sampler_init_dist(params.sampling.seed));
 
     printf("\nsingle seq run: %s", params.prompt.c_str());
 
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index b8e003be9730e..6c55d65c01330 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -175,7 +175,7 @@ struct server_slot {
     // sampling
     json json_schema;
 
-    struct common_sampler_params sparams;
+    struct common_params_sampling sparams;
     struct common_sampler * smpl = nullptr;
 
     llama_token sampled;
@@ -687,7 +687,7 @@ struct server_context {
 
             SLT_INF(slot, "new slot n_ctx_slot = %d\n", slot.n_ctx);
 
-            slot.sparams = params.sparams;
+            slot.sparams = params.sampling;
 
             slot.callback_on_release = [this](int) {
                 queue_tasks.pop_deferred_task();
@@ -743,7 +743,7 @@ struct server_context {
                 }
 
                 // length of the Longest Common Subsequence between the current slot's prompt and the input prompt
-                int cur_lcs_len = longest_common_subsequence(slot.cache_tokens, task.prompt_tokens);
+                int cur_lcs_len = common_lcs(slot.cache_tokens, task.prompt_tokens);
 
                 // fraction of the common subsequence length compared to the current slot's prompt length
                 float cur_similarity = static_cast<float>(cur_lcs_len) / static_cast<int>(slot.cache_tokens.size());
@@ -788,7 +788,7 @@ struct server_context {
     bool launch_slot_with_task(server_slot & slot, const server_task & task) {
         slot_params default_params;
         // Sampling parameter defaults are loaded from the global server context (but individual requests can still override them)
-        auto default_sparams = params.sparams;
+        auto default_sparams = params.sampling;
         const auto & data = task.data;
 
         if (data.count("__oaicompat") != 0) {
@@ -1960,7 +1960,7 @@ struct server_context {
 
                             if (slot.params.cache_prompt) {
                                 // reuse any previously computed tokens that are common with the new prompt
-                                slot.n_past = longest_common_prefix(slot.cache_tokens, prompt_tokens);
+                                slot.n_past = common_lcp(slot.cache_tokens, prompt_tokens);
 
                                 // reuse chunks from the cached prompt by shifting their KV cache in the new position
                                 if (params.n_cache_reuse > 0) {
diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp
index c47ed3e47a76d..1665e9dc37db6 100644
--- a/examples/server/utils.hpp
+++ b/examples/server/utils.hpp
@@ -24,7 +24,6 @@
 #define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
 
 using json = nlohmann::ordered_json;
-using llama_tokens = std::vector<llama_token>;
 
 #define SLT_INF(slot, fmt, ...) LOG_INF("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
 #define SLT_WRN(slot, fmt, ...) LOG_WRN("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
@@ -439,62 +438,6 @@ static std::string gen_chatcmplid() {
 // other common utils
 //
 
-static size_t longest_common_prefix(const llama_tokens & a, const llama_tokens & b) {
-    size_t i;
-    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
-
-    return i;
-}
-
-static size_t longest_common_subsequence(const llama_tokens & a, const llama_tokens & b) {
-    // check for empty sequences
-    if (a.empty() || b.empty()) {
-        return 0;
-    }
-
-    // get the lengths of the input sequences
-    size_t a_len = a.size();
-    size_t b_len = b.size();
-
-    // initialize the maximum length of the longest common subsequence (LCS)
-    size_t max_length = 0;
-
-    // use two rows instead of a 2D matrix to optimize space
-    std::vector<size_t> prev_row(b_len + 1, 0);
-    std::vector<size_t> curr_row(b_len + 1, 0);
-
-    // iterate through the elements of a
-    for (size_t i = 1; i <= a_len; i++) {
-        // iterate through the elements of b
-        for (size_t j = 1; j <= b_len; j++) {
-            // if elements at the current positions match
-            if (a[i - 1] == b[j - 1]) {
-                // if it's the first element of either sequences, set LCS length to 1
-                if (i == 1 || j == 1) {
-                    curr_row[j] = 1;
-                } else {
-                    // increment LCS length by 1 compared to the previous element
-                    curr_row[j] = prev_row[j - 1] + 1;
-                }
-
-                // update max_length if necessary
-                if (curr_row[j] > max_length) {
-                    max_length = curr_row[j];
-                }
-            } else {
-                // reset LCS length if elements don't match
-                curr_row[j] = 0;
-            }
-        }
-
-        // update the previous row for the next iteration
-        prev_row = curr_row;
-    }
-
-    // return the maximum length of the LCS
-    return max_length;
-}
-
 static bool ends_with(const std::string & str, const std::string & suffix) {
     return str.size() >= suffix.size() && 0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
 }
diff --git a/examples/speculative-simple/CMakeLists.txt b/examples/speculative-simple/CMakeLists.txt
new file mode 100644
index 0000000000000..7a3a141c27994
--- /dev/null
+++ b/examples/speculative-simple/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET llama-speculative-simple)
+add_executable(${TARGET} speculative-simple.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/examples/speculative-simple/README.md b/examples/speculative-simple/README.md
new file mode 100644
index 0000000000000..e3a6c6b4aa0bf
--- /dev/null
+++ b/examples/speculative-simple/README.md
@@ -0,0 +1,12 @@
+# llama.cpp/examples/speculative-simple
+
+Demonstration of basic greedy speculative decoding
+
+```bash
+./bin/llama-speculative-simple \
+    -m  ../models/qwen2.5-32b-coder-instruct/ggml-model-q8_0.gguf \
+    -md ../models/qwen2.5-1.5b-coder-instruct/ggml-model-q4_0.gguf \
+    -f test.txt -c 0 -ngl 99 --color \
+    --sampling-seq k --top-k 1 -fa --temp 0.0 \
+    -ngld 99 --draft-max 16 --draft-min 5 --draft-p-min 0.9
+```
diff --git a/examples/speculative-simple/speculative-simple.cpp b/examples/speculative-simple/speculative-simple.cpp
new file mode 100644
index 0000000000000..1bc7f428cf3b5
--- /dev/null
+++ b/examples/speculative-simple/speculative-simple.cpp
@@ -0,0 +1,273 @@
+#include "arg.h"
+#include "common.h"
+#include "sampling.h"
+#include "speculative.h"
+#include "log.h"
+#include "llama.h"
+
+#include <cstdio>
+#include <cstring>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
+        return 1;
+    }
+
+    if (params.n_predict < -1) {
+        LOG_ERR("%s: --n-predict must be >= -1\n", __func__);
+        return 1;
+    }
+
+    common_init();
+
+    if (params.speculative.model.empty()) {
+        LOG_ERR("%s: --model-draft is required\n", __func__);
+        return 1;
+    }
+
+    // init llama.cpp
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model_tgt = NULL;
+    llama_model * model_dft = NULL;
+
+    llama_context * ctx_tgt = NULL;
+    llama_context * ctx_dft = NULL;
+
+    // load the target model
+    common_init_result llama_init_tgt = common_init_from_params(params);
+
+    model_tgt = llama_init_tgt.model;
+    ctx_tgt   = llama_init_tgt.context;
+
+    // load the draft model
+    params.model        = params.speculative.model;
+    params.n_ctx        = params.speculative.n_ctx;
+    params.n_batch      = params.speculative.n_ctx > 0 ? params.speculative.n_ctx : params.n_batch;
+    params.n_gpu_layers = params.speculative.n_gpu_layers;
+
+    if (params.speculative.cpuparams.n_threads > 0) {
+        params.cpuparams.n_threads = params.speculative.cpuparams.n_threads;
+    }
+
+    params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
+    common_init_result llama_init_dft = common_init_from_params(params);
+
+    model_dft = llama_init_dft.model;
+    ctx_dft   = llama_init_dft.context;
+
+    if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) {
+        return 1;
+    }
+
+    // Tokenize the prompt
+    std::vector<llama_token> inp;
+    inp = common_tokenize(ctx_tgt, params.prompt, true, true);
+
+    if (llama_n_ctx(ctx_tgt) < (int) inp.size()) {
+        LOG_ERR("%s: the prompt exceeds the context size (%d tokens, ctx %d)\n", __func__, (int) inp.size(), llama_n_ctx(ctx_tgt));
+
+        return 1;
+    }
+
+    if (llama_n_batch(ctx_tgt) < (int) inp.size()) {
+        LOG_ERR("%s: the prompt exceeds the batch size (%d tokens, batch %d)\n", __func__, (int) inp.size(), llama_n_batch(ctx_tgt));
+
+        return 1;
+    }
+
+    LOG("\n\n");
+
+    for (auto id : inp) {
+        LOG("%s", common_token_to_piece(ctx_tgt, id).c_str());
+    }
+
+    // how many tokens to draft each time
+    int n_draft     = params.speculative.n_max;
+    int n_draft_min = params.speculative.n_min;
+
+    float p_min = params.speculative.p_min;
+
+    int n_predict = 0;
+    int n_drafted = 0;
+    int n_accept  = 0;
+
+    // used to determine end of generation
+    bool has_eos = false;
+
+    // ================================================
+    // everything until here is standard initialization
+    // the relevant stuff for speculative decoding starts here
+
+    const auto t_enc_start = ggml_time_us();
+
+    // target model sampling context
+    struct common_sampler * smpl = common_sampler_init(model_tgt, params.sampling);
+
+    // eval the prompt
+    llama_decode(ctx_tgt, llama_batch_get_one(inp.data(), inp.size() - 1));
+
+    // note: keep the last token separate!
+    llama_token id_last = inp.back();
+
+    // all tokens currently in the target context
+    auto prompt_tgt = std::vector<llama_token>(inp.begin(), inp.end() - 1);
+
+    int n_past = inp.size() - 1;
+
+    // init the speculator
+    struct common_speculative_params params_spec;
+    params_spec.n_draft = n_draft;
+    params_spec.n_reuse = llama_n_ctx(ctx_dft) - n_draft;
+    params_spec.p_min   = p_min;
+
+    struct common_speculative * spec = common_speculative_init(ctx_dft);
+
+    llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);
+
+    const auto t_enc_end = ggml_time_us();
+
+    const auto t_dec_start = ggml_time_us();
+
+    while (true) {
+        // optionally, generate draft tokens that can be appended to the target batch
+        //
+        // this is the most important part of the speculation. the more probable tokens that are provided here
+        // the better the performance will be. in theory, this computation can be performed asynchronously and even
+        // offloaded to a remote device. it doesn't even have to be based on an LLM. instead, it can provide tokens
+        // from a cache or lookup tables.
+        //
+        llama_tokens draft = common_speculative_gen_draft(spec, params_spec, prompt_tgt, id_last);
+
+        //LOG_DBG("draft: %s\n", string_from(ctx_dft, draft).c_str());
+
+        // always have a token to evaluate from before - id_last
+        common_batch_clear(batch_tgt);
+        common_batch_add  (batch_tgt, id_last, n_past++, { 0 }, true);
+
+        // evaluate the target model on [id_last, draft0, draft1, ..., draftN-1]
+        {
+            // do not waste time on small drafts
+            if (draft.size() < n_draft_min) {
+                draft.clear();
+            }
+
+            for (size_t i = 0; i < draft.size(); ++i) {
+                common_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
+            }
+
+            //LOG_DBG("target batch: %s\n", string_from(ctx_tgt, batch_tgt).c_str());
+
+            llama_decode(ctx_tgt, batch_tgt);
+        }
+
+        // sample from the full target batch and return the accepted tokens based on the target sampler
+        //
+        // for each token to be accepted, the sampler would have to sample that same token
+        // in such cases, instead of decoding the sampled token as we normally do, we simply continue with the
+        // available logits from the batch and sample the next token until we run out of logits or the sampler
+        // disagrees with the draft
+        //
+        const auto ids = common_sampler_sample_and_accept_n(smpl, ctx_tgt, draft);
+
+        //LOG_DBG("ids: %s\n", string_from(ctx_tgt, ids).c_str());
+
+        GGML_ASSERT(ids.size() > 0); // there will always be at least one accepted token
+
+        n_past    += ids.size() - 1;
+        n_drafted += batch_tgt.n_tokens - 1;
+        n_accept  += ids.size() - 1;
+
+        // process the accepted tokens and update contexts
+        //
+        // this is the standard token post-processing that we normally do
+        // in this case, we do it for a group of accepted tokens at once
+        //
+        {
+            llama_token id;
+            std::string token_str;
+
+            for (size_t i = 0; i < ids.size(); ++i) {
+                id = ids[i];
+
+                ++n_predict;
+
+                if (llama_token_is_eog(model_tgt, id)) {
+                    has_eos = true;
+                    break;
+                }
+
+                token_str = common_token_to_piece(ctx_tgt, id);
+
+                if (params.use_color && i + 1 < ids.size()) {
+                    LOG("\u001b[%dm%s\u001b[37m", (36 - 0 % 6), token_str.c_str());
+                } else {
+                    LOG("%s", token_str.c_str());
+                }
+            }
+
+            if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
+                break;
+            }
+
+            LOG_DBG("accepted %d/%d draft tokens, the last target token is: (%d, '%s')\n", (int) ids.size() - 1, (int) draft.size(), id, token_str.c_str());
+
+            {
+                LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
+
+                llama_kv_cache_seq_rm(ctx_tgt, 0, n_past, -1);
+            }
+
+            prompt_tgt.push_back(id_last);
+            prompt_tgt.insert(prompt_tgt.end(), ids.begin(), ids.end() - 1);
+
+            // remember the last accepted token for the next iteration
+            id_last = id;
+        }
+    }
+
+    auto t_dec_end = ggml_time_us();
+
+    const int n_input = inp.size();
+
+    LOG("\n\n");
+
+    LOG_INF("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input,   (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
+    LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
+
+    LOG_INF("\n");
+    LOG_INF("n_draft   = %d\n", n_draft);
+    LOG_INF("n_predict = %d\n", n_predict);
+    LOG_INF("n_drafted = %d\n", n_drafted);
+    LOG_INF("n_accept  = %d\n", n_accept);
+    LOG_INF("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
+
+    LOG_INF("\n");
+    LOG_INF("draft:\n\n");
+
+    llama_perf_context_print(ctx_dft);
+
+    LOG_INF("\n");
+    LOG_INF("target:\n\n");
+    common_perf_print(ctx_tgt, smpl);
+
+    common_sampler_free(smpl);
+    common_speculative_free(spec);
+
+    llama_free(ctx_tgt);
+    llama_free_model(model_tgt);
+
+    llama_free(ctx_dft);
+    llama_free_model(model_dft);
+
+    llama_backend_free();
+
+    LOG("\n\n");
+
+    return 0;
+}
diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp
index 6cafd8a837992..eb8bb2de54a26 100644
--- a/examples/speculative/speculative.cpp
+++ b/examples/speculative/speculative.cpp
@@ -12,7 +12,7 @@
 #include <string>
 #include <vector>
 
-#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  100
+#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
 
 struct seq_draft {
@@ -33,7 +33,7 @@ int main(int argc, char ** argv) {
     common_params params;
 
     // needed to get candidate probs even for temp <= 0.0
-    params.sparams.n_probs = 128;
+    params.sampling.n_probs = 128;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
         return 1;
@@ -46,7 +46,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.model_draft.empty()) {
+    if (params.speculative.model.empty()) {
         LOG_ERR("%s: --model-draft is required\n", __func__);
         return 1;
     }
@@ -55,9 +55,9 @@ int main(int argc, char ** argv) {
     const int n_seq_dft = params.n_parallel;
 
     // probability threshold for splitting a draft branch (only for n_seq_dft > 1)
-    const float p_split  = params.p_split;
+    const float p_draft_split = params.speculative.p_split;
 
-    std::default_random_engine rng(params.sparams.seed == LLAMA_DEFAULT_SEED ? std::random_device()() : params.sparams.seed);
+    std::default_random_engine rng(params.sampling.seed == LLAMA_DEFAULT_SEED ? std::random_device()() : params.sampling.seed);
     std::uniform_real_distribution<> u_dist;
 
     // init llama.cpp
@@ -76,13 +76,13 @@ int main(int argc, char ** argv) {
     ctx_tgt = llama_init_tgt.context;
 
     // load the draft model
-    params.model = params.model_draft;
-    params.n_gpu_layers = params.n_gpu_layers_draft;
-    if (params.draft_cpuparams.n_threads > 0) {
-        params.cpuparams.n_threads = params.draft_cpuparams.n_threads;
+    params.model = params.speculative.model;
+    params.n_gpu_layers = params.speculative.n_gpu_layers;
+    if (params.speculative.cpuparams.n_threads > 0) {
+        params.cpuparams.n_threads = params.speculative.cpuparams.n_threads;
     }
 
-    params.cpuparams_batch.n_threads = params.draft_cpuparams_batch.n_threads;
+    params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
     common_init_result llama_init_dft = common_init_from_params(params);
     model_dft = llama_init_dft.model;
     ctx_dft = llama_init_dft.context;
@@ -170,7 +170,7 @@ int main(int argc, char ** argv) {
     //GGML_ASSERT(n_vocab == llama_n_vocab(model_dft));
 
     // how many tokens to draft each time
-    int n_draft = params.n_draft;
+    int n_draft = params.speculative.n_max;
 
     int n_predict = 0;
     int n_drafted = 0;
@@ -183,14 +183,14 @@ int main(int argc, char ** argv) {
     bool has_eos = false;
 
     // target model sampling context (reuse the llama_context's sampling instance)
-    struct common_sampler * smpl = common_sampler_init(model_tgt, params.sparams);
+    struct common_sampler * smpl = common_sampler_init(model_tgt, params.sampling);
 
     // draft sequence data
     std::vector<seq_draft> drafts(n_seq_dft);
 
     for (int s = 0; s < n_seq_dft; ++s) {
         // allocate llama_sampler for each draft sequence
-        drafts[s].smpl = common_sampler_init(model_dft, params.sparams);
+        drafts[s].smpl = common_sampler_init(model_dft, params.sampling);
     }
 
     llama_batch batch_dft = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
@@ -230,7 +230,7 @@ int main(int argc, char ** argv) {
             // for stochastic sampling, attempt to match the token with the drafted tokens
             {
                 bool accept = false;
-                if (params.sparams.temp > 0) {
+                if (params.sampling.temp > 0) {
                     // stochastic verification
                     common_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft], true);
 
@@ -494,7 +494,7 @@ int main(int argc, char ** argv) {
 
                 // attempt to split the branch if the probability is high enough
                 for (int f = 1; f < 8; ++f) {
-                    if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_split) {
+                    if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_draft_split) {
                         LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur);
 
                         llama_kv_cache_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
diff --git a/tests/test-arg-parser.cpp b/tests/test-arg-parser.cpp
index 3665238b5a2d8..69604b87ceec4 100644
--- a/tests/test-arg-parser.cpp
+++ b/tests/test-arg-parser.cpp
@@ -70,7 +70,7 @@ int main(void) {
 
     // non-existence arg in specific example (--draft cannot be used outside llama-speculative)
     argv = {"binary_name", "--draft", "123"};
-    assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_SERVER));
+    assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_EMBEDDING));
 
 
     printf("test-arg-parser: test valid usage\n\n");
@@ -96,7 +96,7 @@ int main(void) {
     // --draft cannot be used outside llama-speculative
     argv = {"binary_name", "--draft", "123"};
     assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_SPECULATIVE));
-    assert(params.n_draft == 123);
+    assert(params.speculative.n_max == 123);
 
 // skip this part on windows, because setenv is not supported
 #ifdef _WIN32

From 5a8987793f3e7c1fbfa6806bfcd17d578071b6c9 Mon Sep 17 00:00:00 2001
From: Neo Zhang Jianyu <jianyu.zhang@intel.com>
Date: Mon, 25 Nov 2024 17:31:10 +0800
Subject: [PATCH 062/102] [SYCL] Fix building Win package for oneAPI 2025.0
 update (#10483)

* fix build package for 2025.0

* debug

* debug

* fix

* rm debug

---------

Co-authored-by: arthw <14088817+arthw@users.noreply.github.com>
---
 .github/workflows/build.yml | 22 +++++++++++++++-------
 1 file changed, 15 insertions(+), 7 deletions(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 572f91643fbba..abaf2c504c8f4 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -952,7 +952,7 @@ jobs:
 
     env:
       WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b380d914-366b-4b77-a74a-05e3c38b3514/intel-oneapi-base-toolkit-2025.0.0.882_offline.exe
-      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
+      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
     steps:
       - name: Clone
@@ -962,7 +962,8 @@ jobs:
           fetch-depth: 0
 
       - name: Install
-        run:  scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+        run:  |
+          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
 
       - name: Build
         id: cmake_build
@@ -981,27 +982,34 @@ jobs:
             echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
           fi
 
-      - name: Pack artifacts
+      - name: Build the release package
         id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        if: ${{ ( github.event_name == 'pull_request' && github.base_ref == 'master' ) }}
         run: |
           echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
+
           cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
 
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
+
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
+
           echo "cp oneAPI running time dll files to ./build/bin done"
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+      - name: Upload the release package
+        if: ${{ ( github.event_name == 'pull_request' && github.base_ref == 'master' ) }}
         uses: actions/upload-artifact@v4
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip

From b756441104ca8384640d6df22ba4ea6dab7ad799 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 25 Nov 2024 15:08:04 +0200
Subject: [PATCH 063/102] metal : minor code formatting

---
 ggml/src/ggml-metal/ggml-metal.m     | 598 +++++++++++++--------------
 ggml/src/ggml-metal/ggml-metal.metal |  43 +-
 2 files changed, 323 insertions(+), 318 deletions(-)

diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index d1abb3cef0ec4..3a533d7f9c9af 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -1951,316 +1951,316 @@ static void ggml_metal_encode_node(
                         }
 #endif
 
-                        // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
-                        // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-                        if ([device supportsFamily:MTLGPUFamilyApple7] &&
-                                !ggml_is_transposed(src0) &&
-                                !ggml_is_transposed(src1) &&
-                                src1t == GGML_TYPE_F32 &&
-                                ne00 % 32 == 0 && ne00 >= 64 &&
-                                (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) {
-                            //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
-
-                            // some Metal matrix data types require aligned pointers
-                            // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
-                            switch (src0->type) {
-                                case GGML_TYPE_F32:  GGML_ASSERT(nb01 % 16 == 0); break;
-                                case GGML_TYPE_F16:  GGML_ASSERT(nb01 % 8  == 0); break;
-                                case GGML_TYPE_BF16: GGML_ASSERT(nb01 % 8  == 0); break;
-                                default: break;
-                            }
+                // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
+                // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
+                if ([device supportsFamily:MTLGPUFamilyApple7] &&
+                        !ggml_is_transposed(src0) &&
+                        !ggml_is_transposed(src1) &&
+                        src1t == GGML_TYPE_F32 &&
+                        ne00 % 32 == 0 && ne00 >= 64 &&
+                        (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) {
+                    //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
 
-                            id<MTLComputePipelineState> pipeline = nil;
-
-                            switch (src0->type) {
-                                case GGML_TYPE_F32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32    ].pipeline; break;
-                                case GGML_TYPE_F16:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32    ].pipeline; break;
-                                case GGML_TYPE_BF16:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32   ].pipeline; break;
-                                case GGML_TYPE_Q4_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32   ].pipeline; break;
-                                case GGML_TYPE_Q4_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32   ].pipeline; break;
-                                case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32   ].pipeline; break;
-                                case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32   ].pipeline; break;
-                                case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32   ].pipeline; break;
-                                case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32   ].pipeline; break;
-                                case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32   ].pipeline; break;
-                                case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32   ].pipeline; break;
-                                case GGML_TYPE_Q5_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32   ].pipeline; break;
-                                case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32   ].pipeline; break;
-                                case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
-                                case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
-                                case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
-                                case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
-                                case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
-                                default: GGML_ABORT("MUL MAT-MAT not implemented");
-                            }
+                    // some Metal matrix data types require aligned pointers
+                    // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
+                    switch (src0->type) {
+                        case GGML_TYPE_F32:  GGML_ASSERT(nb01 % 16 == 0); break;
+                        case GGML_TYPE_F16:  GGML_ASSERT(nb01 % 8  == 0); break;
+                        case GGML_TYPE_BF16: GGML_ASSERT(nb01 % 8  == 0); break;
+                        default: break;
+                    }
 
-                            ggml_metal_kargs_mul_mm args = {
-                                /*.ne00 =*/ ne00,
-                                /*.ne02 =*/ ne02,
-                                /*.nb01 =*/ nb01,
-                                /*.nb02 =*/ nb02,
-                                /*.nb03 =*/ nb03,
-                                /*.ne12 =*/ ne12,
-                                /*.nb10 =*/ nb10,
-                                /*.nb11 =*/ nb11,
-                                /*.nb12 =*/ nb12,
-                                /*.nb13 =*/ nb13,
-                                /*.ne0  =*/ ne0,
-                                /*.ne1  =*/ ne1,
-                                /*.r2   =*/ r2,
-                                /*.r3   =*/ r3,
-                            };
+                    id<MTLComputePipelineState> pipeline = nil;
+
+                    switch (src0->type) {
+                        case GGML_TYPE_F32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32    ].pipeline; break;
+                        case GGML_TYPE_F16:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32    ].pipeline; break;
+                        case GGML_TYPE_BF16:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32   ].pipeline; break;
+                        case GGML_TYPE_Q4_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32   ].pipeline; break;
+                        case GGML_TYPE_Q4_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32   ].pipeline; break;
+                        case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32   ].pipeline; break;
+                        case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32   ].pipeline; break;
+                        case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32   ].pipeline; break;
+                        case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q5_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32   ].pipeline; break;
+                        case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
+                        case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
+                        case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
+                        case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
+                        case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
+                        default: GGML_ABORT("MUL MAT-MAT not implemented");
+                    }
+
+                    ggml_metal_kargs_mul_mm args = {
+                        /*.ne00 =*/ ne00,
+                        /*.ne02 =*/ ne02,
+                        /*.nb01 =*/ nb01,
+                        /*.nb02 =*/ nb02,
+                        /*.nb03 =*/ nb03,
+                        /*.ne12 =*/ ne12,
+                        /*.nb10 =*/ nb10,
+                        /*.nb11 =*/ nb11,
+                        /*.nb12 =*/ nb12,
+                        /*.nb13 =*/ nb13,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                        /*.r2   =*/ r2,
+                        /*.r3   =*/ r3,
+                    };
 
-                            [encoder setComputePipelineState:pipeline];
-                            [encoder setBytes:&args    length:sizeof(args) atIndex:0];
-                            [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
-                            [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
-                            [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
 
-                            [encoder setThreadgroupMemoryLength:8192 atIndex:0];
-                            [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
-                        } else {
-                            int nth0 = 32;
-                            int nth1 = 1;
-                            int nrows = 1;
-                            //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+                    [encoder setThreadgroupMemoryLength:8192 atIndex:0];
+                    [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                } else {
+                    int nth0 = 32;
+                    int nth1 = 1;
+                    int nrows = 1;
+                    //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
 
-                            id<MTLComputePipelineState> pipeline = nil;
+                    id<MTLComputePipelineState> pipeline = nil;
 
-                            // use custom matrix x vector kernel
-                            switch (src0t) {
-                                case GGML_TYPE_F32:
-                                    {
-                                        GGML_ASSERT(src1t == GGML_TYPE_F32);
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
+                    // use custom matrix x vector kernel
+                    switch (src0t) {
+                        case GGML_TYPE_F32:
+                            {
+                                GGML_ASSERT(src1t == GGML_TYPE_F32);
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
+                                nrows = 4;
+                            } break;
+                        case GGML_TYPE_F16:
+                            {
+                                nth0 = 32;
+                                nth1 = 1;
+                                if (src1t == GGML_TYPE_F32) {
+                                    if (ne11 * ne12 < 4) {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline;
+                                    } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
+                                        nrows = ne11;
+                                    } else {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline;
                                         nrows = 4;
-                                    } break;
-                                case GGML_TYPE_F16:
-                                    {
-                                        nth0 = 32;
-                                        nth1 = 1;
-                                        if (src1t == GGML_TYPE_F32) {
-                                            if (ne11 * ne12 < 4) {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline;
-                                            } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
-                                                nrows = ne11;
-                                            } else {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline;
-                                                nrows = 4;
-                                            }
-                                        } else {
-                                            pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline;
-                                            nrows = 4;
-                                        }
-                                    } break;
-                                case GGML_TYPE_BF16:
-                                    {
-                                        nth0 = 32;
-                                        nth1 = 1;
-                                        if (src1t == GGML_TYPE_F32) {
-                                            if (ne11 * ne12 < 4) {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline;
-                                            } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline;
-                                                nrows = ne11;
-                                            } else {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32].pipeline;
-                                                nrows = 4;
-                                            }
-                                        } else {
-                                            pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16].pipeline;
-                                            nrows = 4;
-                                        }
-                                    } break;
-                                case GGML_TYPE_Q4_0:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q4_1:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q5_0:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q5_1:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q8_0:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q2_K:
-                                    {
-                                        nth0 = 2;
-                                        nth1 = 32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q3_K:
-                                    {
-                                        nth0 = 2;
-                                        nth1 = 32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q4_K:
-                                    {
-                                        nth0 = 4; //1;
-                                        nth1 = 8; //32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q5_K:
-                                    {
-                                        nth0 = 2;
-                                        nth1 = 32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q6_K:
-                                    {
-                                        nth0 = 2;
-                                        nth1 = 32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ2_XXS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ2_XS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ3_XXS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ3_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ2_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ1_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ1_M:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ4_NL:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ4_XS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
-                                    } break;
-                                default:
-                                    {
-                                        GGML_LOG_ERROR("Asserting on type %d\n", (int)src0t);
-                                        GGML_ABORT("not implemented");
                                     }
-                            };
+                                } else {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline;
+                                    nrows = 4;
+                                }
+                            } break;
+                        case GGML_TYPE_BF16:
+                            {
+                                nth0 = 32;
+                                nth1 = 1;
+                                if (src1t == GGML_TYPE_F32) {
+                                    if (ne11 * ne12 < 4) {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline;
+                                    } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline;
+                                        nrows = ne11;
+                                    } else {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32].pipeline;
+                                        nrows = 4;
+                                    }
+                                } else {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16].pipeline;
+                                    nrows = 4;
+                                }
+                            } break;
+                        case GGML_TYPE_Q4_0:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q4_1:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q5_0:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q5_1:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q8_0:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q2_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q3_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q4_K:
+                            {
+                                nth0 = 4; //1;
+                                nth1 = 8; //32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q5_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q6_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ2_XXS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ2_XS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ3_XXS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ3_S:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ2_S:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ1_S:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ1_M:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ4_NL:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ4_XS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
+                            } break;
+                        default:
+                            {
+                                GGML_LOG_ERROR("Asserting on type %d\n", (int)src0t);
+                                GGML_ABORT("not implemented");
+                            }
+                    };
 
-                            ggml_metal_kargs_mul_mv args = {
-                                /*.ne00 =*/ ne00,
-                                /*.ne01 =*/ ne01,
-                                /*.ne02 =*/ ne02,
-                                /*.nb00 =*/ nb00,
-                                /*.nb01 =*/ nb01,
-                                /*.nb02 =*/ nb02,
-                                /*.nb03 =*/ nb03,
-                                /*.ne10 =*/ ne10,
-                                /*.ne11 =*/ ne11,
-                                /*.ne12 =*/ ne12,
-                                /*.nb10 =*/ nb10,
-                                /*.nb11 =*/ nb11,
-                                /*.nb12 =*/ nb12,
-                                /*.nb13 =*/ nb13,
-                                /*.ne0  =*/ ne0,
-                                /*.ne1  =*/ ne1,
-                                /*.r2   =*/ r2,
-                                /*.r3   =*/ r3,
-                            };
+                    ggml_metal_kargs_mul_mv args = {
+                        /*.ne00 =*/ ne00,
+                        /*.ne01 =*/ ne01,
+                        /*.ne02 =*/ ne02,
+                        /*.nb00 =*/ nb00,
+                        /*.nb01 =*/ nb01,
+                        /*.nb02 =*/ nb02,
+                        /*.nb03 =*/ nb03,
+                        /*.ne10 =*/ ne10,
+                        /*.ne11 =*/ ne11,
+                        /*.ne12 =*/ ne12,
+                        /*.nb10 =*/ nb10,
+                        /*.nb11 =*/ nb11,
+                        /*.nb12 =*/ nb12,
+                        /*.nb13 =*/ nb13,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                        /*.r2   =*/ r2,
+                        /*.r3   =*/ r3,
+                    };
 
-                            [encoder setComputePipelineState:pipeline];
-                            [encoder setBytes:&args length:sizeof(args) atIndex:0];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
-                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
 
-                            if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
-                                src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
-                                src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
-                                const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
-                                const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
-                                const int mem_size = 32*sizeof(float);
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_Q4_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_Q3_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_Q5_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_Q6_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            } else {
-                                const int64_t ny = (ne11 + nrows - 1)/nrows;
-                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                        }
+                    if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
+                        src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
+                        src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
+                        const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
+                        const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
+                        const int mem_size = 32*sizeof(float);
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q4_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q3_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q5_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q6_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    } else {
+                        const int64_t ny = (ne11 + nrows - 1)/nrows;
+                        [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                }
             } break;
         case GGML_OP_MUL_MAT_ID:
             {
diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
index 971f5054bce2a..eaca38864bd65 100644
--- a/ggml/src/ggml-metal/ggml-metal.metal
+++ b/ggml/src/ggml-metal/ggml-metal.metal
@@ -5447,12 +5447,12 @@ kernel void kernel_mul_mm(
     const int im = tgpig.z;
 
     // if this block is of 64x32 shape or smaller
-    short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    short n_cols = (args.ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (args.ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    const short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
+    const short n_cols = (args.ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (args.ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
 
     // a thread shouldn't load data outside of the matrix
-    short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+    const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
+    const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
 
     simdgroup_T8x8     ma[4];
     simdgroup_float8x8 mb[2];
@@ -5467,20 +5467,23 @@ kernel void kernel_mul_mm(
     const int i12 = im%args.ne12;
     const int i13 = im/args.ne12;
 
-    uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
-    short    offset1 = il/nl;
+    const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const short    offset1 = il/nl;
+
+    device const block_q * x = (device const block_q *)(src0
+        + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
 
-    device const block_q * x = (device const block_q *)(src0 + (r0*BLOCK_SIZE_M + thread_row)*args.nb01 + offset0) + offset1;
     device const float   * y = (device const float   *)(src1
         + args.nb13*i13
         + args.nb12*i12
-        + args.nb11*(r1 * BLOCK_SIZE_N + thread_col)
+        + args.nb11*(r1*BLOCK_SIZE_N + thread_col)
         + args.nb10*(BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
 
     for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
         // load data and store to threadgroup memory
         T4x4 temp_a;
         dequantize_func(x, il, temp_a);
+
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         #pragma unroll(16)
@@ -5490,44 +5493,46 @@ kernel void kernel_mul_mm(
             +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = temp_a[i/4][i%4];
         }
 
-        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL)*8*32 + 8*(tiitg/THREAD_PER_COL)) = *((device float2x4 *) y);
+        *(threadgroup float2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = *((device float2x4 *) y);
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
-        x  = (il < 2) ? x + (2+nl-1)/nl : x;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
         y += BLOCK_SIZE_K;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         // load matrices from threadgroup memory and conduct outer products
-        threadgroup T     * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
-        threadgroup float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
+        threadgroup const T     * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
+        threadgroup const float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
 
         #pragma unroll(4)
-        for (short ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
+        for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
             #pragma unroll(4)
             for (short i = 0; i < 4; i++) {
                 simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
             }
+
             simdgroup_barrier(mem_flags::mem_none);
+
             #pragma unroll(2)
             for (short i = 0; i < 2; i++) {
                 simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
             }
 
-            lsma += BLOCK_SIZE_M/SG_MAT_ROW * SG_MAT_SIZE;
-            lsmb += BLOCK_SIZE_N/SG_MAT_ROW * SG_MAT_SIZE;
-
             #pragma unroll(8)
             for (short i = 0; i < 8; i++){
                 simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
             }
+
+            lsma += (BLOCK_SIZE_M/SG_MAT_ROW)*SG_MAT_SIZE;
+            lsmb += (BLOCK_SIZE_N/SG_MAT_ROW)*SG_MAT_SIZE;
         }
     }
 
     if ((r0 + 1) * BLOCK_SIZE_M <= args.ne0 && (r1 + 1) * BLOCK_SIZE_N <= args.ne1) {
         device float * C = (device float *) dst +
-            (BLOCK_SIZE_M * r0 + 32 * (sgitg &  1)) + \
-            (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+            (BLOCK_SIZE_M * r0 + 32*(sgitg &  1)) + \
+            (BLOCK_SIZE_N * r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
 
         for (short i = 0; i < 8; i++) {
             simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.ne0 * (i/4), args.ne0);
@@ -5536,7 +5541,7 @@ kernel void kernel_mul_mm(
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
         threadgroup_barrier(mem_flags::mem_threadgroup);
         threadgroup float * temp_str = ((threadgroup float *) shmem) \
-                                      + 32 * (sgitg&1) + (16 * (sgitg>>1))*BLOCK_SIZE_M;
+                                     + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
         for (short i = 0; i < 8; i++) {
             simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
         }

From f6d12e7df8fe64384f1939976871252e6422a01e Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 25 Nov 2024 15:17:32 +0200
Subject: [PATCH 064/102] tests : fix compile warning

---
 tests/test-quantize-fns.cpp | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/tests/test-quantize-fns.cpp b/tests/test-quantize-fns.cpp
index 8d0bf0470fe4e..c77c8ed1388d7 100644
--- a/tests/test-quantize-fns.cpp
+++ b/tests/test-quantize-fns.cpp
@@ -79,9 +79,9 @@ static float dot_product(const float * a1, const float * a2, size_t test_size) {
 }
 
 // Total dot product error
-static float dot_product_error(
-    const ggml_type_traits * qfns, const ggml_type_traits_cpu * qfns_cpu, size_t test_size, const float * test_data1, const float *test_data2
-) {
+static float dot_product_error(const ggml_type_traits * qfns, const ggml_type_traits_cpu * qfns_cpu, size_t test_size, const float * test_data1, const float * test_data2) {
+    GGML_UNUSED(qfns);
+
     std::vector<uint8_t> tmp_q1(2*test_size);
     std::vector<uint8_t> tmp_q2(2*test_size);
 

From 5931c1f233c616083d64e41a228249d58e039aa5 Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Mon, 25 Nov 2024 15:13:39 +0100
Subject: [PATCH 065/102] ggml : add support for dynamic loading of backends
 (#10469)

* ggml : add support for dynamic loading of backends

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---
 Makefile                               |   3 +-
 Package.swift                          |   3 +-
 common/common.cpp                      |   3 +
 examples/CMakeLists.txt                |  27 +--
 examples/eval-callback/CMakeLists.txt  |   3 +-
 examples/llama-bench/llama-bench.cpp   |  15 +-
 examples/main/main.cpp                 |  12 +-
 examples/simple-chat/simple-chat.cpp   |   3 +
 examples/simple/simple.cpp             |   4 +
 ggml/CMakeLists.txt                    |   1 +
 ggml/include/ggml-backend.h            |  15 ++
 ggml/include/ggml-cpu.h                |  38 +---
 ggml/include/ggml.h                    |  31 +++
 ggml/src/CMakeLists.txt                |  41 +++-
 ggml/src/ggml-amx/CMakeLists.txt       |  10 +-
 ggml/src/ggml-amx/ggml-amx.cpp         |   7 +-
 ggml/src/ggml-backend-impl.h           |  44 +++--
 ggml/src/ggml-backend-reg.cpp          | 252 +++++++++++++++++++++++--
 ggml/src/ggml-blas/CMakeLists.txt      |   9 +-
 ggml/src/ggml-blas/ggml-blas.cpp       |   7 +-
 ggml/src/ggml-cann/CMakeLists.txt      |   6 +-
 ggml/src/ggml-cann/ggml-cann.cpp       |  13 +-
 ggml/src/ggml-cpu/CMakeLists.txt       |  19 +-
 ggml/src/ggml-cpu/ggml-cpu.c           |  23 ---
 ggml/src/ggml-cpu/ggml-cpu.cpp         |  50 +++--
 ggml/src/ggml-cuda/CMakeLists.txt      |  11 +-
 ggml/src/ggml-cuda/ggml-cuda.cu        |  71 ++++++-
 ggml/src/ggml-hip/CMakeLists.txt       |  10 +-
 ggml/src/ggml-kompute/CMakeLists.txt   |  10 +-
 ggml/src/ggml-kompute/ggml-kompute.cpp |   7 +-
 ggml/src/ggml-metal/CMakeLists.txt     |   9 +-
 ggml/src/ggml-metal/ggml-metal.m       |  34 +++-
 ggml/src/ggml-musa/CMakeLists.txt      |  10 +-
 ggml/src/ggml-rpc/CMakeLists.txt       |   8 +-
 ggml/src/ggml-rpc/ggml-rpc.cpp         |   7 +-
 ggml/src/ggml-sycl/CMakeLists.txt      |  10 +-
 ggml/src/ggml-sycl/ggml-sycl.cpp       |  12 +-
 ggml/src/ggml-vulkan/CMakeLists.txt    |  12 +-
 ggml/src/ggml-vulkan/ggml-vulkan.cpp   |   7 +-
 ggml/src/ggml.c                        |  23 +++
 pocs/CMakeLists.txt                    |   4 +-
 src/llama.cpp                          |  77 ++++----
 tests/CMakeLists.txt                   |  13 +-
 tests/test-backend-ops.cpp             |  26 +--
 44 files changed, 728 insertions(+), 272 deletions(-)

diff --git a/Makefile b/Makefile
index dd6d864ad513a..14c05e93e7535 100644
--- a/Makefile
+++ b/Makefile
@@ -251,7 +251,7 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS  = -Iggml/include -Iggml/src -Iinclude -Isrc -Icommon
+MK_CPPFLAGS  = -Iggml/include -Iggml/src -Iinclude -Isrc -Icommon -DGGML_USE_CPU
 MK_CFLAGS    = -std=c11   -fPIC
 MK_CXXFLAGS  = -std=c++11 -fPIC
 MK_NVCCFLAGS = -std=c++11
@@ -290,6 +290,7 @@ endif
 # some memory allocation are available on Linux through GNU extensions in libc
 ifeq ($(UNAME_S),Linux)
 	MK_CPPFLAGS += -D_GNU_SOURCE
+	MK_LDFLAGS  += -ldl
 endif
 
 # RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
diff --git a/Package.swift b/Package.swift
index 6b68aecdebec1..d9e8a4e2d21d7 100644
--- a/Package.swift
+++ b/Package.swift
@@ -43,7 +43,8 @@ linkerSettings.append(.linkedFramework("Accelerate"))
 cSettings.append(
     contentsOf: [
         .define("GGML_USE_ACCELERATE"),
-        .define("GGML_USE_METAL")
+        .define("GGML_USE_METAL"),
+        .define("GGML_USE_CPU")
     ]
 )
 #endif
diff --git a/common/common.cpp b/common/common.cpp
index c398329d05bf5..98524f7467ab4 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -377,6 +377,9 @@ void common_init() {
 #endif
 
     LOG_INF("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type);
+
+    // load dynamic backends
+    ggml_backend_load_all();
 }
 
 std::string common_params_get_system_info(const common_params & params) {
diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index 9bd099d4ef8a5..632409d5591b9 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -12,13 +12,10 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
 else()
-    add_subdirectory(cvector-generator)
     add_subdirectory(batched-bench)
     add_subdirectory(batched)
-    add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
-    add_subdirectory(export-lora)
     add_subdirectory(gbnf-validator)
     add_subdirectory(gguf-hash)
     add_subdirectory(gguf-split)
@@ -27,24 +24,16 @@ else()
     add_subdirectory(imatrix)
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
-    add_subdirectory(llava)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
     add_subdirectory(main)
     add_subdirectory(parallel)
     add_subdirectory(passkey)
     add_subdirectory(perplexity)
-    add_subdirectory(quantize-stats)
     add_subdirectory(quantize)
     add_subdirectory(retrieval)
-    if (GGML_RPC)
-        add_subdirectory(rpc)
-    endif()
     if (LLAMA_BUILD_SERVER)
-    add_subdirectory(server)
-    endif()
-    if (GGML_SYCL)
-        add_subdirectory(sycl)
+        add_subdirectory(server)
     endif()
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
@@ -52,4 +41,18 @@ else()
     add_subdirectory(speculative)
     add_subdirectory(speculative-simple)
     add_subdirectory(tokenize)
+    if (NOT GGML_BACKEND_DL)
+        # these examples use the backends directly and cannot be built with dynamic loading
+        add_subdirectory(convert-llama2c-to-ggml)
+        add_subdirectory(cvector-generator)
+        add_subdirectory(export-lora)
+        add_subdirectory(quantize-stats)
+        add_subdirectory(llava)
+        if (GGML_RPC)
+            add_subdirectory(rpc)
+        endif()
+        if (GGML_SYCL)
+            add_subdirectory(sycl)
+        endif()
+    endif()
 endif()
diff --git a/examples/eval-callback/CMakeLists.txt b/examples/eval-callback/CMakeLists.txt
index a48753d38e16e..5d1048aad74b6 100644
--- a/examples/eval-callback/CMakeLists.txt
+++ b/examples/eval-callback/CMakeLists.txt
@@ -5,5 +5,6 @@ target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 
 set(TEST_TARGET test-eval-callback)
-add_test(NAME ${TEST_TARGET} COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
+add_test(NAME ${TEST_TARGET}
+        COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
 set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)
diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp
index 3dc84a75cbec7..bac606f471639 100644
--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
@@ -1477,6 +1477,17 @@ int main(int argc, char ** argv) {
 
     cmd_params params = parse_cmd_params(argc, argv);
 
+    // initialize backends
+    ggml_backend_load_all();
+    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    if (!cpu_dev) {
+        fprintf(stderr, "%s: error: CPU backend is not loaded\n", __func__);
+        return 1;
+    }
+    auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+    auto * ggml_threadpool_new_fn = (decltype(ggml_threadpool_new) *) ggml_backend_reg_get_proc_address(cpu_reg, "ggml_threadpool_new");
+    auto * ggml_threadpool_free_fn = (decltype(ggml_threadpool_free) *) ggml_backend_reg_get_proc_address(cpu_reg, "ggml_threadpool_free");
+
     // initialize llama.cpp
     if (!params.verbose) {
         llama_log_set(llama_null_log_callback, NULL);
@@ -1551,7 +1562,7 @@ int main(int argc, char ** argv) {
         tpp.poll       = t.poll;
         tpp.prio       = params.prio;
 
-        struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp);
+        struct ggml_threadpool * threadpool = ggml_threadpool_new_fn(&tpp);
         if (!threadpool) {
             fprintf(stderr, "%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
             exit(1);
@@ -1612,7 +1623,7 @@ int main(int argc, char ** argv) {
 
         llama_free(ctx);
 
-        ggml_threadpool_free(threadpool);
+        ggml_threadpool_free_fn(threadpool);
     }
 
     llama_free_model(lmodel);
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
index 957451af7ce0a..d0c28f317b8c5 100644
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -165,6 +165,10 @@ int main(int argc, char ** argv) {
 
     LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads);
 
+    auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
+    auto * ggml_threadpool_new_fn = (decltype(ggml_threadpool_new) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_new");
+    auto * ggml_threadpool_free_fn = (decltype(ggml_threadpool_free) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_free");
+
     struct ggml_threadpool_params tpp_batch =
             ggml_threadpool_params_from_cpu_params(params.cpuparams_batch);
     struct ggml_threadpool_params tpp =
@@ -174,7 +178,7 @@ int main(int argc, char ** argv) {
 
     struct ggml_threadpool * threadpool_batch = NULL;
     if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
-        threadpool_batch = ggml_threadpool_new(&tpp_batch);
+        threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);
         if (!threadpool_batch) {
             LOG_ERR("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads);
             return 1;
@@ -184,7 +188,7 @@ int main(int argc, char ** argv) {
         tpp.paused = true;
     }
 
-    struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp);
+    struct ggml_threadpool * threadpool = ggml_threadpool_new_fn(&tpp);
     if (!threadpool) {
         LOG_ERR("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
         return 1;
@@ -890,8 +894,8 @@ int main(int argc, char ** argv) {
 
     llama_backend_free();
 
-    ggml_threadpool_free(threadpool);
-    ggml_threadpool_free(threadpool_batch);
+    ggml_threadpool_free_fn(threadpool);
+    ggml_threadpool_free_fn(threadpool_batch);
 
     return 0;
 }
diff --git a/examples/simple-chat/simple-chat.cpp b/examples/simple-chat/simple-chat.cpp
index 5f9973163732d..7f4da666b08ec 100644
--- a/examples/simple-chat/simple-chat.cpp
+++ b/examples/simple-chat/simple-chat.cpp
@@ -62,6 +62,9 @@ int main(int argc, char ** argv) {
         }
     }, nullptr);
 
+    // load dynamic backends
+    ggml_backend_load_all();
+
     // initialize the model
     llama_model_params model_params = llama_model_default_params();
     model_params.n_gpu_layers = ngl;
diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp
index 59760fe95db22..3288c0250a001 100644
--- a/examples/simple/simple.cpp
+++ b/examples/simple/simple.cpp
@@ -74,6 +74,10 @@ int main(int argc, char ** argv) {
         }
     }
 
+    // load dynamic backends
+
+    ggml_backend_load_all();
+
     // initialize the model
 
     llama_model_params model_params = llama_model_default_params();
diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 2d32da1b6d879..70b5cfdf7fbb4 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -33,6 +33,7 @@ else()
 endif()
 
 option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+option(GGML_BACKEND_DL   "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
 
 #
 # option list
diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
index cef164764bb1a..19881a5059f17 100644
--- a/ggml/include/ggml-backend.h
+++ b/ggml/include/ggml-backend.h
@@ -190,6 +190,14 @@ extern "C" {
     typedef void                         (*ggml_backend_set_n_threads_t)(ggml_backend_t backend, int n_threads);
     // Get additional buffer types provided by the device (returns a NULL-terminated array)
     typedef ggml_backend_buffer_type_t * (*ggml_backend_dev_get_extra_bufts_t)(ggml_backend_dev_t device);
+    // Set the abort callback for the backend
+    typedef void                         (*ggml_backend_set_abort_callback_t)(ggml_backend_t backend, ggml_abort_callback abort_callback, void * abort_callback_data);
+    // Get a list of feature flags supported by the backend (returns a NULL-terminated array)
+    struct ggml_backend_feature {
+        const char * name;
+        const char * value;
+    };
+    typedef struct ggml_backend_feature * (*ggml_backend_get_features_t)(ggml_backend_reg_t reg);
 
     //
     // Backend registry
@@ -214,6 +222,13 @@ extern "C" {
     // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU) OR ggml_backend_dev_by_type(CPU), NULL)
     GGML_API ggml_backend_t ggml_backend_init_best(void);
 
+    // Load a backend from a dynamic library and register it
+    GGML_API ggml_backend_reg_t ggml_backend_load(const char * path);
+    // Unload a backend if loaded dynamically and unregister it
+    GGML_API void               ggml_backend_unload(ggml_backend_reg_t reg);
+    // Load all known backends from dynamic libraries
+    GGML_API void               ggml_backend_load_all(void);
+
     //
     // Backend scheduler
     //
diff --git a/ggml/include/ggml-cpu.h b/ggml/include/ggml-cpu.h
index 7571ef9798364..a5358d047a08e 100644
--- a/ggml/include/ggml-cpu.h
+++ b/ggml/include/ggml-cpu.h
@@ -7,29 +7,6 @@
 extern "C" {
 #endif
 
-    // Scheduling priorities
-    enum ggml_sched_priority {
-        GGML_SCHED_PRIO_NORMAL,
-        GGML_SCHED_PRIO_MEDIUM,
-        GGML_SCHED_PRIO_HIGH,
-        GGML_SCHED_PRIO_REALTIME
-    };
-
-    // Threadpool params
-    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
-    struct ggml_threadpool_params {
-        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
-        int                 n_threads;                   // number of threads
-        enum ggml_sched_priority prio;                   // thread priority
-        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
-        bool                strict_cpu;                  // strict cpu placement
-        bool                paused;                      // start in paused state
-    };
-
-    struct ggml_threadpool;     // forward declaration, see ggml.c
-
-    typedef struct ggml_threadpool * ggml_threadpool_t;
-
     // the compute plan that needs to be prepared for ggml_graph_compute()
     // since https://github.com/ggerganov/ggml/issues/287
     struct ggml_cplan {
@@ -75,14 +52,11 @@ extern "C" {
     GGML_BACKEND_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
     GGML_BACKEND_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
 
-    GGML_BACKEND_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
-    GGML_BACKEND_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
-    GGML_BACKEND_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
-    GGML_BACKEND_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
-    GGML_BACKEND_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
-    GGML_BACKEND_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
-    GGML_BACKEND_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
-    GGML_BACKEND_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API struct ggml_threadpool *      ggml_threadpool_new           (struct ggml_threadpool_params  * params);
+    GGML_BACKEND_API void                          ggml_threadpool_free          (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API int                           ggml_threadpool_get_n_threads (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_pause         (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_resume        (struct ggml_threadpool * threadpool);
 
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
@@ -104,10 +78,10 @@ extern "C" {
     GGML_BACKEND_API int ggml_cpu_has_sse3       (void);
     GGML_BACKEND_API int ggml_cpu_has_ssse3      (void);
     GGML_BACKEND_API int ggml_cpu_has_avx        (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx_vnni   (void);
     GGML_BACKEND_API int ggml_cpu_has_avx2       (void);
     GGML_BACKEND_API int ggml_cpu_has_f16c       (void);
     GGML_BACKEND_API int ggml_cpu_has_fma        (void);
-    GGML_BACKEND_API int ggml_cpu_has_avx_vnni   (void);
     GGML_BACKEND_API int ggml_cpu_has_avx512     (void);
     GGML_BACKEND_API int ggml_cpu_has_avx512_vbmi(void);
     GGML_BACKEND_API int ggml_cpu_has_avx512_vnni(void);
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 69e6a24344b97..9843b09fbe83e 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -2215,6 +2215,37 @@ extern "C" {
 
     GGML_API const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type);
 
+    // ggml threadpool
+    // TODO: currently, only a few functions are in the base ggml API, while the rest are in the CPU backend
+    // the goal should be to create an API that other backends can use move everything to the ggml base
+
+    // scheduling priorities
+    enum ggml_sched_priority {
+        GGML_SCHED_PRIO_NORMAL,
+        GGML_SCHED_PRIO_MEDIUM,
+        GGML_SCHED_PRIO_HIGH,
+        GGML_SCHED_PRIO_REALTIME
+    };
+
+    // threadpool params
+    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
+    struct ggml_threadpool_params {
+        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
+        int                 n_threads;                   // number of threads
+        enum ggml_sched_priority prio;                   // thread priority
+        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
+        bool                strict_cpu;                  // strict cpu placement
+        bool                paused;                      // start in paused state
+    };
+
+    struct ggml_threadpool;     // forward declaration, see ggml.c
+
+    typedef struct ggml_threadpool * ggml_threadpool_t;
+
+    GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
+    GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
+    GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
+
 #ifdef  __cplusplus
 }
 #endif
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index 8df0e85c0d092..071508ddae021 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -202,6 +202,10 @@ endif()
 
 # ggml
 
+if (GGML_BACKEND_DL AND NOT BUILD_SHARED_LIBS)
+    message(FATAL_ERROR "GGML_BACKEND_DL requires BUILD_SHARED_LIBS")
+endif()
+
 add_library(ggml-base
             ../include/ggml.h
             ../include/ggml-alloc.h
@@ -226,6 +230,31 @@ add_library(ggml
 
 target_link_libraries(ggml PUBLIC ggml-base)
 
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    target_link_libraries(ggml PRIVATE dl)
+endif()
+
+function(ggml_add_backend_library backend)
+    if (GGML_BACKEND_DL)
+        add_library(${backend} MODULE ${ARGN})
+        # write the shared library to the output directory
+        set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
+        target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
+    else()
+        add_library(${backend} ${ARGN})
+        target_link_libraries(ggml PUBLIC ${backend})
+        install(TARGETS ${backend} LIBRARY)
+    endif()
+
+    target_link_libraries(${backend} PRIVATE ggml-base)
+    target_include_directories(${backend} PRIVATE ..)
+
+    if (${BUILD_SHARED_LIBS})
+        target_compile_definitions(${backend} PRIVATE GGML_BACKEND_BUILD)
+        target_compile_definitions(${backend} PUBLIC  GGML_BACKEND_SHARED)
+    endif()
+endfunction()
+
 function(ggml_add_backend backend)
     string(TOUPPER "GGML_${backend}" backend_id)
     if (${backend_id})
@@ -236,14 +265,10 @@ function(ggml_add_backend backend)
         # however, currently it is necessary for AMX, since it is enabled by default on llama.cpp
         if (${backend_id})
             message(STATUS "Including ${backend} backend")
-            if (${BUILD_SHARED_LIBS})
-                target_compile_definitions(${backend_target} PRIVATE GGML_BACKEND_BUILD)
-                target_compile_definitions(${backend_target} PUBLIC  GGML_BACKEND_SHARED)
+            if (NOT GGML_BACKEND_DL)
+                string(TOUPPER "GGML_USE_${backend}" backend_use)
+                target_compile_definitions(ggml PUBLIC ${backend_use})
             endif()
-            install(TARGETS ${backend_target} LIBRARY)
-            target_link_libraries(ggml PUBLIC ${backend_target})
-            string(TOUPPER "GGML_USE_${backend}" backend_use)
-            target_compile_definitions(ggml PUBLIC ${backend_use})
         endif()
     endif()
 endfunction()
@@ -256,10 +281,10 @@ ggml_add_backend(CUDA)
 ggml_add_backend(HIP)
 ggml_add_backend(Kompute)
 ggml_add_backend(METAL)
+ggml_add_backend(MUSA)
 ggml_add_backend(RPC)
 ggml_add_backend(SYCL)
 ggml_add_backend(Vulkan)
-ggml_add_backend(MUSA)
 
 foreach (target ggml-base ggml)
     target_include_directories(${target} PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)
diff --git a/ggml/src/ggml-amx/CMakeLists.txt b/ggml/src/ggml-amx/CMakeLists.txt
index d6676f3f67b20..cf3ade6f020ed 100644
--- a/ggml/src/ggml-amx/CMakeLists.txt
+++ b/ggml/src/ggml-amx/CMakeLists.txt
@@ -9,12 +9,10 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MA
 
     file(GLOB   GGML_SOURCES_AMX "*.cpp")
 
-    add_library(ggml-amx
-                ${GGML_HEADERS_AMX}
-                ${GGML_SOURCES_AMX})
-
-    target_link_libraries(ggml-amx PRIVATE ggml-base)
-    target_include_directories(ggml-amx PRIVATE . ..)
+    ggml_add_backend_library(ggml-amx
+                             ${GGML_HEADERS_AMX}
+                             ${GGML_SOURCES_AMX}
+                            )
 
     # this is duplicated from the CPU backend, since the AMX backend also depends on the architecture flags
     # TODO: integrate AMX backend into the CPU backend
diff --git a/ggml/src/ggml-amx/ggml-amx.cpp b/ggml/src/ggml-amx/ggml-amx.cpp
index 8568e7965fd2e..6bfb3da274c39 100644
--- a/ggml/src/ggml-amx/ggml-amx.cpp
+++ b/ggml/src/ggml-amx/ggml-amx.cpp
@@ -409,8 +409,9 @@ static const struct ggml_backend_reg_i ggml_backend_amx_reg_i = {
 
 ggml_backend_reg_t ggml_backend_amx_reg(void) {
     static struct ggml_backend_reg ggml_backend_amx_reg = {
-        /* .iface   = */ ggml_backend_amx_reg_i,
-        /* .context = */ NULL,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_amx_reg_i,
+        /* .context     = */ NULL,
     };
 
     return &ggml_backend_amx_reg;
@@ -444,3 +445,5 @@ ggml_backend_reg_t ggml_backend_amx_reg(void) {
 }
 
 #endif
+
+GGML_BACKEND_DL_IMPL(ggml_backend_amx_reg)
diff --git a/ggml/src/ggml-backend-impl.h b/ggml/src/ggml-backend-impl.h
index fa8d5b7fb68c9..dff7749b416dc 100644
--- a/ggml/src/ggml-backend-impl.h
+++ b/ggml/src/ggml-backend-impl.h
@@ -8,6 +8,8 @@
 extern "C" {
 #endif
 
+    #define GGML_BACKEND_API_VERSION 1
+
     //
     // Backend buffer type
     //
@@ -63,20 +65,20 @@ extern "C" {
         enum ggml_backend_buffer_usage usage;
     };
 
-    ggml_backend_buffer_t ggml_backend_buffer_init(
+    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
                    ggml_backend_buffer_type_t buft,
             struct ggml_backend_buffer_i      iface,
                    void *                     context,
                    size_t                     size);
 
     // do not use directly, use ggml_backend_tensor_copy instead
-    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
 
     // multi-buffer
     // buffer that contains a collection of buffers
-    ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
-    bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
-    void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+    GGML_API bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+    GGML_API void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
 
     //
     // Backend (stream)
@@ -199,17 +201,37 @@ extern "C" {
     };
 
     struct ggml_backend_reg {
-        // int api_version; // TODO: for dynamic loading
+        int api_version; // initialize to GGML_BACKEND_API_VERSION
         struct ggml_backend_reg_i iface;
         void * context;
     };
 
-
     // Internal backend registry API
-    void ggml_backend_register(ggml_backend_reg_t reg);
-    void ggml_backend_device_register(ggml_backend_dev_t device);
-    // TODO: backends can be loaded as a dynamic library, in which case it needs to export this function
-    // typedef ggml_backend_register_t * (*ggml_backend_init)(void);
+    GGML_API void ggml_backend_register(ggml_backend_reg_t reg);
+    GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
+
+    // Add backend dynamic loading support to the backend
+    typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
+
+    #ifdef GGML_BACKEND_DL
+        #ifdef __cplusplus
+        #    define GGML_BACKEND_DL_IMPL(reg_fn)                                 \
+                extern "C" {                                                     \
+                    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
+                }                                                                \
+                ggml_backend_reg_t ggml_backend_init(void) {                     \
+                    return reg_fn();                                             \
+                }
+        #else
+        #    define GGML_BACKEND_DL_IMPL(reg_fn)                             \
+                GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
+                ggml_backend_reg_t ggml_backend_init(void) {                 \
+                    return reg_fn();                                         \
+                }
+        #endif
+    #else
+    #    define GGML_BACKEND_DL_IMPL(reg_fn)
+    #endif
 
 #ifdef  __cplusplus
 }
diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
index 63e9d82017457..43d03d7fa7385 100644
--- a/ggml/src/ggml-backend-reg.cpp
+++ b/ggml/src/ggml-backend-reg.cpp
@@ -1,11 +1,29 @@
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
-#include "ggml-cpu.h"
 #include "ggml-impl.h"
+#include <algorithm>
 #include <cstring>
+#include <string>
 #include <vector>
 
+#ifdef _WIN32
+#    define WIN32_LEAN_AND_MEAN
+#    ifndef NOMINMAX
+#        define NOMINMAX
+#    endif
+#    include <windows.h>
+#elif defined(__APPLE__)
+#    include <mach-o/dyld.h>
+#    include <dlfcn.h>
+#else
+#    include <dlfcn.h>
+#    include <unistd.h>
+#endif
+
 // Backend registry
+#ifdef GGML_USE_CPU
+#include "ggml-cpu.h"
+#endif
 
 #ifdef GGML_USE_CUDA
 #include "ggml-cuda.h"
@@ -43,8 +61,13 @@
 #include "ggml-kompute.h"
 #endif
 
+struct ggml_backend_reg_entry {
+    ggml_backend_reg_t reg;
+    void * handle;
+};
+
 struct ggml_backend_registry {
-    std::vector<ggml_backend_reg_t> backends;
+    std::vector<ggml_backend_reg_entry> backends;
     std::vector<ggml_backend_dev_t> devices;
 
     ggml_backend_registry() {
@@ -75,11 +98,19 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_KOMPUTE
         register_backend(ggml_backend_kompute_reg());
 #endif
-
+#ifdef GGML_USE_CPU
         register_backend(ggml_backend_cpu_reg());
+#endif
     }
 
-    void register_backend(ggml_backend_reg_t reg) {
+    ~ggml_backend_registry() {
+        while (!backends.empty()) {
+            // use silent since the log system may have been destroyed at this point
+            unload_backend(backends.back().reg, true);
+        }
+    }
+
+    void register_backend(ggml_backend_reg_t reg, void * handle = nullptr) {
         if (!reg) {
             return;
         }
@@ -88,7 +119,7 @@ struct ggml_backend_registry {
         GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
             __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
 #endif
-        backends.push_back(reg);
+        backends.push_back({ reg, handle });
         for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
             register_device(ggml_backend_reg_dev_get(reg, i));
         }
@@ -100,6 +131,111 @@ struct ggml_backend_registry {
 #endif
         devices.push_back(device);
     }
+
+    ggml_backend_reg_t load_backend(const char * path, bool silent) {
+#ifdef _WIN32
+        // suppress error dialogs for missing DLLs
+        DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+        SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+        HMODULE handle = LoadLibraryA(path);
+
+        if (!handle) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: failed to load %s: %lu\n", __func__, path, GetLastError());
+            }
+            SetErrorMode(old_mode);
+            return nullptr;
+        }
+
+        ggml_backend_init_t backend_init = (ggml_backend_init_t) GetProcAddress(handle, "ggml_backend_init");
+
+        SetErrorMode(old_mode);
+
+        if (!backend_init) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s: %lu\n", __func__, path, GetLastError());
+            }
+            FreeLibrary(handle);
+            return nullptr;
+        }
+#else
+        void * handle = dlopen(path, RTLD_NOW | RTLD_LOCAL);
+
+        if (!handle) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: failed to load %s: %s\n", __func__, path, dlerror());
+            }
+            return nullptr;
+        }
+
+        auto * backend_init = (ggml_backend_init_t) dlsym(handle, "ggml_backend_init");
+
+        if (!backend_init) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s: %s\n", __func__, path, dlerror());
+            }
+            dlclose(handle);
+            return nullptr;
+        }
+#endif
+        ggml_backend_reg_t reg = backend_init();
+
+        if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
+            if (!silent) {
+                if (!reg) {
+                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, path);
+                } else {
+                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
+                                   __func__, path, reg->api_version, GGML_BACKEND_API_VERSION);
+                }
+            }
+#ifdef _WIN32
+            FreeLibrary(handle);
+#else
+            dlclose(handle);
+#endif
+            return nullptr;
+        }
+
+        GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path);
+        register_backend(reg, handle);
+        return reg;
+    }
+
+    void unload_backend(ggml_backend_reg_t reg, bool silent) {
+        auto it = std::find_if(backends.begin(), backends.end(),
+                                [reg](ggml_backend_reg_entry entry) { return entry.reg == reg; });
+
+        if (it == backends.end()) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: backend not found\n", __func__);
+            }
+            return;
+        }
+
+        if (!silent) {
+            GGML_LOG_DEBUG("%s: unloading %s backend\n", __func__, ggml_backend_reg_name(reg));
+        }
+
+        // remove devices
+        devices.erase(
+            std::remove_if(devices.begin(), devices.end(),
+                            [reg](ggml_backend_dev_t dev) { return ggml_backend_dev_backend_reg(dev) == reg; }),
+            devices.end());
+
+        // unload library
+        if (it->handle) {
+#ifdef _WIN32
+            FreeLibrary((HMODULE) it->handle);
+#else
+            dlclose(it->handle);
+#endif
+        }
+
+        // remove backend
+        backends.erase(it);
+    }
 };
 
 static ggml_backend_registry & get_reg() {
@@ -123,7 +259,7 @@ size_t ggml_backend_reg_count() {
 
 ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
     GGML_ASSERT(index < ggml_backend_reg_count());
-    return get_reg().backends[index];
+    return get_reg().backends[index].reg;
 }
 
 ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
@@ -133,7 +269,7 @@ ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
             return reg;
         }
     }
-    return NULL;
+    return nullptr;
 }
 
 // Device enumeration
@@ -153,7 +289,7 @@ ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
             return dev;
         }
     }
-    return NULL;
+    return nullptr;
 }
 
 ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
@@ -163,14 +299,14 @@ ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
             return dev;
         }
     }
-    return NULL;
+    return nullptr;
 }
 
 // Convenience functions
 ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
     ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
     if (!dev) {
-        return NULL;
+        return nullptr;
     }
     return ggml_backend_dev_init(dev, params);
 }
@@ -178,7 +314,7 @@ ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params)
 ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
     ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
     if (!dev) {
-        return NULL;
+        return nullptr;
     }
     return ggml_backend_dev_init(dev, params);
 }
@@ -189,7 +325,97 @@ ggml_backend_t ggml_backend_init_best(void) {
         dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
     }
     if (!dev) {
-        return NULL;
+        return nullptr;
     }
-    return ggml_backend_dev_init(dev, NULL);
+    return ggml_backend_dev_init(dev, nullptr);
+}
+
+// Dynamic loading
+ggml_backend_reg_t ggml_backend_load(const char * path) {
+    return get_reg().load_backend(path, false);
+}
+
+void ggml_backend_unload(ggml_backend_reg_t reg) {
+    get_reg().unload_backend(reg, true);
+}
+
+void ggml_backend_load_all() {
+    std::vector<std::string> search_prefix;
+
+    // add the executable directory to the search path
+    // FIXME: this is convenient for development, but it should probably be disabled in production
+
+#if defined(__APPLE__)
+    // get executable path
+    std::vector<char> path;
+    uint32_t size;
+    while (true) {
+        size = path.size();
+        if (_NSGetExecutablePath(path.data(), &size) == 0) {
+            break;
+        }
+        path.resize(size);
+    }
+    std::string base_path(path.data(), size);
+    // remove executable name
+    auto last_slash = base_path.find_last_of('/');
+    if (last_slash != std::string::npos) {
+        base_path = base_path.substr(0, last_slash);
+    }
+    search_prefix.push_back(base_path + "/");
+#elif defined(__linux__)
+    std::string base_path = ".";
+    std::vector<char> path(1024);
+    while (true) {
+        // get executable path
+        ssize_t len = readlink("/proc/self/exe", path.data(), path.size());
+        if (len == -1) {
+            break;
+        }
+        if (len < (ssize_t) path.size()) {
+            base_path = std::string(path.data(), len);
+            // remove executable name
+            auto last_slash = base_path.find_last_of('/');
+            if (last_slash != std::string::npos) {
+                base_path = base_path.substr(0, last_slash);
+            }
+            break;
+        }
+        path.resize(path.size() * 2);
+    }
+
+    search_prefix.push_back(base_path + "/");
+#endif
+
+    auto & reg = get_reg();
+
+    auto try_load = [&](const std::string & name) {
+        std::string os_name;
+#ifdef _WIN32
+        os_name = "ggml-" + name + ".dll";
+#else
+        os_name = "libggml-" + name + ".so";
+#endif
+        if (reg.load_backend(os_name.c_str(), true)) {
+            return;
+        }
+        for (const auto & prefix : search_prefix) {
+            if (reg.load_backend((prefix + os_name).c_str(), true)) {
+                return;
+            }
+        }
+    };
+
+    try_load("amx");
+    try_load("blas");
+    try_load("cann");
+    try_load("cuda");
+    try_load("hip");
+    try_load("kompute");
+    try_load("metal");
+    try_load("rpc");
+    try_load("sycl");
+    try_load("vulkan");
+    try_load("musa");
+    try_load("cpu");
 }
diff --git a/ggml/src/ggml-blas/CMakeLists.txt b/ggml/src/ggml-blas/CMakeLists.txt
index e2cbabf0dae74..0bf3c05d93a89 100644
--- a/ggml/src/ggml-blas/CMakeLists.txt
+++ b/ggml/src/ggml-blas/CMakeLists.txt
@@ -11,12 +11,9 @@ find_package(BLAS)
 if (BLAS_FOUND)
     message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
 
-    add_library(ggml-blas
-                ggml-blas.cpp
-                )
-
-    target_link_libraries(ggml-blas PRIVATE ggml-base)
-    target_include_directories(ggml-blas PRIVATE . ..)
+    ggml_add_backend_library(ggml-blas
+                             ggml-blas.cpp
+                            )
 
     if (${GGML_BLAS_VENDOR} MATCHES "Apple")
         add_compile_definitions(ACCELERATE_NEW_LAPACK)
diff --git a/ggml/src/ggml-blas/ggml-blas.cpp b/ggml/src/ggml-blas/ggml-blas.cpp
index 648c9d875e346..ec158dfac6e3e 100644
--- a/ggml/src/ggml-blas/ggml-blas.cpp
+++ b/ggml/src/ggml-blas/ggml-blas.cpp
@@ -506,9 +506,12 @@ static const struct ggml_backend_reg_i ggml_backend_blas_reg_i = {
 
 ggml_backend_reg_t ggml_backend_blas_reg(void) {
     static struct ggml_backend_reg ggml_backend_blas_reg = {
-        /* .iface   = */ ggml_backend_blas_reg_i,
-        /* .context = */ NULL,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_blas_reg_i,
+        /* .context     = */ NULL,
     };
 
     return &ggml_backend_blas_reg;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_blas_reg)
diff --git a/ggml/src/ggml-cann/CMakeLists.txt b/ggml/src/ggml-cann/CMakeLists.txt
index 756200b893d02..901327185fb75 100644
--- a/ggml/src/ggml-cann/CMakeLists.txt
+++ b/ggml/src/ggml-cann/CMakeLists.txt
@@ -61,9 +61,9 @@ if (CANN_INSTALL_DIR)
 
     file(GLOB GGML_SOURCES_CANN "*.cpp")
 
-    add_library(ggml-cann ${GGML_SOURCES_CANN})
-    target_link_libraries(ggml-cann PRIVATE ggml-base ${CANN_LIBRARIES})
-    target_include_directories(ggml-cann PRIVATE . .. ${CANN_INCLUDE_DIRS})
+    ggml_add_backend_library(ggml-cann ${GGML_SOURCES_CANN})
+    target_link_libraries(ggml-cann PRIVATE ${CANN_LIBRARIES})
+    target_include_directories(ggml-cann PRIVATE ${CANN_INCLUDE_DIRS})
     target_link_directories(ggml-cann PRIVATE ${CANN_INSTALL_DIR}/lib64)
 
     target_compile_definitions(ggml-cann PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index 776340881434d..d96f65936136d 100644
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -2064,16 +2064,17 @@ ggml_backend_reg_t ggml_backend_cann_reg() {
                 dev_ctx->name = GGML_CANN_NAME + std::to_string(i);
                 ggml_cann_set_device(i);
                 ggml_backend_dev_t dev = new ggml_backend_device {
-                    /* .interface = */ ggml_backend_cann_device_interface,
-                    /* .reg       = */ &reg,
-                    /* .context   = */ dev_ctx
+                    /* .iface   = */ ggml_backend_cann_device_interface,
+                    /* .reg     = */ &reg,
+                    /* .context = */ dev_ctx
                 };
                 ctx->devices.push_back(dev);
             }
 
             reg = ggml_backend_reg {
-                /* .interface = */ ggml_backend_cann_reg_interface,
-                /* .context   = */ ctx
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_cann_reg_interface,
+                /* .context     = */ ctx
             };
         }
 
@@ -2126,3 +2127,5 @@ void ggml_backend_cann_get_device_memory(int32_t device, size_t* free,
     ggml_cann_set_device(device);
     ACL_CHECK(aclrtGetMemInfo(ACL_HBM_MEM, free, total));
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_cann_reg)
diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt
index 2880523331dbd..c2905d1fbf4e8 100644
--- a/ggml/src/ggml-cpu/CMakeLists.txt
+++ b/ggml/src/ggml-cpu/CMakeLists.txt
@@ -1,14 +1,13 @@
-add_library(ggml-cpu
-            ggml-cpu.c
-            ggml-cpu.cpp
-            ggml-cpu-aarch64.c
-            ggml-cpu-aarch64.h
-            ggml-cpu-quants.c
-            ggml-cpu-quants.h
-            )
+ggml_add_backend_library(ggml-cpu
+                         ggml-cpu.c
+                         ggml-cpu.cpp
+                         ggml-cpu-aarch64.c
+                         ggml-cpu-aarch64.h
+                         ggml-cpu-quants.c
+                         ggml-cpu-quants.h
+                        )
 
-target_link_libraries(ggml-cpu PRIVATE ggml-base)
-target_include_directories(ggml-cpu PRIVATE . ..)
+target_include_directories(ggml-cpu PRIVATE .)
 
 if (APPLE AND GGML_ACCELERATE)
     find_library(ACCELERATE_FRAMEWORK Accelerate)
diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
index 4b58254e7d108..c6ede19d9d1c0 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -13578,29 +13578,6 @@ static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int
 
 #endif // GGML_USE_OPENMP
 
-void ggml_threadpool_params_init(struct ggml_threadpool_params * p, int n_threads) {
-    p->n_threads  = n_threads;
-    p->prio       = 0;     // default priority (usually means normal or inherited)
-    p->poll       = 50;    // hybrid-polling enabled
-    p->strict_cpu = false; // no strict placement (all threads share same cpumask)
-    p->paused     = false; // threads are ready to go
-    memset(p->cpumask, 0, GGML_MAX_N_THREADS); // all-zero means use the default affinity (usually inherited)
-}
-
-struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) {
-    struct ggml_threadpool_params p;
-    ggml_threadpool_params_init(&p, n_threads);
-    return p;
-}
-
-bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) {
-    if (p0->n_threads      != p1->n_threads  )    return false;
-    if (p0->prio           != p1->prio       )    return false;
-    if (p0->poll           != p1->poll       )    return false;
-    if (p0->strict_cpu     != p1->strict_cpu )    return false;
-    return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0;
-}
-
 static struct ggml_threadpool * ggml_threadpool_new_impl(
     struct ggml_threadpool_params * tpp,
                struct ggml_cgraph * cgraph,
diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp
index 573b7c5b9b375..febed433ada2b 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.cpp
+++ b/ggml/src/ggml-cpu/ggml-cpu.cpp
@@ -541,16 +541,12 @@ static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg
     return &ggml_backend_cpu_device;
 }
 
-struct ggml_backend_feature {
-    const char * name;
-    const char * value;
-};
-
-// Not used yet
 // This is intended to replace the the ggml_cpu_has_* functions when loading the CPU backend dynamically,
-// and additionally to allow other backends to expose their own list of features that applications can query using the same API.
+// and additionally to allow other backends to expose their own list of features that applications can query using the same API
 static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t reg) {
     static std::vector<ggml_backend_feature> features = []() {
+        ggml_cpu_init();
+
         std::vector<ggml_backend_feature> features;
         if (ggml_cpu_has_sse3()) {
             features.push_back({ "SSE3", "1" });
@@ -561,6 +557,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_avx()) {
             features.push_back({ "AVX", "1" });
         }
+        if (ggml_cpu_has_avx_vnni()) {
+            features.push_back({ "AVX_VNNI", "1" });
+        }
         if (ggml_cpu_has_avx2()) {
             features.push_back({ "AVX2", "1" });
         }
@@ -570,9 +569,6 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_fma()) {
             features.push_back({ "FMA", "1" });
         }
-        if (ggml_cpu_has_avx_vnni()) {
-            features.push_back({ "AVX_VNNI", "1" });
-        }
         if (ggml_cpu_has_avx512()) {
             features.push_back({ "AVX512", "1" });
         }
@@ -619,6 +615,10 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_llamafile()) {
             features.push_back({ "LLAMAFILE", "1" });
         }
+        // TODO: rename this
+    #ifdef GGML_USE_CPU_AARCH64
+        features.push_back({ "AARCH64_REPACK", "1" });
+    #endif
 
         features.push_back({ nullptr, nullptr });
 
@@ -637,6 +637,29 @@ static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const ch
     if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
         return (void *)ggml_backend_cpu_get_extra_bufts;
     }
+    if (strcmp(name, "ggml_backend_get_features") == 0) {
+        return (void *)ggml_backend_cpu_get_features;
+    }
+    if (strcmp(name, "ggml_backend_set_abort_callback") == 0) {
+        return (void *)ggml_backend_cpu_set_abort_callback;
+    }
+    if (strcmp(name, "ggml_backend_cpu_numa_init") == 0) {
+        return (void *)ggml_numa_init;
+    }
+    if (strcmp(name, "ggml_backend_cpu_is_numa") == 0) {
+        return (void *)ggml_is_numa;
+    }
+
+    // threadpool - TODO:  move to ggml-base
+    if (strcmp(name, "ggml_threadpool_new") == 0) {
+        return (void *)ggml_threadpool_new;
+    }
+    if (strcmp(name, "ggml_threadpool_free") == 0) {
+        return (void *)ggml_threadpool_free;
+    }
+    if (strcmp(name, "ggml_backend_cpu_set_threadpool") == 0) {
+        return (void *)ggml_backend_cpu_set_threadpool;
+    }
 
     return NULL;
 
@@ -655,9 +678,12 @@ ggml_backend_reg_t ggml_backend_cpu_reg(void) {
     ggml_cpu_init();
 
     static struct ggml_backend_reg ggml_backend_cpu_reg = {
-        /* .iface   = */ ggml_backend_cpu_reg_i,
-        /* .context = */ NULL,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_cpu_reg_i,
+        /* .context     = */ NULL,
     };
 
     return &ggml_backend_cpu_reg;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_cpu_reg)
diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt
index e1482a269d698..b0cb93e070fd3 100644
--- a/ggml/src/ggml-cuda/CMakeLists.txt
+++ b/ggml/src/ggml-cuda/CMakeLists.txt
@@ -46,13 +46,10 @@ if (CUDAToolkit_FOUND)
         list(APPEND GGML_SOURCES_CUDA ${SRCS})
     endif()
 
-    add_library(ggml-cuda
-                ${GGML_HEADERS_CUDA}
-                ${GGML_SOURCES_CUDA}
-                )
-
-    target_link_libraries(ggml-cuda PRIVATE ggml-base)
-    target_include_directories(ggml-cuda PRIVATE . ..)
+    ggml_add_backend_library(ggml-cuda
+                             ${GGML_HEADERS_CUDA}
+                             ${GGML_SOURCES_CUDA}
+                            )
 
     add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
 
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index dd94ab03d5b6c..2a78a4393d0f7 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -3126,6 +3126,61 @@ static ggml_backend_dev_t ggml_backend_cuda_reg_get_device(ggml_backend_reg_t re
     return ctx->devices[index];
 }
 
+static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t reg) {
+    static std::vector<ggml_backend_feature> features = []() {
+        std::vector<ggml_backend_feature> features;
+    #define _STRINGIFY(...) #__VA_ARGS__
+    #define STRINGIFY(...) _STRINGIFY(__VA_ARGS__)
+
+    #ifdef __CUDA_ARCH_LIST__
+        features.push_back({ "ARCHS", STRINGIFY(__CUDA_ARCH_LIST__) });
+    #endif
+
+    #ifdef GGML_CUDA_FORCE_MMQ
+        features.push_back({ "FORCE_MMQ", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_FORCE_CUBLAS
+        features.push_back({ "FORCE_CUBLAS", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_NO_VMM
+        features.push_back({ "NO_VMM", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_NO_PEER_COPY
+        features.push_back({ "NO_PEER_COPY", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_F16
+        features.push_back({ "F16", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_USE_GRAPHS
+        features.push_back({ "USE_GRAPHS", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
+        features.push_back({ "PEER_MAX_BATCH_SIZE", STRINGIFY(GGML_CUDA_PEER_MAX_BATCH_SIZE) });
+    #endif
+
+    #ifdef GGML_CUDA_FA_ALL_QUANTS
+        features.push_back({ "FA_ALL_QUANTS", "1" });
+    #endif
+
+    #undef _STRINGIFY
+    #undef STRINGIFY
+
+        features.push_back({ nullptr, nullptr });
+
+        return features;
+    }();
+
+    return features.data();
+
+    GGML_UNUSED(reg);
+}
+
 static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
     GGML_UNUSED(reg);
     if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
@@ -3137,6 +3192,9 @@ static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, con
     if (strcmp(name, "ggml_backend_unregister_host_buffer") == 0) {
         return (void *)ggml_backend_cuda_unregister_host_buffer;
     }
+    if (strcmp(name, "ggml_backend_get_features") == 0) {
+        return (void *)ggml_backend_cuda_get_features;
+    }
     return nullptr;
 }
 
@@ -3169,16 +3227,17 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 dev_ctx->description = prop.name;
 
                 ggml_backend_dev_t dev = new ggml_backend_device {
-                    /* .interface = */ ggml_backend_cuda_device_interface,
-                    /* .reg       = */ &reg,
-                    /* .context   = */ dev_ctx
+                    /* .iface   = */ ggml_backend_cuda_device_interface,
+                    /* .reg     = */ &reg,
+                    /* .context = */ dev_ctx
                 };
                 ctx->devices.push_back(dev);
             }
 
             reg = ggml_backend_reg {
-                /* .interface = */ ggml_backend_cuda_reg_interface,
-                /* .context   = */ ctx
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_cuda_reg_interface,
+                /* .context     = */ ctx
             };
         }
 
@@ -3209,3 +3268,5 @@ ggml_backend_t ggml_backend_cuda_init(int device) {
 
     return cuda_backend;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_cuda_reg)
diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt
index fccf8eb8440b8..b15fbd24d6b36 100644
--- a/ggml/src/ggml-hip/CMakeLists.txt
+++ b/ggml/src/ggml-hip/CMakeLists.txt
@@ -64,12 +64,10 @@ else()
     list(APPEND GGML_SOURCES_ROCM ${SRCS})
 endif()
 
-add_library(ggml-hip
-            ${GGML_HEADERS_ROCM}
-            ${GGML_SOURCES_ROCM})
-
-target_link_libraries(ggml-hip PRIVATE ggml-base)
-target_include_directories(ggml-hip PRIVATE . ..)
+ggml_add_backend_library(ggml-hip
+                         ${GGML_HEADERS_ROCM}
+                         ${GGML_SOURCES_ROCM}
+                        )
 
 # TODO: do not use CUDA definitions for HIP
 target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
diff --git a/ggml/src/ggml-kompute/CMakeLists.txt b/ggml/src/ggml-kompute/CMakeLists.txt
index 0bd027c7f537e..dc623926c7685 100644
--- a/ggml/src/ggml-kompute/CMakeLists.txt
+++ b/ggml/src/ggml-kompute/CMakeLists.txt
@@ -6,13 +6,13 @@ if (NOT glslc_executable)
     message(FATAL_ERROR "glslc not found")
 endif()
 
-add_library(ggml-kompute
-            ggml-kompute.cpp
-            ../../include/ggml-kompute.h
-            )
+ggml_add_backend_library(ggml-kompute
+                         ggml-kompute.cpp
+                         ../../include/ggml-kompute.h
+                        )
 
 target_link_libraries(ggml-kompute PRIVATE ggml-base kompute)
-target_include_directories(ggml-kompute PRIVATE . .. ${CMAKE_CURRENT_BINARY_DIR})
+target_include_directories(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
 
 add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
 
diff --git a/ggml/src/ggml-kompute/ggml-kompute.cpp b/ggml/src/ggml-kompute/ggml-kompute.cpp
index 2fea9e4cc8d38..24566404ded0f 100644
--- a/ggml/src/ggml-kompute/ggml-kompute.cpp
+++ b/ggml/src/ggml-kompute/ggml-kompute.cpp
@@ -2176,9 +2176,12 @@ static const struct ggml_backend_reg_i ggml_backend_kompute_reg_i = {
 
 ggml_backend_reg_t ggml_backend_kompute_reg() {
     static ggml_backend_reg reg = {
-        /* .iface   = */ ggml_backend_kompute_reg_i,
-        /* .context = */ nullptr,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_kompute_reg_i,
+        /* .context     = */ nullptr,
     };
 
     return &reg;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_kompute_reg)
diff --git a/ggml/src/ggml-metal/CMakeLists.txt b/ggml/src/ggml-metal/CMakeLists.txt
index b237d79f47ddb..1bad272068244 100644
--- a/ggml/src/ggml-metal/CMakeLists.txt
+++ b/ggml/src/ggml-metal/CMakeLists.txt
@@ -4,19 +4,16 @@ find_library(METALKIT_FRAMEWORK MetalKit   REQUIRED)
 
 message(STATUS "Metal framework found")
 
-add_library(ggml-metal
-            ggml-metal.m
-            )
+ggml_add_backend_library(ggml-metal
+                         ggml-metal.m
+                        )
 
 target_link_libraries(ggml-metal PRIVATE
-                      ggml-base
                       ${FOUNDATION_LIBRARY}
                       ${METAL_FRAMEWORK}
                       ${METALKIT_FRAMEWORK}
                       )
 
-target_include_directories(ggml-metal PRIVATE . ..)
-
 if (GGML_METAL_NDEBUG)
     add_compile_definitions(GGML_METAL_NDEBUG)
 endif()
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index 3a533d7f9c9af..63baaf163581c 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -4372,19 +4372,45 @@ static ggml_backend_dev_t ggml_backend_metal_reg_device_get(ggml_backend_reg_t r
     GGML_UNUSED(index);
 }
 
+static struct ggml_backend_feature g_ggml_backend_metal_features[] = {
+#if defined(GGML_METAL_EMBED_LIBRARY)
+    { "EMBED_LIBRARY", "1" },
+#endif
+#if defined(GGML_METAL_USE_BF16)
+    { "BF16", "1" },
+#endif
+    { nil, nil },
+};
+
+static struct ggml_backend_feature * ggml_backend_metal_get_features(ggml_backend_reg_t reg) {
+    return g_ggml_backend_metal_features;
+
+    GGML_UNUSED(reg);
+}
+
+static void * ggml_backend_metal_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (strcmp(name, "ggml_backend_get_features") == 0) {
+        return (void *)ggml_backend_metal_get_features;
+    }
+
+    return NULL;
+
+    GGML_UNUSED(reg);
+}
 static struct ggml_backend_reg_i ggml_backend_metal_reg_i = {
     /* .get_name         = */ ggml_backend_metal_reg_get_name,
     /* .device_count     = */ ggml_backend_metal_reg_device_count,
     /* .device_get       = */ ggml_backend_metal_reg_device_get,
-    /* .get_proc_address = */ NULL,
+    /* .get_proc_address = */ ggml_backend_metal_get_proc_address,
 };
 
 ggml_backend_reg_t ggml_backend_metal_reg(void) {
     // TODO: make this thread-safe somehow?
     {
         g_ggml_backend_metal_reg = (struct ggml_backend_reg) {
-            /* .iface   = */ ggml_backend_metal_reg_i,
-            /* .context = */ NULL,
+            /* .api_version = */ GGML_BACKEND_API_VERSION,
+            /* .iface       = */ ggml_backend_metal_reg_i,
+            /* .context     = */ NULL,
         };
 
         g_ggml_backend_metal_device = (struct ggml_backend_device) {
@@ -4396,3 +4422,5 @@ ggml_backend_reg_t ggml_backend_metal_reg(void) {
 
     return &g_ggml_backend_metal_reg;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_metal_reg)
diff --git a/ggml/src/ggml-musa/CMakeLists.txt b/ggml/src/ggml-musa/CMakeLists.txt
index f3c0136920540..e1a69186e669f 100644
--- a/ggml/src/ggml-musa/CMakeLists.txt
+++ b/ggml/src/ggml-musa/CMakeLists.txt
@@ -47,12 +47,10 @@ if (MUSAToolkit_FOUND)
         set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22")
     endforeach()
 
-    add_library(ggml-musa
-                ${GGML_HEADERS_MUSA}
-                ${GGML_SOURCES_MUSA})
-
-    target_link_libraries(ggml-musa PRIVATE ggml-base)
-    target_include_directories(ggml-musa PRIVATE . ..)
+    ggml_add_backend_library(ggml-musa
+                             ${GGML_HEADERS_MUSA}
+                             ${GGML_SOURCES_MUSA}
+                            )
 
     # TODO: do not use CUDA definitions for MUSA
     target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
diff --git a/ggml/src/ggml-rpc/CMakeLists.txt b/ggml/src/ggml-rpc/CMakeLists.txt
index a2d6770eb053f..f5acb8ec2cb28 100644
--- a/ggml/src/ggml-rpc/CMakeLists.txt
+++ b/ggml/src/ggml-rpc/CMakeLists.txt
@@ -1,10 +1,8 @@
 message(STATUS "Using RPC backend")
 
-add_library(ggml-rpc
-            ggml-rpc.cpp)
-
-target_link_libraries(ggml-rpc PRIVATE ggml-base)
-target_include_directories(ggml-rpc PRIVATE . ..)
+ggml_add_backend_library(ggml-rpc
+                         ggml-rpc.cpp
+                        )
 
 if (WIN32)
     target_link_libraries(ggml-rpc PRIVATE ws2_32)
diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
index 47357daabdf54..43108242639a3 100644
--- a/ggml/src/ggml-rpc/ggml-rpc.cpp
+++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
@@ -1369,8 +1369,9 @@ static const struct ggml_backend_reg_i ggml_backend_rpc_reg_i = {
 
 ggml_backend_reg_t ggml_backend_rpc_reg(void) {
     static struct ggml_backend_reg ggml_backend_rpc_reg = {
-        /* .iface   = */ ggml_backend_rpc_reg_i,
-        /* .context = */ NULL,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_rpc_reg_i,
+        /* .context     = */ NULL,
     };
 
     return &ggml_backend_rpc_reg;
@@ -1401,3 +1402,5 @@ ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint) {
 
     return dev;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_rpc_reg)
diff --git a/ggml/src/ggml-sycl/CMakeLists.txt b/ggml/src/ggml-sycl/CMakeLists.txt
index d1d0ff83d636c..83f223fd7b6fc 100644
--- a/ggml/src/ggml-sycl/CMakeLists.txt
+++ b/ggml/src/ggml-sycl/CMakeLists.txt
@@ -16,12 +16,10 @@ endif()
 message(STATUS "SYCL found")
 #todo: AOT
 
-add_library(ggml-sycl
-            ggml-sycl.cpp
-            ../../include/ggml-sycl.h)
-
-target_link_libraries(ggml-sycl PRIVATE ggml-base)
-target_include_directories(ggml-sycl PRIVATE . ..)
+ggml_add_backend_library(ggml-sycl
+                         ggml-sycl.cpp
+                         ../../include/ggml-sycl.h
+                        )
 
 if (GGML_SYCL_F16)
     if (GGML_SYCL_TARGET STREQUAL "AMD")
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index 255bc64c6badd..b6392ed8dcc6a 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -4637,16 +4637,17 @@ ggml_backend_reg_t ggml_backend_sycl_reg() {
                 dev_ctx->description = prop.get_name();
 
                 ggml_backend_dev_t dev = new ggml_backend_device {
-                    /* .interface = */ ggml_backend_sycl_device_interface,
-                    /* .reg       = */ &reg,
-                    /* .context   = */ dev_ctx
+                    /* .iface       = */ ggml_backend_sycl_device_interface,
+                    /* .reg         = */ &reg,
+                    /* .context     = */ dev_ctx
                 };
                 ctx->devices.push_back(dev);
             }
 
             reg = ggml_backend_reg {
-                /* .interface = */ ggml_backend_sycl_reg_interface,
-                /* .context   = */ ctx
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_sycl_reg_interface,
+                /* .context     = */ ctx
             };
         }
 
@@ -4678,3 +4679,4 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
     return sycl_backend;
 }
 
+GGML_BACKEND_DL_IMPL(ggml_backend_sycl_reg)
diff --git a/ggml/src/ggml-vulkan/CMakeLists.txt b/ggml/src/ggml-vulkan/CMakeLists.txt
index 1e85dd15b7ab1..ae0485e04255d 100644
--- a/ggml/src/ggml-vulkan/CMakeLists.txt
+++ b/ggml/src/ggml-vulkan/CMakeLists.txt
@@ -3,13 +3,13 @@ find_package(Vulkan COMPONENTS glslc REQUIRED)
 if (Vulkan_FOUND)
     message(STATUS "Vulkan found")
 
-    add_library(ggml-vulkan
-                ggml-vulkan.cpp
-                ../../include/ggml-vulkan.h
-                )
+    ggml_add_backend_library(ggml-vulkan
+                             ggml-vulkan.cpp
+                             ../../include/ggml-vulkan.h
+                            )
 
-    target_link_libraries(ggml-vulkan PRIVATE ggml-base Vulkan::Vulkan)
-    target_include_directories(ggml-vulkan PRIVATE . .. ${CMAKE_CURRENT_BINARY_DIR})
+    target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
+    target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
 
     # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
     # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index ca71da2f7b7f5..49527fdf40e94 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -6738,8 +6738,9 @@ static const struct ggml_backend_reg_i ggml_backend_vk_reg_i = {
 
 ggml_backend_reg_t ggml_backend_vk_reg() {
     static ggml_backend_reg reg = {
-        /* .iface   = */ ggml_backend_vk_reg_i,
-        /* .context = */ nullptr,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_vk_reg_i,
+        /* .context     = */ nullptr,
     };
 
     return &reg;
@@ -7365,3 +7366,5 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
     VK_LOG_DEBUG("END ggml_vk_check_results_1(" << tensor->name << ")");
 }
 #endif
+
+GGML_BACKEND_DL_IMPL(ggml_backend_vk_reg)
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 78e7874dee04d..1a2318cb188c4 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -7571,3 +7571,26 @@ void ggml_log_set(ggml_log_callback log_callback, void * user_data) {
     g_logger_state.log_callback = log_callback ? log_callback : ggml_log_callback_default;
     g_logger_state.log_callback_user_data = user_data;
 }
+
+void ggml_threadpool_params_init(struct ggml_threadpool_params * p, int n_threads) {
+    p->n_threads  = n_threads;
+    p->prio       = 0;     // default priority (usually means normal or inherited)
+    p->poll       = 50;    // hybrid-polling enabled
+    p->strict_cpu = false; // no strict placement (all threads share same cpumask)
+    p->paused     = false; // threads are ready to go
+    memset(p->cpumask, 0, GGML_MAX_N_THREADS); // all-zero means use the default affinity (usually inherited)
+}
+
+struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) {
+    struct ggml_threadpool_params p;
+    ggml_threadpool_params_init(&p, n_threads);
+    return p;
+}
+
+bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) {
+    if (p0->n_threads      != p1->n_threads  )    return false;
+    if (p0->prio           != p1->prio       )    return false;
+    if (p0->poll           != p1->poll       )    return false;
+    if (p0->strict_cpu     != p1->strict_cpu )    return false;
+    return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0;
+}
diff --git a/pocs/CMakeLists.txt b/pocs/CMakeLists.txt
index 03e1d2c04be65..d49d14dee4351 100644
--- a/pocs/CMakeLists.txt
+++ b/pocs/CMakeLists.txt
@@ -8,5 +8,7 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
 else()
-    add_subdirectory(vdot)
+    if (NOT GGML_BACKEND_DL)
+        add_subdirectory(vdot)
+    endif()
 endif()
diff --git a/src/llama.cpp b/src/llama.cpp
index 20df09b133bfb..83bbc10a57a43 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -4866,7 +4866,9 @@ struct llama_model_loader {
             mappings.reserve(files.size());
             mmaps_used.reserve(files.size());
             for (const auto & file : files) {
-                std::unique_ptr<llama_mmap> mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, ggml_is_numa()));
+                auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
+                auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa");
+                std::unique_ptr<llama_mmap> mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, is_numa_fn()));
                 mmaps_used.emplace_back(mapping->size, 0);
                 if (mlock_mmaps) {
                     std::unique_ptr<llama_mlock> mlock_mmap(new llama_mlock());
@@ -9190,7 +9192,7 @@ static bool llm_load_tensors(
         ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
         if (!dev) {
             // FIXME: workaround for CPU backend buft having a NULL device
-            dev = ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0);
+            dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
         }
         ggml_backend_dev_props props;
         ggml_backend_dev_get_props(dev, &props);
@@ -17443,8 +17445,9 @@ static enum ggml_status llama_graph_compute(
                     int   n_threads,
         ggml_threadpool * threadpool) {
     if (lctx.backend_cpu != nullptr) {
-        ggml_backend_cpu_set_threadpool(lctx.backend_cpu, threadpool);
-        ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data);
+        auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(lctx.backend_cpu));
+        auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool");
+        set_threadpool_fn(lctx.backend_cpu, threadpool);
     }
 
     // set the number of threads for all the backends
@@ -19478,7 +19481,11 @@ void llama_backend_init(void) {
 
 void llama_numa_init(enum ggml_numa_strategy numa) {
     if (numa != GGML_NUMA_STRATEGY_DISABLED) {
-        ggml_numa_init(numa);
+        auto * dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        GGML_ASSERT(dev && "CPU backend is not loaded");
+        auto * reg = ggml_backend_dev_backend_reg(dev);
+        auto * numa_init_fn = (decltype(ggml_numa_init) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_numa_init");
+        numa_init_fn(numa);
     }
 }
 
@@ -19752,9 +19759,6 @@ struct llama_context * llama_new_context_with_model(
                 __func__, n_ctx_per_seq, hparams.n_ctx_train);
     }
 
-    ctx->abort_callback      = params.abort_callback;
-    ctx->abort_callback_data = params.abort_callback_data;
-
     ctx->logits_all = params.logits_all;
 
     // build worst-case graph for encoder if a model contains encoder
@@ -19803,7 +19807,7 @@ struct llama_context * llama_new_context_with_model(
         }
 
         // add CPU backend
-        ctx->backend_cpu = ggml_backend_cpu_init();
+        ctx->backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
         if (ctx->backend_cpu == nullptr) {
             LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__);
             llama_free(ctx);
@@ -19823,6 +19827,8 @@ struct llama_context * llama_new_context_with_model(
             }
         }
 
+        llama_set_abort_callback(ctx, params.abort_callback, params.abort_callback_data);
+
         if (!llama_kv_cache_init(ctx->kv_self, ctx, type_k, type_v, kv_size, cparams.offload_kqv)) {
             LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
@@ -19868,7 +19874,8 @@ struct llama_context * llama_new_context_with_model(
             std::vector<ggml_backend_t> backend_ptrs;
             for (auto & backend : ctx->backends) {
                 auto * buft = ggml_backend_get_default_buffer_type(backend.get());
-                if (ggml_backend_is_cpu(backend.get()) && !model->devices.empty()) {
+                auto backend_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));
+                if (backend_type == GGML_BACKEND_DEVICE_TYPE_CPU && !model->devices.empty()) {
                     // use the host buffer of the first device CPU for faster transfer of the intermediate state
                     auto * dev = model->devices[0];
                     auto * host_buft = ggml_backend_dev_host_buffer_type(dev);
@@ -19896,7 +19903,8 @@ struct llama_context * llama_new_context_with_model(
             // pipeline parallelism requires support for async compute and events in all devices
             if (pipeline_parallel) {
                 for (auto & backend : ctx->backends) {
-                    if (ggml_backend_is_cpu(backend.get())) {
+                    auto dev_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));
+                    if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU) {
                         // ignore CPU backend
                         continue;
                     }
@@ -21450,6 +21458,14 @@ int32_t llama_n_threads_batch(struct llama_context * ctx) {
 void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) {
     ctx->abort_callback      = abort_callback;
     ctx->abort_callback_data = abort_callback_data;
+
+    for (auto & backend : ctx->backends) {
+        auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get()));
+        auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback");
+        if (set_abort_callback_fn) {
+            set_abort_callback_fn(backend.get(), ctx->abort_callback, ctx->abort_callback_data);
+        }
+    }
 }
 
 void llama_set_embeddings(struct llama_context * ctx, bool embeddings) {
@@ -22191,32 +22207,23 @@ int llama_split_prefix(char * dest, size_t maxlen, const char * split_path, int
 }
 
 const char * llama_print_system_info(void) {
-    ggml_cpu_init(); // some ARM features are detected at runtime
-
     static std::string s;
 
-    s  = "";
-    s += "AVX = "         + std::to_string(ggml_cpu_has_avx())         + " | ";
-    s += "AVX_VNNI = "    + std::to_string(ggml_cpu_has_avx_vnni())    + " | ";
-    s += "AVX2 = "        + std::to_string(ggml_cpu_has_avx2())        + " | ";
-    s += "AVX512 = "      + std::to_string(ggml_cpu_has_avx512())      + " | ";
-    s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | ";
-    s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | ";
-    s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | ";
-    s += "AMX_INT8 = "    + std::to_string(ggml_cpu_has_amx_int8())    + " | ";
-    s += "FMA = "         + std::to_string(ggml_cpu_has_fma())         + " | ";
-    s += "NEON = "        + std::to_string(ggml_cpu_has_neon())        + " | ";
-    s += "SVE = "         + std::to_string(ggml_cpu_has_sve())         + " | ";
-    s += "ARM_FMA = "     + std::to_string(ggml_cpu_has_arm_fma())     + " | ";
-    s += "F16C = "        + std::to_string(ggml_cpu_has_f16c())        + " | ";
-    s += "FP16_VA = "     + std::to_string(ggml_cpu_has_fp16_va())     + " | ";
-    s += "RISCV_VECT = "  + std::to_string(ggml_cpu_has_riscv_v())     + " | ";
-    s += "WASM_SIMD = "   + std::to_string(ggml_cpu_has_wasm_simd())   + " | ";
-    s += "SSE3 = "        + std::to_string(ggml_cpu_has_sse3())        + " | ";
-    s += "SSSE3 = "       + std::to_string(ggml_cpu_has_ssse3())       + " | ";
-    s += "VSX = "         + std::to_string(ggml_cpu_has_vsx())         + " | ";
-    s += "MATMUL_INT8 = " + std::to_string(ggml_cpu_has_matmul_int8()) + " | ";
-    s += "LLAMAFILE = "   + std::to_string(ggml_cpu_has_llamafile())   + " | ";
+    for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
+        auto * reg = ggml_backend_reg_get(i);
+        auto * get_features_fn = (ggml_backend_get_features_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_get_features");
+        if (get_features_fn) {
+            ggml_backend_feature * features = get_features_fn(reg);
+            s += ggml_backend_reg_name(reg);
+            s += " : ";
+            for (; features->name; features++) {
+                s += features->name;
+                s += " = ";
+                s += features->value;
+                s += " | ";
+            }
+        }
+    }
 
     return s.c_str();
 }
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index b06f122e89873..82373ff4e1862 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -110,23 +110,26 @@ llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-llama-spm ARGS ${CMAKE_CU
 # llama_target_and_test(test-double-float.cpp) # SLOW
 llama_target_and_test(test-log.cpp)
 llama_target_and_test(test-arg-parser.cpp)
-llama_target_and_test(test-quantize-fns.cpp)
-llama_target_and_test(test-quantize-perf.cpp)
 llama_target_and_test(test-sampling.cpp)
 llama_target_and_test(test-chat-template.cpp)
 
 llama_target_and_test(test-grammar-parser.cpp)
 llama_target_and_test(test-grammar-integration.cpp)
 llama_target_and_test(test-llama-grammar.cpp)
-llama_target_and_test(test-barrier.cpp)
 # llama_target_and_test(test-opt.cpp) # SLOW
 llama_target_and_test(test-backend-ops.cpp)
 
-llama_target_and_test(test-rope.cpp)
-
 llama_target_and_test(test-model-load-cancel.cpp  LABEL "model")
 llama_target_and_test(test-autorelease.cpp        LABEL "model")
 
+if (NOT GGML_BACKEND_DL)
+    # these tests use the backends directly and cannot be built with dynamic loading
+    llama_target_and_test(test-barrier.cpp)
+    llama_target_and_test(test-quantize-fns.cpp)
+    llama_target_and_test(test-quantize-perf.cpp)
+    llama_target_and_test(test-rope.cpp)
+endif()
+
 # TODO: disabled on loongarch64 because the ggml-ci node lacks Python 3.8
 if (NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
     llama_target_and_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..)
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index b2b5705243ea2..6376b0e4c66cf 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -16,7 +16,6 @@
 
 
 #include <ggml.h>
-#include <ggml-cpu.h>
 #include <ggml-alloc.h>
 #include <ggml-backend.h>
 
@@ -26,7 +25,6 @@
 #include <cstdint>
 #include <cstring>
 #include <cinttypes>
-#include <functional>
 #include <memory>
 #include <random>
 #include <stdio.h>
@@ -639,19 +637,20 @@ struct test_case {
 
         // determine number of runs
         int n_runs;
+        bool is_cpu = ggml_backend_dev_type(ggml_backend_get_device(backend)) == GGML_BACKEND_DEVICE_TYPE_CPU;
         if (op_flops(out) > 0) {
             // based on flops
             const uint64_t GFLOP = 1000 * 1000 * 1000;
             const uint64_t target_flops_cpu =   8ULL * GFLOP;
             const uint64_t target_flops_gpu = 100ULL * GFLOP;
-            uint64_t target_flops = ggml_backend_is_cpu(backend) ? target_flops_cpu : target_flops_gpu;
+            uint64_t target_flops = is_cpu ? target_flops_cpu : target_flops_gpu;
             n_runs = std::min<int>(ggml_graph_size(gf) - ggml_graph_n_nodes(gf), target_flops / op_flops(out)) + 1;
         } else {
             // based on memory size
             const size_t GB = 1ULL << 30;
             const size_t target_size_cpu =  8 * GB;
             const size_t target_size_gpu = 32 * GB;
-            size_t target_size = ggml_backend_is_cpu(backend) ? target_size_cpu : target_size_gpu;
+            size_t target_size = is_cpu ? target_size_cpu : target_size_gpu;
             n_runs = std::min<int>(ggml_graph_size(gf) - ggml_graph_n_nodes(gf), target_size / op_size(out)) + 1;
         }
 
@@ -3873,7 +3872,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
 static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name) {
     if (mode == MODE_TEST) {
         auto test_cases = make_test_cases_eval();
-        ggml_backend_t backend_cpu = ggml_backend_cpu_init();
+        ggml_backend_t backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, NULL);
+        if (backend_cpu == NULL) {
+            printf("  Failed to initialize CPU backend\n");
+            return false;
+        }
 
         size_t n_ok = 0;
         for (auto & test : test_cases) {
@@ -3953,7 +3956,9 @@ int main(int argc, char ** argv) {
         }
     }
 
-    // enumerate backends
+    // load and enumerate backends
+    ggml_backend_load_all();
+
     printf("Testing %zu devices\n\n", ggml_backend_dev_count());
 
     size_t n_ok = 0;
@@ -3969,16 +3974,15 @@ int main(int argc, char ** argv) {
             continue;
         }
 
-        ggml_backend_t backend = ggml_backend_dev_init(dev, NULL);
-        GGML_ASSERT(backend != NULL);
-
-        if (backend_filter == NULL && ggml_backend_is_cpu(backend) && mode != MODE_GRAD) {
+        if (backend_filter == NULL && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU && mode != MODE_GRAD) {
             printf("  Skipping CPU backend\n");
-            ggml_backend_free(backend);
             n_ok++;
             continue;
         }
 
+        ggml_backend_t backend = ggml_backend_dev_init(dev, NULL);
+        GGML_ASSERT(backend != NULL);
+
         ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
         auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
         if (ggml_backend_set_n_threads_fn) {

From 9ca2e677626fce759d5d95c407c03677b9c87a26 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 25 Nov 2024 16:31:38 +0200
Subject: [PATCH 066/102] server : add speculative decoding support (#10455)

* server : add speculative decoding support

ggml-ci

* server : add helper function slot.can_speculate()

ggml-ci
---
 examples/server/server.cpp | 441 +++++++++++++++++++++++++------------
 1 file changed, 300 insertions(+), 141 deletions(-)

diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index 6c55d65c01330..f9d20fee5a6f6 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -2,10 +2,11 @@
 
 #include "arg.h"
 #include "common.h"
-#include "log.h"
-#include "sampling.h"
 #include "json-schema-to-grammar.h"
 #include "llama.h"
+#include "log.h"
+#include "sampling.h"
+#include "speculative.h"
 
 // Change JSON_ASSERT from assert() to GGML_ASSERT:
 #define JSON_ASSERT GGML_ASSERT
@@ -121,12 +122,21 @@ struct slot_params {
     int64_t t_max_predict_ms = -1; // if positive, limit the generation phase to this time limit
 
     std::vector<std::string> antiprompt;
+
+    struct common_params_sampling sampling;
+    struct common_params_speculative speculative;
 };
 
 struct server_slot {
     int id;
     int id_task = -1;
 
+    llama_batch batch_spec;
+
+    llama_context * ctx_dft = nullptr;
+
+    common_speculative * spec = nullptr;
+
     // the index relative to completion multi-task request
     size_t index = 0;
 
@@ -175,7 +185,6 @@ struct server_slot {
     // sampling
     json json_schema;
 
-    struct common_params_sampling sparams;
     struct common_sampler * smpl = nullptr;
 
     llama_token sampled;
@@ -212,7 +221,7 @@ struct server_slot {
         generated_token_probs.clear();
     }
 
-    bool has_budget(common_params &global_params) {
+    bool has_budget(const common_params & global_params) {
         if (params.n_predict == -1 && global_params.n_predict == -1) {
             return true; // limitless
         }
@@ -232,6 +241,10 @@ struct server_slot {
         return state != SLOT_STATE_IDLE;
     }
 
+    bool can_speculate() const {
+        return ctx_dft && params.speculative.n_max > 0 && params.cache_prompt;
+    }
+
     void add_token(const completion_token_output & token) {
         if (!is_processing()) {
             SLT_WRN(*this, "%s", "slot is not processing\n");
@@ -591,11 +604,14 @@ struct server_response {
 };
 
 struct server_context {
+    common_params params_base;
+
     llama_model * model = nullptr;
     llama_context * ctx = nullptr;
     std::vector<common_lora_adapter_container> loras;
 
-    common_params params;
+    llama_model * model_dft = nullptr;
+    llama_context_params cparams_dft;
 
     llama_batch batch = {};
 
@@ -628,27 +644,41 @@ struct server_context {
             model = nullptr;
         }
 
+        if (model_dft) {
+            llama_free_model(model_dft);
+            model_dft = nullptr;
+        }
+
         // Clear any sampling context
         for (server_slot & slot : slots) {
-            if (slot.smpl != nullptr) {
-                common_sampler_free(slot.smpl);
-            }
+            common_sampler_free(slot.smpl);
+            slot.smpl = nullptr;
+
+            llama_free(slot.ctx_dft);
+            slot.ctx_dft = nullptr;
+
+            common_speculative_free(slot.spec);
+            slot.spec = nullptr;
+
+            llama_batch_free(slot.batch_spec);
         }
 
         llama_batch_free(batch);
     }
 
-    bool load_model(const common_params & params_) {
-        params = params_;
+    bool load_model(const common_params & params) {
+        SRV_INF("loading model '%s'\n", params.model.c_str());
 
-        common_init_result llama_init = common_init_from_params(params);
+        params_base = params;
+
+        common_init_result llama_init = common_init_from_params(params_base);
 
         model = llama_init.model;
         ctx   = llama_init.context;
         loras = llama_init.lora_adapters;
 
         if (model == nullptr) {
-            SRV_ERR("failed to load model, '%s'\n", params.model.c_str());
+            SRV_ERR("failed to load model, '%s'\n", params_base.model.c_str());
             return false;
         }
 
@@ -657,6 +687,40 @@ struct server_context {
         add_bos_token = llama_add_bos_token(model);
         has_eos_token = !llama_add_eos_token(model);
 
+        if (!params_base.speculative.model.empty()) {
+            SRV_INF("loading draft model '%s'\n", params_base.speculative.model.c_str());
+
+            auto params_dft = params_base;
+
+            params_dft.model        = params_base.speculative.model;
+            params_dft.n_ctx        = params_base.speculative.n_ctx;
+            params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
+
+            common_init_result llama_init_dft = common_init_from_params(params_dft);
+
+            model_dft = llama_init_dft.model;
+
+            if (model_dft == nullptr) {
+                SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.c_str());
+                return false;
+            }
+
+            if (!common_speculative_are_compatible(ctx, llama_init_dft.context)) {
+                SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.c_str(), params_base.model.c_str());
+
+                llama_free      (llama_init_dft.context);
+                llama_free_model(llama_init_dft.model);
+
+                return false;
+            }
+
+            cparams_dft = common_context_params_to_llama(params_base);
+            cparams_dft.n_batch = llama_n_ctx(llama_init_dft.context);
+
+            // the context is not needed - we will create one for each slot
+            llama_free(llama_init_dft.context);
+        }
+
         return true;
     }
 
@@ -674,20 +738,36 @@ struct server_context {
     }
 
     void init() {
-        const int32_t n_ctx_slot = n_ctx / params.n_parallel;
+        const int32_t n_ctx_slot = n_ctx / params_base.n_parallel;
 
-        SRV_INF("initializing slots, n_slots = %d\n", params.n_parallel);
+        SRV_INF("initializing slots, n_slots = %d\n", params_base.n_parallel);
 
-        for (int i = 0; i < params.n_parallel; i++) {
+        for (int i = 0; i < params_base.n_parallel; i++) {
             server_slot slot;
 
             slot.id = i;
             slot.n_ctx = n_ctx_slot;
-            slot.n_predict = params.n_predict;
+            slot.n_predict = params_base.n_predict;
+
+            if (model_dft) {
+                slot.batch_spec = llama_batch_init(params_base.speculative.n_max + 1, 0, 1);
+
+                slot.ctx_dft = llama_new_context_with_model(model_dft, cparams_dft);
+                if (slot.ctx_dft == nullptr) {
+                    SRV_ERR("%s", "failed to create draft context\n");
+                    return;
+                }
+
+                slot.spec = common_speculative_init(slot.ctx_dft);
+                if (slot.spec == nullptr) {
+                    SRV_ERR("%s", "failed to create speculator\n");
+                    return;
+                }
+            }
 
             SLT_INF(slot, "new slot n_ctx_slot = %d\n", slot.n_ctx);
 
-            slot.sparams = params.sampling;
+            slot.params.sampling = params_base.sampling;
 
             slot.callback_on_release = [this](int) {
                 queue_tasks.pop_deferred_task();
@@ -707,7 +787,7 @@ struct server_context {
             const int32_t n_batch = llama_n_batch(ctx);
 
             // only a single seq_id per token is needed
-            batch = llama_batch_init(std::max(n_batch, params.n_parallel), 0, 1);
+            batch = llama_batch_init(std::max(n_batch, params_base.n_parallel), 0, 1);
         }
 
         metrics.init();
@@ -786,9 +866,11 @@ struct server_context {
     }
 
     bool launch_slot_with_task(server_slot & slot, const server_task & task) {
-        slot_params default_params;
         // Sampling parameter defaults are loaded from the global server context (but individual requests can still override them)
-        auto default_sparams = params.sampling;
+        slot_params defaults;
+        defaults.sampling    = params_base.sampling;
+        defaults.speculative = params_base.speculative;
+
         const auto & data = task.data;
 
         if (data.count("__oaicompat") != 0) {
@@ -799,42 +881,48 @@ struct server_context {
             slot.oaicompat_model = "";
         }
 
-        slot.params.stream              = json_value(data, "stream",             false);
-        slot.params.cache_prompt        = json_value(data, "cache_prompt",       false);
-        slot.params.n_predict           = json_value(data, "n_predict",          json_value(data, "max_tokens", default_params.n_predict));
-        slot.params.n_indent            = json_value(data, "n_indent",           default_params.n_indent);
-        slot.sparams.top_k              = json_value(data, "top_k",              default_sparams.top_k);
-        slot.sparams.top_p              = json_value(data, "top_p",              default_sparams.top_p);
-        slot.sparams.min_p              = json_value(data, "min_p",              default_sparams.min_p);
-        slot.sparams.xtc_probability    = json_value(data, "xtc_probability",    default_sparams.xtc_probability);
-        slot.sparams.xtc_threshold      = json_value(data, "xtc_threshold",      default_sparams.xtc_threshold);
-        slot.sparams.typ_p              = json_value(data, "typical_p",          default_sparams.typ_p);
-        slot.sparams.temp               = json_value(data, "temperature",        default_sparams.temp);
-        slot.sparams.dynatemp_range     = json_value(data, "dynatemp_range",     default_sparams.dynatemp_range);
-        slot.sparams.dynatemp_exponent  = json_value(data, "dynatemp_exponent",  default_sparams.dynatemp_exponent);
-        slot.sparams.penalty_last_n     = json_value(data, "repeat_last_n",      default_sparams.penalty_last_n);
-        slot.sparams.penalty_repeat     = json_value(data, "repeat_penalty",     default_sparams.penalty_repeat);
-        slot.sparams.penalty_freq       = json_value(data, "frequency_penalty",  default_sparams.penalty_freq);
-        slot.sparams.penalty_present    = json_value(data, "presence_penalty",   default_sparams.penalty_present);
-        slot.sparams.dry_multiplier     = json_value(data, "dry_multiplier",     default_sparams.dry_multiplier);
-        slot.sparams.dry_base           = json_value(data, "dry_base",           default_sparams.dry_base);
-        slot.sparams.dry_allowed_length = json_value(data, "dry_allowed_length", default_sparams.dry_allowed_length);
-        slot.sparams.dry_penalty_last_n = json_value(data, "dry_penalty_last_n", default_sparams.dry_penalty_last_n);
-        slot.sparams.mirostat           = json_value(data, "mirostat",           default_sparams.mirostat);
-        slot.sparams.mirostat_tau       = json_value(data, "mirostat_tau",       default_sparams.mirostat_tau);
-        slot.sparams.mirostat_eta       = json_value(data, "mirostat_eta",       default_sparams.mirostat_eta);
-        slot.sparams.penalize_nl        = json_value(data, "penalize_nl",        default_sparams.penalize_nl);
-        slot.params.n_keep              = json_value(data, "n_keep",             default_params.n_keep);
-        slot.params.n_discard           = json_value(data, "n_discard",          default_params.n_discard);
-        slot.sparams.seed               = json_value(data, "seed",               default_sparams.seed);
-        slot.sparams.n_probs            = json_value(data, "n_probs",            default_sparams.n_probs);
-        slot.sparams.min_keep           = json_value(data, "min_keep",           default_sparams.min_keep);
-      //slot.params.t_max_prompt_ms     = json_value(data, "t_max_prompt_ms",    default_params.t_max_prompt_ms); // TODO: implement
-        slot.params.t_max_predict_ms    = json_value(data, "t_max_predict_ms",   default_params.t_max_predict_ms);
-
-        if (slot.sparams.dry_base < 1.0f)
-        {
-           slot.sparams.dry_base = default_sparams.dry_base;
+        slot.params.stream           = json_value(data, "stream",             false);
+        slot.params.cache_prompt     = json_value(data, "cache_prompt",       false);
+        slot.params.n_predict        = json_value(data, "n_predict",          json_value(data, "max_tokens", defaults.n_predict));
+        slot.params.n_indent         = json_value(data, "n_indent",           defaults.n_indent);
+        slot.params.n_keep           = json_value(data, "n_keep",             defaults.n_keep);
+        slot.params.n_discard        = json_value(data, "n_discard",          defaults.n_discard);
+      //slot.params.t_max_prompt_ms  = json_value(data, "t_max_prompt_ms",    defaults.t_max_prompt_ms); // TODO: implement
+        slot.params.t_max_predict_ms = json_value(data, "t_max_predict_ms",   defaults.t_max_predict_ms);
+
+        slot.params.sampling.top_k              = json_value(data, "top_k",              defaults.sampling.top_k);
+        slot.params.sampling.top_p              = json_value(data, "top_p",              defaults.sampling.top_p);
+        slot.params.sampling.min_p              = json_value(data, "min_p",              defaults.sampling.min_p);
+        slot.params.sampling.xtc_probability    = json_value(data, "xtc_probability",    defaults.sampling.xtc_probability);
+        slot.params.sampling.xtc_threshold      = json_value(data, "xtc_threshold",      defaults.sampling.xtc_threshold);
+        slot.params.sampling.typ_p              = json_value(data, "typical_p",          defaults.sampling.typ_p);
+        slot.params.sampling.temp               = json_value(data, "temperature",        defaults.sampling.temp);
+        slot.params.sampling.dynatemp_range     = json_value(data, "dynatemp_range",     defaults.sampling.dynatemp_range);
+        slot.params.sampling.dynatemp_exponent  = json_value(data, "dynatemp_exponent",  defaults.sampling.dynatemp_exponent);
+        slot.params.sampling.penalty_last_n     = json_value(data, "repeat_last_n",      defaults.sampling.penalty_last_n);
+        slot.params.sampling.penalty_repeat     = json_value(data, "repeat_penalty",     defaults.sampling.penalty_repeat);
+        slot.params.sampling.penalty_freq       = json_value(data, "frequency_penalty",  defaults.sampling.penalty_freq);
+        slot.params.sampling.penalty_present    = json_value(data, "presence_penalty",   defaults.sampling.penalty_present);
+        slot.params.sampling.dry_multiplier     = json_value(data, "dry_multiplier",     defaults.sampling.dry_multiplier);
+        slot.params.sampling.dry_base           = json_value(data, "dry_base",           defaults.sampling.dry_base);
+        slot.params.sampling.dry_allowed_length = json_value(data, "dry_allowed_length", defaults.sampling.dry_allowed_length);
+        slot.params.sampling.dry_penalty_last_n = json_value(data, "dry_penalty_last_n", defaults.sampling.dry_penalty_last_n);
+        slot.params.sampling.mirostat           = json_value(data, "mirostat",           defaults.sampling.mirostat);
+        slot.params.sampling.mirostat_tau       = json_value(data, "mirostat_tau",       defaults.sampling.mirostat_tau);
+        slot.params.sampling.mirostat_eta       = json_value(data, "mirostat_eta",       defaults.sampling.mirostat_eta);
+        slot.params.sampling.penalize_nl        = json_value(data, "penalize_nl",        defaults.sampling.penalize_nl);
+        slot.params.sampling.seed               = json_value(data, "seed",               defaults.sampling.seed);
+        slot.params.sampling.n_probs            = json_value(data, "n_probs",            defaults.sampling.n_probs);
+        slot.params.sampling.min_keep           = json_value(data, "min_keep",           defaults.sampling.min_keep);
+
+        slot.params.speculative.n_min = json_value(data, "speculative.n_min", defaults.speculative.n_min);
+        slot.params.speculative.n_max = json_value(data, "speculative.n_max", defaults.speculative.n_max);
+        slot.params.speculative.p_min = json_value(data, "speculative.p_min", defaults.speculative.p_min);
+
+        slot.params.speculative.n_min = std::min(slot.params.speculative.n_max, slot.params.speculative.n_min);
+
+        if (slot.params.sampling.dry_base < 1.0f) {
+           slot.params.sampling.dry_base = defaults.sampling.dry_base;
         }
 
         // sequence breakers for DRY
@@ -843,8 +931,8 @@ struct server_context {
             // Ref: https://github.com/oobabooga/text-generation-webui/blob/d1af7a41ade7bd3c3a463bfa640725edb818ebaf/extensions/openai/typing.py#L39
 
             if (data.contains("dry_sequence_breakers")) {
-                slot.sparams.dry_sequence_breakers = json_value(data, "dry_sequence_breakers", std::vector<std::string>());
-                if (slot.sparams.dry_sequence_breakers.empty()) {
+                slot.params.sampling.dry_sequence_breakers = json_value(data, "dry_sequence_breakers", std::vector<std::string>());
+                if (slot.params.sampling.dry_sequence_breakers.empty()) {
                     send_error(task, "Error: dry_sequence_breakers must be a non-empty array of strings", ERROR_TYPE_INVALID_REQUEST);
                     return false;
                 }
@@ -858,14 +946,14 @@ struct server_context {
         }
         if (data.contains("json_schema") && !data.contains("grammar")) {
             try {
-                auto schema          = json_value(data, "json_schema", json::object());
-                slot.sparams.grammar = json_schema_to_grammar(schema);
+                auto schema                  = json_value(data, "json_schema", json::object());
+                slot.params.sampling.grammar = json_schema_to_grammar(schema);
             } catch (const std::exception & e) {
                 send_error(task, std::string("\"json_schema\": ") + e.what(), ERROR_TYPE_INVALID_REQUEST);
                 return false;
             }
         } else {
-            slot.sparams.grammar = json_value(data, "grammar", default_sparams.grammar);
+            slot.params.sampling.grammar = json_value(data, "grammar", defaults.sampling.grammar);
         }
 
         if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) {
@@ -875,10 +963,10 @@ struct server_context {
         }
 
         {
-            slot.sparams.logit_bias.clear();
+            slot.params.sampling.logit_bias.clear();
 
             if (json_value(data, "ignore_eos", false) && has_eos_token) {
-                slot.sparams.logit_bias.push_back({llama_token_eos(model), -INFINITY});
+                slot.params.sampling.logit_bias.push_back({llama_token_eos(model), -INFINITY});
             }
 
             const auto & logit_bias = data.find("logit_bias");
@@ -899,12 +987,12 @@ struct server_context {
                         if (el[0].is_number_integer()) {
                             llama_token tok = el[0].get<llama_token>();
                             if (tok >= 0 && tok < n_vocab) {
-                                slot.sparams.logit_bias.push_back({tok, bias});
+                                slot.params.sampling.logit_bias.push_back({tok, bias});
                             }
                         } else if (el[0].is_string()) {
                             auto toks = common_tokenize(model, el[0].get<std::string>(), false);
                             for (auto tok : toks) {
-                                slot.sparams.logit_bias.push_back({tok, bias});
+                                slot.params.sampling.logit_bias.push_back({tok, bias});
                             }
                         }
                     }
@@ -935,16 +1023,16 @@ struct server_context {
                             sampler_names.emplace_back(name);
                         }
                     }
-                    slot.sparams.samplers = common_sampler_types_from_names(sampler_names, false);
+                    slot.params.sampling.samplers = common_sampler_types_from_names(sampler_names, false);
                 } else if (samplers->is_string()){
                     std::string sampler_string;
                     for (const auto & name : *samplers) {
                         sampler_string += name;
                     }
-                    slot.sparams.samplers = common_sampler_types_from_chars(sampler_string);
+                    slot.params.sampling.samplers = common_sampler_types_from_chars(sampler_string);
                 }
             } else {
-                slot.sparams.samplers = default_sparams.samplers;
+                slot.params.sampling.samplers = defaults.sampling.samplers;
             }
         }
 
@@ -953,7 +1041,7 @@ struct server_context {
                 common_sampler_free(slot.smpl);
             }
 
-            slot.smpl = common_sampler_init(model, slot.sparams);
+            slot.smpl = common_sampler_init(model, slot.params.sampling);
             if (slot.smpl == nullptr) {
                 // for now, the only error that may happen here is invalid grammar
                 send_error(task, "Failed to parse grammar", ERROR_TYPE_INVALID_REQUEST);
@@ -961,6 +1049,12 @@ struct server_context {
             }
         }
 
+        if (slot.ctx_dft) {
+            llama_batch_free(slot.batch_spec);
+
+            slot.batch_spec = llama_batch_init(slot.params.speculative.n_max + 1, 0, 1);
+        }
+
         slot.state = SLOT_STATE_STARTED;
 
         SLT_INF(slot, "%s", "processing task\n");
@@ -978,7 +1072,7 @@ struct server_context {
 
     bool process_token(completion_token_output & result, server_slot & slot) {
         // remember which tokens were sampled - used for repetition penalties during sampling
-        const std::string token_str = common_token_to_piece(ctx, result.tok, params.special);
+        const std::string token_str = common_token_to_piece(ctx, result.tok, params_base.special);
         slot.sampled = result.tok;
 
         // search stop word and delete it
@@ -1043,7 +1137,7 @@ struct server_context {
         }
 
         // check the limits
-        if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params)) {
+        if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params_base)) {
             slot.stopped_limit  = true;
             slot.has_next_token = false;
 
@@ -1136,50 +1230,54 @@ struct server_context {
 
     json get_formated_generation(const server_slot & slot) const {
         std::vector<std::string> samplers;
-        samplers.reserve(slot.sparams.samplers.size());
-        for (const auto & sampler : slot.sparams.samplers) {
+        samplers.reserve(slot.params.sampling.samplers.size());
+        for (const auto & sampler : slot.params.sampling.samplers) {
             samplers.emplace_back(common_sampler_type_to_str(sampler));
         }
 
         return json {
             {"n_ctx",                     slot.n_ctx},
             {"n_predict",                 slot.n_predict},     // Server configured n_predict
-            {"model",                     params.model_alias},
-            {"seed",                      slot.sparams.seed},
+            {"model",                     params_base.model_alias},
+            {"seed",                      slot.params.sampling.seed},
             {"seed_cur",                  slot.smpl ? common_sampler_get_seed(slot.smpl) : 0},
-            {"temperature",               slot.sparams.temp},
-            {"dynatemp_range",            slot.sparams.dynatemp_range},
-            {"dynatemp_exponent",         slot.sparams.dynatemp_exponent},
-            {"top_k",                     slot.sparams.top_k},
-            {"top_p",                     slot.sparams.top_p},
-            {"min_p",                     slot.sparams.min_p},
-            {"xtc_probability",           slot.sparams.xtc_probability},
-            {"xtc_threshold",             slot.sparams.xtc_threshold},
-            {"typical_p",                 slot.sparams.typ_p},
-            {"repeat_last_n",             slot.sparams.penalty_last_n},
-            {"repeat_penalty",            slot.sparams.penalty_repeat},
-            {"presence_penalty",          slot.sparams.penalty_present},
-            {"frequency_penalty",         slot.sparams.penalty_freq},
-            {"dry_multiplier",            slot.sparams.dry_multiplier},
-            {"dry_base",                  slot.sparams.dry_base},
-            {"dry_allowed_length",        slot.sparams.dry_allowed_length},
-            {"dry_penalty_last_n",        slot.sparams.dry_penalty_last_n},
-            {"dry_sequence_breakers",     slot.sparams.dry_sequence_breakers},
-            {"mirostat",                  slot.sparams.mirostat},
-            {"mirostat_tau",              slot.sparams.mirostat_tau},
-            {"mirostat_eta",              slot.sparams.mirostat_eta},
-            {"penalize_nl",               slot.sparams.penalize_nl},
+            {"temperature",               slot.params.sampling.temp},
+            {"dynatemp_range",            slot.params.sampling.dynatemp_range},
+            {"dynatemp_exponent",         slot.params.sampling.dynatemp_exponent},
+            {"top_k",                     slot.params.sampling.top_k},
+            {"top_p",                     slot.params.sampling.top_p},
+            {"min_p",                     slot.params.sampling.min_p},
+            {"xtc_probability",           slot.params.sampling.xtc_probability},
+            {"xtc_threshold",             slot.params.sampling.xtc_threshold},
+            {"typical_p",                 slot.params.sampling.typ_p},
+            {"repeat_last_n",             slot.params.sampling.penalty_last_n},
+            {"repeat_penalty",            slot.params.sampling.penalty_repeat},
+            {"presence_penalty",          slot.params.sampling.penalty_present},
+            {"frequency_penalty",         slot.params.sampling.penalty_freq},
+            {"dry_multiplier",            slot.params.sampling.dry_multiplier},
+            {"dry_base",                  slot.params.sampling.dry_base},
+            {"dry_allowed_length",        slot.params.sampling.dry_allowed_length},
+            {"dry_penalty_last_n",        slot.params.sampling.dry_penalty_last_n},
+            {"dry_sequence_breakers",     slot.params.sampling.dry_sequence_breakers},
+            {"mirostat",                  slot.params.sampling.mirostat},
+            {"mirostat_tau",              slot.params.sampling.mirostat_tau},
+            {"mirostat_eta",              slot.params.sampling.mirostat_eta},
+            {"penalize_nl",               slot.params.sampling.penalize_nl},
             {"stop",                      slot.params.antiprompt},
             {"max_tokens",                slot.params.n_predict}, // User configured n_predict
             {"n_keep",                    slot.params.n_keep},
             {"n_discard",                 slot.params.n_discard},
-            {"ignore_eos",                slot.sparams.ignore_eos},
+            {"ignore_eos",                slot.params.sampling.ignore_eos},
             {"stream",                    slot.params.stream},
-          //{"logit_bias",                slot.sparams.logit_bias},
-            {"n_probs",                   slot.sparams.n_probs},
-            {"min_keep",                  slot.sparams.min_keep},
-            {"grammar",                   slot.sparams.grammar},
+          //{"logit_bias",                slot.params.sampling.logit_bias},
+            {"n_probs",                   slot.params.sampling.n_probs},
+            {"min_keep",                  slot.params.sampling.min_keep},
+            {"grammar",                   slot.params.sampling.grammar},
             {"samplers",                  samplers},
+            {"speculative",               slot.can_speculate()},
+            {"speculative.n_max",         slot.params.speculative.n_max},
+            {"speculative.n_min",         slot.params.speculative.n_min},
+            {"speculative.p_min",         slot.params.speculative.p_min},
         };
     }
 
@@ -1216,7 +1314,7 @@ struct server_context {
             {"index",      slot.index},
         };
 
-        if (slot.sparams.n_probs > 0) {
+        if (slot.params.sampling.n_probs > 0) {
             const llama_tokens to_send_toks = common_tokenize(ctx, tkn.text_to_send, false);
             const size_t probs_pos      = std::min(slot.n_sent_token_probs,                       slot.generated_token_probs.size());
             const size_t probs_stop_pos = std::min(slot.n_sent_token_probs + to_send_toks.size(), slot.generated_token_probs.size());
@@ -1249,7 +1347,7 @@ struct server_context {
             {"content",             !slot.params.stream ? slot.generated_text : ""},
             {"id_slot",             slot.id},
             {"stop",                true},
-            {"model",               params.model_alias},
+            {"model",               params_base.model_alias},
             {"tokens_predicted",    slot.n_decoded},
             {"tokens_evaluated",    slot.n_prompt_tokens},
             {"generation_settings", get_formated_generation(slot)},
@@ -1265,7 +1363,7 @@ struct server_context {
             {"index",               slot.index},
         };
 
-        if (slot.sparams.n_probs > 0) {
+        if (slot.params.sampling.n_probs > 0) {
             std::vector<completion_token_output> probs;
             if (!slot.params.stream && slot.stopped_word) {
                 const llama_tokens stop_word_toks = common_tokenize(ctx, slot.stopping_word, false);
@@ -1422,10 +1520,10 @@ struct server_context {
                             data.at("input_prefix"),
                             data.at("input_suffix"),
                             data.at("input_extra"),
-                            params.n_batch,
-                            params.n_predict,
+                            params_base.n_batch,
+                            params_base.n_predict,
                             slots[0].n_ctx, // TODO: there should be a better way
-                            params.spm_infill,
+                            params_base.spm_infill,
                             tokenized_prompts[i]
                         );
                         create_task(data, tokens);
@@ -1798,7 +1896,7 @@ struct server_context {
         // TODO: simplify and improve
         for (server_slot & slot : slots) {
             if (slot.is_processing() && slot.n_past + 1 >= slot.n_ctx) {
-                if (!params.ctx_shift) {
+                if (!params_base.ctx_shift) {
                     // this check is redundant (for good)
                     // we should never get here, because generation should already stopped in process_token()
                     slot.release();
@@ -1864,7 +1962,7 @@ struct server_context {
         int32_t batch_type = batch.n_tokens > 0 ? 0 : -1;
 
         // next, batch any pending prompts without exceeding n_batch
-        if (params.cont_batching || batch.n_tokens == 0) {
+        if (params_base.cont_batching || batch.n_tokens == 0) {
             for (auto & slot : slots) {
                 // this slot still has a prompt to be processed
                 if (slot.state == SLOT_STATE_PROCESSING_PROMPT || slot.state == SLOT_STATE_STARTED) {
@@ -1917,7 +2015,7 @@ struct server_context {
                                 continue;
                             }
                         } else {
-                            if (!params.ctx_shift) {
+                            if (!params_base.ctx_shift) {
                                 // if context shift is disabled, we make sure prompt size is smaller than KV size
                                 // TODO: there should be a separate parameter that control prompt truncation
                                 //       context shift should be applied only during the generation phase
@@ -1963,11 +2061,11 @@ struct server_context {
                                 slot.n_past = common_lcp(slot.cache_tokens, prompt_tokens);
 
                                 // reuse chunks from the cached prompt by shifting their KV cache in the new position
-                                if (params.n_cache_reuse > 0) {
+                                if (params_base.n_cache_reuse > 0) {
                                     size_t head_c = slot.n_past; // cache
                                     size_t head_p = slot.n_past; // current prompt
 
-                                    SLT_DBG(slot, "trying to reuse chunks with size > %d, slot.n_past = %d\n", params.n_cache_reuse, slot.n_past);
+                                    SLT_DBG(slot, "trying to reuse chunks with size > %d, slot.n_past = %d\n", params_base.n_cache_reuse, slot.n_past);
 
                                     while (head_c < slot.cache_tokens.size() &&
                                            head_p < prompt_tokens.size()) {
@@ -1980,7 +2078,7 @@ struct server_context {
                                             n_match++;
                                         }
 
-                                        if (n_match >= (size_t) params.n_cache_reuse) {
+                                        if (n_match >= (size_t) params_base.n_cache_reuse) {
                                             SLT_INF(slot, "reusing chunk with size %zu, shifting KV cache [%zu, %zu) -> [%zu, %zu)\n", n_match, head_c, head_c + n_match, head_p, head_p + n_match);
                                             //for (size_t i = head_p; i < head_p + n_match; i++) {
                                             //    SLT_DBG(slot, "cache token %3zu: %6d '%s'\n", i, prompt_tokens[i], common_token_to_piece(ctx, prompt_tokens[i]).c_str());
@@ -2168,38 +2266,99 @@ struct server_context {
                     continue; // continue loop of slots
                 }
 
-                completion_token_output result;
-                const llama_token id = common_sampler_sample(slot.smpl, ctx, slot.i_batch - i);
+                llama_token id;
+
+                {
+                    completion_token_output result;
+
+                    id = common_sampler_sample(slot.smpl, ctx, slot.i_batch - i);
+
+                    slot.i_batch = -1;
+
+                    common_sampler_accept(slot.smpl, id, true);
+
+                    slot.n_decoded += 1;
+                    if (slot.n_decoded == 1) {
+                        slot.t_start_generation = ggml_time_us();
+                        slot.t_prompt_processing = (slot.t_start_generation - slot.t_start_process_prompt) / 1e3;
+                        metrics.on_prompt_eval(slot);
+                    }
+
+                    result.tok = id;
+
+                    const auto * cur_p = common_sampler_get_candidates(slot.smpl);
 
-                common_sampler_accept(slot.smpl, id, true);
+                    for (size_t i = 0; i < (size_t) slot.params.sampling.n_probs; ++i) {
+                        result.probs.push_back({
+                            cur_p->data[i].id,
+                                i >= cur_p->size ? 0.0f : cur_p->data[i].p,
+                        });
+                    }
+
+                    if (!process_token(result, slot)) {
+                        // release slot because of stop condition
+                        slot.release();
+                        slot.print_timings();
+                        send_final_response(slot);
+                        metrics.on_prediction(slot);
+                        continue;
+                    }
+                }
 
-                slot.n_decoded += 1;
-                if (slot.n_decoded == 1) {
-                    slot.t_start_generation = ggml_time_us();
-                    slot.t_prompt_processing = (slot.t_start_generation - slot.t_start_process_prompt) / 1e3;
-                    metrics.on_prompt_eval(slot);
+                // check if the slot supports speculative decoding
+                if (!slot.can_speculate()) {
+                    continue;
                 }
 
-                result.tok = id;
+                struct common_speculative_params params_spec;
+                params_spec.n_draft   = slot.params.speculative.n_max;
+                params_spec.n_reuse   = llama_n_ctx(slot.ctx_dft) - slot.params.speculative.n_max;
+                params_spec.p_min     = slot.params.speculative.p_min;
 
-                const auto * cur_p = common_sampler_get_candidates(slot.smpl);
+                llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);
 
-                for (size_t i = 0; i < (size_t) slot.sparams.n_probs; ++i) {
-                    result.probs.push_back({
-                        cur_p->data[i].id,
-                        i >= cur_p->size ? 0.0f : cur_p->data[i].p,
-                    });
+                // ignore small drafts
+                if (slot.params.speculative.n_min > (int) draft.size()) {
+                    continue;
                 }
 
-                if (!process_token(result, slot)) {
-                    // release slot because of stop condition
-                    slot.release();
-                    slot.print_timings();
-                    send_final_response(slot);
-                    metrics.on_prediction(slot);
+                // construct the speculation batch
+                common_batch_clear(slot.batch_spec);
+                common_batch_add  (slot.batch_spec, id, slot.n_past, { slot.id }, true);
+
+                for (size_t i = 0; i < draft.size(); ++i) {
+                    common_batch_add(slot.batch_spec, draft[i], slot.n_past + 1 + i, { slot.id }, true);
+                }
+
+                llama_decode(ctx, slot.batch_spec);
+
+                // the accepted tokens from the speculation
+                const auto ids = common_sampler_sample_and_accept_n(slot.smpl, ctx, draft);
+
+                slot.n_past    += ids.size();
+                slot.n_decoded += ids.size();
+
+                slot.cache_tokens.push_back(id);
+                slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);
+
+                llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1);
+
+                for (size_t i = 0; i < ids.size(); ++i) {
+                    completion_token_output result;
+
+                    result.tok = ids[i];
+
+                    if (!process_token(result, slot)) {
+                        // release slot because of stop condition
+                        slot.release();
+                        slot.print_timings();
+                        send_final_response(slot);
+                        metrics.on_prediction(slot);
+                        break;
+                    }
                 }
 
-                slot.i_batch = -1;
+                SRV_DBG("accepted %d/%d draft tokens\n", (int) ids.size() - 1, (int) draft.size());
             }
         }
 
@@ -2697,7 +2856,7 @@ int main(int argc, char ** argv) {
     const auto handle_props = [&ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
         json data = {
             { "default_generation_settings", ctx_server.default_generation_settings_for_props },
-            { "total_slots",                 ctx_server.params.n_parallel },
+            { "total_slots",                 ctx_server.params_base.n_parallel },
             { "chat_template",               llama_get_chat_template(ctx_server.model) },
         };
 
@@ -2705,7 +2864,7 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_props_change = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
-        if (!ctx_server.params.endpoint_props) {
+        if (!ctx_server.params_base.endpoint_props) {
             res_error(res, format_error_response("This server does not support changing global properties. Start it with `--props`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
@@ -2718,7 +2877,7 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](server_task_inf_type inf_type, json & data, httplib::Response & res) {
-        if (ctx_server.params.embedding) {
+        if (ctx_server.params_base.embedding) {
             res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
@@ -2824,7 +2983,7 @@ int main(int argc, char ** argv) {
 
     // TODO: maybe merge this function with "handle_completions_generic"
     const auto handle_chat_completions = [&ctx_server, &params, &res_error, &res_ok, verbose](const httplib::Request & req, httplib::Response & res) {
-        if (ctx_server.params.embedding) {
+        if (ctx_server.params_base.embedding) {
             res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
@@ -3001,7 +3160,7 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
-        if (!ctx_server.params.reranking || ctx_server.params.embedding) {
+        if (!ctx_server.params_base.reranking || ctx_server.params_base.embedding) {
             res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking` and without `--embedding`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }

From a9a678a6b2264e5895533217b0cf8f250534cc58 Mon Sep 17 00:00:00 2001
From: brucepro <git@brucepro.net>
Date: Mon, 25 Nov 2024 08:11:55 -0800
Subject: [PATCH 067/102] Add download chat feature to server chat (#10481)

* Add download chat feature to server chat

Add a download feature next to the delete chat feature in the server vue chat interface.

* code style

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
---
 examples/server/public/index.html | 25 ++++++++++++++++++++++++-
 1 file changed, 24 insertions(+), 1 deletion(-)

diff --git a/examples/server/public/index.html b/examples/server/public/index.html
index 6216c08410a28..c54260867d2c8 100644
--- a/examples/server/public/index.html
+++ b/examples/server/public/index.html
@@ -81,7 +81,13 @@ <h2 class="font-bold mb-4 ml-4">Conversations</h2>
               <path d="M14.5 3a1 1 0 0 1-1 1H13v9a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2V4h-.5a1 1 0 0 1-1-1V2a1 1 0 0 1 1-1H6a1 1 0 0 1 1-1h2a1 1 0 0 1 1 1h3.5a1 1 0 0 1 1 1zM4.118 4 4 4.059V13a1 1 0 0 0 1 1h6a1 1 0 0 0 1-1V4.059L11.882 4zM2.5 3h11V2h-11z"/>
             </svg>
           </button>
-
+          <button v-if="messages.length > 0" class="btn mr-1" @click="downloadConv(viewingConvId)" :disabled="isGenerating">
+              <!-- download conversation button -->
+              <svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" class="bi bi-download" viewBox="0 0 16 16">
+                  <path d="M.5 9.9a.5.5 0 0 1 .5.5v2.5a1 1 0 0 0 1 1h12a1 1 0 0 0 1-1v-2.5a.5.5 0 0 1 1 0v2.5a2 2 0 0 1-2 2H2a2 2 0 0 1-2-2v-2.5a.5.5 0 0 1 .5-.5"/>
+                  <path d="M7.646 11.854a.5.5 0 0 0 .708 0l3-3a.5.5 0 0 0-.708-.708L8.5 10.293V1.5a.5.5 0 0 0-1 0v8.793L5.354 8.146a.5.5 0 1 0-.708.708z"/>
+            </svg>
+          </button>
           <button class="btn" @click="showConfigDialog = true" :disabled="isGenerating">
             <!-- edit config button -->
             <svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" class="bi bi-gear" viewBox="0 0 16 16">
@@ -526,6 +532,23 @@ <h3 class="text-lg font-bold mb-6">Settings</h3>
             this.fetchMessages();
           }
         },
+        downloadConv(convId) {
+          const conversation = StorageUtils.getOneConversation(convId);
+          if (!conversation) {
+            alert('Conversation not found.');
+            return;
+          }
+          const conversationJson = JSON.stringify(conversation, null, 2);
+          const blob = new Blob([conversationJson], { type: 'application/json' });
+          const url = URL.createObjectURL(blob);
+          const a = document.createElement('a');
+          a.href = url;
+          a.download = `conversation_${convId}.json`;
+          document.body.appendChild(a);
+          a.click();
+          document.body.removeChild(a);
+          URL.revokeObjectURL(url);
+        },
         async sendMessage() {
           if (!this.inputMsg) return;
           const currConvId = this.viewingConvId;

From 1f922254f0c984a8fb9fbaa0c390d7ffae49aedb Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Mon, 25 Nov 2024 19:18:37 +0100
Subject: [PATCH 068/102] Github: update issue templates [no ci] (#10489)

---
 .../ISSUE_TEMPLATE/010-bug-compilation.yml    | 12 ++++++----
 .github/ISSUE_TEMPLATE/011-bug-results.yml    | 15 +++++++-----
 .github/ISSUE_TEMPLATE/019-bug-misc.yml       | 23 +++++++++++--------
 3 files changed, 30 insertions(+), 20 deletions(-)

diff --git a/.github/ISSUE_TEMPLATE/010-bug-compilation.yml b/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
index 550ee1b498ab9..f10b3a2b22648 100644
--- a/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
+++ b/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
@@ -24,7 +24,8 @@ body:
   - type: dropdown
     id: operating-system
     attributes:
-      label: Which operating systems do you know to be affected?
+      label: Operating systems
+      description: Which operating systems do you know to be affected?
       multiple: true
       options:
         - Linux
@@ -41,14 +42,17 @@ body:
         description: Which GGML backends do you know to be affected?
         options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
         multiple: true
+    validations:
+      required: true
   - type: textarea
-    id: steps_to_reproduce
+    id: info
     attributes:
-      label: Steps to Reproduce
+      label: Problem description & steps to reproduce
       description: >
-        Please tell us how to reproduce the bug and any additional information that you think could be useful for fixing it.
+        Please give us a summary of the problem and tell us how to reproduce it.
         If you can narrow down the bug to specific compile flags, that information would be very much appreciated by us.
       placeholder: >
+        I'm trying to compile llama.cpp with CUDA support on a fresh install of Ubuntu and get error XY.
         Here are the exact commands that I used: ...
     validations:
       required: true
diff --git a/.github/ISSUE_TEMPLATE/011-bug-results.yml b/.github/ISSUE_TEMPLATE/011-bug-results.yml
index 1adb162b79a77..1ccef0793d45e 100644
--- a/.github/ISSUE_TEMPLATE/011-bug-results.yml
+++ b/.github/ISSUE_TEMPLATE/011-bug-results.yml
@@ -26,7 +26,8 @@ body:
   - type: dropdown
     id: operating-system
     attributes:
-      label: Which operating systems do you know to be affected?
+      label: Operating systems
+      description: Which operating systems do you know to be affected?
       multiple: true
       options:
         - Linux
@@ -43,6 +44,8 @@ body:
         description: Which GGML backends do you know to be affected?
         options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
         multiple: true
+    validations:
+      required: true
   - type: textarea
     id: hardware
     attributes:
@@ -55,20 +58,20 @@ body:
   - type: textarea
     id: model
     attributes:
-      label: Model
+      label: Models
       description: >
-        Which model at which quantization were you using when encountering the bug?
+        Which model(s) at which quantization were you using when encountering the bug?
         If you downloaded a GGUF file off of Huggingface, please provide a link.
       placeholder: >
         e.g. Meta LLaMA 3.1 Instruct 8b q4_K_M
     validations:
       required: false
   - type: textarea
-    id: steps_to_reproduce
+    id: info
     attributes:
-      label: Steps to Reproduce
+      label: Problem description & steps to reproduce
       description: >
-        Please tell us how to reproduce the bug and any additional information that you think could be useful for fixing it.
+        Please give us a summary of the problem and tell us how to reproduce it.
         If you can narrow down the bug to specific hardware, compile flags, or command line arguments,
         that information would be very much appreciated by us.
       placeholder: >
diff --git a/.github/ISSUE_TEMPLATE/019-bug-misc.yml b/.github/ISSUE_TEMPLATE/019-bug-misc.yml
index 124cdee919089..d157ea307e50f 100644
--- a/.github/ISSUE_TEMPLATE/019-bug-misc.yml
+++ b/.github/ISSUE_TEMPLATE/019-bug-misc.yml
@@ -14,7 +14,7 @@ body:
     id: version
     attributes:
       label: Name and Version
-      description: Which version of our software are you running? (use `--version` to get a version string)
+      description: Which version of our software is affected? (You can use `--version` to get a version string.)
       placeholder: |
         $./llama-cli --version
         version: 2999 (42b4109e)
@@ -24,7 +24,8 @@ body:
   - type: dropdown
     id: operating-system
     attributes:
-      label: Which operating systems do you know to be affected?
+      label: Operating systems
+      description: Which operating systems do you know to be affected?
       multiple: true
       options:
         - Linux
@@ -33,28 +34,30 @@ body:
         - BSD
         - Other? (Please let us know in description)
     validations:
-      required: true
+      required: false
   - type: dropdown
     id: module
     attributes:
       label: Which llama.cpp modules do you know to be affected?
       multiple: true
       options:
+        - Documentation/Github
         - libllama (core library)
         - llama-cli
         - llama-server
         - llama-bench
         - llama-quantize
         - Python/Bash scripts
+        - Test code
         - Other (Please specify in the next section)
     validations:
-      required: true
+      required: false
   - type: textarea
-    id: steps_to_reproduce
+    id: info
     attributes:
-      label: Steps to Reproduce
+      label: Problem description & steps to reproduce
       description: >
-        Please tell us how to reproduce the bug and any additional information that you think could be useful for fixing it.
+        Please give us a summary of the problem and tell us how to reproduce it (if applicable).
     validations:
       required: true
   - type: textarea
@@ -62,7 +65,7 @@ body:
     attributes:
       label: First Bad Commit
       description: >
-        If the bug was not present on an earlier version: when did it start appearing?
+        If the bug was not present on an earlier version and it's not trivial to track down: when did it start appearing?
         If possible, please do a git bisect and identify the exact commit that introduced the bug.
     validations:
       required: false
@@ -71,8 +74,8 @@ body:
     attributes:
       label: Relevant log output
       description: >
-          Please copy and paste any relevant log output, including the command that you entered and any generated text.
+          If applicable, please copy and paste any relevant log output, including the command that you entered and any generated text.
           This will be automatically formatted into code, so no need for backticks.
       render: shell
     validations:
-      required: true
+      required: false

From 10bce0450f0c4d80087e06312b9dbbab3e87f16b Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Mon, 25 Nov 2024 19:30:06 +0100
Subject: [PATCH 069/102] llama : accept a list of devices to use to offload a
 model (#10497)

* llama : accept a list of devices to use to offload a model

* accept `--dev none` to completely disable offloading

* fix dev list with dl backends

* rename env parameter to LLAMA_ARG_DEVICE for consistency
---
 common/arg.cpp                                | 60 +++++++++++++++++--
 common/common.cpp                             |  8 +--
 common/common.h                               | 14 +++--
 examples/server/server.cpp                    |  1 +
 .../speculative-simple/speculative-simple.cpp |  1 +
 examples/speculative/speculative.cpp          |  1 +
 ggml/src/ggml-backend-reg.cpp                 | 13 +++-
 include/llama.h                               |  3 +
 src/llama.cpp                                 | 30 ++++++----
 9 files changed, 104 insertions(+), 27 deletions(-)

diff --git a/common/arg.cpp b/common/arg.cpp
index 32240f21f2469..272492e50df15 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -298,6 +298,27 @@ static void common_params_print_usage(common_params_context & ctx_arg) {
     print_options(specific_options);
 }
 
+static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & value) {
+    std::vector<ggml_backend_dev_t> devices;
+    auto dev_names = string_split<std::string>(value, ',');
+    if (dev_names.empty()) {
+        throw std::invalid_argument("no devices specified");
+    }
+    if (dev_names.size() == 1 && dev_names[0] == "none") {
+        devices.push_back(nullptr);
+    } else {
+        for (const auto & device : dev_names) {
+            auto * dev = ggml_backend_dev_by_name(device.c_str());
+            if (!dev || ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_GPU) {
+                throw std::invalid_argument(string_format("invalid device: %s", device.c_str()));
+            }
+            devices.push_back(dev);
+        }
+        devices.push_back(nullptr);
+    }
+    return devices;
+}
+
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
     auto ctx_arg = common_params_parser_init(params, ex, print_usage);
     const common_params params_org = ctx_arg.params; // the example can modify the default params
@@ -324,6 +345,9 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
 }
 
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
+    // load dynamic backends
+    ggml_backend_load_all();
+
     common_params_context ctx_arg(params);
     ctx_arg.print_usage = print_usage;
     ctx_arg.ex          = ex;
@@ -1312,6 +1336,30 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             else { throw std::invalid_argument("invalid value"); }
         }
     ).set_env("LLAMA_ARG_NUMA"));
+    add_opt(common_arg(
+        {"-dev", "--device"}, "<dev1,dev2,..>",
+        "comma-separated list of devices to use for offloading (none = don't offload)\n"
+        "use --list-devices to see a list of available devices",
+        [](common_params & params, const std::string & value) {
+            params.devices = parse_device_list(value);
+        }
+    ).set_env("LLAMA_ARG_DEVICE"));
+    add_opt(common_arg(
+        {"--list-devices"},
+        "print list of available devices and exit",
+        [](common_params &) {
+            printf("Available devices:\n");
+            for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+                auto * dev = ggml_backend_dev_get(i);
+                if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
+                    size_t free, total;
+                    ggml_backend_dev_memory(dev, &free, &total);
+                    printf("  %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
+                }
+            }
+            exit(0);
+        }
+    ));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -1336,10 +1384,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             } else if (arg_next == "layer") {
                 params.split_mode = LLAMA_SPLIT_MODE_LAYER;
             } else if (arg_next == "row") {
-#ifdef GGML_USE_SYCL
-                fprintf(stderr, "warning: The split mode value:[row] is not supported by llama.cpp with SYCL. It's developing.\nExit!\n");
-                exit(1);
-#endif // GGML_USE_SYCL
                 params.split_mode = LLAMA_SPLIT_MODE_ROW;
             } else {
                 throw std::invalid_argument("invalid value");
@@ -2042,6 +2086,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.n_ctx = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-devd", "--device-draft"}, "<dev1,dev2,..>",
+        "comma-separated list of devices to use for offloading the draft model (none = don't offload)\n"
+        "use --list-devices to see a list of available devices",
+        [](common_params & params, const std::string & value) {
+            params.speculative.devices = parse_device_list(value);
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-ngld", "--gpu-layers-draft", "--n-gpu-layers-draft"}, "N",
         "number of layers to store in VRAM for the draft model",
diff --git a/common/common.cpp b/common/common.cpp
index 98524f7467ab4..09ec9f2388afb 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -377,9 +377,6 @@ void common_init() {
 #endif
 
     LOG_INF("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type);
-
-    // load dynamic backends
-    ggml_backend_load_all();
 }
 
 std::string common_params_get_system_info(const common_params & params) {
@@ -982,9 +979,12 @@ void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_l
     }
 }
 
-struct llama_model_params common_model_params_to_llama(const common_params & params) {
+struct llama_model_params common_model_params_to_llama(common_params & params) {
     auto mparams = llama_model_default_params();
 
+    if (!params.devices.empty()) {
+        mparams.devices = params.devices.data();
+    }
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
diff --git a/common/common.h b/common/common.h
index 5c579b5abfe03..286642db24158 100644
--- a/common/common.h
+++ b/common/common.h
@@ -156,6 +156,7 @@ struct common_params_sampling {
 };
 
 struct common_params_speculative {
+    std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
     int32_t n_ctx        =     0; // draft context size
     int32_t n_max        =    16; // maximum number of tokens to draft during speculative decoding
     int32_t n_min        =     5; // minimum number of draft tokens to use for speculative decoding
@@ -178,9 +179,6 @@ struct common_params {
     int32_t n_chunks              =    -1; // max number of chunks to process (-1 = unlimited)
     int32_t n_parallel            =     1; // number of parallel sequences to decode
     int32_t n_sequences           =     1; // number of sequences to decode
-    int32_t n_gpu_layers          =    -1; // number of layers to store in VRAM (-1 - use default)
-    int32_t main_gpu              =     0; // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]     =   {0}; // how split tensors should be distributed across GPUs
     int32_t grp_attn_n            =     1; // group-attention factor
     int32_t grp_attn_w            =   512; // group-attention width
     int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
@@ -193,6 +191,13 @@ struct common_params {
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
     float   defrag_thold          =  0.1f; // KV cache defragmentation threshold
 
+    // offload params
+    std::vector<ggml_backend_dev_t> devices;         // devices to use for offloading
+    int32_t n_gpu_layers                    =    -1; // number of layers to store in VRAM (-1 - use default)
+    int32_t main_gpu                        =     0; // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]               =   {0}; // how split tensors should be distributed across GPUs
+    enum llama_split_mode        split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
+
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
@@ -201,7 +206,6 @@ struct common_params {
 
     ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
 
-    enum llama_split_mode        split_mode        = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
     enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
     enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
@@ -462,7 +466,7 @@ struct common_init_result {
 
 struct common_init_result     common_init_from_params(common_params & params);
 
-struct llama_model_params     common_model_params_to_llama  (const common_params & params);
+struct llama_model_params     common_model_params_to_llama  (      common_params & params);
 struct llama_context_params   common_context_params_to_llama(const common_params & params);
 struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
 
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index f9d20fee5a6f6..8684771e2d0bb 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -692,6 +692,7 @@ struct server_context {
 
             auto params_dft = params_base;
 
+            params_dft.devices      = params_base.speculative.devices;
             params_dft.model        = params_base.speculative.model;
             params_dft.n_ctx        = params_base.speculative.n_ctx;
             params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
diff --git a/examples/speculative-simple/speculative-simple.cpp b/examples/speculative-simple/speculative-simple.cpp
index 1bc7f428cf3b5..7bf9056bf6db1 100644
--- a/examples/speculative-simple/speculative-simple.cpp
+++ b/examples/speculative-simple/speculative-simple.cpp
@@ -46,6 +46,7 @@ int main(int argc, char ** argv) {
     ctx_tgt   = llama_init_tgt.context;
 
     // load the draft model
+    params.devices      = params.speculative.devices;
     params.model        = params.speculative.model;
     params.n_ctx        = params.speculative.n_ctx;
     params.n_batch      = params.speculative.n_ctx > 0 ? params.speculative.n_ctx : params.n_batch;
diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp
index eb8bb2de54a26..d4ad9751e813a 100644
--- a/examples/speculative/speculative.cpp
+++ b/examples/speculative/speculative.cpp
@@ -76,6 +76,7 @@ int main(int argc, char ** argv) {
     ctx_tgt = llama_init_tgt.context;
 
     // load the draft model
+    params.devices = params.speculative.devices;
     params.model = params.speculative.model;
     params.n_gpu_layers = params.speculative.n_gpu_layers;
     if (params.speculative.cpuparams.n_threads > 0) {
diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
index 43d03d7fa7385..a0e0e2c5852f7 100644
--- a/ggml/src/ggml-backend-reg.cpp
+++ b/ggml/src/ggml-backend-reg.cpp
@@ -253,6 +253,15 @@ void ggml_backend_device_register(ggml_backend_dev_t device) {
 }
 
 // Backend (reg) enumeration
+static bool striequals(const char * a, const char * b) {
+    for (; *a && *b; a++, b++) {
+        if (std::tolower(*a) != std::tolower(*b)) {
+            return false;
+        }
+    }
+    return *a == *b;
+}
+
 size_t ggml_backend_reg_count() {
     return get_reg().backends.size();
 }
@@ -265,7 +274,7 @@ ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
 ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
     for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
         ggml_backend_reg_t reg = ggml_backend_reg_get(i);
-        if (std::strcmp(ggml_backend_reg_name(reg), name) == 0) {
+        if (striequals(ggml_backend_reg_name(reg), name)) {
             return reg;
         }
     }
@@ -285,7 +294,7 @@ ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
 ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
     for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
         ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-        if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
+        if (striequals(ggml_backend_dev_name(dev), name)) {
             return dev;
         }
     }
diff --git a/include/llama.h b/include/llama.h
index 90791d5f5ea12..ab5e376e6c7f2 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -272,6 +272,9 @@ extern "C" {
     };
 
     struct llama_model_params {
+        // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used)
+        ggml_backend_dev_t * devices;
+
         int32_t n_gpu_layers; // number of layers to store in VRAM
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
diff --git a/src/llama.cpp b/src/llama.cpp
index 83bbc10a57a43..571cb68e28854 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -19364,6 +19364,7 @@ void llama_lora_adapter_free(struct llama_lora_adapter * adapter) {
 //
 struct llama_model_params llama_model_default_params() {
     struct llama_model_params result = {
+        /*.devices                     =*/ nullptr,
         /*.n_gpu_layers                =*/ 0,
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
         /*.main_gpu                    =*/ 0,
@@ -19576,19 +19577,24 @@ struct llama_model * llama_load_model_from_file(
     }
 
     // create list of devices to use with this model
-    // currently, we use all available devices
-    // TODO: rework API to give user more control over device selection
-    for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-        switch (ggml_backend_dev_type(dev)) {
-            case GGML_BACKEND_DEVICE_TYPE_CPU:
-            case GGML_BACKEND_DEVICE_TYPE_ACCEL:
-                // skip CPU backends since they are handled separately
-                break;
+    if (params.devices) {
+        for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) {
+            model->devices.push_back(*dev);
+        }
+    } else {
+        // use all available devices
+        for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+            ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+            switch (ggml_backend_dev_type(dev)) {
+                case GGML_BACKEND_DEVICE_TYPE_CPU:
+                case GGML_BACKEND_DEVICE_TYPE_ACCEL:
+                    // skip CPU backends since they are handled separately
+                    break;
 
-            case GGML_BACKEND_DEVICE_TYPE_GPU:
-                model->devices.push_back(dev);
-                break;
+                case GGML_BACKEND_DEVICE_TYPE_GPU:
+                    model->devices.push_back(dev);
+                    break;
+            }
         }
     }
 

From 80acb7b430d826a7685326603a07342e9abc1b45 Mon Sep 17 00:00:00 2001
From: Shane A <shanea@allenai.org>
Date: Mon, 25 Nov 2024 10:36:09 -0800
Subject: [PATCH 070/102] Rename Olmo1124 to Olmo2 (#10500)

---
 convert_hf_to_gguf.py          |  6 +++---
 gguf-py/gguf/constants.py      |  6 +++---
 gguf-py/gguf/tensor_mapping.py | 28 ++++++++++++++--------------
 src/llama.cpp                  | 18 +++++++++---------
 4 files changed, 29 insertions(+), 29 deletions(-)

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 80a179b86af7e..b931049d11e2d 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -3040,9 +3040,9 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         return [(self.map_tensor_name(name), data_torch)]
 
 
-@Model.register("Olmo1124ForCausalLM")
-class Olmo1124Model(Model):
-    model_arch = gguf.MODEL_ARCH.OLMO_1124
+@Model.register("Olmo2ForCausalLM")
+class Olmo2Model(Model):
+    model_arch = gguf.MODEL_ARCH.OLMO2
 
 
 @Model.register("OlmoeForCausalLM")
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index d83b72f761951..7df23371cc100 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -243,7 +243,7 @@ class MODEL_ARCH(IntEnum):
     COMMAND_R    = auto()
     DBRX         = auto()
     OLMO         = auto()
-    OLMO_1124    = auto()
+    OLMO2        = auto()
     OLMOE        = auto()
     OPENELM      = auto()
     ARCTIC       = auto()
@@ -405,7 +405,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.COMMAND_R:      "command-r",
     MODEL_ARCH.DBRX:           "dbrx",
     MODEL_ARCH.OLMO:           "olmo",
-    MODEL_ARCH.OLMO_1124:      "olmo_1124",
+    MODEL_ARCH.OLMO2:          "olmo2",
     MODEL_ARCH.OLMOE:          "olmoe",
     MODEL_ARCH.OPENELM:        "openelm",
     MODEL_ARCH.ARCTIC:         "arctic",
@@ -1071,7 +1071,7 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
-    MODEL_ARCH.OLMO_1124: [
+    MODEL_ARCH.OLMO2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
         MODEL_TENSOR.OUTPUT,
diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py
index 4cbd39e03e942..1b6a3f4add875 100644
--- a/gguf-py/gguf/tensor_mapping.py
+++ b/gguf-py/gguf/tensor_mapping.py
@@ -13,7 +13,7 @@ class TensorNameMap:
             "transformer.wte",                           # gpt2 gpt-j mpt refact qwen dbrx jais exaone
             "transformer.word_embeddings",               # falcon
             "word_embeddings",                           # bloom
-            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo_1124
+            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo2
             "tok_embeddings",                            # llama-pth
             "embeddings.word_embeddings",                # bert nomic-bert
             "language_model.embedding.word_embeddings",  # persimmon
@@ -54,7 +54,7 @@ class TensorNameMap:
         # Output
         MODEL_TENSOR.OUTPUT: (
             "embed_out",                 # gptneox
-            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo_1124
+            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo2
             "output",                    # llama-pth bloom internlm2
             "word_embeddings_for_head",  # persimmon
             "lm_head.linear",            # phi2
@@ -66,7 +66,7 @@ class TensorNameMap:
         MODEL_TENSOR.OUTPUT_NORM: (
             "gpt_neox.final_layer_norm",               # gptneox
             "transformer.ln_f",                        # gpt2 gpt-j falcon jais exaone
-            "model.norm",                              # llama-hf baichuan internlm2 olmoe olmo_1124
+            "model.norm",                              # llama-hf baichuan internlm2 olmoe olmo2
             "norm",                                    # llama-pth
             "transformer.norm_f",                      # mpt dbrx
             "ln_f",                                    # refact bloom qwen gpt2
@@ -145,7 +145,7 @@ class TensorNameMap:
 
         # Attention query
         MODEL_TENSOR.ATTN_Q: (
-            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron olmoe olmo_1124
+            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron olmoe olmo2
             "layers.{bid}.attention.wq",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.query",                  # bert
             "transformer.h.{bid}.attn.q_proj",                           # gpt-j
@@ -157,7 +157,7 @@ class TensorNameMap:
 
         # Attention key
         MODEL_TENSOR.ATTN_K: (
-            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron olmoe olmo_1124
+            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron olmoe olmo2
             "layers.{bid}.attention.wk",                               # llama-pth
             "encoder.layer.{bid}.attention.self.key",                  # bert
             "transformer.h.{bid}.attn.k_proj",                         # gpt-j
@@ -170,7 +170,7 @@ class TensorNameMap:
 
         # Attention value
         MODEL_TENSOR.ATTN_V: (
-            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron olmoe olmo_1124
+            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron olmoe olmo2
             "layers.{bid}.attention.wv",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.value",                  # bert
             "transformer.h.{bid}.attn.v_proj",                           # gpt-j
@@ -188,7 +188,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.attn.out_proj",                       # mpt
             "transformer.h.{bid}.self_attention.dense",                     # falcon
             "h.{bid}.self_attention.dense",                                 # bloom
-            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron olmoe olmo_1124
+            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron olmoe olmo2
             "layers.{bid}.attention.wo",                                    # llama-pth
             "encoder.layer.{bid}.attention.output.dense",                   # bert
             "transformer.h.{bid}.attn.out_proj",                            # gpt-j
@@ -215,7 +215,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.ATTN_POST_NORM: (
-            "model.layers.{bid}.post_attention_layernorm",     # gemma2 olmo_1124
+            "model.layers.{bid}.post_attention_layernorm",     # gemma2 olmo2
         ),
 
         # Rotary embeddings
@@ -250,7 +250,7 @@ class TensorNameMap:
 
         # Post feed-forward norm
         MODEL_TENSOR.FFN_POST_NORM: (
-            "model.layers.{bid}.post_feedforward_layernorm", # gemma2 olmo_1124
+            "model.layers.{bid}.post_feedforward_layernorm", # gemma2 olmo2
         ),
 
         MODEL_TENSOR.FFN_GATE_INP: (
@@ -273,7 +273,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.ffn.up_proj",                   # mpt
             "transformer.h.{bid}.mlp.dense_h_to_4h",                  # falcon
             "h.{bid}.mlp.dense_h_to_4h",                              # bloom
-            "model.layers.{bid}.mlp.up_proj",                         # llama-hf refact nemotron olmo_1124
+            "model.layers.{bid}.mlp.up_proj",                         # llama-hf refact nemotron olmo2
             "layers.{bid}.feed_forward.w3",                           # llama-pth
             "encoder.layer.{bid}.intermediate.dense",                 # bert
             "transformer.h.{bid}.mlp.fc_in",                          # gpt-j
@@ -314,7 +314,7 @@ class TensorNameMap:
 
         # Feed-forward gate
         MODEL_TENSOR.FFN_GATE: (
-            "model.layers.{bid}.mlp.gate_proj",           # llama-hf refact olmo_1124
+            "model.layers.{bid}.mlp.gate_proj",           # llama-hf refact olmo2
             "layers.{bid}.feed_forward.w1",               # llama-pth
             "transformer.h.{bid}.mlp.w2",                 # qwen
             "transformer.h.{bid}.mlp.c_fc2",              # jais
@@ -346,7 +346,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.ffn.down_proj",                 # mpt
             "transformer.h.{bid}.mlp.dense_4h_to_h",                  # falcon
             "h.{bid}.mlp.dense_4h_to_h",                              # bloom
-            "model.layers.{bid}.mlp.down_proj",                       # llama-hf nemotron olmo_1124
+            "model.layers.{bid}.mlp.down_proj",                       # llama-hf nemotron olmo2
             "layers.{bid}.feed_forward.w2",                           # llama-pth
             "encoder.layer.{bid}.output.dense",                       # bert
             "transformer.h.{bid}.mlp.fc_out",                         # gpt-j
@@ -383,7 +383,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_Q_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
             "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon olmo_1124
+            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon olmo2
             "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_q",                # jina-bert-v2
             "transformer.layers.{bid}.attn.q_norm",                           # openelm
@@ -392,7 +392,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_K_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
             "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon olmo_1124
+            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon olmo2
             "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_k",                # jina-bert-v2
             "transformer.layers.{bid}.attn.k_norm",                           # openelm
diff --git a/src/llama.cpp b/src/llama.cpp
index 571cb68e28854..af5e686e07eda 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -179,7 +179,7 @@ enum llm_arch {
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_DBRX,
     LLM_ARCH_OLMO,
-    LLM_ARCH_OLMO_1124,
+    LLM_ARCH_OLMO2,
     LLM_ARCH_OLMOE,
     LLM_ARCH_OPENELM,
     LLM_ARCH_ARCTIC,
@@ -233,7 +233,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COMMAND_R,       "command-r"    },
     { LLM_ARCH_DBRX,            "dbrx"         },
     { LLM_ARCH_OLMO,            "olmo"         },
-    { LLM_ARCH_OLMO_1124,       "olmo_1124"    },
+    { LLM_ARCH_OLMO2,           "olmo2"        },
     { LLM_ARCH_OLMOE,           "olmoe"        },
     { LLM_ARCH_OPENELM,         "openelm"      },
     { LLM_ARCH_ARCTIC,          "arctic"       },
@@ -1210,7 +1210,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
         },
     },
     {
-        LLM_ARCH_OLMO_1124,
+        LLM_ARCH_OLMO2,
         {
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
             { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
@@ -5900,7 +5900,7 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
-        case LLM_ARCH_OLMO_1124:
+        case LLM_ARCH_OLMO2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
@@ -8593,7 +8593,7 @@ static bool llm_load_tensors(
                         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
                     }
                 } break;
-            case LLM_ARCH_OLMO_1124:
+            case LLM_ARCH_OLMO2:
                 {
                     model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -14483,7 +14483,7 @@ struct llm_build_context {
         return gf;
     }
 
-    struct ggml_cgraph * build_olmo_1124() {
+    struct ggml_cgraph * build_olmo2() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         // mutable variable, needed during the last layer of the computation to skip unused tokens
@@ -16799,9 +16799,9 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_olmo();
             } break;
-        case LLM_ARCH_OLMO_1124:
+        case LLM_ARCH_OLMO2:
             {
-                result = llm.build_olmo_1124();
+                result = llm.build_olmo2();
             } break;
         case LLM_ARCH_OLMOE:
             {
@@ -20084,7 +20084,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_QWEN:
         case LLM_ARCH_QWEN2:
         case LLM_ARCH_QWEN2MOE:
-        case LLM_ARCH_OLMO_1124:
+        case LLM_ARCH_OLMO2:
         case LLM_ARCH_OLMOE:
         case LLM_ARCH_PHI2:
         case LLM_ARCH_PHI3:

From 106964e3d266740f571b5aad7b57545b4a901ac9 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 25 Nov 2024 21:49:31 +0200
Subject: [PATCH 071/102] metal : enable mat-vec kernels for bs <= 4 (#10491)

---
 ggml/src/ggml-metal/ggml-metal.m | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index 63baaf163581c..c47f07a9e0416 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -1927,7 +1927,7 @@ static void ggml_metal_encode_node(
 
                 // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                 // to the matrix-vector kernel
-                int ne11_mm_min = 1;
+                int ne11_mm_min = 4;
 
 #if 0
                 // the numbers below are measured on M2 Ultra for 7B and 13B models

From 47f931c8f9a26c072d71224bc8013cc66ea9e445 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 25 Nov 2024 21:50:07 +0200
Subject: [PATCH 072/102] server : enable cache_prompt by default (#10501)

ggml-ci
---
 examples/server/README.md  | 2 +-
 examples/server/server.cpp | 4 ++--
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/examples/server/README.md b/examples/server/README.md
index 0936e0b7ba13a..877768c8b0bd2 100644
--- a/examples/server/README.md
+++ b/examples/server/README.md
@@ -412,7 +412,7 @@ node index.js
 
     `id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot.  Default: `-1`
 
-    `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `false`
+    `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `true`
 
     `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["dry", "top_k", "typ_p", "top_p", "min_p", "xtc", "temperature"]` - these are all the available values.
 
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index 8684771e2d0bb..c0ea4faf77d42 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -111,7 +111,7 @@ struct server_static_file {
 
 struct slot_params {
     bool stream       = true;
-    bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt
+    bool cache_prompt = true; // remember the prompt to avoid reprocessing all prompt
 
     int32_t n_keep    =  0; // number of tokens to keep from initial prompt
     int32_t n_discard =  0; // number of tokens after n_keep that may be discarded when shifting context, 0 defaults to half
@@ -883,7 +883,7 @@ struct server_context {
         }
 
         slot.params.stream           = json_value(data, "stream",             false);
-        slot.params.cache_prompt     = json_value(data, "cache_prompt",       false);
+        slot.params.cache_prompt     = json_value(data, "cache_prompt",       true);
         slot.params.n_predict        = json_value(data, "n_predict",          json_value(data, "max_tokens", defaults.n_predict));
         slot.params.n_indent         = json_value(data, "n_indent",           defaults.n_indent);
         slot.params.n_keep           = json_value(data, "n_keep",             defaults.n_keep);

From 9fd8c2687f5aa2f095ac6e12a376e1c0583888e8 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 25 Nov 2024 22:28:27 +0200
Subject: [PATCH 073/102] server : add more information about error (#10455)

---
 common/speculative.cpp | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/common/speculative.cpp b/common/speculative.cpp
index fe315a2703e9c..e559675c436ef 100644
--- a/common/speculative.cpp
+++ b/common/speculative.cpp
@@ -90,9 +90,10 @@ bool common_speculative_are_compatible(
     if (llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
         llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
         llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
-        llama_token_eos(model_tgt) != llama_token_eos(model_dft)
-    ) {
+        llama_token_eos(model_tgt) != llama_token_eos(model_dft)) {
         LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
+        LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_tgt), llama_add_bos_token(model_tgt), llama_token_eos(model_tgt), llama_add_eos_token(model_tgt));
+        LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_dft), llama_add_bos_token(model_dft), llama_token_eos(model_dft), llama_add_eos_token(model_dft));
         return false;
     }
 

From 50d5cecbda3b0d03344eed326287adc1f6c7f3ef Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Mon, 25 Nov 2024 22:05:39 +0100
Subject: [PATCH 074/102] ci : build docker images only once daily (#10503)

---
 .github/workflows/docker.yml | 11 ++++-------
 1 file changed, 4 insertions(+), 7 deletions(-)

diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml
index a953cdac907ae..9cef283d946e1 100644
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -10,12 +10,10 @@
 name: Publish Docker image
 
 on:
-  #pull_request:
-  push:
-    branches:
-      - master
-    paths: ['.github/workflows/docker.yml', '.devops/*.Dockerfile', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
-  workflow_dispatch: # allows manual triggering, useful for debugging
+  workflow_dispatch: # allows manual triggering
+  schedule:
+    # Rebuild daily rather than on every push because it is expensive
+    - cron: '12 4 * * *'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -29,7 +27,6 @@ permissions:
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
-    #if: github.event.pull_request.draft == false
 
     runs-on: ubuntu-latest
     env:

From 0cc63754b831d3a6c37bc5d721d12ce9540ffe76 Mon Sep 17 00:00:00 2001
From: Eric Curtin <ecurtin@redhat.com>
Date: Mon, 25 Nov 2024 16:56:24 -0500
Subject: [PATCH 075/102] Introduce llama-run (#10291)

It's like simple-chat but it uses smart pointers to avoid manual
memory cleanups. Less memory leaks in the code now. Avoid printing
multiple dots. Split code into smaller functions. Uses no exception
handling.

Signed-off-by: Eric Curtin <ecurtin@redhat.com>
---
 CMakeLists.txt              |   7 +-
 Makefile                    |   6 +
 examples/CMakeLists.txt     |   1 +
 examples/run/CMakeLists.txt |   5 +
 examples/run/README.md      |   7 +
 examples/run/run.cpp        | 409 ++++++++++++++++++++++++++++++++++++
 include/llama-cpp.h         |  25 +++
 7 files changed, 458 insertions(+), 2 deletions(-)
 create mode 100644 examples/run/CMakeLists.txt
 create mode 100644 examples/run/README.md
 create mode 100644 examples/run/run.cpp
 create mode 100644 include/llama-cpp.h

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 994e61e45fedd..e7d91a5b5c9ac 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -163,8 +163,11 @@ if (GGML_TARGET_DEFINES)
     list(APPEND GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES})
 endif()
 get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
-
-set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h)
+# all public headers
+set(LLAMA_PUBLIC_HEADERS
+    ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/include/llama-cpp.h)
+set_target_properties(llama PROPERTIES PUBLIC_HEADER "${LLAMA_PUBLIC_HEADERS}")
 install(TARGETS llama LIBRARY PUBLIC_HEADER)
 
 configure_package_config_file(
diff --git a/Makefile b/Makefile
index 14c05e93e7535..cfc74c1dc71b2 100644
--- a/Makefile
+++ b/Makefile
@@ -34,6 +34,7 @@ BUILD_TARGETS = \
 	llama-server \
 	llama-simple \
 	llama-simple-chat \
+	llama-run \
 	llama-speculative \
 	llama-tokenize \
 	llama-vdot \
@@ -1167,6 +1168,11 @@ llama-infill: examples/infill/infill.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+llama-run: examples/run/run.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 llama-simple: examples/simple/simple.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index 632409d5591b9..21db1f3c2f199 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -36,6 +36,7 @@ else()
         add_subdirectory(server)
     endif()
     add_subdirectory(save-load-state)
+    add_subdirectory(run)
     add_subdirectory(simple)
     add_subdirectory(simple-chat)
     add_subdirectory(speculative)
diff --git a/examples/run/CMakeLists.txt b/examples/run/CMakeLists.txt
new file mode 100644
index 0000000000000..084f1e92d77b0
--- /dev/null
+++ b/examples/run/CMakeLists.txt
@@ -0,0 +1,5 @@
+set(TARGET llama-run)
+add_executable(${TARGET} run.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
diff --git a/examples/run/README.md b/examples/run/README.md
new file mode 100644
index 0000000000000..6e926811f3cff
--- /dev/null
+++ b/examples/run/README.md
@@ -0,0 +1,7 @@
+# llama.cpp/example/run
+
+The purpose of this example is to demonstrate a minimal usage of llama.cpp for running models.
+
+```bash
+./llama-run Meta-Llama-3.1-8B-Instruct.gguf
+...
diff --git a/examples/run/run.cpp b/examples/run/run.cpp
new file mode 100644
index 0000000000000..cac2faefcc256
--- /dev/null
+++ b/examples/run/run.cpp
@@ -0,0 +1,409 @@
+#if defined(_WIN32)
+#include <windows.h>
+#else
+#include <unistd.h>
+#endif
+
+#include <climits>
+#include <cstdio>
+#include <cstring>
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "llama-cpp.h"
+
+typedef std::unique_ptr<char[]> char_array_ptr;
+
+struct Argument {
+    std::string flag;
+    std::string help_text;
+};
+
+struct Options {
+    std::string model_path, prompt_non_interactive;
+    int ngl = 99;
+    int n_ctx = 2048;
+};
+
+class ArgumentParser {
+   public:
+    ArgumentParser(const char * program_name) : program_name(program_name) {}
+
+    void add_argument(const std::string & flag, std::string & var, const std::string & help_text = "") {
+        string_args[flag] = &var;
+        arguments.push_back({flag, help_text});
+    }
+
+    void add_argument(const std::string & flag, int & var, const std::string & help_text = "") {
+        int_args[flag] = &var;
+        arguments.push_back({flag, help_text});
+    }
+
+    int parse(int argc, const char ** argv) {
+        for (int i = 1; i < argc; ++i) {
+            std::string arg = argv[i];
+            if (string_args.count(arg)) {
+                if (i + 1 < argc) {
+                    *string_args[arg] = argv[++i];
+                } else {
+                    fprintf(stderr, "error: missing value for %s\n", arg.c_str());
+                    print_usage();
+                    return 1;
+                }
+            } else if (int_args.count(arg)) {
+                if (i + 1 < argc) {
+                    if (parse_int_arg(argv[++i], *int_args[arg]) != 0) {
+                        fprintf(stderr, "error: invalid value for %s: %s\n", arg.c_str(), argv[i]);
+                        print_usage();
+                        return 1;
+                    }
+                } else {
+                    fprintf(stderr, "error: missing value for %s\n", arg.c_str());
+                    print_usage();
+                    return 1;
+                }
+            } else {
+                fprintf(stderr, "error: unrecognized argument %s\n", arg.c_str());
+                print_usage();
+                return 1;
+            }
+        }
+
+        if (string_args["-m"]->empty()) {
+            fprintf(stderr, "error: -m is required\n");
+            print_usage();
+            return 1;
+        }
+
+        return 0;
+    }
+
+   private:
+    const char * program_name;
+    std::unordered_map<std::string, std::string *> string_args;
+    std::unordered_map<std::string, int *> int_args;
+    std::vector<Argument> arguments;
+
+    int parse_int_arg(const char * arg, int & value) {
+        char * end;
+        const long val = std::strtol(arg, &end, 10);
+        if (*end == '\0' && val >= INT_MIN && val <= INT_MAX) {
+            value = static_cast<int>(val);
+            return 0;
+        }
+        return 1;
+    }
+
+    void print_usage() const {
+        printf("\nUsage:\n");
+        printf("  %s [OPTIONS]\n\n", program_name);
+        printf("Options:\n");
+        for (const auto & arg : arguments) {
+            printf("  %-10s %s\n", arg.flag.c_str(), arg.help_text.c_str());
+        }
+
+        printf("\n");
+    }
+};
+
+class LlamaData {
+   public:
+    llama_model_ptr model;
+    llama_sampler_ptr sampler;
+    llama_context_ptr context;
+    std::vector<llama_chat_message> messages;
+
+    int init(const Options & opt) {
+        model = initialize_model(opt.model_path, opt.ngl);
+        if (!model) {
+            return 1;
+        }
+
+        context = initialize_context(model, opt.n_ctx);
+        if (!context) {
+            return 1;
+        }
+
+        sampler = initialize_sampler();
+        return 0;
+    }
+
+   private:
+    // Initializes the model and returns a unique pointer to it
+    llama_model_ptr initialize_model(const std::string & model_path, const int ngl) {
+        llama_model_params model_params = llama_model_default_params();
+        model_params.n_gpu_layers = ngl;
+
+        llama_model_ptr model(llama_load_model_from_file(model_path.c_str(), model_params));
+        if (!model) {
+            fprintf(stderr, "%s: error: unable to load model\n", __func__);
+        }
+
+        return model;
+    }
+
+    // Initializes the context with the specified parameters
+    llama_context_ptr initialize_context(const llama_model_ptr & model, const int n_ctx) {
+        llama_context_params ctx_params = llama_context_default_params();
+        ctx_params.n_ctx = n_ctx;
+        ctx_params.n_batch = n_ctx;
+
+        llama_context_ptr context(llama_new_context_with_model(model.get(), ctx_params));
+        if (!context) {
+            fprintf(stderr, "%s: error: failed to create the llama_context\n", __func__);
+        }
+
+        return context;
+    }
+
+    // Initializes and configures the sampler
+    llama_sampler_ptr initialize_sampler() {
+        llama_sampler_ptr sampler(llama_sampler_chain_init(llama_sampler_chain_default_params()));
+        llama_sampler_chain_add(sampler.get(), llama_sampler_init_min_p(0.05f, 1));
+        llama_sampler_chain_add(sampler.get(), llama_sampler_init_temp(0.8f));
+        llama_sampler_chain_add(sampler.get(), llama_sampler_init_dist(LLAMA_DEFAULT_SEED));
+
+        return sampler;
+    }
+};
+
+// Add a message to `messages` and store its content in `owned_content`
+static void add_message(const char * role, const std::string & text, LlamaData & llama_data,
+                        std::vector<char_array_ptr> & owned_content) {
+    char_array_ptr content(new char[text.size() + 1]);
+    std::strcpy(content.get(), text.c_str());
+    llama_data.messages.push_back({role, content.get()});
+    owned_content.push_back(std::move(content));
+}
+
+// Function to apply the chat template and resize `formatted` if needed
+static int apply_chat_template(const LlamaData & llama_data, std::vector<char> & formatted, const bool append) {
+    int result = llama_chat_apply_template(llama_data.model.get(), nullptr, llama_data.messages.data(),
+                                           llama_data.messages.size(), append, formatted.data(), formatted.size());
+    if (result > static_cast<int>(formatted.size())) {
+        formatted.resize(result);
+        result = llama_chat_apply_template(llama_data.model.get(), nullptr, llama_data.messages.data(),
+                                           llama_data.messages.size(), append, formatted.data(), formatted.size());
+    }
+
+    return result;
+}
+
+// Function to tokenize the prompt
+static int tokenize_prompt(const llama_model_ptr & model, const std::string & prompt,
+                           std::vector<llama_token> & prompt_tokens) {
+    const int n_prompt_tokens = -llama_tokenize(model.get(), prompt.c_str(), prompt.size(), NULL, 0, true, true);
+    prompt_tokens.resize(n_prompt_tokens);
+    if (llama_tokenize(model.get(), prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true,
+                       true) < 0) {
+        GGML_ABORT("failed to tokenize the prompt\n");
+    }
+
+    return n_prompt_tokens;
+}
+
+// Check if we have enough space in the context to evaluate this batch
+static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) {
+    const int n_ctx = llama_n_ctx(ctx.get());
+    const int n_ctx_used = llama_get_kv_cache_used_cells(ctx.get());
+    if (n_ctx_used + batch.n_tokens > n_ctx) {
+        printf("\033[0m\n");
+        fprintf(stderr, "context size exceeded\n");
+        return 1;
+    }
+
+    return 0;
+}
+
+// convert the token to a string
+static int convert_token_to_string(const llama_model_ptr & model, const llama_token token_id, std::string & piece) {
+    char buf[256];
+    int n = llama_token_to_piece(model.get(), token_id, buf, sizeof(buf), 0, true);
+    if (n < 0) {
+        GGML_ABORT("failed to convert token to piece\n");
+    }
+
+    piece = std::string(buf, n);
+    return 0;
+}
+
+static void print_word_and_concatenate_to_response(const std::string & piece, std::string & response) {
+    printf("%s", piece.c_str());
+    fflush(stdout);
+    response += piece;
+}
+
+// helper function to evaluate a prompt and generate a response
+static int generate(LlamaData & llama_data, const std::string & prompt, std::string & response) {
+    std::vector<llama_token> prompt_tokens;
+    const int n_prompt_tokens = tokenize_prompt(llama_data.model, prompt, prompt_tokens);
+    if (n_prompt_tokens < 0) {
+        return 1;
+    }
+
+    // prepare a batch for the prompt
+    llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
+    llama_token new_token_id;
+    while (true) {
+        check_context_size(llama_data.context, batch);
+        if (llama_decode(llama_data.context.get(), batch)) {
+            GGML_ABORT("failed to decode\n");
+        }
+
+        // sample the next token, check is it an end of generation?
+        new_token_id = llama_sampler_sample(llama_data.sampler.get(), llama_data.context.get(), -1);
+        if (llama_token_is_eog(llama_data.model.get(), new_token_id)) {
+            break;
+        }
+
+        std::string piece;
+        if (convert_token_to_string(llama_data.model, new_token_id, piece)) {
+            return 1;
+        }
+
+        print_word_and_concatenate_to_response(piece, response);
+
+        // prepare the next batch with the sampled token
+        batch = llama_batch_get_one(&new_token_id, 1);
+    }
+
+    return 0;
+}
+
+static int parse_arguments(const int argc, const char ** argv, Options & opt) {
+    ArgumentParser parser(argv[0]);
+    parser.add_argument("-m", opt.model_path, "model");
+    parser.add_argument("-p", opt.prompt_non_interactive, "prompt");
+    parser.add_argument("-c", opt.n_ctx, "context_size");
+    parser.add_argument("-ngl", opt.ngl, "n_gpu_layers");
+    if (parser.parse(argc, argv)) {
+        return 1;
+    }
+
+    return 0;
+}
+
+static int read_user_input(std::string & user) {
+    std::getline(std::cin, user);
+    return user.empty();  // Indicate an error or empty input
+}
+
+// Function to generate a response based on the prompt
+static int generate_response(LlamaData & llama_data, const std::string & prompt, std::string & response) {
+    // Set response color
+    printf("\033[33m");
+    if (generate(llama_data, prompt, response)) {
+        fprintf(stderr, "failed to generate response\n");
+        return 1;
+    }
+
+    // End response with color reset and newline
+    printf("\n\033[0m");
+    return 0;
+}
+
+// Helper function to apply the chat template and handle errors
+static int apply_chat_template_with_error_handling(const LlamaData & llama_data, std::vector<char> & formatted,
+                                                   const bool is_user_input, int & output_length) {
+    const int new_len = apply_chat_template(llama_data, formatted, is_user_input);
+    if (new_len < 0) {
+        fprintf(stderr, "failed to apply the chat template\n");
+        return -1;
+    }
+
+    output_length = new_len;
+    return 0;
+}
+
+// Helper function to handle user input
+static bool handle_user_input(std::string & user_input, const std::string & prompt_non_interactive) {
+    if (!prompt_non_interactive.empty()) {
+        user_input = prompt_non_interactive;
+        return true;  // No need for interactive input
+    }
+
+    printf("\033[32m> \033[0m");
+    return !read_user_input(user_input);  // Returns false if input ends the loop
+}
+
+// Function to tokenize the prompt
+static int chat_loop(LlamaData & llama_data, std::string & prompt_non_interactive) {
+    std::vector<char_array_ptr> owned_content;
+    std::vector<char> fmtted(llama_n_ctx(llama_data.context.get()));
+    int prev_len = 0;
+
+    while (true) {
+        // Get user input
+        std::string user_input;
+        if (!handle_user_input(user_input, prompt_non_interactive)) {
+            break;
+        }
+
+        add_message("user", prompt_non_interactive.empty() ? user_input : prompt_non_interactive, llama_data,
+                    owned_content);
+
+        int new_len;
+        if (apply_chat_template_with_error_handling(llama_data, fmtted, true, new_len) < 0) {
+            return 1;
+        }
+
+        std::string prompt(fmtted.begin() + prev_len, fmtted.begin() + new_len);
+        std::string response;
+        if (generate_response(llama_data, prompt, response)) {
+            return 1;
+        }
+    }
+    return 0;
+}
+
+static void log_callback(const enum ggml_log_level level, const char * text, void *) {
+    if (level == GGML_LOG_LEVEL_ERROR) {
+        fprintf(stderr, "%s", text);
+    }
+}
+
+static bool is_stdin_a_terminal() {
+#if defined(_WIN32)
+    HANDLE hStdin = GetStdHandle(STD_INPUT_HANDLE);
+    DWORD mode;
+    return GetConsoleMode(hStdin, &mode);
+#else
+    return isatty(STDIN_FILENO);
+#endif
+}
+
+static std::string read_pipe_data() {
+    std::ostringstream result;
+    result << std::cin.rdbuf();  // Read all data from std::cin
+    return result.str();
+}
+
+int main(int argc, const char ** argv) {
+    Options opt;
+    if (parse_arguments(argc, argv, opt)) {
+        return 1;
+    }
+
+    if (!is_stdin_a_terminal()) {
+        if (!opt.prompt_non_interactive.empty()) {
+            opt.prompt_non_interactive += "\n\n";
+        }
+
+        opt.prompt_non_interactive += read_pipe_data();
+    }
+
+    llama_log_set(log_callback, nullptr);
+    LlamaData llama_data;
+    if (llama_data.init(opt)) {
+        return 1;
+    }
+
+    if (chat_loop(llama_data, opt.prompt_non_interactive)) {
+        return 1;
+    }
+
+    return 0;
+}
diff --git a/include/llama-cpp.h b/include/llama-cpp.h
new file mode 100644
index 0000000000000..daa04d4d84ac2
--- /dev/null
+++ b/include/llama-cpp.h
@@ -0,0 +1,25 @@
+#pragma once
+
+#ifndef __cplusplus
+#error "This header is for C++ only"
+#endif
+
+#include <memory>
+
+#include "llama.h"
+
+struct llama_model_deleter {
+    void operator()(llama_model * model) { llama_free_model(model); }
+};
+
+struct llama_context_deleter {
+    void operator()(llama_context * context) { llama_free(context); }
+};
+
+struct llama_sampler_deleter {
+    void operator()(llama_sampler * sampler) { llama_sampler_free(sampler); }
+};
+
+typedef std::unique_ptr<llama_model, llama_model_deleter> llama_model_ptr;
+typedef std::unique_ptr<llama_context, llama_context_deleter> llama_context_ptr;
+typedef std::unique_ptr<llama_sampler, llama_sampler_deleter> llama_sampler_ptr;

From 0eb4e12beebabae46d37b78742f4c5d4dbe52dc1 Mon Sep 17 00:00:00 2001
From: Junil Kim <logyourself@gmail.com>
Date: Tue, 26 Nov 2024 10:47:20 +0900
Subject: [PATCH 076/102] vulkan: Fix a vulkan-shaders-gen arugment parsing
 error (#10484)

The vulkan-shaders-gen was not parsing the --no-clean argument correctly.
Because the previous code was parsing the arguments which have a value only
and the --no-clean argument does not have a value, it was not being parsed
correctly. This commit can now correctly parse arguments that don't have values.
---
 .../vulkan-shaders/vulkan-shaders-gen.cpp            | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index 6bbe8e96edd9f..5c317b68b7ea0 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -474,9 +474,15 @@ void write_output_files() {
 
 int main(int argc, char** argv) {
     std::map<std::string, std::string> args;
-    for (int i = 1; i < argc; i += 2) {
-        if (i + 1 < argc) {
-            args[argv[i]] = argv[i + 1];
+    for (int i = 1; i < argc; ++i) {
+        std::string arg = argv[i];
+        if (arg.rfind("--", 0) == 0) {
+            if (i + 1 < argc && argv[i + 1][0] != '-') {
+                args[arg] = argv[i + 1];
+                ++i;
+            } else {
+                args[arg] = "";
+            }
         }
     }
 

From 7066b4cce2898993e943ad6af5d8f1de5840c8e9 Mon Sep 17 00:00:00 2001
From: Chenguang Li <87689256+noemotiovon@users.noreply.github.com>
Date: Tue, 26 Nov 2024 17:31:05 +0800
Subject: [PATCH 077/102] CANN: RoPE and CANCAT operator optimization (#10488)

Co-authored-by: noemotiovon <noemotiovon@gmail.com>
---
 ggml/src/ggml-cann/aclnn_ops.cpp | 225 +++++++------------------------
 ggml/src/ggml-cann/ggml-cann.cpp |  68 ++++++++--
 2 files changed, 106 insertions(+), 187 deletions(-)

diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index 1f4ee986ceb89..6113b59f45c2a 100644
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -21,6 +21,7 @@
  */
 
 #include "aclnn_ops.h"
+#include "ggml-impl.h"
 
 #include <aclnnop/aclnn_avgpool2d.h>
 #include <aclnnop/aclnn_cast.h>
@@ -241,10 +242,14 @@ void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
-    int64_t concat_dim = 1;
+    const int32_t dim = ggml_get_op_params_i32(dst, 0);
+
+    GGML_ASSERT(dim >= 0 && dim < 4);
+    int32_t acl_dim = 3 - dim;
+
     aclTensor* tensors[] = {acl_src0, acl_src1};
     aclTensorList* tensorList = aclCreateTensorList(tensors, 2);
-    aclnn_concat(ctx, tensorList, acl_dst, concat_dim);
+    aclnn_concat(ctx, tensorList, acl_dst, acl_dim);
 
     ACL_CHECK(aclDestroyTensorList(tensorList));
     ACL_CHECK(aclDestroyTensor(acl_dst));
@@ -1437,10 +1442,6 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src0 = dst->src[0];  // kernel
     ggml_tensor* src1 = dst->src[1];  // input
 
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
-
     GGML_TENSOR_BINARY_OP_LOCALS;
 
     // aclnnIm2col only works on 2D. set s1, p1, d1 to 1 to perform 2D
@@ -1462,9 +1463,6 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     const int64_t OH = is_2D ? ne2 : 1;
     const int64_t OW = ne1;
 
-    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nb10 == sizeof(float));
-
     // memory allocated increased to 3x when is_2D == false
     const int64_t n_bytes_factor = is_2D ? 1 : 3;
 
@@ -2859,15 +2857,27 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     ACL_CHECK(aclDestroyTensor(acl_cos_tensor));
 }
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+aclnnStatus aclnnRotaryPositionEmbeddingGetWorkspaceSize(
+    const aclTensor* x, const aclTensor* cos, const aclTensor* sin,
+    int64_t mode, const aclTensor* yOut, uint64_t* workspaceSize,
+    aclOpExecutor** executor);
+aclnnStatus aclnnRotaryPositionEmbedding(void* workspace,
+                                         uint64_t workspaceSize,
+                                         aclOpExecutor* executor,
+                                         aclrtStream stream);
+#ifdef __cplusplus
+}
+#endif
+
 void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     // TODO: use ascendc
     // Only test with LLAMA model.
     ggml_tensor* src0 = dst->src[0];  // input
     ggml_tensor* src2 = dst->src[2];  // freq_factors
 
-    // TODO: with freq_factors
-    GGML_ASSERT(src2 == NULL);
-
     // param
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
     // const int n_past     = ((int32_t *) dst->op_params)[0];
@@ -2885,13 +2895,19 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     memcpy(&beta_fast, (int32_t*)dst->op_params + 9, sizeof(float));
     memcpy(&beta_slow, (int32_t*)dst->op_params + 10, sizeof(float));
 
-    GGML_ASSERT(n_dims <= ne0);
+    // TODO: with freq_factors
+    GGML_ASSERT(src2 == NULL);
+    // TODO: attn_factor != 1
+    GGML_ASSERT(attn_factor == 1);
+    // TODO: n_dims <= ne0
+    GGML_ASSERT(n_dims == ne0);
     GGML_ASSERT(n_dims % 2 == 0);
-
     // TODO: ext_factor != 0
     GGML_ASSERT(ext_factor == 0);
     // TODO: freq_scale != 1
     GGML_ASSERT(freq_scale == 1);
+    // TODO: type == GGML_TYPE_F16
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
 
     const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
@@ -2924,177 +2940,30 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     aclnn_cache_init(ctx, dst, acl_cos_reshape_tensor, acl_sin_reshape_tensor,
                      theta_scale, is_neox);
 
-    // roll input
-    void* input_roll_buffer;
-    aclTensor* acl_minus_one_tensor;
-    void* minus_one_scale_buffer = nullptr;
-    ggml_cann_pool_alloc roll_allocator(ctx.pool(), ggml_nbytes(src0));
-    ggml_cann_pool_alloc minus_one_scale_allocator(
-        ctx.pool(), sizeof(float_t) * src0->ne[0]);
-    if (!is_neox) {
-        // roll input: [q0,q1,q2,q3,...] -> [q1,q0,q3,q2,...]
-        input_roll_buffer = roll_allocator.get();
-        int64_t input_roll_ne[4] = {2, src0->ne[1] * (src0->ne[0] / 2),
-                                    src0->ne[2], src0->ne[3]};
-        size_t input_roll_nb[GGML_MAX_DIMS];
-        input_roll_nb[0] = ggml_type_size(src0->type);
-        for (int i = 1; i < GGML_MAX_DIMS; i++) {
-            input_roll_nb[i] = input_roll_nb[i - 1] * input_roll_ne[i - 1];
-        }
-        aclTensor* acl_input_roll_tensor = ggml_cann_create_tensor(
-            input_roll_buffer, ggml_cann_type_mapping(src0->type),
-            ggml_type_size(src0->type), input_roll_ne, input_roll_nb,
-            GGML_MAX_DIMS);
-        aclTensor* acl_input_tensor = ggml_cann_create_tensor(
-            src0->data, ggml_cann_type_mapping(src0->type),
-            ggml_type_size(src0->type), input_roll_ne, input_roll_nb,
-            GGML_MAX_DIMS);
-
-        int64_t shifts[] = {1};
-        int64_t dims[] = {3};
-        aclnn_roll(ctx, acl_input_tensor, acl_input_roll_tensor, shifts, dims);
-        ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor));
-        ACL_CHECK(aclDestroyTensor(acl_input_tensor));
-
-        // init [-1, 1, -1, 1, ...]
-        minus_one_scale_buffer = minus_one_scale_allocator.get();
-
-        int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
-        size_t minus_one_nb[GGML_MAX_DIMS];
-        minus_one_nb[0] = sizeof(float_t);
-        for (int i = 1; i < GGML_MAX_DIMS; i++) {
-            minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
-        }
-        acl_minus_one_tensor = aclnn_ones(
-            ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
-            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
-        int64_t dim = 3;
-        int64_t* index = new int64_t[src0->ne[0]];
-        for (int i = 0; i < src0->ne[0]; i++) {
-            index[i] = i / 2 * 2;
-        }
-        int64_t index_num = src0->ne[0];
-        float value = -1;
-        aclnn_index_fill_tensor(ctx, acl_minus_one_tensor, dim, index,
-                                index_num, value);
-    } else {
-        // roll input: [q0,q1,q2,...] ->
-        // [q_half,q_half+1,...,q_end,q0,q1,...q_half-1]
-        input_roll_buffer = roll_allocator.get();
-        aclTensor* acl_input_roll_tensor = ggml_cann_create_tensor(
-            input_roll_buffer, ggml_cann_type_mapping(src0->type),
-            ggml_type_size(src0->type), src0->ne, src0->nb, GGML_MAX_DIMS);
-        aclTensor* acl_input_tensor = ggml_cann_create_tensor(src0);
-
-        int64_t shifts[] = {src0->ne[0] / 2};
-        int64_t dims[] = {3};
-        aclnn_roll(ctx, acl_input_tensor, acl_input_roll_tensor, shifts, dims);
-
-        ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor));
-        ACL_CHECK(aclDestroyTensor(acl_input_tensor));
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
 
-        // init [-1, -1, -1, 1, 1，1，...]
-        minus_one_scale_buffer = minus_one_scale_allocator.get();
+    void* workspaceAddr = nullptr;
 
-        int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
-        size_t minus_one_nb[GGML_MAX_DIMS];
-        minus_one_nb[0] = sizeof(float_t);
-        for (int i = 1; i < GGML_MAX_DIMS; i++) {
-            minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
-        }
-        acl_minus_one_tensor = aclnn_ones(
-            ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
-            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
-        // -1 * first half
-        int64_t first_half_ne[4] = {src0->ne[0] / 2, 1, 1, 1};
-        size_t first_half_nb[GGML_MAX_DIMS];
-        first_half_nb[0] = sizeof(float_t);
-        for (int i = 1; i < GGML_MAX_DIMS; i++) {
-            first_half_nb[i] = first_half_nb[i - 1] * first_half_ne[i - 1];
-        }
-        aclTensor* acl_first_half_tensor = ggml_cann_create_tensor(
-            minus_one_scale_buffer, ACL_FLOAT, sizeof(float_t), first_half_ne,
-            first_half_nb, GGML_MAX_DIMS);
-        bool inplace = true;
-        float scale = -1;
-        aclnn_muls(ctx, acl_first_half_tensor, scale, nullptr, inplace);
-        ACL_CHECK(aclDestroyTensor(acl_first_half_tensor));
-    }
-
-    // TODO: n_dims < ne0
-    GGML_ASSERT(n_dims == src0->ne[0]);
-
-    // input * scale
-    ggml_cann_pool_alloc roll_mul_scale_allocator(ctx.pool(),
-                                                  ggml_nbytes(src0));
-    void* input_roll_mul_scale_buffer = roll_mul_scale_allocator.get();
-    size_t input_nb[GGML_MAX_DIMS];
-    input_nb[0] = ggml_type_size(src0->type);
-    for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        input_nb[i] = input_nb[i - 1] * src0->ne[i - 1];
+    int acl_mode = mode;
+    if (mode == 0) {
+        acl_mode = 1;
     }
-    aclTensor* acl_input_roll_mul_scale_tensor = ggml_cann_create_tensor(
-        input_roll_mul_scale_buffer, ggml_cann_type_mapping(src0->type),
-        ggml_type_size(src0->type), src0->ne, input_nb, GGML_MAX_DIMS);
-    aclTensor* acl_input_roll_reshape_tensor = ggml_cann_create_tensor(
-        input_roll_buffer, ggml_cann_type_mapping(src0->type),
-        ggml_type_size(src0->type), src0->ne, input_nb, GGML_MAX_DIMS);
 
-    aclnn_mul(ctx, acl_input_roll_reshape_tensor, acl_minus_one_tensor,
-              acl_input_roll_mul_scale_tensor);
-
-    // output
-    aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
+    aclTensor* acl_x = ggml_cann_create_tensor(src0);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
-    void* output_fp32_buffer;
-    if (src0->type == GGML_TYPE_F32) {
-        aclnn_inplace_mul(ctx, acl_src0, acl_cos_reshape_tensor);
-        aclnn_inplace_mul(ctx, acl_input_roll_mul_scale_tensor,
-                          acl_sin_reshape_tensor);
-        aclnn_add(ctx, acl_src0, acl_input_roll_mul_scale_tensor, acl_dst);
-        // TODO: ne0 != n_dims in mode2
-    } else if (src0->type == GGML_TYPE_F16) {
-        size_t input_fp32_nb[GGML_MAX_DIMS];
-        input_fp32_nb[0] = sizeof(float_t);
-        for (int i = 1; i < GGML_MAX_DIMS; i++) {
-            input_fp32_nb[i] = input_fp32_nb[i - 1] * dst->ne[i - 1];
-        }
-        ggml_cann_pool_alloc fp32_allocator1(
-            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
-        void* input_fp32_buffer1 = fp32_allocator1.get();
-        aclTensor* input_fp32_tensor1 = ggml_cann_create_tensor(
-            input_fp32_buffer1, ACL_FLOAT, sizeof(float_t), dst->ne,
-            input_fp32_nb, GGML_MAX_DIMS);
-        ggml_cann_pool_alloc fp32_allocator2(
-            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
-        void* input_fp32_buffer2 = fp32_allocator2.get();
-        aclTensor* input_fp32_tensor2 = ggml_cann_create_tensor(
-            input_fp32_buffer2, ACL_FLOAT, sizeof(float_t), dst->ne,
-            input_fp32_nb, GGML_MAX_DIMS);
-
-        ggml_cann_pool_alloc fp32_allocator(
-            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
-        output_fp32_buffer = fp32_allocator.get();
-        aclTensor* output_fp32_tensor = ggml_cann_create_tensor(
-            output_fp32_buffer, ACL_FLOAT, sizeof(float_t), dst->ne,
-            input_fp32_nb, GGML_MAX_DIMS);
-        aclnn_mul(ctx, acl_src0, acl_cos_reshape_tensor, input_fp32_tensor1);
-        aclnn_mul(ctx, acl_input_roll_mul_scale_tensor, acl_sin_reshape_tensor,
-                  input_fp32_tensor2);
-        aclnn_add(ctx, input_fp32_tensor1, input_fp32_tensor2,
-                  output_fp32_tensor);
-        aclnn_cast(ctx, output_fp32_tensor, acl_dst, ACL_FLOAT16);
-
-        ACL_CHECK(aclDestroyTensor(input_fp32_tensor1));
-        ACL_CHECK(aclDestroyTensor(input_fp32_tensor2));
-        ACL_CHECK(aclDestroyTensor(output_fp32_tensor));
+    ACL_CHECK(aclnnRotaryPositionEmbeddingGetWorkspaceSize(
+        acl_x, acl_cos_reshape_tensor, acl_sin_reshape_tensor, acl_mode, acl_dst, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
     }
 
-    ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+    ACL_CHECK(aclnnRotaryPositionEmbedding(workspaceAddr, workspaceSize,
+                                           executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyTensor(acl_x));
     ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_minus_one_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_input_roll_mul_scale_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_input_roll_reshape_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_src0));
+    ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
     ACL_CHECK(aclDestroyTensor(acl_dst));
 }
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index d96f65936136d..2ef5b590a8ecb 100644
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -1669,12 +1669,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             }
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
-                case GGML_TYPE_F16:
-                case GGML_TYPE_F32:
                 case GGML_TYPE_Q8_0:
-                    // TODO: fix me
                     // Current groupsize should not be greater than k-1 in
-                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
+                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize
+                    if (op->src[0]->ne[0] <= QK8_0) {
+                        return false;
+                    }
+                case GGML_TYPE_F16:
+                case GGML_TYPE_F32:
                 case GGML_TYPE_Q4_0:
                     return true;
                 default:
@@ -1706,9 +1708,61 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return false;
             }
         }
+        case GGML_OP_CONT: {
+            // TODO: support GGML_TYPE_BF16
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                    return true;
+                default:
+                    return false;
+            }
+        }
+        case GGML_OP_ROPE: {
+            // TODO: with ops-test v == 1
+            float * freq_scale = (float*)((int32_t*)op->op_params + 6);
+            float * ext_factor = (float*)((int32_t*)op->op_params + 7);
+            float * attn_factor = (float*)((int32_t*)op->op_params + 8);
+            // TODO: with freq_factors
+            if (op->src[2] != NULL) {
+                return false;
+            }
+            // TODO: n_dims <= ne0
+            if (op->src[0]->ne[0] != op->op_params[1]) {
+                return false;
+            }
+            // TODO: ext_factor != 0
+            if (*ext_factor != 0) {
+                return false;
+            }
+            // TODO: freq_scale != 1
+            if (*freq_scale != 1) {
+                return false;
+            }
+            // TODO: attn_factor != 1
+            if (*attn_factor != 1) {
+                return false;
+            }
+            //TODO: type == GGML_TYPE_F16
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                    return true;
+                default:
+                    return false;
+            }
+        }
+        case GGML_OP_UPSCALE: {
+            // aclnnUpsampleNearest2dGetWorkspaceSize not support
+            // selfDimN[2]/outDimN[2] or selfDimC[3]/outDimC[3] not equal
+            if (op->src[0]->ne[2] * op->ne[3] != op->src[0]->ne[3] * op->ne[2]) {
+                return false;
+            }
+            return true;
+        }
+        case GGML_OP_IM2COL:
+        case GGML_OP_CONCAT:
         case GGML_OP_DUP:
         case GGML_OP_REPEAT:
-        case GGML_OP_CONCAT:
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
@@ -1722,17 +1776,13 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_CLAMP:
-        case GGML_OP_CONT:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
-        case GGML_OP_ROPE:
-        case GGML_OP_IM2COL:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
         case GGML_OP_GROUP_NORM:
-        case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:

From 9a4b79bcfa4338b922fa8cf903bd5ac058aaf46f Mon Sep 17 00:00:00 2001
From: Shanshan Shen <467638484@qq.com>
Date: Tue, 26 Nov 2024 18:08:37 +0800
Subject: [PATCH 078/102] CANN: Improve the Inferencing Performance for Ascend
 NPU Device (#10454)

* improve inferencing performance for ascend npu.

Co-authored-by: Frank Mai <thxCode@thxcode0824@gmail.com>

* some modification after review

* some modifications after review

* restore some modifications

* restore some modifications

---------

Co-authored-by: shanshan shen <shanshanshen333@gmail.com>
Co-authored-by: Frank Mai <thxCode@thxcode0824@gmail.com>
---
 ggml/src/ggml-cann/aclnn_ops.cpp | 299 +++++++++++++++++++++++--------
 ggml/src/ggml-cann/common.h      |   9 +-
 ggml/src/ggml-cann/ggml-cann.cpp |  58 +++---
 3 files changed, 265 insertions(+), 101 deletions(-)

diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index 6113b59f45c2a..d7472ee3a55c7 100644
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -33,6 +33,8 @@
 #include <aclnnop/aclnn_group_norm.h>
 #include <aclnnop/aclnn_index_fill_tensor.h>
 #include <aclnnop/aclnn_layer_norm.h>
+#include <aclnnop/aclnn_mm.h>
+#include <aclnnop/aclnn_batch_matmul.h>
 #include <aclnnop/aclnn_matmul.h>
 #include <aclnnop/aclnn_max_pool.h>
 #include <aclnnop/aclnn_permute.h>
@@ -2423,7 +2425,6 @@ static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input,
                           aclTensor* acl_weight, aclTensor* acl_dst) {
     int8_t cube_math_type = 1;  // ALLOW_FP32_DOWN_PRECISION, when input is
                                 // fp32, atlas a2 will transpose it to HFLOAT32.
-
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
@@ -2441,6 +2442,80 @@ static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input,
         aclnnMatmul(workspaceAddr, workspaceSize, executor, ctx.stream()));
 }
 
+/**
+ * @brief Performs matrix multiplication of two 2D tensors.
+ *
+ * This function computes the matrix multiplication of the input tensor
+ * `acl_input` and the weight tensor `acl_weight`, and stores the result in the
+ * destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst}=\text {acl_input@acl_weight}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_input The input tensor for the matrix multiplication.
+ * @param acl_weight The weight tensor for the matrix multiplication.
+ * @param acl_dst The destination tensor where the result of the matrix
+ * multiplication will be stored.
+ */
+static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx, aclTensor* acl_input,
+                             aclTensor* acl_weight, aclTensor* acl_dst) {
+    int8_t cube_math_type = 2;
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnMmGetWorkspaceSize(acl_input, acl_weight, acl_dst,
+                                      cube_math_type, &workspaceSize,
+                                      &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnMm(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+/**
+ * @brief Performs matrix multiplication of two 3D tensors.
+ *
+ * This function computes the matrix multiplication of the input tensor
+ * `acl_input` and the weight tensor `acl_weight`, and stores the result in the
+ * destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst}=\text {acl_input@acl_weight}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_input The input tensor for the matrix multiplication.
+ * @param acl_weight The weight tensor for the matrix multiplication.
+ * @param acl_dst The destination tensor where the result of the matrix
+ * multiplication will be stored.
+ */
+static void aclnn_mat_mul_3d(ggml_backend_cann_context& ctx, aclTensor* acl_input,
+                             aclTensor* acl_weight, aclTensor* acl_dst) {
+    int8_t cube_math_type = 2;
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnBatchMatMulGetWorkspaceSize(acl_input, acl_weight, acl_dst,
+                                               cube_math_type, &workspaceSize,
+                                               &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnBatchMatMul(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
 /**
  * @brief Performs matrix multiplication with floating-point precision on
  * tensors using the CANN backend.
@@ -2462,20 +2537,43 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
     // broadcast, when weight ne2 or ne3 is not 1, weight need repeat.
     BCAST_MUL_MAT_SHAPE(input, weight, dst);
 
-    // transpose weight: [1,2,3,4] -> [1,2,4,3]
-    int64_t transpose_ne[] = {bcast_weight_ne[1], bcast_weight_ne[0],
-                              bcast_weight_ne[2], bcast_weight_ne[3],
-                              bcast_weight_ne[4], bcast_weight_ne[5]};
-    size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0],
-                             bcast_weight_nb[2], bcast_weight_nb[3],
-                             bcast_weight_nb[4], bcast_weight_nb[5]};
+    int64_t n_dims = bcast_dims;
+    if (bcast_input_ne[3] == bcast_weight_ne[3] && bcast_input_ne[3] == 1) {
+        if (bcast_input_ne[2] == 1 && bcast_weight_ne[2] == 1) {
+            n_dims = 2;
+        } else if (bcast_input_ne[2] == 1) {
+            n_dims = 3;
+        }
+    }
 
-    aclTensor* acl_weight_tensor =
-        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, bcast_dims);
     aclTensor* acl_input_tensor =
-        ggml_cann_create_tensor(input, BCAST_MUL_MAT_PARAM(input));
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst, BCAST_MUL_MAT_PARAM(dst));
-    aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+        ggml_cann_create_tensor(input, bcast_input_ne, bcast_input_nb, n_dims);
+    int64_t transpose_ne[] = {
+        bcast_weight_ne[1], bcast_weight_ne[0],
+        bcast_weight_ne[2], bcast_weight_ne[3],
+        bcast_weight_ne[4], bcast_weight_ne[5]
+    };
+    size_t transpose_nb[] = {
+        bcast_weight_nb[1], bcast_weight_nb[0],
+        bcast_weight_nb[2], bcast_weight_nb[3],
+        bcast_weight_nb[4], bcast_weight_nb[5]
+    };
+    aclTensor* acl_weight_tensor =
+        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims);
+    aclTensor* acl_dst =
+        ggml_cann_create_tensor(dst, bcast_dst_ne, bcast_dst_nb, n_dims);
+
+    switch (n_dims) {
+    case 2:
+        aclnn_mat_mul_2d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+        break;
+    case 3:
+        aclnn_mat_mul_3d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+        break;
+    default:
+        aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+        break;
+    }
 
     ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
     ACL_CHECK(aclDestroyTensor(acl_input_tensor));
@@ -2501,46 +2599,40 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
     ggml_tensor* src0 = dst->src[0];  // weight
     ggml_tensor* src1 = dst->src[1];  // input
 
-    // The shape of the weight is NCHW. Matrix multiplication uses HW dims. HC
-    // is regarded as batch. weight need transpose.
-    int64_t weight_ne[] = {src0->ne[1], src0->ne[0]};
+    // The shape of the weight is NCHW.
+    // Matrix multiplication uses HW dims.
+    // HC is regarded as batch.
+    // weight need transpose.
     float weight_elem_size;
     if (type == GGML_TYPE_Q4_0) {
         weight_elem_size = float(sizeof(uint8_t)) / 2;
-    }
-    else if (type == GGML_TYPE_Q8_0) {
+    } else if (type == GGML_TYPE_Q8_0) {
         weight_elem_size = float(sizeof(uint8_t));
-    }
-    else {
+    } else {
         GGML_ABORT("Only support Q4_0 and Q8_0 MUL_MAT");
     }
-    float weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
-
-    // size of one matrix is element_size * height * width.
-    size_t weight_stride = weight_elem_size * src0->ne[0] * src0->ne[1];
+    float weight_nb[] = {src0->ne[0] * weight_elem_size, weight_elem_size};
+    size_t weight_stride = src0->ne[1] * src0->ne[0] * weight_elem_size;
     size_t weight_size = weight_stride * src0->ne[2] * src0->ne[3];
 
     // scale stored at the end of weight. Also need transpose.
-    GGML_ASSERT(QK4_0 == QK8_0);
-    int64_t scale_ne[] = {src0->ne[1], src0->ne[0] / QK8_0};
     size_t scale_elem_size = sizeof(uint16_t);
-    size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size,
-                         scale_elem_size};
-    size_t scale_stride = scale_elem_size * src0->ne[0] * src0->ne[1] / QK8_0;
+    size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size, scale_elem_size};
+    size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
     char* scale_offset = (char*)src0->data + weight_size;
 
     // input
-    void* input_buffer;
     size_t input_elem_size = sizeof(uint16_t);
     int64_t input_ne[] = {src1->ne[0], src1->ne[1]};
-    size_t input_nb[] = {input_elem_size, input_elem_size * src1->ne[0]};
-    size_t input_stride = input_elem_size * src1->ne[0] * src1->ne[1];
-
+    size_t input_nb[] = {input_elem_size,  input_ne[0] * input_elem_size};
+    size_t input_stride = input_ne[0] * input_ne[1] * input_elem_size;
     ggml_cann_pool_alloc input_alloctor(ctx.pool());
+    void* input_buffer = src1->data;
+
+    // case in
     if (src1->type != GGML_TYPE_F16) {
         aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1);
-        input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
-        input_buffer = input_alloctor.get();
+        input_buffer = input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
 
         int64_t* input_cast_ne = src1->ne;
         size_t input_cast_nb[GGML_MAX_DIMS];
@@ -2550,88 +2642,139 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
         }
 
         aclTensor* acl_input_tensor = ggml_cann_create_tensor(
-            input_buffer, ACL_FLOAT16, input_elem_size, input_cast_ne,
-            input_cast_nb, GGML_MAX_DIMS);
+            input_buffer,
+            ACL_FLOAT16,
+            input_elem_size, input_cast_ne, input_cast_nb, GGML_MAX_DIMS);
         aclnn_cast(ctx, acl_src1_tensor, acl_input_tensor, ACL_FLOAT16);
+
         ACL_CHECK(aclDestroyTensor(acl_input_tensor));
         ACL_CHECK(aclDestroyTensor(acl_src1_tensor));
-    } else {
-        input_buffer = src1->data;
     }
 
     // output
     size_t output_elem_size = sizeof(uint16_t);
-    int64_t output_ne[] = {dst->ne[0], dst->ne[1]};
-    size_t output_nb[] = {output_elem_size, output_elem_size * dst->ne[0]};
-    ggml_cann_pool_alloc output_alloctor(
-        ctx.pool(), ggml_nelements(dst) * output_elem_size);
-    void* output_buffer = output_alloctor.get();
-    size_t output_stride = output_elem_size * dst->ne[0] * dst->ne[1];
+    size_t output_nb[] = {output_elem_size, dst->ne[0] * output_elem_size};
+    ggml_cann_pool_alloc output_allocator(ctx.pool());
+    void* output_buffer = output_allocator.alloc(ggml_nelements(dst) * output_elem_size);
+    size_t output_stride = dst->ne[0] * dst->ne[1] * output_elem_size;
 
     // aclnn
+    int64_t max_elem_size = 65535;
+    int64_t split_size = (src0->ne[1] / max_elem_size) + 1;
+    ggml_cann_pool_alloc workspace_allocator(ctx.pool());
+    aclOpExecutor* executor = nullptr;
     uint64_t workspaceSize = 0;
-    aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
-
     for (int64_t n1 = 0; n1 < src1->ne[3]; n1++) {
         for (int64_t c1 = 0; c1 < src1->ne[2]; c1++) {
             int64_t n0 = n1 / (src1->ne[3] / src0->ne[3]);
             int64_t c0 = c1 / (src1->ne[2] / src0->ne[2]);
 
-            int64_t batch1 = n1 * src1->ne[2] + c1;
-            int64_t batch0 = n0 * src0->ne[2] + c0;
+            int64_t batch1 = (n1 * src1->ne[2]) + c1;
+            int64_t batch0 = (n0 * src0->ne[2]) + c0;
 
             aclTensor* acl_input_tensor = ggml_cann_create_tensor(
                 (char*)input_buffer + batch1 * input_stride, ACL_FLOAT16,
                 input_elem_size, input_ne, input_nb, 2);
+
+            // first split
+            int64_t weight_ne_offset = 0;
+            int64_t weight_ne[2] = {max_elem_size > src0->ne[1] ? src0->ne[1] : max_elem_size, src0->ne[0]};
+            int64_t scale_ne_offset = 0;
+            int64_t scale_ne[2] = {weight_ne[0], weight_ne[1] / QK8_0};
+            int64_t output_ne_offset = 0;
+            int64_t output_ne[2] = {weight_ne[0], dst->ne[1]};
+
             aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
                 (char*)src0->data + batch0 * weight_stride,
-                ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
-                weight_nb, 2);
+                ggml_cann_type_mapping(type),
+                weight_elem_size, weight_ne, weight_nb, 2,
+                ACL_FORMAT_ND, weight_ne_offset);
             aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
-                scale_offset + batch0 * scale_stride, ACL_FLOAT16,
-                scale_elem_size, scale_ne, scale_nb, 2);
+                scale_offset + batch0 * scale_stride,
+                ACL_FLOAT16,
+                scale_elem_size, scale_ne, scale_nb, 2,
+                ACL_FORMAT_ND, scale_ne_offset);
             aclTensor* acl_output_tensor = ggml_cann_create_tensor(
-                (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
-                output_elem_size, output_ne, output_nb, 2);
+                (char*)output_buffer + batch1 * output_stride,
+                ACL_FLOAT16,
+                output_elem_size, output_ne, output_nb, 2,
+                ACL_FORMAT_ND, output_ne_offset);
 
             ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
-                acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr,
-                nullptr, nullptr, nullptr, QK8_0, acl_output_tensor,
-                &workspaceSize, &executor));
-
-            if (workspaceSize > 0 && workspaceAddr == nullptr) {
-                ggml_cann_pool_alloc workspace_allocator(ctx.pool(),
-                                                         workspaceSize);
-                workspaceAddr = workspace_allocator.get();
+                acl_input_tensor, acl_weight_tensor, acl_scale_tensor,
+                nullptr, nullptr, nullptr, nullptr, QK8_0,
+                acl_output_tensor, &workspaceSize, &executor));
+            if (workspaceAddr == nullptr) {
+                workspaceAddr = workspace_allocator.alloc(workspaceSize);
             }
-
             ACL_CHECK(aclnnWeightQuantBatchMatmulV2(
                 workspaceAddr, workspaceSize, executor, ctx.stream()));
 
-            ACL_CHECK(aclDestroyTensor(acl_input_tensor));
             ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
             ACL_CHECK(aclDestroyTensor(acl_scale_tensor));
             ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+
+            // other splits
+            for (int64_t split = 1; split < split_size; split++) {
+                weight_ne_offset += weight_elem_size * weight_ne[0] * weight_ne[1];
+                weight_ne[0] = max_elem_size * (split + 1) > src0->ne[1] ? src0->ne[1] - (max_elem_size * split) : max_elem_size;
+                scale_ne_offset += scale_elem_size * scale_ne[0] * scale_ne[1];
+                scale_ne[0] = weight_ne[0];
+                output_ne_offset += output_elem_size * output_ne[0] * output_ne[1];
+                output_ne[0] = weight_ne[0];
+
+                acl_weight_tensor = ggml_cann_create_tensor(
+                    (char*)src0->data + batch0 * weight_stride,
+                    ggml_cann_type_mapping(type),
+                    weight_elem_size, weight_ne, weight_nb, 2,
+                    ACL_FORMAT_ND, weight_ne_offset);
+                acl_scale_tensor = ggml_cann_create_tensor(
+                    scale_offset + batch0 * scale_stride,
+                    ACL_FLOAT16,
+                    scale_elem_size, scale_ne, scale_nb, 2,
+                    ACL_FORMAT_ND, scale_ne_offset);
+                acl_output_tensor = ggml_cann_create_tensor(
+                    (char*)output_buffer + batch1 * output_stride,
+                    ACL_FLOAT16,
+                    output_elem_size, output_ne, output_nb, 2,
+                    ACL_FORMAT_ND, output_ne_offset);
+
+                ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
+                    acl_input_tensor, acl_weight_tensor, acl_scale_tensor,
+                    nullptr, nullptr, nullptr, nullptr, QK8_0,
+                    acl_output_tensor, &workspaceSize, &executor));
+                ACL_CHECK(aclnnWeightQuantBatchMatmulV2(
+                    workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+                ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
+                ACL_CHECK(aclDestroyTensor(acl_scale_tensor));
+                ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+            }
+
+            ACL_CHECK(aclDestroyTensor(acl_input_tensor));
         }
     }
 
     // cast out
-    int64_t* output_cast_ne = dst->ne;
-    size_t output_cast_nb[GGML_MAX_DIMS];
-    output_cast_nb[0] = sizeof(uint16_t);
-    for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        output_cast_nb[i] = output_cast_nb[i - 1] * output_cast_ne[i - 1];
-    }
+    if (dst->type != GGML_TYPE_F16) {
+        int64_t* output_cast_ne = dst->ne;
+        size_t output_cast_nb[GGML_MAX_DIMS];
+        output_cast_nb[0] = sizeof(uint16_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            output_cast_nb[i] = output_cast_nb[i - 1] * output_cast_ne[i - 1];
+        }
 
-    aclTensor* acl_output_tensor =
-        ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, output_elem_size,
-                                output_cast_ne, output_cast_nb, GGML_MAX_DIMS);
-    aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
-    aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, ACL_FLOAT);
+        aclTensor* acl_output_tensor = ggml_cann_create_tensor(
+            output_buffer,
+            ACL_FLOAT16,
+            output_elem_size, output_cast_ne, output_cast_nb, GGML_MAX_DIMS);
+        aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
+        aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, ggml_cann_type_mapping(dst->type));
 
-    ACL_CHECK(aclDestroyTensor(acl_output_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_dst_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_dst_tensor));
+    }
 }
 
 void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
diff --git a/ggml/src/ggml-cann/common.h b/ggml/src/ggml-cann/common.h
index edfa496148ff2..5164cb74ec92e 100644
--- a/ggml/src/ggml-cann/common.h
+++ b/ggml/src/ggml-cann/common.h
@@ -211,17 +211,20 @@ struct ggml_cann_pool_alloc {
 struct ggml_backend_cann_context {
     int32_t device;                  /**< Device ID. */
     std::string name;                /**< Name of the device. */
+    std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
 
-    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {
-        {nullptr}}; /**< Array of streams for the device. */
+    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
     /**
      * @brief Constructor for initializing the context with a given device.
      * @param device Device ID.
      */
     explicit ggml_backend_cann_context(int device)
-        : device(device), name("CANN" + std::to_string(device)) {}
+        : device(device), name("CANN" + std::to_string(device)) {
+        ggml_cann_set_device(device);
+        description = aclrtGetSocName();
+    }
 
     /**
      * @brief Destructor for cleaning up resources.
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index 2ef5b590a8ecb..c7a3419c796de 100644
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -122,6 +122,10 @@ static ggml_cann_device_info ggml_cann_init() {
         ACL_CHECK(aclrtMemGetAllocationGranularity(
             &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
             &info.devices[id].vmm_granularity));
+
+        size_t free, total;
+        ggml_backend_cann_get_device_memory(id, &free, &total);
+        info.devices[id].total_vram = free;
     }
 
     // TODO: add more device info later.
@@ -208,6 +212,11 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      * @return A pointer to the allocated buffer.
      */
     void* alloc(size_t size, size_t* actual_size) override {
+        const size_t alignment = 128;
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
+        }
 #ifdef DEBUG_CANN_MALLOC
         int nnz = 0;
         size_t max_size = 0;
@@ -246,13 +255,11 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
             return ptr;
         }
         void* ptr;
-        size_t look_ahead_size = (size_t)(1.05 * size);
-        look_ahead_size = 256 * ((look_ahead_size + 255) / 256);
         ggml_cann_set_device(device);
         ACL_CHECK(
-            aclrtMalloc(&ptr, look_ahead_size, ACL_MEM_MALLOC_HUGE_FIRST));
-        *actual_size = look_ahead_size;
-        pool_size += look_ahead_size;
+            aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        *actual_size = size;
+        pool_size += size;
 #ifdef DEBUG_CANN_MALLOC
         GGML_LOG_INFO(
             "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
@@ -296,7 +303,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
     /**
      * @brief The maximum size of the virtual memory pool (32 GB).
      */
-    static const size_t CANN_POOL_VMM_MAX_SIZE = 1ull << 35;  // 32 GB
+    size_t max_size;
 
     /**
      * @brief The device ID associated with this buffer pool.
@@ -341,7 +348,11 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
      */
     explicit ggml_cann_pool_vmm(int device)
         : device(device),
-          granularity(ggml_cann_info().devices[device].vmm_granularity) {}
+          granularity(ggml_cann_info().devices[device].vmm_granularity) {
+        auto dev = ggml_cann_info().devices[device];
+        granularity = dev.vmm_granularity;
+        max_size = dev.total_vram;
+    }
 
     /**
      * @brief Destructor to free all buffers in the virtual memory pool.
@@ -370,17 +381,19 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
         // round up the allocation size to the alignment to ensure that all
         // allocations are aligned for all data types
         const size_t alignment = 128;
-        size = alignment * ((size + alignment - 1) / alignment);
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
+        }
 
         size_t avail = pool_size - pool_used;
 
         if (size > avail) {
             // round up to the next multiple of the granularity
             size_t reserve_size = size - avail;
-            reserve_size =
-                granularity * ((reserve_size + granularity - 1) / granularity);
+            reserve_size = GGML_PAD(reserve_size, granularity);
 
-            GGML_ASSERT(pool_size + reserve_size <= CANN_POOL_VMM_MAX_SIZE);
+            GGML_ASSERT(pool_size + reserve_size <= max_size);
 
             // allocate more physical memory
             aclrtPhysicalMemProp prop = {};
@@ -396,7 +409,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
             // reserve virtual address space (if not already reserved)
             if (pool_addr == 0) {
                 ACL_CHECK(aclrtReserveMemAddress(
-                    &pool_addr, CANN_POOL_VMM_MAX_SIZE, 0, NULL, 1));
+                    &pool_addr, max_size, 0, NULL, 1));
             }
 
             // map at the end of the pool
@@ -409,10 +422,11 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
             // add to the pool
             pool_size += reserve_size;
 
-            // GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (
-            // reserved %llu MB)\n",
-            //       device, (unsigned long long) (pool_size/1024/1024),
-            //       (unsigned long long) (reserve_size/1024/1024));
+#ifdef DEBUG_CANN_MALLOC
+             GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (reserved %llu MB)\n",
+                   device, (unsigned long long) (pool_size/1024/1024),
+                   (unsigned long long) (reserve_size/1024/1024));
+#endif
         }
 
         GGML_ASSERT(pool_addr != 0);
@@ -457,7 +471,6 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
  */
 std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
     int device) {
-    // return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_leg(device));
     return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
 }
 
@@ -1130,10 +1143,10 @@ ggml_backend_cann_buffer_type(int32_t device) {
     static bool ggml_backend_cann_buffer_type_initialized = false;
 
     if (!ggml_backend_cann_buffer_type_initialized) {
-        for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) {
+        for (int32_t i = 0; i < ggml_cann_info().device_count; i++) {
             ggml_backend_cann_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_cann_buffer_type_interface,
-                /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
+                /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), i),
                 /* .context  = */
                  new ggml_backend_cann_buffer_type_context{
                     i, "CANN" + std::to_string(i)},
@@ -1199,10 +1212,15 @@ static void * ggml_cann_host_malloc(size_t size) {
         return nullptr;
     }
 
+    const size_t alignment = 128;
+    size = GGML_PAD(size, alignment);
+    if (size == 0) {
+        size = alignment;
+    }
+
     void * hostPtr = nullptr;
     aclError err = aclrtMallocHost((void **) &hostPtr, size);
     if (err != ACL_SUCCESS) {
-
         GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
                            size / 1024.0 / 1024.0, aclGetRecentErrMsg());
         return nullptr;

From 811872a59daefb25fc0c4326bcb6d8ae893c2f7c Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Tue, 26 Nov 2024 12:29:38 +0200
Subject: [PATCH 079/102] speculative : simplify the implementation (#10504)

ggml-ci
---
 .../speculative-simple/speculative-simple.cpp | 57 ++++++++-----------
 1 file changed, 24 insertions(+), 33 deletions(-)

diff --git a/examples/speculative-simple/speculative-simple.cpp b/examples/speculative-simple/speculative-simple.cpp
index 7bf9056bf6db1..2ea49d47c433e 100644
--- a/examples/speculative-simple/speculative-simple.cpp
+++ b/examples/speculative-simple/speculative-simple.cpp
@@ -117,7 +117,8 @@ int main(int argc, char ** argv) {
     llama_token id_last = inp.back();
 
     // all tokens currently in the target context
-    auto prompt_tgt = std::vector<llama_token>(inp.begin(), inp.end() - 1);
+    llama_tokens prompt_tgt(inp.begin(), inp.end() - 1);
+    prompt_tgt.reserve(llama_n_ctx(ctx_tgt));
 
     int n_past = inp.size() - 1;
 
@@ -181,54 +182,44 @@ int main(int argc, char ** argv) {
         GGML_ASSERT(ids.size() > 0); // there will always be at least one accepted token
 
         n_past    += ids.size() - 1;
-        n_drafted += batch_tgt.n_tokens - 1;
+        n_drafted += draft.size(); // note: we ignore the discarded small drafts
         n_accept  += ids.size() - 1;
+        n_predict += ids.size();
 
         // process the accepted tokens and update contexts
         //
         // this is the standard token post-processing that we normally do
         // in this case, we do it for a group of accepted tokens at once
         //
-        {
-            llama_token id;
-            std::string token_str;
-
-            for (size_t i = 0; i < ids.size(); ++i) {
-                id = ids[i];
-
-                ++n_predict;
-
-                if (llama_token_is_eog(model_tgt, id)) {
-                    has_eos = true;
-                    break;
-                }
-
-                token_str = common_token_to_piece(ctx_tgt, id);
+        for (size_t i = 0; i < ids.size(); ++i) {
+            prompt_tgt.push_back(id_last);
 
-                if (params.use_color && i + 1 < ids.size()) {
-                    LOG("\u001b[%dm%s\u001b[37m", (36 - 0 % 6), token_str.c_str());
-                } else {
-                    LOG("%s", token_str.c_str());
-                }
-            }
+            id_last = ids[i];
 
-            if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
+            if (llama_token_is_eog(model_tgt, id_last)) {
+                has_eos = true;
                 break;
             }
 
-            LOG_DBG("accepted %d/%d draft tokens, the last target token is: (%d, '%s')\n", (int) ids.size() - 1, (int) draft.size(), id, token_str.c_str());
+            const std::string token_str = common_token_to_piece(ctx_tgt, id_last);
 
-            {
-                LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
-
-                llama_kv_cache_seq_rm(ctx_tgt, 0, n_past, -1);
+            if (params.use_color && i + 1 < ids.size()) {
+                LOG("\u001b[%dm%s\u001b[37m", (36 - 0 % 6), token_str.c_str());
+            } else {
+                LOG("%s", token_str.c_str());
             }
+        }
 
-            prompt_tgt.push_back(id_last);
-            prompt_tgt.insert(prompt_tgt.end(), ids.begin(), ids.end() - 1);
+        LOG_DBG("accepted %d/%d draft tokens, the last target token is: (%d)\n", (int) ids.size() - 1, (int) draft.size(), id_last);
+
+        {
+            LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
+
+            llama_kv_cache_seq_rm(ctx_tgt, 0, n_past, -1);
+        }
 
-            // remember the last accepted token for the next iteration
-            id_last = id;
+        if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
+            break;
         }
     }
 

From 84e1c33cde9e0a7aafcda2d4f21ba51c300482d7 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Tue, 26 Nov 2024 13:36:40 +0200
Subject: [PATCH 080/102] server : fix parallel speculative decoding (#10513)

ggml-ci
---
 examples/server/server.cpp | 63 +++++++++++++++++++-------------------
 1 file changed, 31 insertions(+), 32 deletions(-)

diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index c0ea4faf77d42..9c86407c28eba 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -2267,50 +2267,49 @@ struct server_context {
                     continue; // continue loop of slots
                 }
 
-                llama_token id;
+                llama_token id = common_sampler_sample(slot.smpl, ctx, slot.i_batch - i);
 
-                {
-                    completion_token_output result;
-
-                    id = common_sampler_sample(slot.smpl, ctx, slot.i_batch - i);
+                slot.i_batch = -1;
 
-                    slot.i_batch = -1;
+                common_sampler_accept(slot.smpl, id, true);
 
-                    common_sampler_accept(slot.smpl, id, true);
-
-                    slot.n_decoded += 1;
-                    if (slot.n_decoded == 1) {
-                        slot.t_start_generation = ggml_time_us();
-                        slot.t_prompt_processing = (slot.t_start_generation - slot.t_start_process_prompt) / 1e3;
-                        metrics.on_prompt_eval(slot);
-                    }
+                slot.n_decoded += 1;
+                if (slot.n_decoded == 1) {
+                    slot.t_start_generation = ggml_time_us();
+                    slot.t_prompt_processing = (slot.t_start_generation - slot.t_start_process_prompt) / 1e3;
+                    metrics.on_prompt_eval(slot);
+                }
 
-                    result.tok = id;
+                completion_token_output result;
+                result.tok = id;
 
-                    const auto * cur_p = common_sampler_get_candidates(slot.smpl);
+                const auto * cur_p = common_sampler_get_candidates(slot.smpl);
 
-                    for (size_t i = 0; i < (size_t) slot.params.sampling.n_probs; ++i) {
-                        result.probs.push_back({
-                            cur_p->data[i].id,
-                                i >= cur_p->size ? 0.0f : cur_p->data[i].p,
-                        });
-                    }
+                for (size_t i = 0; i < (size_t) slot.params.sampling.n_probs; ++i) {
+                    result.probs.push_back({
+                        cur_p->data[i].id,
+                            i >= cur_p->size ? 0.0f : cur_p->data[i].p,
+                    });
+                }
 
-                    if (!process_token(result, slot)) {
-                        // release slot because of stop condition
-                        slot.release();
-                        slot.print_timings();
-                        send_final_response(slot);
-                        metrics.on_prediction(slot);
-                        continue;
-                    }
+                if (!process_token(result, slot)) {
+                    // release slot because of stop condition
+                    slot.release();
+                    slot.print_timings();
+                    send_final_response(slot);
+                    metrics.on_prediction(slot);
+                    continue;
                 }
+            }
 
-                // check if the slot supports speculative decoding
-                if (!slot.can_speculate()) {
+            // do speculative decoding
+            for (auto & slot : slots) {
+                if (!slot.is_processing() || !slot.can_speculate()) {
                     continue;
                 }
 
+                llama_token id = slot.sampled;
+
                 struct common_speculative_params params_spec;
                 params_spec.n_draft   = slot.params.speculative.n_max;
                 params_spec.n_reuse   = llama_n_ctx(slot.ctx_dft) - slot.params.speculative.n_max;

From 25669aa92caaddff09f39b54a5173e5cb2680fa3 Mon Sep 17 00:00:00 2001
From: Charles Xu <charles.xu@arm.com>
Date: Tue, 26 Nov 2024 12:37:05 +0100
Subject: [PATCH 081/102] ggml-cpu: cmake add arm64 cpu feature check for macos
 (#10487)

* ggml-cpu: cmake add arm64 cpu feature check for macos

* use vmmlaq_s32 for compile option i8mm check
---
 ggml/src/ggml-cpu/CMakeLists.txt | 33 ++++++++++++++++++++++++++++++++
 1 file changed, 33 insertions(+)

diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt
index c2905d1fbf4e8..ddc05ecef6f5c 100644
--- a/ggml/src/ggml-cpu/CMakeLists.txt
+++ b/ggml/src/ggml-cpu/CMakeLists.txt
@@ -96,6 +96,39 @@ if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
         endif ()
 
         set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
+    elseif (APPLE)
+        if (GGML_NATIVE)
+            set(USER_PROVIDED_MARCH FALSE)
+            foreach(flag_var IN ITEMS CMAKE_C_FLAGS CMAKE_CXX_FLAGS CMAKE_REQUIRED_FLAGS)
+                if ("${${flag_var}}" MATCHES "-march=[a-zA-Z0-9+._-]+")
+                    set(USER_PROVIDED_MARCH TRUE)
+                    break()
+                endif()
+            endforeach()
+
+            if (NOT USER_PROVIDED_MARCH)
+                set(MARCH_FLAGS "-march=armv8.2a")
+
+                check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+                if (GGML_COMPILER_SUPPORT_DOTPROD)
+                    set(MARCH_FLAGS "${MARCH_FLAGS}+dotprod")
+                    add_compile_definitions(__ARM_FEATURE_DOTPROD)
+                endif ()
+
+                set(TEST_I8MM_FLAGS "-march=armv8.2a+i8mm")
+
+                set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+                set(CMAKE_REQUIRED_FLAGS "${CMAKE_REQUIRED_FLAGS} ${TEST_I8MM_FLAGS}")
+                check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+                if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+                    set(MARCH_FLAGS "${MARCH_FLAGS}+i8mm")
+                    add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
+                endif ()
+                set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
+
+                list(APPEND ARCH_FLAGS "${MARCH_FLAGS}")
+            endif ()
+        endif ()
     else()
         check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
         if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")

From c6807b3f28cc3dbfda3ec390bcb87e69fb5634e2 Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Tue, 26 Nov 2024 13:05:07 +0100
Subject: [PATCH 082/102] ci : add ubuntu cuda build, build with one arch on
 windows (#10456)

---
 .github/labeler.yml               | 15 +++-----
 .github/workflows/build.yml       | 59 ++++++++++++++++++++++++++++++-
 .github/workflows/nix-ci.yml      |  2 ++
 .github/workflows/python-lint.yml |  9 ++++-
 4 files changed, 73 insertions(+), 12 deletions(-)

diff --git a/.github/labeler.yml b/.github/labeler.yml
index 89436740d1ffb..1b47bc96885c4 100644
--- a/.github/labeler.yml
+++ b/.github/labeler.yml
@@ -3,19 +3,18 @@ Kompute:
     - changed-files:
         - any-glob-to-any-file:
             - ggml/include/ggml-kompute.h
-            - ggml/src/ggml-kompute.cpp
+            - ggml/src/ggml-kompute/**
             - README-kompute.md
 Apple Metal:
     - changed-files:
         - any-glob-to-any-file:
             - ggml/include/ggml-metal.h
-            - ggml/src/ggml-metal.cpp
+            - ggml/src/ggml-metal/**
             - README-metal.md
 SYCL:
     - changed-files:
         - any-glob-to-any-file:
             - ggml/include/ggml-sycl.h
-            - ggml/src/ggml-sycl.cpp
             - ggml/src/ggml-sycl/**
             - docs/backend/SYCL.md
             - examples/sycl/**
@@ -27,8 +26,8 @@ Nvidia GPU:
 Vulkan:
     - changed-files:
         - any-glob-to-any-file:
-            - ggml/ggml_vk_generate_shaders.py
-            - ggml/src/ggml-vulkan*
+            - ggml/include/ggml-vulkan.h
+            - ggml/src/ggml-vulkan/**
 documentation:
     - changed-files:
         - any-glob-to-any-file:
@@ -75,11 +74,7 @@ server:
 ggml:
     - changed-files:
         - any-glob-to-any-file:
-            - ggml/include/ggml*.h
-            - ggml/src/ggml*.c
-            - ggml/src/ggml*.cpp
-            - ggml/src/ggml*.h
-            - ggml-cuda/**
+            - ggml/**
 nix:
     - changed-files:
         - any-glob-to-any-file:
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index abaf2c504c8f4..6281663ecb347 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -871,8 +871,65 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
           name: llama-bin-win-${{ matrix.build }}.zip
 
+  ubuntu-latest-cmake-cuda:
+    runs-on: ubuntu-latest
+    container: nvidia/cuda:12.6.2-devel-ubuntu24.04
+
+    steps:
+        - name: Clone
+          id: checkout
+          uses: actions/checkout@v4
+
+        - name: Install dependencies
+          env:
+            DEBIAN_FRONTEND: noninteractive
+          run: |
+              apt update
+              apt install -y cmake build-essential ninja-build libgomp1 git
+
+        - name: Build with CMake
+          run: |
+            cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=89-real -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined -DLLAMA_FATAL_WARNINGS=ON
+            cmake --build build
+
   windows-latest-cmake-cuda:
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        cuda: ['12.6.2']
+        build: ['cuda']
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Install CUDA toolkit
+        id: cuda-toolkit
+        uses: Jimver/cuda-toolkit@v0.2.19
+        with:
+          cuda: ${{ matrix.cuda }}
+          method: 'network'
+          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
+
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
+      - name: Build
+        id: cmake_build
+        shell: cmd
+        run: |
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON -DCMAKE_CUDA_ARCHITECTURES=89-real
+          cmake --build build --config Release -t ggml-cuda
+          cmake --build build --config Release
+
+  windows-2019-cmake-cuda:
     runs-on: windows-2019
+    if: ${{ github.event == 'push' && github.ref == 'refs/heads/master' }}
 
     strategy:
       matrix:
@@ -1173,7 +1230,7 @@ jobs:
       - macOS-latest-make
       - macOS-latest-cmake
       - windows-latest-cmake
-      - windows-latest-cmake-cuda
+      - windows-2019-cmake-cuda
       - windows-latest-cmake-hip-release
       - macOS-latest-cmake-arm64
       - macOS-latest-cmake-x64
diff --git a/.github/workflows/nix-ci.yml b/.github/workflows/nix-ci.yml
index 8ecbbe53b4ed1..3fe94157644ad 100644
--- a/.github/workflows/nix-ci.yml
+++ b/.github/workflows/nix-ci.yml
@@ -5,8 +5,10 @@ on:
   push:
     branches:
       - master
+    paths: ['.github/workflows/nix-ci.yml', '**/flake.nix', '**/flake.lock', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
   pull_request:
     types: [opened, synchronize, reopened]
+    paths: ['.github/workflows/nix-ci.yml', '**/flake.nix', '**/flake.lock', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
diff --git a/.github/workflows/python-lint.yml b/.github/workflows/python-lint.yml
index a8d46f31dd4f5..ddfdf73b8fce2 100644
--- a/.github/workflows/python-lint.yml
+++ b/.github/workflows/python-lint.yml
@@ -1,6 +1,13 @@
 name: flake8 Lint
 
-on: [push, pull_request]
+on:
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/python-lint.yml', '**/*.py']
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: ['.github/workflows/python-lint.yml', '**/*.py']
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}

From 7db3846a94ce7683b3e8120abe427457edf840c9 Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Tue, 26 Nov 2024 13:05:20 +0100
Subject: [PATCH 083/102] ci : publish the docker images created during
 scheduled runs (#10515)

---
 .github/workflows/docker.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml
index 9cef283d946e1..bc2e5020de25c 100644
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -114,7 +114,7 @@ jobs:
           swap-storage: true
 
       - name: Build and push Docker image (tagged + versioned)
-        if: github.event_name == 'push'
+        if: ${{ github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
         uses: docker/build-push-action@v6
         with:
           context: .

From ab96610b1e58684bc5e8b810130c4cf6d8252e21 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Tue, 26 Nov 2024 14:18:08 +0200
Subject: [PATCH 084/102] cmake : enable warnings in llama (#10474)

* cmake : enable warnings in llama

ggml-ci

* cmake : add llama_get_flags and respect LLAMA_FATAL_WARNINGS

* cmake : get_flags -> ggml_get_flags

* speculative-simple : fix warnings

* cmake : reuse ggml_get_flags

ggml-ci

* speculative-simple : fix compile warning

ggml-ci
---
 CMakeLists.txt                                |  1 +
 cmake/common.cmake                            | 33 +++++++++++++++++++
 common/CMakeLists.txt                         |  2 ++
 examples/CMakeLists.txt                       |  4 +++
 .../speculative-simple/speculative-simple.cpp |  6 ++--
 ggml/src/CMakeLists.txt                       |  5 +--
 ggml/src/ggml-cuda/CMakeLists.txt             |  2 +-
 src/CMakeLists.txt                            |  2 ++
 8 files changed, 49 insertions(+), 6 deletions(-)
 create mode 100644 cmake/common.cmake

diff --git a/CMakeLists.txt b/CMakeLists.txt
index e7d91a5b5c9ac..0d389dccbfb42 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -82,6 +82,7 @@ option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
 
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
 
 # override ggml options
 set(GGML_SANITIZE_THREAD    ${LLAMA_SANITIZE_THREAD})
diff --git a/cmake/common.cmake b/cmake/common.cmake
new file mode 100644
index 0000000000000..0f54871e4143d
--- /dev/null
+++ b/cmake/common.cmake
@@ -0,0 +1,33 @@
+function(llama_add_compile_flags)
+    if (LLAMA_FATAL_WARNINGS)
+        if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+            list(APPEND C_FLAGS   -Werror)
+            list(APPEND CXX_FLAGS -Werror)
+        elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+            add_compile_options(/WX)
+        endif()
+    endif()
+
+    if (LLAMA_ALL_WARNINGS)
+        if (NOT MSVC)
+            list(APPEND C_FLAGS -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
+                                -Werror=implicit-int -Werror=implicit-function-declaration)
+
+            list(APPEND CXX_FLAGS -Wmissing-declarations -Wmissing-noreturn)
+
+            list(APPEND WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+
+            list(APPEND C_FLAGS   ${WARNING_FLAGS})
+            list(APPEND CXX_FLAGS ${WARNING_FLAGS})
+
+            ggml_get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
+
+            add_compile_options("$<$<COMPILE_LANGUAGE:C>:${C_FLAGS};${GF_C_FLAGS}>"
+                                "$<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS};${GF_CXX_FLAGS}>")
+        else()
+            # todo : msvc
+            set(C_FLAGS   "" PARENT_SCOPE)
+            set(CXX_FLAGS "" PARENT_SCOPE)
+        endif()
+    endif()
+endfunction()
diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt
index 62a8a7db5652f..2231748848bf5 100644
--- a/common/CMakeLists.txt
+++ b/common/CMakeLists.txt
@@ -2,6 +2,8 @@
 
 find_package(Threads REQUIRED)
 
+llama_add_compile_flags()
+
 # Build info header
 #
 
diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt
index 21db1f3c2f199..9210e9fea5eb3 100644
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -6,6 +6,10 @@ find_package(Threads REQUIRED)
 
 # ...
 
+# flags
+
+llama_add_compile_flags()
+
 # examples
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
diff --git a/examples/speculative-simple/speculative-simple.cpp b/examples/speculative-simple/speculative-simple.cpp
index 2ea49d47c433e..8ca84f7afacbc 100644
--- a/examples/speculative-simple/speculative-simple.cpp
+++ b/examples/speculative-simple/speculative-simple.cpp
@@ -70,13 +70,13 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> inp;
     inp = common_tokenize(ctx_tgt, params.prompt, true, true);
 
-    if (llama_n_ctx(ctx_tgt) < (int) inp.size()) {
+    if (llama_n_ctx(ctx_tgt) < (uint32_t) inp.size()) {
         LOG_ERR("%s: the prompt exceeds the context size (%d tokens, ctx %d)\n", __func__, (int) inp.size(), llama_n_ctx(ctx_tgt));
 
         return 1;
     }
 
-    if (llama_n_batch(ctx_tgt) < (int) inp.size()) {
+    if (llama_n_batch(ctx_tgt) < (uint32_t) inp.size()) {
         LOG_ERR("%s: the prompt exceeds the batch size (%d tokens, batch %d)\n", __func__, (int) inp.size(), llama_n_batch(ctx_tgt));
 
         return 1;
@@ -155,7 +155,7 @@ int main(int argc, char ** argv) {
         // evaluate the target model on [id_last, draft0, draft1, ..., draftN-1]
         {
             // do not waste time on small drafts
-            if (draft.size() < n_draft_min) {
+            if (draft.size() < (size_t) n_draft_min) {
                 draft.clear();
             }
 
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index 071508ddae021..9022aa3ae197d 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -24,7 +24,7 @@ if (NOT MSVC)
     endif()
 endif()
 
-function(get_flags CCID CCVER)
+function(ggml_get_flags CCID CCVER)
     set(C_FLAGS "")
     set(CXX_FLAGS "")
 
@@ -41,6 +41,7 @@ function(get_flags CCID CCVER)
     elseif (CCID STREQUAL "GNU")
         set(C_FLAGS   -Wdouble-promotion)
         set(CXX_FLAGS -Wno-array-bounds)
+
         if (CCVER VERSION_GREATER_EQUAL 8.1.0)
             list(APPEND CXX_FLAGS -Wextra-semi)
         endif()
@@ -69,7 +70,7 @@ if (GGML_ALL_WARNINGS)
         list(APPEND C_FLAGS   ${WARNING_FLAGS})
         list(APPEND CXX_FLAGS ${WARNING_FLAGS})
 
-        get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
+        ggml_get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
 
         add_compile_options("$<$<COMPILE_LANGUAGE:C>:${C_FLAGS};${GF_C_FLAGS}>"
                             "$<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS};${GF_CXX_FLAGS}>")
diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt
index b0cb93e070fd3..14761650f8a8e 100644
--- a/ggml/src/ggml-cuda/CMakeLists.txt
+++ b/ggml/src/ggml-cuda/CMakeLists.txt
@@ -132,7 +132,7 @@ if (CUDAToolkit_FOUND)
 
         message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
 
-        get_flags(${CUDA_CCID} ${CUDA_CCVER})
+        ggml_get_flags(${CUDA_CCID} ${CUDA_CCVER})
         list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
     endif()
 
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index a86624750305e..2f581b9218728 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -5,6 +5,8 @@ if (WIN32)
     endif()
 endif()
 
+llama_add_compile_flags()
+
 #
 # libraries
 #

From 0bbd2262a3263f37385297b30de37941836e57f7 Mon Sep 17 00:00:00 2001
From: Neo Zhang Jianyu <jianyu.zhang@intel.com>
Date: Tue, 26 Nov 2024 21:43:47 +0800
Subject: [PATCH 085/102] restore the condistion to build & update pacakge when
 merge (#10507)

Co-authored-by: arthw <14088817+arthw@users.noreply.github.com>
---
 .github/workflows/build.yml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 6281663ecb347..c6aecec6e94d5 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -1041,7 +1041,7 @@ jobs:
 
       - name: Build the release package
         id: pack_artifacts
-        if: ${{ ( github.event_name == 'pull_request' && github.base_ref == 'master' ) }}
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
           echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
 
@@ -1066,7 +1066,7 @@ jobs:
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload the release package
-        if: ${{ ( github.event_name == 'pull_request' && github.base_ref == 'master' ) }}
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: actions/upload-artifact@v4
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip

From 45abe0f74ee281aea6e5283c1e738061256cfcae Mon Sep 17 00:00:00 2001
From: Xuan Son Nguyen <thichthat@gmail.com>
Date: Tue, 26 Nov 2024 16:20:18 +0100
Subject: [PATCH 086/102] server : replace behave with pytest (#10416)

* server : replace behave with pytest

* fix test on windows

* misc

* add more tests

* more tests

* styling

* log less, fix embd test

* added all sequential tests

* fix coding style

* fix save slot test

* add parallel completion test

* fix parallel test

* remove feature files

* update test docs

* no cache_prompt for some tests

* add test_cache_vs_nocache_prompt
---
 .devops/nix/python-scripts.nix                |    2 +-
 .github/workflows/server.yml                  |    9 +-
 examples/server/tests/.gitignore              |    1 +
 examples/server/tests/README.md               |   33 +-
 examples/server/tests/conftest.py             |   15 +
 .../server/tests/features/ctx_shift.feature   |   66 -
 .../server/tests/features/embeddings.feature  |  113 --
 examples/server/tests/features/environment.py |   71 -
 examples/server/tests/features/infill.feature |   36 -
 examples/server/tests/features/issues.feature |    5 -
 examples/server/tests/features/lora.feature   |   36 -
 .../server/tests/features/parallel.feature    |  131 --
 .../server/tests/features/passkey.feature     |   56 -
 examples/server/tests/features/rerank.feature |   42 -
 .../server/tests/features/results.feature     |  118 --
 .../server/tests/features/security.feature    |   68 -
 examples/server/tests/features/server.feature |  120 --
 .../server/tests/features/slotsave.feature    |   58 -
 examples/server/tests/features/steps/steps.py | 1518 -----------------
 .../tests/features/wrong_usages.feature       |   25 -
 examples/server/tests/requirements.txt        |    2 +-
 examples/server/tests/tests.sh                |    5 +-
 examples/server/tests/unit/test_basic.py      |   34 +
 .../server/tests/unit/test_chat_completion.py |  129 ++
 examples/server/tests/unit/test_completion.py |  223 +++
 examples/server/tests/unit/test_ctx_shift.py  |   67 +
 examples/server/tests/unit/test_embedding.py  |   99 ++
 examples/server/tests/unit/test_infill.py     |   35 +
 examples/server/tests/unit/test_lora.py       |   42 +
 examples/server/tests/unit/test_rerank.py     |   38 +
 examples/server/tests/unit/test_security.py   |   83 +
 examples/server/tests/unit/test_slot_save.py  |   98 ++
 examples/server/tests/unit/test_tokenize.py   |   59 +
 examples/server/tests/utils.py                |  377 ++++
 34 files changed, 1317 insertions(+), 2497 deletions(-)
 create mode 100644 examples/server/tests/conftest.py
 delete mode 100644 examples/server/tests/features/ctx_shift.feature
 delete mode 100644 examples/server/tests/features/embeddings.feature
 delete mode 100644 examples/server/tests/features/environment.py
 delete mode 100644 examples/server/tests/features/infill.feature
 delete mode 100644 examples/server/tests/features/issues.feature
 delete mode 100644 examples/server/tests/features/lora.feature
 delete mode 100644 examples/server/tests/features/parallel.feature
 delete mode 100644 examples/server/tests/features/passkey.feature
 delete mode 100644 examples/server/tests/features/rerank.feature
 delete mode 100644 examples/server/tests/features/results.feature
 delete mode 100644 examples/server/tests/features/security.feature
 delete mode 100644 examples/server/tests/features/server.feature
 delete mode 100644 examples/server/tests/features/slotsave.feature
 delete mode 100644 examples/server/tests/features/steps/steps.py
 delete mode 100644 examples/server/tests/features/wrong_usages.feature
 create mode 100644 examples/server/tests/unit/test_basic.py
 create mode 100644 examples/server/tests/unit/test_chat_completion.py
 create mode 100644 examples/server/tests/unit/test_completion.py
 create mode 100644 examples/server/tests/unit/test_ctx_shift.py
 create mode 100644 examples/server/tests/unit/test_embedding.py
 create mode 100644 examples/server/tests/unit/test_infill.py
 create mode 100644 examples/server/tests/unit/test_lora.py
 create mode 100644 examples/server/tests/unit/test_rerank.py
 create mode 100644 examples/server/tests/unit/test_security.py
 create mode 100644 examples/server/tests/unit/test_slot_save.py
 create mode 100644 examples/server/tests/unit/test_tokenize.py
 create mode 100644 examples/server/tests/utils.py

diff --git a/.devops/nix/python-scripts.nix b/.devops/nix/python-scripts.nix
index 392e9ffe41bf5..56ea182788764 100644
--- a/.devops/nix/python-scripts.nix
+++ b/.devops/nix/python-scripts.nix
@@ -34,7 +34,7 @@ let
 
     # server tests
     openai
-    behave
+    pytest
     prometheus-client
   ];
 in
diff --git a/.github/workflows/server.yml b/.github/workflows/server.yml
index 699ac095d6c83..2e8e3348f4292 100644
--- a/.github/workflows/server.yml
+++ b/.github/workflows/server.yml
@@ -122,14 +122,14 @@ jobs:
         id: server_integration_tests
         run: |
           cd examples/server/tests
-          PORT=8888 ./tests.sh
+          ./tests.sh
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
           cd examples/server/tests
-          PORT=8888 ./tests.sh --stop --no-skipped --no-capture --tags slow
+          SLOW_TESTS=1 ./tests.sh
 
 
   server-windows:
@@ -180,11 +180,12 @@ jobs:
         run: |
           cd examples/server/tests
           $env:PYTHONIOENCODING = ":replace"
-          behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
+          pytest -v -x
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
           cd examples/server/tests
-          behave.exe --stop --no-skipped --no-capture --tags slow
+          $env:SLOW_TESTS = "1"
+          pytest -v -x
diff --git a/examples/server/tests/.gitignore b/examples/server/tests/.gitignore
index 1d17dae13b53a..90ee7fe6d971a 100644
--- a/examples/server/tests/.gitignore
+++ b/examples/server/tests/.gitignore
@@ -1 +1,2 @@
 .venv
+tmp
diff --git a/examples/server/tests/README.md b/examples/server/tests/README.md
index 10f22c4471ea7..2930a2e0dea0f 100644
--- a/examples/server/tests/README.md
+++ b/examples/server/tests/README.md
@@ -1,19 +1,9 @@
 # Server tests
 
-Python based server tests scenario using [BDD](https://en.wikipedia.org/wiki/Behavior-driven_development)
-and [behave](https://behave.readthedocs.io/en/latest/):
-
-* [issues.feature](./features/issues.feature) Pending issues scenario
-* [parallel.feature](./features/parallel.feature) Scenario involving multi slots and concurrent requests
-* [security.feature](./features/security.feature) Security, CORS and API Key
-* [server.feature](./features/server.feature) Server base scenario: completion, embedding, tokenization, etc...
+Python based server tests scenario using [pytest](https://docs.pytest.org/en/stable/).
 
 Tests target GitHub workflows job runners with 4 vCPU.
 
-Requests are
-using [aiohttp](https://docs.aiohttp.org/en/stable/client_reference.html), [asyncio](https://docs.python.org/fr/3/library/asyncio.html)
-based http client.
-
 Note: If the host architecture inference speed is faster than GitHub runners one, parallel scenario may randomly fail.
 To mitigate it, you can increase values in `n_predict`, `kv_size`.
 
@@ -39,26 +29,19 @@ It's possible to override some scenario steps values with environment variables:
 |--------------------------|------------------------------------------------------------------------------------------------|
 | `PORT`                   | `context.server_port` to set the listening port of the server during scenario, default: `8080` |
 | `LLAMA_SERVER_BIN_PATH`  | to change the server binary path, default: `../../../build/bin/llama-server`                         |
-| `DEBUG`                  | "ON" to enable steps and server verbose mode `--verbose`                                       |
+| `DEBUG`                  | to enable steps and server verbose mode `--verbose`                                       |
 | `N_GPU_LAYERS`           | number of model layers to offload to VRAM `-ngl --n-gpu-layers`                                |
 
-### Run @bug, @wip or @wrong_usage annotated scenario
-
-Feature or Scenario must be annotated with `@llama.cpp` to be included in the default scope.
-
-- `@bug` annotation aims to link a scenario with a GitHub issue.
-- `@wrong_usage` are meant to show user issue that are actually an expected behavior
-- `@wip` to focus on a scenario working in progress
-- `@slow` heavy test, disabled by default
-
-To run a scenario annotated with `@bug`, start:
+To run slow tests:
 
 ```shell
-DEBUG=ON ./tests.sh --no-skipped --tags bug --stop
+SLOW_TESTS=1 ./tests.sh
 ```
 
-After changing logic in `steps.py`, ensure that `@bug` and `@wrong_usage` scenario are updated.
+To run with stdout/stderr display in real time (verbose output, but useful for debugging):
 
 ```shell
-./tests.sh --no-skipped --tags bug,wrong_usage || echo "should failed but compile"
+DEBUG=1 ./tests.sh -s -v -x
 ```
+
+To see all available arguments, please refer to [pytest documentation](https://docs.pytest.org/en/stable/how-to/usage.html)
diff --git a/examples/server/tests/conftest.py b/examples/server/tests/conftest.py
new file mode 100644
index 0000000000000..017d1bb841efd
--- /dev/null
+++ b/examples/server/tests/conftest.py
@@ -0,0 +1,15 @@
+import pytest
+from utils import *
+
+
+# ref: https://stackoverflow.com/questions/22627659/run-code-before-and-after-each-test-in-py-test
+@pytest.fixture(autouse=True)
+def stop_server_after_each_test():
+    # do nothing before each test
+    yield
+    # stop all servers after each test
+    instances = set(
+        server_instances
+    )  # copy the set to prevent 'Set changed size during iteration'
+    for server in instances:
+        server.stop()
diff --git a/examples/server/tests/features/ctx_shift.feature b/examples/server/tests/features/ctx_shift.feature
deleted file mode 100644
index ae6c6b01b0221..0000000000000
--- a/examples/server/tests/features/ctx_shift.feature
+++ /dev/null
@@ -1,66 +0,0 @@
-@llama.cpp
-@ctx_shift
-Feature: llama.cpp server
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
-    And   a model file test-model.gguf
-    And   a model alias tinyllama-2
-    And   BOS token is 1
-    And   42 as server seed
-    And   256 KV cache size
-    And   32 as batch size
-    And   2 slots
-
-    # the prompt is 301 tokens
-    # the slot context is 256/2 = 128 tokens
-    # the prompt is truncated to keep the last 109 tokens
-    # 64 tokens are generated thanks to shifting the context when it gets full
-  Scenario: Inference with context shift
-    And   64 server max tokens to predict
-    Then  the server is starting
-    Then  the server is healthy
-    Given a prompt:
-    """
-    Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
-    Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.
-    Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
-    Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
-    """
-    And   a completion request with no api error
-    Then  64 tokens are predicted matching fun|Annaks|popcorns|pictry|bowl
-    And   the completion is  truncated
-    And   109 prompt tokens are processed
-
-  Scenario Outline: Inference without context shift
-    And   <n_predict> server max tokens to predict
-    And   disable context shifting
-    Then  the server is starting
-    Then  the server is healthy
-    Given a prompt:
-    """
-    Hi how are you
-    """
-    And   a completion request with no api error
-    Then  <n_token_output> tokens are predicted matching twind|Anna
-    And   the completion is <truncated> truncated
-    And   8 prompt tokens are processed
-    Examples:
-      | n_predict | n_token_output | truncated |
-      | 64        | 64             | not       |
-      | -1        | 120            |           |
-
-  Scenario: Inference without context shift (expected error: prompt too long)
-    And   disable context shifting
-    Then  the server is starting
-    Then  the server is healthy
-    Given a prompt:
-    """
-    Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
-    Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.
-    Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
-    Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
-    """
-    And   a completion request with 400 api error
-
diff --git a/examples/server/tests/features/embeddings.feature b/examples/server/tests/features/embeddings.feature
deleted file mode 100644
index f4fe2ee4335ff..0000000000000
--- a/examples/server/tests/features/embeddings.feature
+++ /dev/null
@@ -1,113 +0,0 @@
-@llama.cpp
-@embeddings
-Feature: llama.cpp server
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model url https://huggingface.co/ggml-org/models/resolve/main/bert-bge-small/ggml-model-f16.gguf
-    And   a model file bert-bge-small.gguf
-    And   a model alias bert-bge-small
-    And   42 as server seed
-    And   2 slots
-    # the bert-bge-small model has context size of 512
-    # since the generated prompts are as big as the batch size, we need to set the batch size to <= 512
-    # ref: https://huggingface.co/BAAI/bge-small-en-v1.5/blob/5c38ec7c405ec4b44b94cc5a9bb96e735b38267a/config.json#L20
-    And   128 as batch size
-    And   128 as ubatch size
-    And   512 KV cache size
-    And   enable embeddings endpoint
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario: Embedding
-    When embeddings are computed for:
-    """
-    What is the capital of Bulgaria ?
-    """
-    Then embeddings are generated
-
-  Scenario: Embedding (error: prompt too long)
-    When embeddings are computed for:
-    """
-    Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
-    Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.
-    Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
-    Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
-    Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
-    Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.
-    Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
-    Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
-    """
-    And  embeddings request with 500 api error
-
-  Scenario: OAI Embeddings compatibility
-    Given a model bert-bge-small
-    When an OAI compatible embeddings computation request for:
-    """
-    What is the capital of Spain ?
-    """
-    Then embeddings are generated
-
-  Scenario: OAI Embeddings compatibility with multiple inputs
-    Given a model bert-bge-small
-    Given a prompt:
-      """
-      In which country Paris is located ?
-      """
-    And a prompt:
-      """
-      Is Madrid the capital of Spain ?
-      """
-    When an OAI compatible embeddings computation request for multiple inputs
-    Then embeddings are generated
-
-  Scenario: Multi users embeddings
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And a prompt:
-      """
-      Write a very long poem.
-      """
-    And a prompt:
-      """
-      Write a very long joke.
-      """
-    Given concurrent embedding requests
-    Then the server is busy
-    Then the server is idle
-    Then all embeddings are generated
-
-  Scenario: Multi users OAI compatibility embeddings
-    Given a prompt:
-      """
-      In which country Paris is located ?
-      """
-    And a prompt:
-      """
-      Is Madrid the capital of Spain ?
-      """
-    And a prompt:
-      """
-      What is the biggest US city ?
-      """
-    And a prompt:
-      """
-      What is the capital of Bulgaria ?
-      """
-    And   a model bert-bge-small
-    Given concurrent OAI embedding requests
-    Then the server is busy
-    Then the server is idle
-    Then all embeddings are generated
-
-  Scenario: All embeddings should be the same
-    Given 10 fixed prompts
-    And   a model bert-bge-small
-    Given concurrent OAI embedding requests
-    Then all embeddings are the same
diff --git a/examples/server/tests/features/environment.py b/examples/server/tests/features/environment.py
deleted file mode 100644
index e7845dc2f51fc..0000000000000
--- a/examples/server/tests/features/environment.py
+++ /dev/null
@@ -1,71 +0,0 @@
-import os
-import signal
-import socket
-import sys
-import time
-import traceback
-from contextlib import closing
-from subprocess import TimeoutExpired
-
-
-def before_scenario(context, scenario):
-    context.debug = 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON'
-    if context.debug:
-        print("DEBUG=ON")
-    print(f"\x1b[33;42mStarting new scenario: {scenario.name}!\x1b[0m")
-    port = 8080
-    if 'PORT' in os.environ:
-        port = int(os.environ['PORT'])
-    if is_server_listening("localhost", port):
-        assert False, "Server already started"
-
-
-def after_scenario(context, scenario):
-    try:
-        if 'server_process' not in context or context.server_process is None:
-            return
-        if scenario.status == "failed":
-            if 'GITHUB_ACTIONS' in os.environ:
-                print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n")
-                if os.path.isfile('llama.log'):
-                    with closing(open('llama.log', 'r')) as f:
-                        for line in f:
-                            print(line)
-            if not is_server_listening(context.server_fqdn, context.server_port):
-                print("\x1b[33;101mERROR: Server stopped listening\x1b[0m")
-
-        if context.server_process.poll() is not None:
-            assert False, f"Server not running pid={context.server_process.pid} ..."
-
-        server_graceful_shutdown(context)  # SIGINT
-
-        try:
-            context.server_process.wait(0.5)
-        except TimeoutExpired:
-            print(f"server still alive after 500ms, force-killing pid={context.server_process.pid} ...")
-            context.server_process.kill()  # SIGKILL
-            context.server_process.wait()
-
-        while is_server_listening(context.server_fqdn, context.server_port):
-            time.sleep(0.1)
-    except Exception:
-        print("ignoring error in after_scenario:")
-        traceback.print_exc(file=sys.stdout)
-
-
-def server_graceful_shutdown(context):
-    print(f"shutting down server pid={context.server_process.pid} ...")
-    if os.name == 'nt':
-        interrupt = signal.CTRL_C_EVENT
-    else:
-        interrupt = signal.SIGINT
-    context.server_process.send_signal(interrupt)
-
-
-def is_server_listening(server_fqdn, server_port):
-    with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
-        result = sock.connect_ex((server_fqdn, server_port))
-        _is_server_listening = result == 0
-        if _is_server_listening:
-            print(f"server is listening on {server_fqdn}:{server_port}...")
-        return _is_server_listening
diff --git a/examples/server/tests/features/infill.feature b/examples/server/tests/features/infill.feature
deleted file mode 100644
index a0bbfef77707b..0000000000000
--- a/examples/server/tests/features/infill.feature
+++ /dev/null
@@ -1,36 +0,0 @@
-@llama.cpp
-@infill
-Feature: llama.cpp server
-
-  # The current model is made by adding FIM tokens to the existing stories260K
-  # We may want to use a better model in the future, maybe something like SmolLM 360M
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K-infill.gguf from HF repo ggml-org/models
-    And   a model file test-model-infill.gguf
-    And   a model alias tinyllama-infill
-    And   42 as server seed
-    And   1024 as batch size
-    And   1024 as ubatch size
-    And   2048 KV cache size
-    And   64 max tokens to predict
-    And   0.0 temperature
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario: Infill without input_extra
-    Given a prompt "Complete this"
-    And   an infill input extra none none
-    And   an infill input prefix "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_"
-    And   an infill input suffix "}\n"
-    And   an infill request with no api error
-    Then  64 tokens are predicted matching One|day|she|saw|big|scary|bird
-
-  Scenario: Infill with input_extra
-    Given a prompt "Complete this"
-    And   an infill input extra "llama.h" "LLAMA_API int32_t llama_n_threads();\n"
-    And   an infill input prefix "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_"
-    And   an infill input suffix "}\n"
-    And   an infill request with no api error
-    Then  64 tokens are predicted matching cuts|Jimmy|mom|came|into|the|room"
diff --git a/examples/server/tests/features/issues.feature b/examples/server/tests/features/issues.feature
deleted file mode 100644
index 7b13e44cad395..0000000000000
--- a/examples/server/tests/features/issues.feature
+++ /dev/null
@@ -1,5 +0,0 @@
-# List of ongoing issues
-# run with: DEBUG=ON ./tests.sh --no-skipped --tags bug
-@bug
-Feature: Issues
-  # No confirmed issue at the moment
diff --git a/examples/server/tests/features/lora.feature b/examples/server/tests/features/lora.feature
deleted file mode 100644
index 7b85988ac6e87..0000000000000
--- a/examples/server/tests/features/lora.feature
+++ /dev/null
@@ -1,36 +0,0 @@
-@llama.cpp
-@lora
-Feature: llama.cpp server
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model url https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/stories15M_MOE-F16.gguf
-    And   a model file stories15M_MOE-F16.gguf
-    And   a model alias stories15M_MOE
-    And   a lora adapter file from https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe_shakespeare15M.gguf
-    And   42 as server seed
-    And   1024 as batch size
-    And   1024 as ubatch size
-    And   2048 KV cache size
-    And   64 max tokens to predict
-    And   0.0 temperature
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario: Completion LoRA disabled
-    Given switch off lora adapter 0
-    Given a prompt:
-    """
-    Look in thy glass
-    """
-    And   a completion request with no api error
-    Then  64 tokens are predicted matching little|girl|three|years|old
-
-  Scenario: Completion LoRA enabled
-    Given switch on lora adapter 0
-    Given a prompt:
-    """
-    Look in thy glass
-    """
-    And   a completion request with no api error
-    Then  64 tokens are predicted matching eye|love|glass|sun
diff --git a/examples/server/tests/features/parallel.feature b/examples/server/tests/features/parallel.feature
deleted file mode 100644
index 423d0f1d42f55..0000000000000
--- a/examples/server/tests/features/parallel.feature
+++ /dev/null
@@ -1,131 +0,0 @@
-@llama.cpp
-@parallel
-Feature: Parallel
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/split/stories15M-00001-of-00003.gguf from HF repo ggml-org/models
-    And   a model file test-model-00001-of-00003.gguf
-    And   42 as server seed
-    And   128 as batch size
-    And   256 KV cache size
-    And   2 slots
-    And   continuous batching
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario Outline: Multi users completion
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And <n_predict> max tokens to predict
-    Given concurrent completion requests
-    Then the server is busy
-    Then the server is idle
-    And  all slots are idle
-    Then all prompts are predicted with <n_predict> tokens
-    Examples:
-      | n_predict |
-      | 128       |
-
-  Scenario Outline: Multi users OAI completions compatibility
-    Given a system prompt You are a writer.
-    And   a model tinyllama-2
-    Given a prompt:
-      """
-      Write a very long book.
-      """
-    And a prompt:
-      """
-      Write another a poem.
-      """
-    And <n_predict> max tokens to predict
-    And streaming is <streaming>
-    Given concurrent OAI completions requests
-    Then the server is busy
-    Then the server is idle
-    Then all prompts are predicted with <n_predict> tokens
-    Examples:
-      | streaming | n_predict |
-      | disabled  | 128       |
-      | enabled   | 64        |
-
-  Scenario Outline: Multi users OAI completions compatibility no v1
-    Given a system prompt You are a writer.
-    And   a model tinyllama-2
-    Given a prompt:
-      """
-      Write a very long book.
-      """
-    And a prompt:
-      """
-      Write another a poem.
-      """
-    And <n_predict> max tokens to predict
-    And streaming is <streaming>
-    Given concurrent OAI completions requests no v1
-    Then the server is busy
-    Then the server is idle
-    Then all prompts are predicted with <n_predict> tokens
-    Examples:
-      | streaming | n_predict |
-      | disabled  | 128       |
-      | enabled   | 64        |
-
-  Scenario Outline: Multi users with number of prompts exceeding number of slots
-    Given a system prompt You are a writer.
-    And   a model tinyllama-2
-    Given a prompt:
-      """
-      Write a very long book.
-      """
-    And a prompt:
-      """
-      Write another a poem.
-      """
-    And a prompt:
-      """
-      What is LLM?
-      """
-    And a prompt:
-      """
-      The sky is blue and I love it.
-      """
-    And <n_predict> max tokens to predict
-    And streaming is <streaming>
-    Given concurrent OAI completions requests
-    Then the server is busy
-    Then the server is idle
-    Then all prompts are predicted with <n_predict> tokens
-    Examples:
-      | streaming | n_predict |
-      | disabled  | 128       |
-      | enabled   | 64        |
-
-  Scenario:  Multi users with total number of tokens to predict exceeds the KV Cache size #3969
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And a prompt:
-      """
-      Write a very long poem.
-      """
-    And a prompt:
-      """
-      Write a very long joke.
-      """
-    And 128 max tokens to predict
-    Given concurrent completion requests
-    Then the server is busy
-    Then the server is idle
-    Then all prompts are predicted
diff --git a/examples/server/tests/features/passkey.feature b/examples/server/tests/features/passkey.feature
deleted file mode 100644
index ff0a82cc46581..0000000000000
--- a/examples/server/tests/features/passkey.feature
+++ /dev/null
@@ -1,56 +0,0 @@
-# run with: ./tests.sh --no-skipped --tags passkey
-@passkey
-@slow
-Feature: Passkey / Self-extend with context shift
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-
-  # Generates a long text of junk and inserts a secret passkey number inside it.
-  # Then we query the LLM for the secret passkey.
-  # see #3856 and #4810
-  Scenario Outline: Passkey
-    Given a model file <hf_file> from HF repo <hf_repo>
-    And   <n_batch> as batch size
-    And   <n_junk> as number of junk
-    And   <n_predicted> server max tokens to predict
-    And   42 as seed
-    And   0.0 temperature
-    And   <n_ctx> KV cache size
-    And   1 slots
-    And   <n_ga> group attention factor to extend context size through self-extend
-    And   <n_ga_w> group attention width to extend context size through self-extend
-    # Can be override with N_GPU_LAYERS
-    And   <ngl> GPU offloaded layers
-    Then  the server is starting
-    # Higher timeout because the model may need to be downloaded from the internet
-    Then  the server is healthy with timeout 120 seconds
-    Given available models
-    Then  model 0 is trained on <n_ctx_train> tokens context
-    Given a prefix prompt:
-    """
-    here is an important info hidden inside a lot of irrelevant text. Find it and memorize them. I will quiz you about the important information there.
-    """
-    And a passkey prompt template:
-    """
-    The pass key is <passkey> Remember it. <passkey> is the pass key.
-    """
-    And a junk suffix prompt:
-    """
-    The grass is green. The sky is blue. The sun is yellow. Here we go. There and back again.
-    """
-    And a suffix prompt:
-    """
-    What is the pass key? The pass key is
-    """
-    Given a "<passkey>" passkey challenge prompt with the passkey inserted every <i_pos> junk
-    And  a completion request with no api error
-    Then <n_predicted> tokens are predicted matching <re_content>
-
-    Examples:
-      | hf_repo                         | hf_file                     | n_ctx_train | ngl | n_ctx | n_batch | n_ga | n_ga_w | n_junk | i_pos | passkey | n_predicted | re_content     |
-      | TheBloke/phi-2-GGUF             | phi-2.Q4_K_M.gguf           | 2048        | 5   | 8192  | 512     | 4    | 512    | 250    | 50    | 42      | 1           | 42             |
-      | TheBloke/phi-2-GGUF             | phi-2.Q4_K_M.gguf           | 2048        | 5   | 8192  | 512     | 2    | 512    | 250    | 50    | 42      | 1           | \b((?!42)\w)+\b  |
-      #| TheBloke/Llama-2-7B-GGUF        | llama-2-7b.Q2_K.gguf        | 4096        | 3   | 16384 | 512     | 4    | 512    | 500    | 300   | 1234    | 5           | 1234           |
-      #| TheBloke/Mixtral-8x7B-v0.1-GGUF | mixtral-8x7b-v0.1.Q2_K.gguf | 32768       | 2   | 16384 | 512     | 4    | 512    | 500    | 100   | 0987    | 5           | 0
-      # 987           |
diff --git a/examples/server/tests/features/rerank.feature b/examples/server/tests/features/rerank.feature
deleted file mode 100644
index c36cc8e215fa6..0000000000000
--- a/examples/server/tests/features/rerank.feature
+++ /dev/null
@@ -1,42 +0,0 @@
-@llama.cpp
-@rerank
-Feature: llama.cpp server
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model url https://huggingface.co/ggml-org/models/resolve/main/jina-reranker-v1-tiny-en/ggml-model-f16.gguf
-    And   a model file jina-reranker-v1-tiny-en.gguf
-    And   a model alias jina-reranker-v1-tiny-en
-    And   42 as server seed
-    And   2 slots
-    And   512 as batch size
-    And   512 as ubatch size
-    And   512 KV cache size
-    And   enable reranking endpoint
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario: Rerank
-    Given a rerank query:
-      """
-      Machine learning is
-      """
-    And   a rerank document:
-      """
-      A machine is a physical system that uses power to apply forces and control movement to perform an action. The term is commonly applied to artificial devices, such as those employing engines or motors, but also to natural biological macromolecules, such as molecular machines.
-      """
-    And   a rerank document:
-      """
-      Learning is the process of acquiring new understanding, knowledge, behaviors, skills, values, attitudes, and preferences. The ability to learn is possessed by humans, non-human animals, and some machines; there is also evidence for some kind of learning in certain plants.
-      """
-    And   a rerank document:
-      """
-      Machine learning is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions.
-      """
-    And   a rerank document:
-      """
-      Paris, capitale de la France, est une grande ville européenne et un centre mondial de l'art, de la mode, de la gastronomie et de la culture. Son paysage urbain du XIXe siècle est traversé par de larges boulevards et la Seine.
-      """
-    When  reranking request
-    Then  reranking results are returned
-    Then  reranking highest score is index 2 and lowest score is index 3
diff --git a/examples/server/tests/features/results.feature b/examples/server/tests/features/results.feature
deleted file mode 100644
index e8e1b54147b05..0000000000000
--- a/examples/server/tests/features/results.feature
+++ /dev/null
@@ -1,118 +0,0 @@
-@llama.cpp
-@results
-Feature: Results
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/split/stories15M-00001-of-00003.gguf from HF repo ggml-org/models
-    And   a model file test-model-00001-of-00003.gguf
-    And   128 as batch size
-    And   1024 KV cache size
-    And   128 max tokens to predict
-    And   continuous batching
-
-  Scenario Outline: consistent results with same seed
-    Given <n_slots> slots
-    And   1.0 temperature
-    Then  the server is starting
-    Then  the server is healthy
-
-    Given 4 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 42
-
-    Given concurrent completion requests
-    Then the server is busy
-    Then the server is idle
-    And  all slots are idle
-    Then all predictions are equal
-    Examples:
-      | n_slots |
-      | 1       |
-      # FIXME: unified KV cache nondeterminism
-      # | 2       |
-
-  Scenario Outline: different results with different seed
-    Given <n_slots> slots
-    And   1.0 temperature
-    Then  the server is starting
-    Then  the server is healthy
-
-    Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 42
-    Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 43
-    Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 44
-    Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 45
-
-    Given concurrent completion requests
-    Then the server is busy
-    Then the server is idle
-    And  all slots are idle
-    Then all predictions are different
-    Examples:
-      | n_slots |
-      | 1       |
-      | 2       |
-
-  Scenario Outline: consistent results with same seed and varying batch size
-    Given 4 slots
-    And   <temp> temperature
-    # And   0 as draft
-    Then  the server is starting
-    Then  the server is healthy
-
-    Given 1 prompts "Write a very long story about AI." with seed 42
-    And   concurrent completion requests
-    # Then the server is busy # Not all slots will be utilized.
-    Then  the server is idle
-    And   all slots are idle
-
-    Given <n_parallel> prompts "Write a very long story about AI." with seed 42
-    And   concurrent completion requests
-    # Then the server is busy # Not all slots will be utilized.
-    Then the server is idle
-    And  all slots are idle
-
-    Then all predictions are equal
-    Examples:
-      | n_parallel | temp |
-      | 1          | 0.0  |
-      | 1          | 1.0  |
-      # FIXME: unified KV cache nondeterminism
-      # See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227
-      # and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574
-      # and https://github.com/ggerganov/llama.cpp/pull/7347 .
-      # | 2          | 0.0  |
-      # | 4          | 0.0  |
-      # | 2          | 1.0  |
-      # | 4          | 1.0  |
-
-  Scenario Outline: consistent token probs with same seed and prompt
-    Given <n_slots> slots
-    And   <n_kv> KV cache size
-    And   1.0 temperature
-    And   <n_predict> max tokens to predict
-    Then  the server is starting
-    Then  the server is healthy
-
-    Given 1 prompts "The meaning of life is" with seed 42
-    And   concurrent completion requests
-    # Then the server is busy # Not all slots will be utilized.
-    Then  the server is idle
-    And   all slots are idle
-
-    Given <n_parallel> prompts "The meaning of life is" with seed 42
-    And   concurrent completion requests
-    # Then the server is busy # Not all slots will be utilized.
-    Then the server is idle
-    And  all slots are idle
-
-    Then all token probabilities are equal
-    Examples:
-      | n_slots | n_kv | n_predict | n_parallel |
-      | 4       | 1024 | 1         | 1          |
-      # FIXME: unified KV cache nondeterminism
-      # See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227
-      # and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574
-      # and https://github.com/ggerganov/llama.cpp/pull/7347 .
-      # | 4       | 1024 | 1         | 4          |
-      # | 4       | 1024 | 100       | 1          |
-      # This test still fails even the above patches; the first token probabilities are already different.
-      # | 4       | 1024 | 100       | 4          |
diff --git a/examples/server/tests/features/security.feature b/examples/server/tests/features/security.feature
deleted file mode 100644
index ef30007c3eddb..0000000000000
--- a/examples/server/tests/features/security.feature
+++ /dev/null
@@ -1,68 +0,0 @@
-@llama.cpp
-@security
-Feature: Security
-
-  Background: Server startup with an api key defined
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
-    And   a server api key THIS_IS_THE_KEY
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario Outline: Completion with some user api key
-    Given a prompt test
-    And   a user api key <api_key>
-    And   4 max tokens to predict
-    And   a completion request with <api_error> api error
-
-    Examples: Prompts
-      | api_key         | api_error |
-      | THIS_IS_THE_KEY | no        |
-      | THIS_IS_THE_KEY | no        |
-      | hackeme         | raised    |
-      |                 | raised    |
-
-  Scenario Outline: OAI Compatibility
-    Given a system prompt test
-    And   a user prompt test
-    And   a model test
-    And   2 max tokens to predict
-    And   streaming is disabled
-    And   a user api key <api_key>
-    Given an OAI compatible chat completions request with <api_error> api error
-
-    Examples: Prompts
-      | api_key         | api_error |
-      | THIS_IS_THE_KEY | no        |
-      | THIS_IS_THE_KEY | no        |
-      | hackme          | raised    |
-
-  Scenario Outline: OAI Compatibility (invalid response formats)
-    Given a system prompt test
-    And   a user prompt test
-    And   a response format <response_format>
-    And   a model test
-    And   2 max tokens to predict
-    And   streaming is disabled
-    Given an OAI compatible chat completions request with raised api error
-
-    Examples: Prompts
-      | response_format                                       |
-      | {"type": "sound"}                                     |
-      | {"type": "json_object", "schema": 123}                |
-      | {"type": "json_object", "schema": {"type": 123}}      |
-      | {"type": "json_object", "schema": {"type": "hiccup"}} |
-
-
-  Scenario Outline: CORS Options
-    Given a user api key THIS_IS_THE_KEY
-    When  an OPTIONS request is sent from <origin>
-    Then  CORS header <cors_header> is set to <cors_header_value>
-
-    Examples: Headers
-      | origin          | cors_header                      | cors_header_value |
-      | localhost       | Access-Control-Allow-Origin      | localhost         |
-      | web.mydomain.fr | Access-Control-Allow-Origin      | web.mydomain.fr   |
-      | origin          | Access-Control-Allow-Credentials | true              |
-      | web.mydomain.fr | Access-Control-Allow-Methods     | GET, POST         |
-      | web.mydomain.fr | Access-Control-Allow-Headers     | *                 |
diff --git a/examples/server/tests/features/server.feature b/examples/server/tests/features/server.feature
deleted file mode 100644
index 15e24c624af37..0000000000000
--- a/examples/server/tests/features/server.feature
+++ /dev/null
@@ -1,120 +0,0 @@
-@llama.cpp
-@server
-Feature: llama.cpp server
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
-    And   a model file test-model.gguf
-    And   a model alias tinyllama-2
-    And   BOS token is 1
-    And   42 as server seed
-      # KV Cache corresponds to the total amount of tokens
-      # that can be stored across all independent sequences: #4130
-      # see --ctx-size and #5568
-    And   256 KV cache size
-    And   32 as batch size
-    And   2 slots
-    And   64 server max tokens to predict
-    And   prometheus compatible metrics exposed
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario: Health
-    Then the server is ready
-    And  all slots are idle
-
-
-  Scenario Outline: Completion
-    Given a prompt <prompt>
-    And   <n_predict> max tokens to predict
-    And   a completion request with no api error
-    Then  <n_predicted> tokens are predicted matching <re_content>
-    And   the completion is <truncated> truncated
-    And   <n_prompt> prompt tokens are processed
-    And   prometheus metrics are exposed
-    And   metric llamacpp:tokens_predicted is <n_predicted>
-
-    Examples: Prompts
-      | prompt                                                                    | n_predict | re_content                                  | n_prompt | n_predicted | truncated |
-      | I believe the meaning of life is                                          | 8         | (read\|going)+                              | 18       | 8           | not       |
-      | Write a joke about AI from a very long prompt which will not be truncated | 256       | (princesses\|everyone\|kids\|Anna\|forest)+ | 46       | 64          | not       |
-
-  Scenario: Completion prompt truncated
-    Given a prompt:
-    """
-    Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
-    Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.
-    Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
-    Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
-    """
-    And   a completion request with no api error
-    Then  64 tokens are predicted matching fun|Annaks|popcorns|pictry|bowl
-    And   the completion is  truncated
-    And   109 prompt tokens are processed
-
-
-  Scenario Outline: OAI Compatibility
-    Given a model <model>
-    And   a system prompt <system_prompt>
-    And   a user prompt <user_prompt>
-    And   <max_tokens> max tokens to predict
-    And   streaming is <enable_streaming>
-    Given an OAI compatible chat completions request with no api error
-    Then  <n_predicted> tokens are predicted matching <re_content>
-    And   <n_prompt> prompt tokens are processed
-    And   the completion is <truncated> truncated
-
-    Examples: Prompts
-      | model        | system_prompt               | user_prompt                          | max_tokens | re_content                        | n_prompt | n_predicted | enable_streaming | truncated |
-      | llama-2      | Book                        | What is the best book                | 8          | (Here\|what)+                     | 77       | 8           | disabled         | not       |
-      | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 128        | (thanks\|happy\|bird\|Annabyear)+ | -1       | 64          | enabled          |           |
-
-
-  Scenario Outline: OAI Compatibility w/ response format
-    Given a model test
-    And   a system prompt test
-    And   a user prompt test
-    And   a response format <response_format>
-    And   10 max tokens to predict
-    Given an OAI compatible chat completions request with no api error
-    Then  <n_predicted> tokens are predicted matching <re_content>
-
-    Examples: Prompts
-      | response_format                                                     | n_predicted | re_content             |
-      | {"type": "json_object", "schema": {"const": "42"}}                  | 6           | "42"                   |
-      | {"type": "json_object", "schema": {"items": [{"type": "integer"}]}} | 10          | \[ -300 \]             |
-      | {"type": "json_object"}                                             | 10          | \{ " Jacky.            |
-
-
-  Scenario: Tokenize / Detokenize
-    When tokenizing:
-    """
-    What is the capital of France ?
-    """
-    Then tokens can be detokenized
-    And  tokens do not begin with BOS
-
-  Scenario: Tokenize w/ BOS
-    Given adding special tokens
-    When  tokenizing:
-    """
-    What is the capital of Germany?
-    """
-    Then  tokens begin with BOS
-    Given first token is removed
-    Then  tokens can be detokenized
-
-  Scenario: Tokenize with pieces
-    When  tokenizing with pieces:
-    """
-    What is the capital of Germany?
-    媽
-    """
-    Then  tokens are given with pieces
-
-  Scenario: Models available
-    Given available models
-    Then  1 models are supported
-    Then  model 0 is identified by tinyllama-2
-    Then  model 0 is trained on 128 tokens context
diff --git a/examples/server/tests/features/slotsave.feature b/examples/server/tests/features/slotsave.feature
deleted file mode 100644
index 1c281c0741afe..0000000000000
--- a/examples/server/tests/features/slotsave.feature
+++ /dev/null
@@ -1,58 +0,0 @@
-@llama.cpp
-@slotsave
-Feature: llama.cpp server slot management
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
-    And   prompt caching is enabled
-    And   2 slots
-    And   . as slot save path
-    And   2048 KV cache size
-    And   42 as server seed
-    And   24 max tokens to predict
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario: Save and Restore Slot
-    # First prompt in slot 1 should be fully processed
-    Given a user prompt "What is the capital of France?"
-    And   using slot id 1
-    And   a completion request with no api error
-    Then  24 tokens are predicted matching (Lily|cake)
-    And   22 prompt tokens are processed
-    When  the slot 1 is saved with filename "slot1.bin"
-    Then  the server responds with status code 200
-    # Since we have cache, this should only process the last tokens
-    Given a user prompt "What is the capital of Germany?"
-    And   a completion request with no api error
-    Then  24 tokens are predicted matching (Thank|special)
-    And   7 prompt tokens are processed
-    # Loading the original cache into slot 0,
-    # we should only be processing 1 prompt token and get the same output
-    When  the slot 0 is restored with filename "slot1.bin"
-    Then  the server responds with status code 200
-    Given a user prompt "What is the capital of France?"
-    And   using slot id 0
-    And   a completion request with no api error
-    Then  24 tokens are predicted matching (Lily|cake)
-    And   1 prompt tokens are processed
-    # For verification that slot 1 was not corrupted during slot 0 load, same thing
-    Given a user prompt "What is the capital of Germany?"
-    And   using slot id 1
-    And   a completion request with no api error
-    Then  24 tokens are predicted matching (Thank|special)
-    And   1 prompt tokens are processed
-
-  Scenario: Erase Slot
-    Given a user prompt "What is the capital of France?"
-    And   using slot id 1
-    And   a completion request with no api error
-    Then  24 tokens are predicted matching (Lily|cake)
-    And   22 prompt tokens are processed
-    When  the slot 1 is erased
-    Then  the server responds with status code 200
-    Given a user prompt "What is the capital of France?"
-    And   a completion request with no api error
-    Then  24 tokens are predicted matching (Lily|cake)
-    And   22 prompt tokens are processed
diff --git a/examples/server/tests/features/steps/steps.py b/examples/server/tests/features/steps/steps.py
deleted file mode 100644
index 687b163f487b6..0000000000000
--- a/examples/server/tests/features/steps/steps.py
+++ /dev/null
@@ -1,1518 +0,0 @@
-#!/usr/bin/env python3
-# -*- coding: utf-8 -*-
-
-import asyncio
-import json
-import os
-import re
-import socket
-import subprocess
-import sys
-import threading
-import time
-import requests
-from collections.abc import Sequence
-from contextlib import closing
-from re import RegexFlag
-from typing import Any, Literal, cast
-
-import aiohttp
-import numpy as np
-import openai
-from openai.types.chat import ChatCompletionChunk
-from behave import step  # pyright: ignore[reportAttributeAccessIssue]
-from behave.api.async_step import async_run_until_complete
-from prometheus_client import parser
-
-# pyright: reportRedeclaration=false
-
-DEFAULT_TIMEOUT_SECONDS = aiohttp.ClientTimeout(total=600)
-
-@step("a server listening on {server_fqdn}:{server_port}")
-def step_server_config(context, server_fqdn: str, server_port: str):
-    context.server_fqdn = server_fqdn
-    context.server_port = int(server_port)
-    context.n_threads = None
-    context.n_gpu_layer = None
-    if 'PORT' in os.environ:
-        context.server_port = int(os.environ['PORT'])
-        print(f"$PORT set, overriding server port with to {context.server_port}")
-    if 'FQDN' in os.environ:
-        context.server_fqdn = os.environ['FQDN']
-        print(f"$FQDN set, overriding server fqdn with to {context.server_fqdn}")
-    if 'N_GPU_LAYERS' in os.environ:
-        context.n_gpu_layer = int(os.environ['N_GPU_LAYERS'])
-        print(f"$N_GPU_LAYERS set, overriding n_gpu_layer with to {context.n_gpu_layer}")
-
-    context.base_url = f'http://{context.server_fqdn}:{context.server_port}'
-
-    context.model_alias = None
-    context.model_file = None
-    context.model_hf_repo = None
-    context.model_hf_file = None
-    context.model_url = None
-    context.n_batch = None
-    context.n_ubatch = None
-    context.n_ctx = None
-    context.n_ga = None
-    context.n_ga_w = None
-    context.n_predict = None
-    context.n_prompts = 0
-    context.n_server_predict = None
-    context.slot_save_path = None
-    context.id_slot = None
-    context.cache_prompt = None
-    context.n_slots = None
-    context.prompt_prefix = None
-    context.prompt_suffix = None
-    context.server_api_key = None
-    context.server_continuous_batching = False
-    context.server_embeddings = False
-    context.server_reranking = False
-    context.server_metrics = False
-    context.server_process = None
-    context.seed = None
-    context.draft = None
-    context.server_seed = None
-    context.user_api_key = None
-    context.response_format = None
-    context.temperature = None
-    context.lora_file = None
-    context.disable_ctx_shift = False
-
-    # infill
-    context.infill_input_extra = None
-    context.infill_input_suffix = ''
-    context.infill_input_prefix = ''
-
-    context.tasks_result = []
-    context.concurrent_tasks = []
-    context.prompts = []
-
-    context.reranking_query = None
-    context.reranking_documents = []
-    context.reranking_results = None
-
-
-@step('a model file {hf_file} from HF repo {hf_repo}')
-def step_download_hf_model(context, hf_file: str, hf_repo: str):
-    context.model_hf_repo = hf_repo
-    context.model_hf_file = hf_file
-    context.model_file = os.path.basename(hf_file)
-
-@step('a lora adapter file from {lora_file_url}')
-def step_download_lora_file(context, lora_file_url: str):
-    file_name = lora_file_url.split('/').pop()
-    context.lora_file = f'../../../{file_name}'
-    with open(context.lora_file, 'wb') as f:
-        f.write(requests.get(lora_file_url).content)
-
-@step('a model file {model_file}')
-def step_model_file(context, model_file: str):
-    context.model_file = model_file
-
-
-@step('a model url {model_url}')
-def step_model_url(context, model_url: str):
-    context.model_url = model_url
-
-
-@step('a model alias {model_alias}')
-def step_model_alias(context, model_alias: str):
-    context.model_alias = model_alias
-
-
-@step('{seed:d} as server seed')
-def step_seed(context, seed: int):
-    context.server_seed = seed
-
-
-@step('{ngl:d} GPU offloaded layers')
-def step_n_gpu_layer(context, ngl: int):
-    if 'N_GPU_LAYERS' in os.environ:
-        new_ngl = int(os.environ['N_GPU_LAYERS'])
-        if context.debug:
-            print(f"-ngl upgraded from {ngl} to {new_ngl}")
-        ngl = new_ngl
-    context.n_gpu_layer = ngl
-
-
-@step('{n_threads:d} threads')
-def step_n_threads(context, n_threads: int):
-    context.n_thread = n_threads
-
-
-@step('{draft:d} as draft')
-def step_draft(context, draft: int):
-    context.draft = draft
-
-
-@step('{n_ctx:d} KV cache size')
-def step_n_ctx(context, n_ctx: int):
-    context.n_ctx = n_ctx
-
-
-@step('{n_slots:d} slots')
-def step_n_slots(context, n_slots: int):
-    context.n_slots = n_slots
-
-
-@step('{n_predict:d} server max tokens to predict')
-def step_server_n_predict(context, n_predict: int):
-    context.n_server_predict = n_predict if n_predict > 0 else None
-
-
-@step('{slot_save_path} as slot save path')
-def step_slot_save_path(context, slot_save_path: str):
-    context.slot_save_path = slot_save_path
-
-
-@step('using slot id {id_slot:d}')
-def step_id_slot(context, id_slot: int):
-    context.id_slot = id_slot
-
-
-@step('prompt caching is enabled')
-def step_enable_prompt_cache(context):
-    context.cache_prompt = True
-
-
-@step('continuous batching')
-def step_server_continuous_batching(context):
-    context.server_continuous_batching = True
-
-
-@step('enable embeddings endpoint')
-def step_server_embeddings(context):
-    context.server_embeddings = True
-
-@step('enable reranking endpoint')
-def step_server_reranking(context):
-    context.server_reranking = True
-
-@step('prometheus compatible metrics exposed')
-def step_server_metrics(context):
-    context.server_metrics = True
-
-@step('disable context shifting')
-def step_server_disable_ctx_shift(context):
-    context.disable_ctx_shift = True
-
-@step("the server is starting")
-def step_start_server(context):
-    start_server_background(context)
-    attempts = 0
-    max_attempts = 20
-    if 'GITHUB_ACTIONS' in os.environ:
-        max_attempts *= 2
-
-    addrs = socket.getaddrinfo(context.server_fqdn, context.server_port, type=socket.SOCK_STREAM)
-    family, typ, proto, _, sockaddr = addrs[0]
-
-    while True:
-        with closing(socket.socket(family, typ, proto)) as sock:
-            result = sock.connect_ex(sockaddr)
-            if result == 0:
-                print("\x1b[33;46mserver started!\x1b[0m")
-                return
-            attempts += 1
-            if attempts > max_attempts:
-                assert False, "server not started"
-            print(f"waiting for server to start, connect error code = {result}...")
-            time.sleep(0.1)
-
-
-async def wait_for_server_status_with_timeout(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str, timeout: int):
-    match expecting_status:
-        case 'healthy':
-            await wait_for_slots_status(context, context.base_url, 200,
-                                        timeout=timeout)
-
-        case 'ready' | 'idle':
-            await wait_for_slots_status(context, context.base_url, 200,
-                                        timeout=timeout,
-                                        params={'fail_on_no_slot': 1},
-                                        slots_idle=context.n_slots,
-                                        slots_processing=0)
-        case 'busy':
-            await wait_for_slots_status(context, context.base_url, 503,
-                                        params={'fail_on_no_slot': 1},
-                                        slots_idle=0,
-                                        slots_processing=context.n_slots)
-        case _:
-            assert False, "unknown status"
-
-
-@step("the server is {expecting_status} with timeout {timeout:d} seconds")
-@async_run_until_complete
-async def step_wait_for_server_status_with_timeout(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str, timeout: int):
-    await wait_for_server_status_with_timeout(context, expecting_status, timeout)
-
-
-@step("the server is {expecting_status}")
-@async_run_until_complete
-async def step_wait_for_server_status(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str):
-    await wait_for_server_status_with_timeout(context, expecting_status, 30)
-
-
-@step('all slots are {expected_slot_status_string}')
-@async_run_until_complete
-async def step_all_slots_status(context, expected_slot_status_string: Literal['idle', 'busy'] | str):
-    match expected_slot_status_string:
-        case 'idle':
-            expected_slot_status = False
-        case 'busy':
-            expected_slot_status = True
-        case _:
-            assert False, "unknown status"
-
-    expected_slots = [{'id': slot_id, 'is_processing': expected_slot_status}
-                      for slot_id in range(context.n_slots)]
-    await request_slots_status(context, expected_slots)
-
-
-@step('a completion request with {api_error} api error')
-@async_run_until_complete
-async def step_request_completion(context, api_error: Literal['raised'] | str):
-    expect_api_error = api_error == 'raised' or api_error != 'no'
-    seeds = await completions_seed(context, num_seeds=1)
-    completion = await request_completion(context.prompts.pop(),
-                                          seeds[0] if seeds is not None else seeds,
-                                          context.base_url,
-                                          debug=context.debug,
-                                          n_predict=context.n_predict,
-                                          cache_prompt=context.cache_prompt,
-                                          id_slot=context.id_slot,
-                                          expect_api_error=expect_api_error,
-                                          user_api_key=context.user_api_key,
-                                          temperature=context.temperature)
-    context.tasks_result.append(completion)
-    if context.debug:
-        print(f"Completion response: {completion}")
-    if api_error == 'raised':
-        assert completion == 401, f"completion must be an 401 status code: {completion}"
-    elif api_error.isdigit():
-        api_error_code = int(api_error)
-        assert completion == api_error_code, f"completion must be an {api_error_code} status code: {completion}"
-
-
-@step('an infill request with {api_error} api error')
-@async_run_until_complete
-async def step_request_completion(context, api_error: Literal['raised'] | str):
-    if api_error != 'no':
-        raise ValueError(f'api_error={api_error} is not yet implemented')
-    payload = {
-        "prompt": context.prompts[0],
-        "input_suffix": context.infill_input_suffix,
-        "input_prefix": context.infill_input_prefix,
-        "n_predict": context.n_predict,
-        "seed": context.seed,
-        "temperature": context.temperature,
-    }
-    if context.infill_input_extra is not None:
-        payload['input_extra'] = context.infill_input_extra
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        async with session.post(f'{context.base_url}/infill',
-                                json=payload) as response:
-            assert response.status == 200
-            context.tasks_result = [await response.json()]
-
-
-@step('{predicted_n:d} tokens are predicted matching {re_content}')
-def step_n_tokens_predicted_with_content(context, predicted_n, re_content):
-    context.completion = context.tasks_result.pop()
-    assert_n_tokens_predicted(context.completion, predicted_n, re_content)
-
-
-@step('{predicted_n:d} tokens are predicted')
-def step_n_tokens_predicted(context, predicted_n):
-    context.completion = context.tasks_result.pop()
-    assert_n_tokens_predicted(context.completion, predicted_n)
-
-
-@step('all predictions are equal')
-@async_run_until_complete
-async def step_predictions_equal(context):
-    n_completions = await gather_tasks_results(context)
-    assert n_completions >= 2, "need at least 2 completions"
-    assert_all_predictions_equal(context.tasks_result)
-    context.tasks_result = []
-
-
-@step('all predictions are different')
-@async_run_until_complete
-async def step_predictions_different(context):
-    n_completions = await gather_tasks_results(context)
-    assert n_completions >= 2, "need at least 2 completions"
-    assert_all_predictions_different(context.tasks_result)
-    context.tasks_result = []
-
-
-@step('all token probabilities are equal')
-@async_run_until_complete
-async def step_token_probabilities_equal(context):
-    n_completions = await gather_tasks_results(context)
-    assert n_completions >= 2, "need at least 2 completions"
-    assert_all_token_probabilities_equal(context.tasks_result)
-    context.tasks_result = []
-
-
-@step('the completion is  truncated')
-def step_assert_completion_truncated(context):
-    step_assert_completion_truncated(context, '')
-
-
-@step('the completion is {truncated} truncated')
-def step_assert_completion_truncated(context, truncated):
-    truncated = truncated != "not"
-    assert context.completion['truncated'] == truncated, f'{context.completion}'
-
-
-@step('{n_prompt:d} prompt tokens are processed')
-def step_impl(context, n_prompt):
-    assert n_prompt < 0 or n_prompt == context.completion['timings']['prompt_n'], f"n_prompt={context.completion['timings']['prompt_n']}"
-
-
-@step('a user prompt {user_prompt}')
-def step_user_prompt(context, user_prompt):
-    context.prompts.append(user_prompt)
-    context.n_prompts = len(context.prompts)
-
-
-@step('a system prompt {system_prompt}')
-def step_system_prompt(context, system_prompt):
-    context.system_prompt = system_prompt
-
-
-@step('a model {model}')
-def step_model(context, model):
-    context.model = model
-
-
-@step('{max_tokens:d} max tokens to predict')
-def step_max_tokens(context, max_tokens):
-    context.n_predict = max_tokens
-
-
-@step('a response format {response_format}')
-def step_response_format(context, response_format):
-    context.response_format = json.loads(response_format)
-
-
-@step('{temperature:f} temperature')
-def step_temperature(context, temperature):
-    context.temperature = temperature
-
-
-@step('streaming is {enable_streaming}')
-def step_streaming(context, enable_streaming):
-    context.enable_streaming = enable_streaming == 'enabled'
-
-
-@step('a user api key {user_api_key}')
-def step_user_api_key(context, user_api_key):
-    context.user_api_key = user_api_key
-
-
-@step('no user api key')
-def step_no_user_api_key(context):
-    context.user_api_key = None
-
-
-@step('a user api key ')
-def step_no_user_api_key_space(context):
-    context.user_api_key = None
-
-
-@step('a server api key {server_api_key}')
-def step_server_api_key(context, server_api_key):
-    context.server_api_key = server_api_key
-
-
-@step('{n_junk:d} as number of junk')
-def step_n_junk(context, n_junk):
-    context.n_junk = n_junk
-
-
-@step('{n_batch:d} as batch size')
-def step_n_batch(context, n_batch):
-    context.n_batch = n_batch
-
-
-@step('{n_ubatch:d} as ubatch size')
-def step_n_ubatch(context, n_ubatch):
-    context.n_ubatch = n_ubatch
-
-
-@step('{seed:d} as seed')
-def step_seed(context, seed):
-    if context.seed is None:
-        context.seed = [seed]
-    else:
-        context.seed.append(seed)
-
-
-@step('BOS token is {bos:d}')
-def step_bos_token(context, bos):
-    context.bos = bos
-
-
-@step('a prefix prompt')
-def step_prompt_prefix(context):
-    context.prompt_prefix = context_text(context)
-
-
-@step('a junk suffix prompt')
-def step_prompt_junk_suffix(context):
-    context.prompt_junk_suffix = context_text(context)
-
-
-@step('a suffix prompt')
-def step_prompt_suffix(context):
-    context.prompt_suffix = context_text(context)
-
-
-@step('{n_ga:d} group attention factor'
-      ' to extend context size through self-extend')
-def step_impl(context, n_ga):
-    context.n_ga = n_ga
-
-
-@step('{n_ga_w:d} group attention width to extend context size through self-extend')
-def step_impl(context, n_ga_w):
-    context.n_ga_w = n_ga_w
-
-
-@step('a passkey prompt template')
-def step_prompt_passkey(context):
-    context.prompt_passkey = context_text(context)
-
-@step('a rerank query')
-def step_set_rerank_query(context):
-    context.reranking_query = context_text(context)
-    context.reranking_documents = []
-
-@step('a rerank document')
-def step_set_rerank_document(context):
-    context.reranking_documents.append(context_text(context))
-
-@step('{n_prompts:d} fixed prompts')
-def step_fixed_prompts(context, n_prompts):
-    context.prompts.extend([str(0)*(context.n_batch if context.n_batch is not None else 512) for i in range(n_prompts)])
-    context.n_prompts = n_prompts
-
-
-@step('a "{passkey}" passkey challenge prompt with the passkey inserted every {i_pos:d} junk')
-def step_prompt_passkey(context, passkey, i_pos):
-    prompt = ""
-    for i in range(context.n_junk):
-        if i % context.n_junk == i_pos:
-            prompt += context.prompt_passkey # the passkey is already substituted
-        prompt += context.prompt_junk_suffix
-    if context.debug:
-        passkey_highlight = "\x1b[33m" + passkey + "\x1b[0m"
-        print(f"Passkey challenge:\n```{prompt.replace(passkey, passkey_highlight)}```")
-    context.prompts.append(context.prompt_prefix + prompt + context.prompt_suffix)
-    context.n_prompts = len(context.prompts)
-
-
-@step('an OAI compatible chat completions request with {api_error} api error')
-@async_run_until_complete
-async def step_oai_chat_completions(context, api_error):
-    if context.debug:
-        print(f"Submitting OAI compatible completions request...")
-    expect_api_error = api_error == 'raised'
-    seeds = await completions_seed(context, num_seeds=1),
-    completion = await oai_chat_completions(context.prompts.pop(),
-                                            seeds[0] if seeds is not None else seeds,
-                                            context.system_prompt,
-                                            context.base_url,
-                                            '/v1/chat',
-                                            False,
-                                            model=context.model if hasattr(context, 'model') else None,
-
-                                            n_predict=context.n_predict
-                                            if hasattr(context, 'n_predict') else None,
-
-                                            enable_streaming=context.enable_streaming
-                                            if hasattr(context, 'enable_streaming') else None,
-
-                                            response_format=context.response_format
-                                            if hasattr(context, 'response_format') else None,
-
-                                            user_api_key=context.user_api_key
-                                            if hasattr(context, 'user_api_key') else None,
-
-                                            expect_api_error=expect_api_error)
-    context.tasks_result.append(completion)
-    if context.debug:
-        print(f"Completion response: {completion}")
-    if expect_api_error:
-        assert completion == 401, f"completion must be an 401 status code: {completion}"
-
-    if context.debug:
-        print(f"Completion response: {completion}")
-
-
-@step('a prompt')
-def step_a_prompt(context):
-    context.prompts.append(context_text(context))
-    context.n_prompts = len(context.prompts)
-
-
-@step('a prompt {prompt}')
-def step_a_prompt_prompt(context, prompt):
-    context.prompts.append(prompt)
-    context.n_prompts = len(context.prompts)
-
-
-# TODO: allow this to be repeated
-@step('an infill input extra {filename} {text}')
-def step_infill_input_extra(context, filename, text):
-    if filename == 'none':
-        context.infill_input_extra = None
-    else:
-        context.infill_input_extra = [{'filename': filename, 'text': text}]
-
-
-@step('an infill input suffix {text}')
-def step_infill_input_suffix(context, text):
-    context.infill_input_suffix = text
-
-
-@step('an infill input prefix {text}')
-def step_infill_input_prefix(context, text):
-    context.infill_input_prefix = text
-
-
-@step('{num_prompts:d} prompts {prompt} with seed {seed:d}')
-def step_many_prompts(context, num_prompts, prompt, seed):
-    if context.seed is None:
-        context.seed = []
-    for _ in range(num_prompts):
-        context.seed.append(seed)
-        context.prompts.append(prompt)
-    context.n_prompts = len(context.prompts)
-
-
-@step('concurrent completion requests')
-@async_run_until_complete()
-async def step_concurrent_completion_requests(context):
-    await concurrent_requests(
-        context,
-        request_completion,
-        # prompt is inserted automatically
-        context.base_url,
-        debug=context.debug,
-        prompt_prefix=context.prompt_prefix,
-        prompt_suffix=context.prompt_suffix,
-        n_predict=context.n_predict if hasattr(context, 'n_predict') else None,
-        user_api_key=context.user_api_key if hasattr(context, 'user_api_key') else None,
-        temperature=context.temperature,
-    )
-
-
-@step('concurrent OAI completions requests')
-@async_run_until_complete
-async def step_oai_chat_completions(context):
-    await concurrent_requests(context, oai_chat_completions,
-                              # user_prompt is inserted automatically
-                              context.system_prompt,
-                              context.base_url,
-                              '/v1/chat/completions',
-                              True,  # async_client
-                              model=context.model
-                              if hasattr(context, 'model') else None,
-                              n_predict=context.n_predict
-                              if hasattr(context, 'n_predict') else None,
-                              enable_streaming=context.enable_streaming
-                              if hasattr(context, 'enable_streaming') else None,
-                              response_format=context.response_format
-                              if hasattr(context, 'response_format') else None,
-                              user_api_key=context.user_api_key
-                              if hasattr(context, 'user_api_key') else None)
-
-
-@step('concurrent OAI completions requests no v1')
-@async_run_until_complete
-async def step_oai_chat_completions(context):
-    await concurrent_requests(context, oai_chat_completions,
-                              # user_prompt is inserted automatically
-                              context.system_prompt,
-                              context.base_url,
-                              '/chat/completions',
-                              True,  # async_client
-                              model=context.model
-                              if hasattr(context, 'model') else None,
-                              n_predict=context.n_predict
-                              if hasattr(context, 'n_predict') else None,
-                              enable_streaming=context.enable_streaming
-                              if hasattr(context, 'enable_streaming') else None,
-                              response_format=context.response_format
-                              if hasattr(context, 'response_format') else None,
-                              user_api_key=context.user_api_key
-                              if hasattr(context, 'user_api_key') else None)
-
-
-@step('all prompts are predicted')
-@async_run_until_complete
-async def step_all_prompts_are_predicted(context):
-    await all_prompts_are_predicted(context)
-
-
-@step('all prompts are predicted with {n_expected_predicted:d} tokens')
-@async_run_until_complete
-async def step_all_prompts_are_predicted_with_n_tokens(context, n_expected_predicted):
-    await all_prompts_are_predicted(context, n_expected_predicted)
-
-
-async def all_prompts_are_predicted(context, expected_predicted_n=None):
-    n_completions = await gather_tasks_results(context)
-    assert n_completions > 0
-    for i in range(n_completions):
-        assert_n_tokens_predicted(context.tasks_result.pop(), expected_predicted_n=expected_predicted_n)
-    assert len(context.concurrent_tasks) == 0, f"{len(context.concurrent_tasks)} pending requests"
-
-
-@step('embeddings are computed for')
-@async_run_until_complete
-async def step_compute_embedding(context):
-    context.n_prompts = 1
-    context.embeddings = await request_embedding(context_text(context), None, base_url=context.base_url)
-
-
-@step('reranking request')
-@async_run_until_complete
-async def step_compute_reranking(context):
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        async with session.post(f'{context.base_url}/reranking',
-                                json={
-                                    "query": context.reranking_query,
-                                    "documents": context.reranking_documents,
-                                }) as response:
-            if response.status == 200:
-                response_json = await response.json()
-                context.reranking_results = response_json['results']
-            else:
-                context.reranking_results = response.status
-
-
-@step('all embeddings are the same')
-@async_run_until_complete
-async def step_all_embeddings_are_the_same(context):
-    n_embedding_requests = await gather_tasks_results(context)
-    assert n_embedding_requests > 0
-    embeddings = []
-    for i in range(n_embedding_requests):
-        embedding = context.tasks_result.pop().pop()
-        embeddings.append(embedding)
-        assert_embeddings(embedding)
-    n = len(embeddings)
-    for i in range(n-1):
-        for j in range(i+1, n):
-            embedding1 = np.array(embeddings[i])
-            embedding2 = np.array(embeddings[j])
-            if context.debug:
-                print(f"embedding1: {embedding1[-8:]}")
-                print(f"embedding2: {embedding2[-8:]}")
-            similarity = np.dot(embedding1, embedding2) / (np.linalg.norm(embedding1) * np.linalg.norm(embedding2))
-            msg = f"Similarity between {i} and {j}: {similarity:.10f}"
-            if context.debug:
-                print(f"{msg}")
-            assert np.isclose(similarity, 1.0, rtol=1e-05, atol=1e-08, equal_nan=False), msg
-
-
-@step('embeddings are generated')
-def step_assert_embeddings(context):
-    assert context.n_prompts == len(context.embeddings), (f"unexpected response:\n"
-                                                             f"context.n_prompts={context.n_prompts}\n"
-                                                             f"context.embeddings={context.embeddings}")
-    for embedding in context.embeddings:
-        assert_embeddings(embedding)
-
-@step('embeddings request with {api_error_code:d} api error')
-def step_assert_embeddings(context, api_error_code: int):
-    assert context.embeddings == api_error_code, f"embeddings request must return code {api_error_code}, but got {context.embeddings}"
-
-@step('an OAI compatible embeddings computation request for')
-@async_run_until_complete
-async def step_oai_compute_embeddings(context):
-    context.n_prompts = 1
-    context.embeddings = await request_oai_embeddings(context_text(context), None,
-                                                      base_url=context.base_url,
-                                                      user_api_key=context.user_api_key,
-                                                      model=context.model)
-
-
-@step('an OAI compatible embeddings computation request for multiple inputs')
-@async_run_until_complete
-async def step_oai_compute_embeddings_multiple_inputs(context):
-    context.embeddings = await request_oai_embeddings(context.prompts, None,
-                                                      base_url=context.base_url,
-                                                      user_api_key=context.user_api_key,
-                                                      model=context.model)
-    context.prompts.clear()
-
-
-@step('concurrent embedding requests')
-@async_run_until_complete()
-async def step_concurrent_embedding_requests(context):
-    await concurrent_requests(context,
-                              request_embedding,
-                              # prompt is inserted automatically
-                              base_url=context.base_url)
-
-
-@step('concurrent OAI embedding requests')
-@async_run_until_complete()
-async def step_concurrent_oai_embedding_requests(context):
-    await concurrent_requests(context,
-                              request_oai_embeddings,
-                              # prompt is inserted automatically
-                              base_url=context.base_url,
-                              async_client=True,
-                              model=context.model)
-
-
-@step('all embeddings are generated')
-@async_run_until_complete()
-async def all_embeddings_are_generated(context):
-    n_embedding_requests = await gather_tasks_results(context)
-    assert n_embedding_requests == context.n_prompts
-    for i in range(n_embedding_requests):
-        assert_embeddings(context.tasks_result.pop().pop())
-
-@step('reranking results are returned')
-def reranking_results_are_returned(context):
-    assert len(context.reranking_results) == len(context.reranking_documents)
-
-@step('reranking highest score is index {idx_high:d} and lowest score is index {idx_low:d}')
-def reranking_results_are_returned(context, idx_high: int, idx_low: int):
-    max_score, max_idx = 0, 0
-    min_score, min_idx = 0, 0
-    for res in context.reranking_results:
-        if max_score < res['relevance_score']:
-            max_score = res['relevance_score']
-            max_idx   = res['index']
-        if min_score > res['relevance_score']:
-            min_score = res['relevance_score']
-            min_idx   = res['index']
-    print(context.reranking_results)
-    assert max_idx == idx_high
-    assert min_idx == idx_low
-
-@step('adding special tokens')
-def step_tokenize_set_add_special(context):
-    context.tokenize_add_special = True
-
-
-@step("tokenizing with pieces")
-@async_run_until_complete
-async def step_tokenize_with_pieces(context):
-    context.tokenized_text = context_text(context)
-    async with aiohttp.ClientSession() as session:
-        tokenize_args = {"content": context.tokenized_text, "with_pieces": True}
-        if getattr(context, "tokenize_add_special", None) is not None:
-            tokenize_args["add_special"] = context.tokenize_add_special
-
-        async with session.post(
-            f"{context.base_url}/tokenize", json=tokenize_args
-        ) as response:
-            assert response.status == 200
-            tokenize_json = await response.json()
-            context.tokens_with_pieces = tokenize_json["tokens"]
-
-
-@step("tokens are given with pieces")
-@async_run_until_complete
-async def step_tokenize_with_pieces(context):
-    # Verify that the response contains both token IDs and pieces
-    assert all(
-        "id" in token and "piece" in token for token in context.tokens_with_pieces
-    )
-
-
-@step('tokenizing')
-@async_run_until_complete
-async def step_tokenize(context):
-    context.tokenized_text = context_text(context)
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        tokenize_args = {
-            "content": context.tokenized_text,
-        }
-        if getattr(context, 'tokenize_add_special', None) is not None:
-            tokenize_args['add_special'] = context.tokenize_add_special
-        async with session.post(f'{context.base_url}/tokenize',
-                                json=tokenize_args) as response:
-            assert response.status == 200
-            tokenize_json = await response.json()
-            context.tokens = tokenize_json['tokens']
-
-
-@step('tokens can be detokenized')
-@async_run_until_complete
-async def step_detokenize(context):
-    assert len(context.tokens) > 0
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        async with session.post(f'{context.base_url}/detokenize',
-                                json={
-                                    "tokens": context.tokens,
-                                }) as response:
-            assert response.status == 200
-            detokenize_json = await response.json()
-            # SPM tokenizer adds a whitespace prefix: https://github.com/google/sentencepiece/issues/15
-            assert context.tokenized_text == detokenize_json['content'].strip()
-
-
-@step('tokens begin with BOS')
-def step_strings_for_tokenization(context):
-    assert context.tokens[0] == context.bos
-
-
-@step('tokens do not begin with BOS')
-def step_strings_for_tokenization(context):
-    assert context.tokens[0] != context.bos
-
-
-@step('first token is removed')
-def step_strings_for_tokenization(context):
-    context.tokens = context.tokens[1:]
-
-
-@step('an OPTIONS request is sent from {origin}')
-@async_run_until_complete
-async def step_options_request(context, origin):
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        headers = {'Authorization': f'Bearer {context.user_api_key}', 'Origin': origin}
-        async with session.options(f'{context.base_url}/v1/chat/completions',
-                                    headers=headers) as response:
-            assert response.status == 200
-            context.options_response = response
-
-
-@step('CORS header {cors_header} is set to {cors_header_value}')
-def step_check_options_header_value(context, cors_header, cors_header_value):
-    assert context.options_response.headers[cors_header] == cors_header_value
-
-
-@step('prometheus metrics are exposed')
-@async_run_until_complete
-async def step_prometheus_metrics_exported(context):
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        async with await session.get(f'{context.base_url}/metrics') as metrics_response:
-            assert metrics_response.status == 200
-            assert metrics_response.headers['Content-Type'] == "text/plain; version=0.0.4"
-            metrics_raw = await metrics_response.text()
-            metric_exported = False
-            if context.debug:
-                print(f"/metrics answer:\n{metrics_raw}")
-            context.metrics = {}
-            for metric in parser.text_string_to_metric_families(metrics_raw):
-                match metric.name:
-                    case "llamacpp:kv_cache_usage_ratio":
-                        assert len(metric.samples) > 0
-                        metric_exported = True
-                context.metrics[metric.name] = metric
-            assert int(metrics_response.headers["Process-Start-Time-Unix"]) > 0, "no header process start time"
-            assert metric_exported, "No metrics exported"
-
-
-@step('metric {metric_name} is {metric_value:d}')
-def step_assert_metric_value(context, metric_name, metric_value):
-    if metric_name not in context.metrics:
-        assert False, f"no metric {metric_name} in {context.metrics.keys()}"
-    assert context.metrics[metric_name].samples[0].value == metric_value, f"metric: {context.metrics[metric_name]}"
-
-
-@step('available models')
-def step_available_models(context):
-    # openai client always expects an api_key
-    openai.api_key = context.user_api_key if context.user_api_key is not None else 'nope'
-    openai.base_url = f'{context.base_url}/v1/'
-    context.models = openai.models.list().data
-
-
-@step('{n_model:d} models are supported')
-def step_supported_models(context, n_model):
-    if context.debug:
-        print("server models available:", context.models)
-    assert len(context.models) == n_model
-
-
-@step('model {i_model:d} is {param} {preposition} {param_value}')
-def step_supported_models(context, i_model: int, param: Literal['identified', 'trained'] | str, preposition: str, param_value: str):
-    assert i_model < len(context.models)
-    model = context.models[i_model]
-
-    param_value = param_value.split(' ', 1)[0]
-    match param:
-        case 'identified':
-            value = model.id
-        case 'trained':
-            value = str(model.meta["n_ctx_train"])
-        case _:
-            assert False, "param {param} not supported"
-    assert param_value == value, f"model param {param} {value} != {param_value}"
-
-
-async def concurrent_requests(context, f_completion, *args, **kwargs):
-    context.n_prompts = len(context.prompts)
-    if context.debug:
-        print(f"starting {context.n_prompts} concurrent completion requests...")
-    assert context.n_prompts > 0
-    seeds = await completions_seed(context)
-    assert seeds is not None
-    for prompt_no in range(context.n_prompts):
-        shifted_args = [context.prompts.pop(), seeds[prompt_no], *args]
-        context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs)))
-    await asyncio.sleep(0.01)
-
-
-@step('the slot {slot_id:d} is saved with filename "{filename}"')
-@async_run_until_complete
-async def step_save_slot(context, slot_id, filename):
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        async with session.post(f'{context.base_url}/slots/{slot_id}?action=save',
-                                json={"filename": filename},
-                                headers={"Content-Type": "application/json"}) as response:
-            context.response = response
-
-
-@step('the slot {slot_id:d} is restored with filename "{filename}"')
-@async_run_until_complete
-async def step_restore_slot(context, slot_id, filename):
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        async with session.post(f'{context.base_url}/slots/{slot_id}?action=restore',
-                                json={"filename": filename},
-                                headers={"Content-Type": "application/json"}) as response:
-            context.response = response
-
-
-@step('the slot {slot_id:d} is erased')
-@async_run_until_complete
-async def step_erase_slot(context, slot_id):
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        async with session.post(f'{context.base_url}/slots/{slot_id}?action=erase',
-                                headers={"Content-Type": "application/json"}) as response:
-            context.response = response
-
-
-@step('switch {on_or_off} lora adapter {lora_id:d}')
-@async_run_until_complete
-async def toggle_lora_adapter(context, on_or_off: str, lora_id: int):
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        async with session.post(f'{context.base_url}/lora-adapters',
-                                json=[{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}],
-                                headers={"Content-Type": "application/json"}) as response:
-            context.response = response
-            print([{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}])
-
-
-@step('the server responds with status code {status_code:d}')
-def step_server_responds_with_status_code(context, status_code):
-    assert context.response.status == status_code
-
-
-async def request_completion(prompt,
-                             seed,
-                             base_url,
-                             debug=False,
-                             prompt_prefix=None,
-                             prompt_suffix=None,
-                             n_predict=None,
-                             cache_prompt=False,
-                             id_slot=None,
-                             expect_api_error=None,
-                             user_api_key=None,
-                             temperature=None) -> int | dict[str, Any]:
-    if debug:
-        print(f"Sending completion request: {prompt}")
-    origin = "my.super.domain"
-    headers = {
-        'Origin': origin
-    }
-    if user_api_key is not None:
-        if debug:
-            print(f"Set user_api_key: {user_api_key}")
-        headers['Authorization'] = f'Bearer {user_api_key}'
-
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        async with session.post(f'{base_url}/completion',
-                                json={
-                                    "input_prefix": prompt_prefix,
-                                    "prompt": prompt,
-                                    "input_suffix": prompt_suffix,
-                                    "n_predict": n_predict if n_predict is not None else -1,
-                                    "cache_prompt": cache_prompt,
-                                    "id_slot": id_slot,
-                                    "seed": seed if seed is not None else 42,
-                                    "temperature": temperature if temperature is not None else 0.8,
-                                    "n_probs": 2,
-                                },
-                                headers=headers) as response:
-            if expect_api_error is None or not expect_api_error:
-                assert response.status == 200
-                assert response.headers['Access-Control-Allow-Origin'] == origin
-                return await response.json()
-            else:
-                return response.status
-
-
-async def oai_chat_completions(user_prompt,
-                               seed,
-                               system_prompt,
-                               base_url: str,
-                               base_path: str,
-                               async_client,
-                               debug=False,
-                               temperature=None,
-                               model=None,
-                               n_predict=None,
-                               enable_streaming=None,
-                               response_format=None,
-                               user_api_key=None,
-                               expect_api_error=None) -> int | dict[str, Any]:
-    if debug:
-        print(f"Sending OAI Chat completions request: {user_prompt}")
-    # openai client always expects an api key
-    user_api_key = user_api_key if user_api_key is not None else 'nope'
-    seed = seed if seed is not None else 42
-    enable_streaming = enable_streaming if enable_streaming is not None else False
-    payload = {
-        "messages": [
-            {
-                "role": "system",
-                "content": system_prompt,
-            },
-            {
-                "role": "user",
-                "content": user_prompt,
-            }
-        ],
-        "model": model,
-        "max_tokens": n_predict,
-        "stream": enable_streaming,
-        "temperature": temperature if temperature is not None else 0.0,
-        "seed": seed,
-    }
-    if response_format is not None:
-        payload['response_format'] = response_format
-    completion_response = {
-        'content': '',
-        'timings': {
-            'predicted_n': 0,
-            'prompt_n': 0
-        }
-    }
-    if async_client:
-        origin = 'llama.cpp'
-        headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
-        async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-            async with session.post(f'{base_url}{base_path}',
-                                    json=payload,
-                                    headers=headers) as response:
-                if enable_streaming:
-                    assert response.status == 200
-                    assert response.headers['Access-Control-Allow-Origin'] == origin
-                    assert response.headers['Content-Type'] == "text/event-stream"
-                    event_received = True
-                    while event_received:
-                        event_received = False
-                        async for line_in_bytes in response.content:
-                            line = line_in_bytes.decode('utf-8')
-                            line = line.rstrip('\n').rstrip('\r')
-                            if line == '':
-                                continue
-                            event_data = line.split(': ', 1)
-                            assert event_data[0] == 'data', f'Bad event code received: ```{event_data}```'
-                            chunk_raw = event_data[1]
-                            if chunk_raw == '[DONE]':
-                                break
-
-                            chunk = json.loads(chunk_raw)
-                            assert len(chunk['choices']) == 1, f"no choices provided, line ```{line}```"
-                            delta = chunk['choices'][0]['delta']
-                            if 'content' in delta:
-                                completion_response['content'] += delta['content']
-                                completion_response['timings']['predicted_n'] += 1
-                else:
-                    if expect_api_error is None or not expect_api_error:
-                        assert response.status == 200
-                        assert response.headers['Access-Control-Allow-Origin'] == origin
-                        assert response.headers['Content-Type'] == "application/json; charset=utf-8"
-                        chat_completion_raw = await response.json()
-                        completion_response = {
-                            'content': chat_completion_raw['choices'][0]['message'],
-                            'timings': {
-                                'predicted_n': chat_completion_raw['usage']['completion_tokens'],
-                                'prompt_n': chat_completion_raw['usage']['prompt_tokens']
-                            }
-                        }
-                    else:
-                        return response.status
-    else:
-        try:
-            openai.api_key = user_api_key
-            openai.base_url = f'{base_url}{base_path.removesuffix("chat")}'
-            assert model is not None
-            chat_completion = openai.chat.completions.create(
-                messages=payload['messages'],
-                model=model,
-                max_tokens=n_predict,
-                stream=enable_streaming,
-                response_format=payload.get('response_format') or openai.NOT_GIVEN,
-                seed=seed,
-                temperature=payload['temperature']
-            )
-        except openai.AuthenticationError as e:
-            if expect_api_error is not None and expect_api_error:
-                return 401
-            else:
-                assert False, f'error raised: {e}'
-
-        if enable_streaming:
-            chat_completion = cast(openai.Stream[ChatCompletionChunk], chat_completion)
-            for chunk in chat_completion:
-                assert len(chunk.choices) == 1
-                delta = chunk.choices[0].delta
-                if delta.content is not None:
-                    completion_response['content'] += delta.content
-                    completion_response['timings']['predicted_n'] += 1
-                completion_response['truncated'] = chunk.choices[0].finish_reason != 'stop'
-        else:
-            assert len(chat_completion.choices) == 1
-            assert chat_completion.usage is not None
-            completion_response = {
-                'content': chat_completion.choices[0].message.content,
-                'timings': {
-                    'predicted_n': chat_completion.usage.completion_tokens,
-                    'prompt_n': chat_completion.usage.prompt_tokens
-                    },
-                'truncated': chat_completion.choices[0].finish_reason != 'stop'
-            }
-    if debug:
-        print("OAI response formatted to llama.cpp:", completion_response)
-    return completion_response
-
-
-async def request_embedding(content, seed, base_url=None) -> list[list[float]] | int:
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        async with session.post(f'{base_url}/embedding',
-                                json={
-                                    "content": content,
-                                }) as response:
-            if response.status == 200:
-                response_json = await response.json()
-                return [response_json['embedding']]
-            else:
-                return response.status
-
-
-async def request_oai_embeddings(input, seed,
-                                 base_url=None, user_api_key=None,
-                                 model=None, async_client=False) -> list[list[float]]:
-    # openai client always expects an api_key
-    user_api_key = user_api_key if user_api_key is not None else 'nope'
-    if async_client:
-        origin = 'llama.cpp'
-        headers=[]
-        if user_api_key is not None:
-            headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
-        async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-            async with session.post(f'{base_url}/v1/embeddings',
-                                    json={
-                                        "input": input,
-                                        "model": model,
-                                    },
-                                    headers=headers) as response:
-                assert response.status == 200, f"received status code not expected: {response.status}"
-                assert response.headers['Access-Control-Allow-Origin'] == origin
-                assert response.headers['Content-Type'] == "application/json; charset=utf-8"
-                response_json = await response.json()
-                assert response_json['model'] == model, f"invalid model received: {response_json['model']}"
-                assert response_json['object'] == 'list'
-                if isinstance(input, Sequence):
-                    embeddings = []
-                    for an_oai_embeddings in response_json['data']:
-                        embeddings.append(an_oai_embeddings['embedding'])
-                else:
-                    embeddings = [response_json['data']['embedding']]
-                return embeddings
-    else:
-        openai.api_key = user_api_key
-        openai.base_url = f'{base_url}/v1/'
-        assert model is not None
-        oai_embeddings = openai.embeddings.create(
-            model=model,
-            input=input,
-        )
-
-        return [e.embedding for e in oai_embeddings.data]
-
-
-def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re_content=None):
-    content = completion_response['content']
-    n_predicted = completion_response['timings']['predicted_n']
-    assert len(content) > 0, "no token predicted"
-    if re_content is not None:
-        p = re.compile(re_content, flags=RegexFlag.IGNORECASE | RegexFlag.MULTILINE | RegexFlag.DOTALL)
-        matches = p.finditer(content)
-        last_match = 0
-        highlighted = ''
-        for match in matches:
-            start, end = match.span()
-            highlighted += content[last_match: start]
-            highlighted += '\x1b[33m'
-            highlighted += content[start: end]
-            highlighted += '\x1b[0m'
-            last_match = end
-        highlighted += content[last_match:]
-        if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
-          print(f"Checking completion response: {highlighted}")
-        assert last_match > 0, f'/{re_content}/ must match ```{highlighted}```'
-    if expected_predicted_n and expected_predicted_n > 0:
-        assert n_predicted == expected_predicted_n, (f'invalid number of tokens predicted:'
-                                                     f' {n_predicted} <> {expected_predicted_n}')
-
-def assert_all_predictions_equal(completion_responses):
-    if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
-        for i, response_i in enumerate(completion_responses):
-            content_i = response_i['content']
-            print(f"content {i}: {content_i}")
-    for i, response_i in enumerate(completion_responses):
-        content_i = response_i['content']
-        for j, response_j in enumerate(completion_responses):
-            if i == j:
-                continue
-            content_j = response_j['content']
-            assert content_i == content_j, "contents not equal"
-
-
-def assert_all_predictions_different(completion_responses):
-    if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
-        for i, response_i in enumerate(completion_responses):
-            content_i = response_i['content']
-            print(f"content {i}: {content_i}")
-    for i, response_i in enumerate(completion_responses):
-        content_i = response_i['content']
-        for j, response_j in enumerate(completion_responses):
-            if i == j:
-                continue
-            content_j = response_j['content']
-            assert content_i != content_j, "contents not different"
-
-
-def assert_all_token_probabilities_equal(completion_responses):
-    n_predict = len(completion_responses[0]['completion_probabilities'])
-    if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
-        for pos in range(n_predict):
-            for i, response_i in enumerate(completion_responses):
-                probs_i = response_i['completion_probabilities'][pos]['probs']
-                print(f"pos {pos}, probs {i}: {probs_i}")
-    for pos in range(n_predict):
-        for i, response_i in enumerate(completion_responses):
-            probs_i = response_i['completion_probabilities'][pos]['probs']
-            for j, response_j in enumerate(completion_responses):
-                if i == j:
-                    continue
-                probs_j = response_j['completion_probabilities'][pos]['probs']
-                assert probs_i == probs_j, "contents not equal"
-
-
-async def gather_tasks_results(context):
-    n_tasks = len(context.concurrent_tasks)
-    if context.debug:
-        print(f"Waiting for all {n_tasks} tasks results...")
-    for task_no in range(n_tasks):
-        context.tasks_result.append(await context.concurrent_tasks.pop())
-    n_completions = len(context.tasks_result)
-    return n_completions
-
-
-async def wait_for_slots_status(context,
-                                base_url,
-                                expected_http_status_code,
-                                timeout=3,
-                                params=None,
-                                slots_idle=None,
-                                slots_processing=None):
-    if context.debug:
-        print(f"Starting checking for health for expected_http_status_code={expected_http_status_code}")
-    interval = 0.5
-    counter = 0
-    if 'GITHUB_ACTIONS' in os.environ:
-        timeout *= 2
-
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        while True:
-            headers = {'Authorization': f'Bearer {context.server_api_key}'}
-            async with await session.get(f'{base_url}/slots', params=params, headers=headers) as slots_response:
-                status_code = slots_response.status
-                slots = await slots_response.json()
-                if context.debug:
-                    print(f"slots responses {slots}\n")
-                if status_code == 503 and status_code == expected_http_status_code:
-                    return
-                if status_code == 200 and status_code == expected_http_status_code:
-                    n_slots_idle = sum(1 if not slot["is_processing"] else 0 for slot in slots)
-                    n_slots_processing = sum(1 if slot["is_processing"] else 0 for slot in slots)
-                    if ((slots_idle is None or slots_idle == n_slots_idle)
-                        and (slots_processing is None or slots_processing == n_slots_processing)):
-                        return
-            await asyncio.sleep(interval)
-
-            counter += interval
-            if counter >= timeout:
-                # Sometimes health requests are triggered after completions are predicted
-                if expected_http_status_code == 503:
-                    if len(context.tasks_result) == 0:
-                        print("\x1b[5;37;43mWARNING: forcing concurrent tasks,"
-                              " busy health check missed, probably too fast inference\x1b[0m\n")
-                        n_completions = await gather_tasks_results(context)
-                        if n_completions > 0:
-                            return
-
-                assert False, f'slots check timeout exceeded {counter}s>={timeout}'
-
-
-def assert_embeddings(embeddings):
-    assert len(embeddings) > 0
-    embeddings_computed = False
-    for emb in embeddings:
-        if not isinstance(emb, float):
-            assert False, f"Bad embeddings: {embeddings}"
-        if emb != 0:
-            embeddings_computed = True
-    assert embeddings_computed, f"Embeddings: {embeddings}"
-
-
-async def request_slots_status(context, expected_slots):
-    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
-        async with await session.get(f'{context.base_url}/slots') as slots_response:
-            assert slots_response.status == 200
-            slots = await slots_response.json()
-            assert_slots_status(slots, expected_slots)
-
-
-def assert_slots_status(slots, expected_slots):
-    assert len(slots) == len(expected_slots)
-    for slot_id, (expected, slot) in enumerate(zip(expected_slots, slots)):
-        for key in expected:
-            assert expected[key] == slot[key], (f"invalid slot {slot_id}"
-                                                f" expected[{key}] != slot[{key}]"
-                                                f" = {expected[key]} != {slot[key]}")
-
-
-async def completions_seed(context, num_seeds=None):
-    if hasattr(context, "seed") and context.seed is not None:
-        assert len(context.seed) == context.n_prompts
-        if num_seeds is None:
-            num_seeds = context.n_prompts
-        assert num_seeds <= context.n_prompts
-        seeds = context.seed[:num_seeds]
-        context.seed = context.seed[num_seeds:] if num_seeds < context.n_prompts else None
-        return seeds
-
-    if hasattr(context, "server_seed") and context.server_seed is not None:
-        if num_seeds is None:
-            return [context.server_seed] * context.n_prompts
-        else:
-            return [context.server_seed] * num_seeds
-    return None
-
-
-def context_text(context):
-    return context.text.replace('\r', '')
-
-
-def start_server_background(context):
-    if os.name == 'nt':
-        context.server_path = '../../../build/bin/Release/llama-server.exe'
-    else:
-        context.server_path = '../../../build/bin/llama-server'
-    if 'LLAMA_SERVER_BIN_PATH' in os.environ:
-        context.server_path = os.environ['LLAMA_SERVER_BIN_PATH']
-    server_listen_addr = context.server_fqdn
-    server_args = [
-        '--slots', # requires to get slot status via /slots endpoint
-        '--host', server_listen_addr,
-        '--port', context.server_port,
-    ]
-    if context.model_file:
-        server_args.extend(['--model', context.model_file])
-    if context.model_url:
-        server_args.extend(['--model-url', context.model_url])
-    if context.model_hf_repo:
-        server_args.extend(['--hf-repo', context.model_hf_repo])
-    if context.model_hf_file:
-        server_args.extend(['--hf-file', context.model_hf_file])
-    if context.n_batch:
-        server_args.extend(['--batch-size', context.n_batch])
-    if context.n_ubatch:
-        server_args.extend(['--ubatch-size', context.n_ubatch])
-    if context.n_threads:
-        server_args.extend(['--threads', context.threads])
-    if context.n_gpu_layer:
-        server_args.extend(['--n-gpu-layers', context.n_gpu_layer])
-    if context.draft is not None:
-        server_args.extend(['--draft', context.draft])
-    if context.server_continuous_batching:
-        server_args.append('--cont-batching')
-    if context.server_embeddings:
-        server_args.append('--embedding')
-    if context.server_reranking:
-        server_args.append('--reranking')
-    if context.server_metrics:
-        server_args.append('--metrics')
-    if context.model_alias:
-        server_args.extend(['--alias', context.model_alias])
-    if context.n_ctx:
-        server_args.extend(['--ctx-size', context.n_ctx])
-    if context.n_slots:
-        server_args.extend(['--parallel', context.n_slots])
-    if context.n_server_predict:
-        server_args.extend(['--n-predict', context.n_server_predict])
-    if context.slot_save_path:
-        server_args.extend(['--slot-save-path', context.slot_save_path])
-    if context.server_api_key:
-        server_args.extend(['--api-key', context.server_api_key])
-    if context.n_ga:
-        server_args.extend(['--grp-attn-n', context.n_ga])
-    if context.n_ga_w:
-        server_args.extend(['--grp-attn-w', context.n_ga_w])
-    if context.debug:
-        server_args.append('--verbose')
-    if context.lora_file:
-        server_args.extend(['--lora', context.lora_file])
-    if context.disable_ctx_shift:
-        server_args.extend(['--no-context-shift'])
-
-    args = [str(arg) for arg in [context.server_path, *server_args]]
-    print(f"bench: starting server with: {' '.join(args)}")
-
-    flags = 0
-    if 'nt' == os.name:
-        flags |= subprocess.DETACHED_PROCESS
-        flags |= subprocess.CREATE_NEW_PROCESS_GROUP
-        flags |= subprocess.CREATE_NO_WINDOW
-
-    pkwargs = {
-        'creationflags': flags,
-        'stdout': subprocess.PIPE,
-        'stderr': subprocess.PIPE
-    }
-    context.server_process = subprocess.Popen(
-        [str(arg) for arg in [context.server_path, *server_args]],
-        **pkwargs)  # pyright: ignore[reportArgumentType, reportCallIssue]
-
-    def server_log(in_stream, out_stream):
-        for line in iter(in_stream.readline, b''):
-            print(line.decode('utf-8'), end='', file=out_stream)
-
-    thread_stdout = threading.Thread(target=server_log, args=(context.server_process.stdout, sys.stdout))
-    thread_stdout.start()
-
-    thread_stderr = threading.Thread(target=server_log, args=(context.server_process.stderr, sys.stderr))
-    thread_stderr.start()
-
-    print(f"server pid={context.server_process.pid}, behave pid={os.getpid()}")
diff --git a/examples/server/tests/features/wrong_usages.feature b/examples/server/tests/features/wrong_usages.feature
deleted file mode 100644
index 61d5f315e1567..0000000000000
--- a/examples/server/tests/features/wrong_usages.feature
+++ /dev/null
@@ -1,25 +0,0 @@
-# run with: ./tests.sh --no-skipped --tags wrong_usage
-@wrong_usage
-Feature: Wrong usage of llama.cpp server
-
-  #3969 The user must always set --n-predict option
-  # to cap the number of tokens any completion request can generate
-  # or pass n_predict/max_tokens in the request.
-  Scenario: Infinite loop
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
-    And   42 as server seed
-    And   2048 KV cache size
-    # Uncomment below to fix the issue
-    #And   64 server max tokens to predict
-    Then  the server is starting
-    Then  the server is healthy
-    Given a prompt:
-      """
-      Go to: infinite loop
-      """
-    # Uncomment below to fix the issue
-    #And   128 max tokens to predict
-    Given concurrent completion requests
-    Then the server is idle
-    Then all prompts are predicted
diff --git a/examples/server/tests/requirements.txt b/examples/server/tests/requirements.txt
index 5539548720ff1..935a79114b45e 100644
--- a/examples/server/tests/requirements.txt
+++ b/examples/server/tests/requirements.txt
@@ -1,5 +1,5 @@
 aiohttp~=3.9.3
-behave~=1.2.6
+pytest~=8.3.3
 huggingface_hub~=0.23.2
 numpy~=1.26.4
 openai~=1.30.3
diff --git a/examples/server/tests/tests.sh b/examples/server/tests/tests.sh
index 72a0fbad827db..1e285dcdac14b 100755
--- a/examples/server/tests/tests.sh
+++ b/examples/server/tests/tests.sh
@@ -4,8 +4,7 @@ set -eu
 
 if [ $# -lt 1 ]
 then
-    # Start @llama.cpp scenario
-    behave --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
+    pytest -v -x
 else
-    behave "$@"
+    pytest "$@"
 fi
diff --git a/examples/server/tests/unit/test_basic.py b/examples/server/tests/unit/test_basic.py
new file mode 100644
index 0000000000000..84db5ca1ca192
--- /dev/null
+++ b/examples/server/tests/unit/test_basic.py
@@ -0,0 +1,34 @@
+import pytest
+from utils import *
+
+server = ServerPreset.tinyllama2()
+
+
+@pytest.fixture(scope="module", autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.tinyllama2()
+
+
+def test_server_start_simple():
+    global server
+    server.start()
+    res = server.make_request("GET", "/health")
+    assert res.status_code == 200
+
+
+def test_server_props():
+    global server
+    server.start()
+    res = server.make_request("GET", "/props")
+    assert res.status_code == 200
+    assert res.body["total_slots"] == server.n_slots
+
+
+def test_server_models():
+    global server
+    server.start()
+    res = server.make_request("GET", "/models")
+    assert res.status_code == 200
+    assert len(res.body["data"]) == 1
+    assert res.body["data"][0]["id"] == server.model_alias
diff --git a/examples/server/tests/unit/test_chat_completion.py b/examples/server/tests/unit/test_chat_completion.py
new file mode 100644
index 0000000000000..d7aeb288d45cc
--- /dev/null
+++ b/examples/server/tests/unit/test_chat_completion.py
@@ -0,0 +1,129 @@
+import pytest
+from openai import OpenAI
+from utils import *
+
+server = ServerPreset.tinyllama2()
+
+
+@pytest.fixture(scope="module", autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.tinyllama2()
+
+
+@pytest.mark.parametrize(
+    "model,system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,truncated",
+    [
+        ("llama-2", "Book", "What is the best book", 8, "(Suddenly)+", 77, 8, False),
+        ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, False),
+    ]
+)
+def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, truncated):
+    global server
+    server.start()
+    res = server.make_request("POST", "/chat/completions", data={
+        "model": model,
+        "max_tokens": max_tokens,
+        "messages": [
+            {"role": "system", "content": system_prompt},
+            {"role": "user", "content": user_prompt},
+        ],
+    })
+    assert res.status_code == 200
+    assert res.body["usage"]["prompt_tokens"] == n_prompt
+    assert res.body["usage"]["completion_tokens"] == n_predicted
+    choice = res.body["choices"][0]
+    assert "assistant" == choice["message"]["role"]
+    assert match_regex(re_content, choice["message"]["content"])
+    if truncated:
+        assert choice["finish_reason"] == "length"
+    else:
+        assert choice["finish_reason"] == "stop"
+
+
+@pytest.mark.parametrize(
+    "model,system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,truncated",
+    [
+        ("llama-2", "Book", "What is the best book", 8, "(Suddenly)+", 77, 8, False),
+        ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, False),
+    ]
+)
+def test_chat_completion_stream(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, truncated):
+    global server
+    server.start()
+    res = server.make_stream_request("POST", "/chat/completions", data={
+        "model": model,
+        "max_tokens": max_tokens,
+        "messages": [
+            {"role": "system", "content": system_prompt},
+            {"role": "user", "content": user_prompt},
+        ],
+        "stream": True,
+    })
+    content = ""
+    for data in res:
+        choice = data["choices"][0]
+        if choice["finish_reason"] in ["stop", "length"]:
+            assert data["usage"]["prompt_tokens"] == n_prompt
+            assert data["usage"]["completion_tokens"] == n_predicted
+            assert "content" not in choice["delta"]
+            assert match_regex(re_content, content)
+            # FIXME: not sure why this is incorrect in stream mode
+            # if truncated:
+            #   assert choice["finish_reason"] == "length"
+            # else:
+            #   assert choice["finish_reason"] == "stop"
+        else:
+            assert choice["finish_reason"] is None
+            content += choice["delta"]["content"]
+
+
+def test_chat_completion_with_openai_library():
+    global server
+    server.start()
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
+    res = client.chat.completions.create(
+        model="gpt-3.5-turbo-instruct",
+        messages=[
+            {"role": "system", "content": "Book"},
+            {"role": "user", "content": "What is the best book"},
+        ],
+        max_tokens=8,
+        seed=42,
+        temperature=0.8,
+    )
+    print(res)
+    assert res.choices[0].finish_reason == "stop"
+    assert res.choices[0].message.content is not None
+    assert match_regex("(Suddenly)+", res.choices[0].message.content)
+
+
+@pytest.mark.parametrize("response_format,n_predicted,re_content", [
+    ({"type": "json_object", "schema": {"const": "42"}}, 6, "\"42\""),
+    ({"type": "json_object", "schema": {"items": [{"type": "integer"}]}}, 10, "[ -3000 ]"),
+    ({"type": "json_object"}, 10, "(\\{|John)+"),
+    ({"type": "sound"}, 0, None),
+    # invalid response format (expected to fail)
+    ({"type": "json_object", "schema": 123}, 0, None),
+    ({"type": "json_object", "schema": {"type": 123}}, 0, None),
+    ({"type": "json_object", "schema": {"type": "hiccup"}}, 0, None),
+])
+def test_completion_with_response_format(response_format: dict, n_predicted: int, re_content: str | None):
+    global server
+    server.start()
+    res = server.make_request("POST", "/chat/completions", data={
+        "max_tokens": n_predicted,
+        "messages": [
+            {"role": "system", "content": "You are a coding assistant."},
+            {"role": "user", "content": "Write an example"},
+        ],
+        "response_format": response_format,
+    })
+    if re_content is not None:
+        assert res.status_code == 200
+        choice = res.body["choices"][0]
+        assert match_regex(re_content, choice["message"]["content"])
+    else:
+        assert res.status_code != 200
+        assert "error" in res.body
+
diff --git a/examples/server/tests/unit/test_completion.py b/examples/server/tests/unit/test_completion.py
new file mode 100644
index 0000000000000..2fa30dd033431
--- /dev/null
+++ b/examples/server/tests/unit/test_completion.py
@@ -0,0 +1,223 @@
+import pytest
+import time
+from utils import *
+
+server = ServerPreset.tinyllama2()
+
+
+@pytest.fixture(scope="module", autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.tinyllama2()
+
+@pytest.mark.parametrize("prompt,n_predict,re_content,n_prompt,n_predicted,truncated", [
+    ("I believe the meaning of life is", 8, "(going|bed)+", 18, 8, False),
+    ("Write a joke about AI from a very long prompt which will not be truncated", 256, "(princesses|everyone|kids|Anna|forest)+", 46, 64, False),
+])
+def test_completion(prompt: str, n_predict: int, re_content: str, n_prompt: int, n_predicted: int, truncated: bool):
+    global server
+    server.start()
+    res = server.make_request("POST", "/completion", data={
+        "n_predict": n_predict,
+        "prompt": prompt,
+    })
+    assert res.status_code == 200
+    assert res.body["timings"]["prompt_n"] == n_prompt
+    assert res.body["timings"]["predicted_n"] == n_predicted
+    assert res.body["truncated"] == truncated
+    assert match_regex(re_content, res.body["content"])
+
+
+@pytest.mark.parametrize("prompt,n_predict,re_content,n_prompt,n_predicted,truncated", [
+    ("I believe the meaning of life is", 8, "(going|bed)+", 18, 8, False),
+    ("Write a joke about AI from a very long prompt which will not be truncated", 256, "(princesses|everyone|kids|Anna|forest)+", 46, 64, False),
+])
+def test_completion_stream(prompt: str, n_predict: int, re_content: str, n_prompt: int, n_predicted: int, truncated: bool):
+    global server
+    server.start()
+    res = server.make_stream_request("POST", "/completion", data={
+        "n_predict": n_predict,
+        "prompt": prompt,
+        "stream": True,
+    })
+    content = ""
+    for data in res:
+        if data["stop"]:
+            assert data["timings"]["prompt_n"] == n_prompt
+            assert data["timings"]["predicted_n"] == n_predicted
+            assert data["truncated"] == truncated
+            assert match_regex(re_content, content)
+        else:
+            content += data["content"]
+
+
+@pytest.mark.parametrize("n_slots", [1, 2])
+def test_consistent_result_same_seed(n_slots: int):
+    global server
+    server.n_slots = n_slots
+    server.start()
+    last_res = None
+    for _ in range(4):
+        res = server.make_request("POST", "/completion", data={
+            "prompt": "I believe the meaning of life is",
+            "seed": 42,
+            "temperature": 1.0,
+            "cache_prompt": False,  # TODO: remove this once test_cache_vs_nocache_prompt is fixed
+        })
+        if last_res is not None:
+            assert res.body["content"] == last_res.body["content"]
+        last_res = res
+
+
+@pytest.mark.parametrize("n_slots", [1, 2])
+def test_different_result_different_seed(n_slots: int):
+    global server
+    server.n_slots = n_slots
+    server.start()
+    last_res = None
+    for seed in range(4):
+        res = server.make_request("POST", "/completion", data={
+            "prompt": "I believe the meaning of life is",
+            "seed": seed,
+            "temperature": 1.0,
+            "cache_prompt": False,  # TODO: remove this once test_cache_vs_nocache_prompt is fixed
+        })
+        if last_res is not None:
+            assert res.body["content"] != last_res.body["content"]
+        last_res = res
+
+
+@pytest.mark.parametrize("n_batch", [16, 32])
+@pytest.mark.parametrize("temperature", [0.0, 1.0])
+def test_consistent_result_different_batch_size(n_batch: int, temperature: float):
+    global server
+    server.n_batch = n_batch
+    server.start()
+    last_res = None
+    for _ in range(4):
+        res = server.make_request("POST", "/completion", data={
+            "prompt": "I believe the meaning of life is",
+            "seed": 42,
+            "temperature": temperature,
+            "cache_prompt": False,  # TODO: remove this once test_cache_vs_nocache_prompt is fixed
+        })
+        if last_res is not None:
+            assert res.body["content"] == last_res.body["content"]
+        last_res = res
+
+
+@pytest.mark.skip(reason="This test fails on linux, need to be fixed")
+def test_cache_vs_nocache_prompt():
+    global server
+    server.start()
+    res_cache = server.make_request("POST", "/completion", data={
+        "prompt": "I believe the meaning of life is",
+        "seed": 42,
+        "temperature": 1.0,
+        "cache_prompt": True,
+    })
+    res_no_cache = server.make_request("POST", "/completion", data={
+        "prompt": "I believe the meaning of life is",
+        "seed": 42,
+        "temperature": 1.0,
+        "cache_prompt": False,
+    })
+    assert res_cache.body["content"] == res_no_cache.body["content"]
+
+
+def test_completion_with_tokens_input():
+    global server
+    server.temperature = 0.0
+    server.start()
+    prompt_str = "I believe the meaning of life is"
+    res = server.make_request("POST", "/tokenize", data={
+        "content": prompt_str,
+        "add_special": True,
+    })
+    assert res.status_code == 200
+    tokens = res.body["tokens"]
+
+    # single completion
+    res = server.make_request("POST", "/completion", data={
+        "prompt": tokens,
+    })
+    assert res.status_code == 200
+    assert type(res.body["content"]) == str
+
+    # batch completion
+    res = server.make_request("POST", "/completion", data={
+        "prompt": [tokens, tokens],
+    })
+    assert res.status_code == 200
+    assert type(res.body) == list
+    assert len(res.body) == 2
+    assert res.body[0]["content"] == res.body[1]["content"]
+
+    # mixed string and tokens
+    res = server.make_request("POST", "/completion", data={
+        "prompt": [tokens, prompt_str],
+    })
+    assert res.status_code == 200
+    assert type(res.body) == list
+    assert len(res.body) == 2
+    assert res.body[0]["content"] == res.body[1]["content"]
+
+    # mixed string and tokens in one sequence
+    res = server.make_request("POST", "/completion", data={
+        "prompt": [1, 2, 3, 4, 5, 6, prompt_str, 7, 8, 9, 10, prompt_str],
+    })
+    assert res.status_code == 200
+    assert type(res.body["content"]) == str
+
+
+@pytest.mark.parametrize("n_slots,n_requests", [
+    (1, 3),
+    (2, 2),
+    (2, 4),
+    (4, 2), # some slots must be idle
+    (4, 6),
+])
+def test_completion_parallel_slots(n_slots: int, n_requests: int):
+    global server
+    server.n_slots = n_slots
+    server.temperature = 0.0
+    server.start()
+
+    PROMPTS = [
+        ("Write a very long book.", "(very|special|big)+"),
+        ("Write another a poem.", "(small|house)+"),
+        ("What is LLM?", "(Dad|said)+"),
+        ("The sky is blue and I love it.", "(climb|leaf)+"),
+        ("Write another very long music lyrics.", "(friends|step|sky)+"),
+        ("Write a very long joke.", "(cat|Whiskers)+"),
+    ]
+    def check_slots_status():
+        should_all_slots_busy = n_requests >= n_slots
+        time.sleep(0.1)
+        res = server.make_request("GET", "/slots")
+        n_busy = sum([1 for slot in res.body if slot["is_processing"]])
+        if should_all_slots_busy:
+            assert n_busy == n_slots
+        else:
+            assert n_busy <= n_slots
+
+    tasks = []
+    for i in range(n_requests):
+        prompt, re_content = PROMPTS[i % len(PROMPTS)]
+        tasks.append((server.make_request, ("POST", "/completion", {
+            "prompt": prompt,
+            "seed": 42,
+            "temperature": 1.0,
+        })))
+    tasks.append((check_slots_status, ()))
+    results = parallel_function_calls(tasks)
+
+    # check results
+    for i in range(n_requests):
+        prompt, re_content = PROMPTS[i % len(PROMPTS)]
+        res = results[i]
+        assert res.status_code == 200
+        assert type(res.body["content"]) == str
+        assert len(res.body["content"]) > 10
+        # FIXME: the result is not deterministic when using other slot than slot 0
+        # assert match_regex(re_content, res.body["content"])
diff --git a/examples/server/tests/unit/test_ctx_shift.py b/examples/server/tests/unit/test_ctx_shift.py
new file mode 100644
index 0000000000000..be93a6d31f410
--- /dev/null
+++ b/examples/server/tests/unit/test_ctx_shift.py
@@ -0,0 +1,67 @@
+import pytest
+from utils import *
+
+server = ServerPreset.tinyllama2()
+
+
+LONG_TEXT = """
+Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
+Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.
+Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
+Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
+""".strip()
+
+@pytest.fixture(scope="module", autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.tinyllama2()
+    server.n_ctx = 256
+    server.n_slots = 2
+
+
+def test_ctx_shift_enabled():
+    # the prompt is 301 tokens
+    # the slot context is 256/2 = 128 tokens
+    # the prompt is truncated to keep the last 109 tokens
+    # 64 tokens are generated thanks to shifting the context when it gets full
+    global server
+    server.start()
+    res = server.make_request("POST", "/completion", data={
+        "n_predict": 64,
+        "prompt": LONG_TEXT,
+    })
+    assert res.status_code == 200
+    assert res.body["timings"]["prompt_n"] == 109
+    assert res.body["timings"]["predicted_n"] == 64
+    assert res.body["truncated"] is True
+
+
+@pytest.mark.parametrize("n_predict,n_token_output,truncated", [
+    (64, 64, False),
+    (-1, 120, True),
+])
+def test_ctx_shift_disabled_short_prompt(n_predict: int, n_token_output: int, truncated: bool):
+    global server
+    server.disable_ctx_shift = True
+    server.n_predict = -1
+    server.start()
+    res = server.make_request("POST", "/completion", data={
+        "n_predict": n_predict,
+        "prompt": "Hi how are you",
+    })
+    assert res.status_code == 200
+    assert res.body["timings"]["predicted_n"] == n_token_output
+    assert res.body["truncated"] == truncated
+
+
+def test_ctx_shift_disabled_long_prompt():
+    global server
+    server.disable_ctx_shift = True
+    server.start()
+    res = server.make_request("POST", "/completion", data={
+        "n_predict": 64,
+        "prompt": LONG_TEXT,
+    })
+    assert res.status_code != 200
+    assert "error" in res.body
+    assert "exceeds the available context size" in res.body["error"]["message"]
diff --git a/examples/server/tests/unit/test_embedding.py b/examples/server/tests/unit/test_embedding.py
new file mode 100644
index 0000000000000..fc7c20064ddfc
--- /dev/null
+++ b/examples/server/tests/unit/test_embedding.py
@@ -0,0 +1,99 @@
+import pytest
+from openai import OpenAI
+from utils import *
+
+server = ServerPreset.bert_bge_small()
+
+EPSILON = 1e-3
+
+@pytest.fixture(scope="module", autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.bert_bge_small()
+
+
+def test_embedding_single():
+    global server
+    server.start()
+    res = server.make_request("POST", "/embeddings", data={
+        "input": "I believe the meaning of life is",
+    })
+    assert res.status_code == 200
+    assert len(res.body['data']) == 1
+    assert 'embedding' in res.body['data'][0]
+    assert len(res.body['data'][0]['embedding']) > 1
+
+    # make sure embedding vector is normalized
+    assert abs(sum([x ** 2 for x in res.body['data'][0]['embedding']]) - 1) < EPSILON
+
+
+def test_embedding_multiple():
+    global server
+    server.start()
+    res = server.make_request("POST", "/embeddings", data={
+        "input": [
+            "I believe the meaning of life is",
+            "Write a joke about AI from a very long prompt which will not be truncated",
+            "This is a test",
+            "This is another test",
+        ],
+    })
+    assert res.status_code == 200
+    assert len(res.body['data']) == 4
+    for d in res.body['data']:
+        assert 'embedding' in d
+        assert len(d['embedding']) > 1
+
+
+def test_embedding_openai_library_single():
+    global server
+    server.start()
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
+    res = client.embeddings.create(model="text-embedding-3-small", input="I believe the meaning of life is")
+    assert len(res.data) == 1
+    assert len(res.data[0].embedding) > 1
+
+
+def test_embedding_openai_library_multiple():
+    global server
+    server.start()
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}")
+    res = client.embeddings.create(model="text-embedding-3-small", input=[
+        "I believe the meaning of life is",
+        "Write a joke about AI from a very long prompt which will not be truncated",
+        "This is a test",
+        "This is another test",
+    ])
+    assert len(res.data) == 4
+    for d in res.data:
+        assert len(d.embedding) > 1
+
+
+def test_embedding_error_prompt_too_long():
+    global server
+    server.start()
+    res = server.make_request("POST", "/embeddings", data={
+        "input": "This is a test " * 512,
+    })
+    assert res.status_code != 200
+    assert "too large" in res.body["error"]["message"]
+
+
+def test_same_prompt_give_same_result():
+    server.start()
+    res = server.make_request("POST", "/embeddings", data={
+        "input": [
+            "I believe the meaning of life is",
+            "I believe the meaning of life is",
+            "I believe the meaning of life is",
+            "I believe the meaning of life is",
+            "I believe the meaning of life is",
+        ],
+    })
+    assert res.status_code == 200
+    assert len(res.body['data']) == 5
+    for i in range(1, len(res.body['data'])):
+        v0 = res.body['data'][0]['embedding']
+        vi = res.body['data'][i]['embedding']
+        for x, y in zip(v0, vi):
+            assert abs(x - y) < EPSILON
diff --git a/examples/server/tests/unit/test_infill.py b/examples/server/tests/unit/test_infill.py
new file mode 100644
index 0000000000000..38ce6c42954ed
--- /dev/null
+++ b/examples/server/tests/unit/test_infill.py
@@ -0,0 +1,35 @@
+import pytest
+from utils import *
+
+server = ServerPreset.tinyllama_infill()
+
+@pytest.fixture(scope="module", autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.tinyllama_infill()
+
+def test_infill_without_input_extra():
+    global server
+    server.start()
+    res = server.make_request("POST", "/infill", data={
+        "prompt": "Complete this",
+        "input_prefix": "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_",
+        "input_suffix": "}\n",
+    })
+    assert res.status_code == 200
+    assert match_regex("(One|day|she|saw|big|scary|bird)+", res.body["content"])
+
+def test_infill_with_input_extra():
+    global server
+    server.start()
+    res = server.make_request("POST", "/infill", data={
+        "prompt": "Complete this",
+        "input_extra": [{
+            "filename": "llama.h",
+            "text": "LLAMA_API int32_t llama_n_threads();\n"
+        }],
+        "input_prefix": "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_",
+        "input_suffix": "}\n",
+    })
+    assert res.status_code == 200
+    assert match_regex("(cuts|Jimmy|mom|came|into|the|room)+", res.body["content"])
diff --git a/examples/server/tests/unit/test_lora.py b/examples/server/tests/unit/test_lora.py
new file mode 100644
index 0000000000000..7496154493917
--- /dev/null
+++ b/examples/server/tests/unit/test_lora.py
@@ -0,0 +1,42 @@
+import pytest
+import os
+from utils import *
+
+server = ServerPreset.stories15m_moe()
+
+LORA_FILE_URL = "https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe_shakespeare15M.gguf"
+
+@pytest.fixture(scope="module", autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.stories15m_moe()
+    # download lora file if needed
+    file_name = LORA_FILE_URL.split('/').pop()
+    lora_file = f'../../../{file_name}'
+    if not os.path.exists(lora_file):
+        print(f"Downloading {LORA_FILE_URL} to {lora_file}")
+        with open(lora_file, 'wb') as f:
+            f.write(requests.get(LORA_FILE_URL).content)
+        print(f"Done downloading lora file")
+    server.lora_files = [lora_file]
+
+
+@pytest.mark.parametrize("scale,re_content", [
+    # without applying lora, the model should behave like a bedtime story generator
+    (0.0, "(little|girl|three|years|old)+"),
+    # with lora, the model should behave like a Shakespearean text generator
+    (1.0, "(eye|love|glass|sun)+"),
+])
+def test_lora(scale: float, re_content: str):
+    global server
+    server.start()
+    res_lora_control = server.make_request("POST", "/lora-adapters", data=[
+        {"id": 0, "scale": scale}
+    ])
+    assert res_lora_control.status_code == 200
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "Look in thy glass",
+    })
+    assert res.status_code == 200
+    assert match_regex(re_content, res.body["content"])
+
diff --git a/examples/server/tests/unit/test_rerank.py b/examples/server/tests/unit/test_rerank.py
new file mode 100644
index 0000000000000..3a49fd3ac6bdf
--- /dev/null
+++ b/examples/server/tests/unit/test_rerank.py
@@ -0,0 +1,38 @@
+import pytest
+from utils import *
+
+server = ServerPreset.jina_reranker_tiny()
+
+
+@pytest.fixture(scope="module", autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.jina_reranker_tiny()
+
+
+def test_rerank():
+    global server
+    server.start()
+    res = server.make_request("POST", "/rerank", data={
+        "query": "Machine learning is",
+        "documents": [
+            "A machine is a physical system that uses power to apply forces and control movement to perform an action. The term is commonly applied to artificial devices, such as those employing engines or motors, but also to natural biological macromolecules, such as molecular machines.",
+            "Learning is the process of acquiring new understanding, knowledge, behaviors, skills, values, attitudes, and preferences. The ability to learn is possessed by humans, non-human animals, and some machines; there is also evidence for some kind of learning in certain plants.",
+            "Machine learning is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions.",
+            "Paris, capitale de la France, est une grande ville européenne et un centre mondial de l'art, de la mode, de la gastronomie et de la culture. Son paysage urbain du XIXe siècle est traversé par de larges boulevards et la Seine."
+        ]
+    })
+    assert res.status_code == 200
+    assert len(res.body["results"]) == 4
+
+    most_relevant = res.body["results"][0]
+    least_relevant = res.body["results"][0]
+    for doc in res.body["results"]:
+        if doc["relevance_score"] > most_relevant["relevance_score"]:
+            most_relevant = doc
+        if doc["relevance_score"] < least_relevant["relevance_score"]:
+            least_relevant = doc
+
+    assert most_relevant["relevance_score"] > least_relevant["relevance_score"]
+    assert most_relevant["index"] == 2
+    assert least_relevant["index"] == 3
diff --git a/examples/server/tests/unit/test_security.py b/examples/server/tests/unit/test_security.py
new file mode 100644
index 0000000000000..620b25376bd81
--- /dev/null
+++ b/examples/server/tests/unit/test_security.py
@@ -0,0 +1,83 @@
+import pytest
+from openai import OpenAI
+from utils import *
+
+server = ServerPreset.tinyllama2()
+
+TEST_API_KEY = "sk-this-is-the-secret-key"
+
+@pytest.fixture(scope="module", autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.tinyllama2()
+    server.api_key = TEST_API_KEY
+
+
+@pytest.mark.parametrize("endpoint", ["/health", "/models"])
+def test_access_public_endpoint(endpoint: str):
+    global server
+    server.start()
+    res = server.make_request("GET", endpoint)
+    assert res.status_code == 200
+    assert "error" not in res.body
+
+
+@pytest.mark.parametrize("api_key", [None, "invalid-key"])
+def test_incorrect_api_key(api_key: str):
+    global server
+    server.start()
+    res = server.make_request("POST", "/completions", data={
+        "prompt": "I believe the meaning of life is",
+    }, headers={
+        "Authorization": f"Bearer {api_key}" if api_key else None,
+    })
+    assert res.status_code == 401
+    assert "error" in res.body
+    assert res.body["error"]["type"] == "authentication_error"
+
+
+def test_correct_api_key():
+    global server
+    server.start()
+    res = server.make_request("POST", "/completions", data={
+        "prompt": "I believe the meaning of life is",
+    }, headers={
+        "Authorization": f"Bearer {TEST_API_KEY}",
+    })
+    assert res.status_code == 200
+    assert "error" not in res.body
+    assert "content" in res.body
+
+
+def test_openai_library_correct_api_key():
+    global server
+    server.start()
+    client = OpenAI(api_key=TEST_API_KEY, base_url=f"http://{server.server_host}:{server.server_port}")
+    res = client.chat.completions.create(
+        model="gpt-3.5-turbo",
+        messages=[
+            {"role": "system", "content": "You are a chatbot."},
+            {"role": "user", "content": "What is the meaning of life?"},
+        ],
+    )
+    assert len(res.choices) == 1
+
+
+@pytest.mark.parametrize("origin,cors_header,cors_header_value", [
+    ("localhost", "Access-Control-Allow-Origin", "localhost"),
+    ("web.mydomain.fr", "Access-Control-Allow-Origin", "web.mydomain.fr"),
+    ("origin", "Access-Control-Allow-Credentials", "true"),
+    ("web.mydomain.fr", "Access-Control-Allow-Methods", "GET, POST"),
+    ("web.mydomain.fr", "Access-Control-Allow-Headers", "*"),
+])
+def test_cors_options(origin: str, cors_header: str, cors_header_value: str):
+    global server
+    server.start()
+    res = server.make_request("OPTIONS", "/completions", headers={
+        "Origin": origin,
+        "Access-Control-Request-Method": "POST",
+        "Access-Control-Request-Headers": "Authorization",
+    })
+    assert res.status_code == 200
+    assert cors_header in res.headers
+    assert res.headers[cors_header] == cors_header_value
diff --git a/examples/server/tests/unit/test_slot_save.py b/examples/server/tests/unit/test_slot_save.py
new file mode 100644
index 0000000000000..38704f5ece35a
--- /dev/null
+++ b/examples/server/tests/unit/test_slot_save.py
@@ -0,0 +1,98 @@
+import pytest
+from utils import *
+
+server = ServerPreset.tinyllama2()
+
+@pytest.fixture(scope="module", autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.tinyllama2()
+    server.slot_save_path = "./tmp"
+    server.temperature = 0.0
+
+
+def test_slot_save_restore():
+    global server
+    server.start()
+
+    # First prompt in slot 1 should be fully processed
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "What is the capital of France?",
+        "id_slot": 1,
+        "cache_prompt": True,
+    })
+    assert res.status_code == 200
+    assert match_regex("(Whiskers|Flana)+", res.body["content"])
+    assert res.body["timings"]["prompt_n"] == 21  # all tokens are processed
+
+    # Save state of slot 1
+    res = server.make_request("POST", "/slots/1?action=save", data={
+        "filename": "slot1.bin",
+    })
+    assert res.status_code == 200
+    assert res.body["n_saved"] == 84
+
+    # Since we have cache, this should only process the last tokens
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "What is the capital of Germany?",
+        "id_slot": 1,
+        "cache_prompt": True,
+    })
+    assert res.status_code == 200
+    assert match_regex("(Jack|said)+", res.body["content"])
+    assert res.body["timings"]["prompt_n"] == 6  # only different part is processed
+
+    # Loading the saved cache into slot 0
+    res = server.make_request("POST", "/slots/0?action=restore", data={
+        "filename": "slot1.bin",
+    })
+    assert res.status_code == 200
+    assert res.body["n_restored"] == 84
+
+    # Since we have cache, slot 0 should only process the last tokens
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "What is the capital of Germany?",
+        "id_slot": 0,
+        "cache_prompt": True,
+    })
+    assert res.status_code == 200
+    assert match_regex("(Jack|said)+", res.body["content"])
+    assert res.body["timings"]["prompt_n"] == 6  # only different part is processed
+
+    # For verification that slot 1 was not corrupted during slot 0 load, same thing should work
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "What is the capital of Germany?",
+        "id_slot": 1,
+        "cache_prompt": True,
+    })
+    assert res.status_code == 200
+    assert match_regex("(Jack|said)+", res.body["content"])
+    assert res.body["timings"]["prompt_n"] == 1
+
+
+def test_slot_erase():
+    global server
+    server.start()
+
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "What is the capital of France?",
+        "id_slot": 1,
+        "cache_prompt": True,
+    })
+    assert res.status_code == 200
+    assert match_regex("(Whiskers|Flana)+", res.body["content"])
+    assert res.body["timings"]["prompt_n"] == 21  # all tokens are processed
+
+    # erase slot 1
+    res = server.make_request("POST", "/slots/1?action=erase")
+    assert res.status_code == 200
+
+    # re-run the same prompt, it should process all tokens again
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "What is the capital of France?",
+        "id_slot": 1,
+        "cache_prompt": True,
+    })
+    assert res.status_code == 200
+    assert match_regex("(Whiskers|Flana)+", res.body["content"])
+    assert res.body["timings"]["prompt_n"] == 21  # all tokens are processed
diff --git a/examples/server/tests/unit/test_tokenize.py b/examples/server/tests/unit/test_tokenize.py
new file mode 100644
index 0000000000000..382457c9d602f
--- /dev/null
+++ b/examples/server/tests/unit/test_tokenize.py
@@ -0,0 +1,59 @@
+import pytest
+from utils import *
+
+server = ServerPreset.tinyllama2()
+
+
+@pytest.fixture(scope="module", autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.tinyllama2()
+
+
+def test_tokenize_detokenize():
+    global server
+    server.start()
+    # tokenize
+    content = "What is the capital of France ?"
+    res_tok = server.make_request("POST", "/tokenize", data={
+        "content": content
+    })
+    assert res_tok.status_code == 200
+    assert len(res_tok.body["tokens"]) > 5
+    # detokenize
+    res_detok = server.make_request("POST", "/detokenize", data={
+        "tokens": res_tok.body["tokens"],
+    })
+    assert res_detok.status_code == 200
+    assert res_detok.body["content"].strip() == content
+
+
+def test_tokenize_with_bos():
+    global server
+    server.start()
+    # tokenize
+    content = "What is the capital of France ?"
+    bosId = 1
+    res_tok = server.make_request("POST", "/tokenize", data={
+        "content": content,
+        "add_special": True,
+    })
+    assert res_tok.status_code == 200
+    assert res_tok.body["tokens"][0] == bosId
+
+
+def test_tokenize_with_pieces():
+    global server
+    server.start()
+    # tokenize
+    content = "This is a test string with unicode 媽 and emoji 🤗"
+    res_tok = server.make_request("POST", "/tokenize", data={
+        "content": content,
+        "with_pieces": True,
+    })
+    assert res_tok.status_code == 200
+    for token in res_tok.body["tokens"]:
+        assert "id" in token
+        assert token["id"] > 0
+        assert "piece" in token
+        assert len(token["piece"]) > 0
diff --git a/examples/server/tests/utils.py b/examples/server/tests/utils.py
new file mode 100644
index 0000000000000..bc590bcb31547
--- /dev/null
+++ b/examples/server/tests/utils.py
@@ -0,0 +1,377 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+# type: ignore[reportUnusedImport]
+
+import subprocess
+import os
+import re
+import json
+import sys
+import threading
+import requests
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from typing import (
+    Any,
+    Callable,
+    ContextManager,
+    Iterable,
+    Iterator,
+    List,
+    Literal,
+    Tuple,
+    Set,
+)
+from re import RegexFlag
+
+
+class ServerResponse:
+    headers: dict
+    status_code: int
+    body: dict | Any
+
+
+class ServerProcess:
+    # default options
+    debug: bool = False
+    server_port: int = 8080
+    server_host: str = "127.0.0.1"
+    model_hf_repo: str = "ggml-org/models"
+    model_hf_file: str = "tinyllamas/stories260K.gguf"
+    model_alias: str = "tinyllama-2"
+    temperature: float = 0.8
+    seed: int = 42
+
+    # custom options
+    model_alias: str | None = None
+    model_url: str | None = None
+    model_file: str | None = None
+    n_threads: int | None = None
+    n_gpu_layer: int | None = None
+    n_batch: int | None = None
+    n_ubatch: int | None = None
+    n_ctx: int | None = None
+    n_ga: int | None = None
+    n_ga_w: int | None = None
+    n_predict: int | None = None
+    n_prompts: int | None = 0
+    slot_save_path: str | None = None
+    id_slot: int | None = None
+    cache_prompt: bool | None = None
+    n_slots: int | None = None
+    server_continuous_batching: bool | None = False
+    server_embeddings: bool | None = False
+    server_reranking: bool | None = False
+    server_metrics: bool | None = False
+    draft: int | None = None
+    api_key: str | None = None
+    response_format: str | None = None
+    lora_files: List[str] | None = None
+    disable_ctx_shift: int | None = False
+
+    # session variables
+    process: subprocess.Popen | None = None
+
+    def __init__(self):
+        if "N_GPU_LAYERS" in os.environ:
+            self.n_gpu_layer = int(os.environ["N_GPU_LAYERS"])
+        if "DEBUG" in os.environ:
+            self.debug = True
+        if "PORT" in os.environ:
+            self.server_port = int(os.environ["PORT"])
+
+    def start(self, timeout_seconds: int = 10) -> None:
+        if "LLAMA_SERVER_BIN_PATH" in os.environ:
+            server_path = os.environ["LLAMA_SERVER_BIN_PATH"]
+        elif os.name == "nt":
+            server_path = "../../../build/bin/Release/llama-server.exe"
+        else:
+            server_path = "../../../build/bin/llama-server"
+        server_args = [
+            "--slots",  # requires to get slot status via /slots endpoint
+            "--host",
+            self.server_host,
+            "--port",
+            self.server_port,
+            "--temp",
+            self.temperature,
+            "--seed",
+            self.seed,
+        ]
+        if self.model_file:
+            server_args.extend(["--model", self.model_file])
+        if self.model_url:
+            server_args.extend(["--model-url", self.model_url])
+        if self.model_hf_repo:
+            server_args.extend(["--hf-repo", self.model_hf_repo])
+        if self.model_hf_file:
+            server_args.extend(["--hf-file", self.model_hf_file])
+        if self.n_batch:
+            server_args.extend(["--batch-size", self.n_batch])
+        if self.n_ubatch:
+            server_args.extend(["--ubatch-size", self.n_ubatch])
+        if self.n_threads:
+            server_args.extend(["--threads", self.n_threads])
+        if self.n_gpu_layer:
+            server_args.extend(["--n-gpu-layers", self.n_gpu_layer])
+        if self.draft is not None:
+            server_args.extend(["--draft", self.draft])
+        if self.server_continuous_batching:
+            server_args.append("--cont-batching")
+        if self.server_embeddings:
+            server_args.append("--embedding")
+        if self.server_reranking:
+            server_args.append("--reranking")
+        if self.server_metrics:
+            server_args.append("--metrics")
+        if self.model_alias:
+            server_args.extend(["--alias", self.model_alias])
+        if self.n_ctx:
+            server_args.extend(["--ctx-size", self.n_ctx])
+        if self.n_slots:
+            server_args.extend(["--parallel", self.n_slots])
+        if self.n_predict:
+            server_args.extend(["--n-predict", self.n_predict])
+        if self.slot_save_path:
+            server_args.extend(["--slot-save-path", self.slot_save_path])
+        if self.n_ga:
+            server_args.extend(["--grp-attn-n", self.n_ga])
+        if self.n_ga_w:
+            server_args.extend(["--grp-attn-w", self.n_ga_w])
+        if self.debug:
+            server_args.append("--verbose")
+        if self.lora_files:
+            for lora_file in self.lora_files:
+                server_args.extend(["--lora", lora_file])
+        if self.disable_ctx_shift:
+            server_args.extend(["--no-context-shift"])
+        if self.api_key:
+            server_args.extend(["--api-key", self.api_key])
+
+        args = [str(arg) for arg in [server_path, *server_args]]
+        print(f"bench: starting server with: {' '.join(args)}")
+
+        flags = 0
+        if "nt" == os.name:
+            flags |= subprocess.DETACHED_PROCESS
+            flags |= subprocess.CREATE_NEW_PROCESS_GROUP
+            flags |= subprocess.CREATE_NO_WINDOW
+
+        self.process = subprocess.Popen(
+            [str(arg) for arg in [server_path, *server_args]],
+            creationflags=flags,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            env={**os.environ, "LLAMA_CACHE": "tmp"},
+        )
+        server_instances.add(self)
+
+        def server_log(in_stream, out_stream):
+            for line in iter(in_stream.readline, b""):
+                print(line.decode("utf-8"), end="", file=out_stream)
+
+        thread_stdout = threading.Thread(
+            target=server_log, args=(self.process.stdout, sys.stdout), daemon=True
+        )
+        thread_stdout.start()
+
+        thread_stderr = threading.Thread(
+            target=server_log, args=(self.process.stderr, sys.stderr), daemon=True
+        )
+        thread_stderr.start()
+
+        print(f"server pid={self.process.pid}, pytest pid={os.getpid()}")
+
+        # wait for server to start
+        start_time = time.time()
+        while time.time() - start_time < timeout_seconds:
+            try:
+                response = self.make_request("GET", "/slots", headers={
+                    "Authorization": f"Bearer {self.api_key}" if self.api_key else None
+                })
+                if response.status_code == 200:
+                    self.ready = True
+                    return  # server is ready
+            except Exception as e:
+                pass
+            print(f"Waiting for server to start...")
+            time.sleep(0.5)
+        raise TimeoutError(f"Server did not start within {timeout_seconds} seconds")
+
+    def stop(self) -> None:
+        server_instances.remove(self)
+        if self.process:
+            print(f"Stopping server with pid={self.process.pid}")
+            self.process.kill()
+            self.process = None
+
+    def make_request(
+        self,
+        method: str,
+        path: str,
+        data: dict | Any | None = None,
+        headers: dict | None = None,
+    ) -> ServerResponse:
+        url = f"http://{self.server_host}:{self.server_port}{path}"
+        parse_body = False
+        if method == "GET":
+            response = requests.get(url, headers=headers)
+            parse_body = True
+        elif method == "POST":
+            response = requests.post(url, headers=headers, json=data)
+            parse_body = True
+        elif method == "OPTIONS":
+            response = requests.options(url, headers=headers)
+        else:
+            raise ValueError(f"Unimplemented method: {method}")
+        result = ServerResponse()
+        result.headers = dict(response.headers)
+        result.status_code = response.status_code
+        result.body = response.json() if parse_body else None
+        print("Response from server", result.body)
+        return result
+
+    def make_stream_request(
+        self,
+        method: str,
+        path: str,
+        data: dict | None = None,
+        headers: dict | None = None,
+    ) -> Iterator[dict]:
+        url = f"http://{self.server_host}:{self.server_port}{path}"
+        if method == "POST":
+            response = requests.post(url, headers=headers, json=data, stream=True)
+        else:
+            raise ValueError(f"Unimplemented method: {method}")
+        for line_bytes in response.iter_lines():
+            line = line_bytes.decode("utf-8")
+            if '[DONE]' in line:
+                break
+            elif line.startswith('data: '):
+                data = json.loads(line[6:])
+                print("Partial response from server", data)
+                yield data
+
+
+server_instances: Set[ServerProcess] = set()
+
+
+class ServerPreset:
+    @staticmethod
+    def tinyllama2() -> ServerProcess:
+        server = ServerProcess()
+        server.model_hf_repo = "ggml-org/models"
+        server.model_hf_file = "tinyllamas/stories260K.gguf"
+        server.model_alias = "tinyllama-2"
+        server.n_ctx = 256
+        server.n_batch = 32
+        server.n_slots = 2
+        server.n_predict = 64
+        server.seed = 42
+        return server
+
+    @staticmethod
+    def bert_bge_small() -> ServerProcess:
+        server = ServerProcess()
+        server.model_hf_repo = "ggml-org/models"
+        server.model_hf_file = "bert-bge-small/ggml-model-f16.gguf"
+        server.model_alias = "bert-bge-small"
+        server.n_ctx = 512
+        server.n_batch = 128
+        server.n_ubatch = 128
+        server.n_slots = 2
+        server.seed = 42
+        server.server_embeddings = True
+        return server
+
+    @staticmethod
+    def tinyllama_infill() -> ServerProcess:
+        server = ServerProcess()
+        server.model_hf_repo = "ggml-org/models"
+        server.model_hf_file = "tinyllamas/stories260K-infill.gguf"
+        server.model_alias = "tinyllama-infill"
+        server.n_ctx = 2048
+        server.n_batch = 1024
+        server.n_slots = 1
+        server.n_predict = 64
+        server.temperature = 0.0
+        server.seed = 42
+        return server
+
+    @staticmethod
+    def stories15m_moe() -> ServerProcess:
+        server = ServerProcess()
+        server.model_hf_repo = "ggml-org/stories15M_MOE"
+        server.model_hf_file = "stories15M_MOE-F16.gguf"
+        server.model_alias = "stories15m-moe"
+        server.n_ctx = 2048
+        server.n_batch = 1024
+        server.n_slots = 1
+        server.n_predict = 64
+        server.temperature = 0.0
+        server.seed = 42
+        return server
+
+    @staticmethod
+    def jina_reranker_tiny() -> ServerProcess:
+        server = ServerProcess()
+        server.model_hf_repo = "ggml-org/models"
+        server.model_hf_file = "jina-reranker-v1-tiny-en/ggml-model-f16.gguf"
+        server.model_alias = "jina-reranker"
+        server.model_file = "./tmp/jina-reranker-v1-tiny-en.gguf"
+        server.n_ctx = 512
+        server.n_batch = 512
+        server.n_slots = 1
+        server.seed = 42
+        server.server_reranking = True
+        return server
+
+
+def parallel_function_calls(function_list: List[Tuple[Callable[..., Any], Tuple[Any, ...]]]) -> List[Any]:
+    """
+    Run multiple functions in parallel and return results in the same order as calls. Equivalent to Promise.all in JS.
+
+    Example usage:
+
+    results = parallel_function_calls([
+        (func1, (arg1, arg2)),
+        (func2, (arg3, arg4)),
+    ])
+    """
+    results = [None] * len(function_list)
+    exceptions = []
+
+    def worker(index, func, args):
+        try:
+            result = func(*args)
+            results[index] = result
+        except Exception as e:
+            exceptions.append((index, str(e)))
+
+    with ThreadPoolExecutor() as executor:
+        futures = []
+        for i, (func, args) in enumerate(function_list):
+            future = executor.submit(worker, i, func, args)
+            futures.append(future)
+
+        # Wait for all futures to complete
+        for future in as_completed(futures):
+            pass
+
+    # Check if there were any exceptions
+    if exceptions:
+        print("Exceptions occurred:")
+        for index, error in exceptions:
+            print(f"Function at index {index}: {error}")
+
+    return results
+
+
+def match_regex(regex: str, text: str) -> bool:
+    return (
+        re.compile(
+            regex, flags=RegexFlag.IGNORECASE | RegexFlag.MULTILINE | RegexFlag.DOTALL
+        ).search(text)
+        is not None
+    )

From 904109ed0d97c9b656a5e8bf612925f739bb8166 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Tue, 26 Nov 2024 09:45:05 -0600
Subject: [PATCH 087/102] vulkan: fix group_norm (#10496)

Fix bad calculation of the end of the range. Add a backend test that
covers the bad case (taken from stable diffusion).

Fixes https://github.com/leejet/stable-diffusion.cpp/issues/439.
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp                | 2 +-
 ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp | 2 +-
 tests/test-backend-ops.cpp                          | 3 ++-
 3 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 49527fdf40e94..da1cfd24eac22 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -7157,7 +7157,7 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         const int32_t max_period = tensor->op_params[1];
         tensor_clone = ggml_timestep_embedding(ggml_ctx, src0_clone, dim, max_period);
     } else if (tensor->op == GGML_OP_POOL_2D) {
-        enum ggml_op_pool op = static_cast<ggml_op_pool>(dst->op_params[0]);
+        enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
         const int32_t k0 = tensor->op_params[1];
         const int32_t k1 = tensor->op_params[2];
         const int32_t s0 = tensor->op_params[3];
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp
index 5ad9b28daffaa..b6a0d56454951 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp
@@ -19,7 +19,7 @@ void main() {
 
     const uint tid = gl_LocalInvocationID.x;
     const uint start = gl_WorkGroupID.x * group_size + tid;
-    const uint end = start + group_size;
+    const uint end = (gl_WorkGroupID.x + 1) * group_size;
 
     tmp[tid] = 0.0f;
 
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 6376b0e4c66cf..da66ed8566ed9 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -3796,7 +3796,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_upscale());
     test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, { 512, 512, 3, 1 }, 2, true));
     test_cases.emplace_back(new test_upscale_ext());
-    test_cases.emplace_back(new test_group_norm());
+    test_cases.emplace_back(new test_group_norm(GGML_TYPE_F32, {64, 64, 320, 1}));
+    test_cases.emplace_back(new test_group_norm(GGML_TYPE_F32, {9, 9, 1280, 1}));
     test_cases.emplace_back(new test_acc());
     test_cases.emplace_back(new test_pad());
     test_cases.emplace_back(new test_arange());

From 249cd93da3df9c8fa78869b0522526d1625aca91 Mon Sep 17 00:00:00 2001
From: R0CKSTAR <xiaodong.ye@mthreads.com>
Date: Wed, 27 Nov 2024 00:00:41 +0800
Subject: [PATCH 088/102] mtgpu: Add MUSA_DOCKER_ARCH in Dockerfiles && update
 cmake and make (#10516)

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
---
 .devops/full-musa.Dockerfile         |  9 ++++++++-
 .devops/llama-cli-musa.Dockerfile    |  9 ++++++++-
 .devops/llama-server-musa.Dockerfile |  9 ++++++++-
 Makefile                             |  9 +++++----
 ggml/src/ggml-musa/CMakeLists.txt    | 11 ++++++++++-
 5 files changed, 39 insertions(+), 8 deletions(-)

diff --git a/.devops/full-musa.Dockerfile b/.devops/full-musa.Dockerfile
index 575e81b486d13..3193fea1e9ae5 100644
--- a/.devops/full-musa.Dockerfile
+++ b/.devops/full-musa.Dockerfile
@@ -6,6 +6,9 @@ ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_V
 
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 
+# MUSA architecture to build for (defaults to all supported archs)
+ARG MUSA_DOCKER_ARCH=default
+
 RUN apt-get update && \
     apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
 
@@ -19,7 +22,11 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+# Use the default MUSA archs if not specified
+RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
+        export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
+    fi && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc) && \
     cp build/bin/* .
 
diff --git a/.devops/llama-cli-musa.Dockerfile b/.devops/llama-cli-musa.Dockerfile
index 3372749bee70e..e7c75af20e265 100644
--- a/.devops/llama-cli-musa.Dockerfile
+++ b/.devops/llama-cli-musa.Dockerfile
@@ -8,6 +8,9 @@ ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU
 
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 
+# MUSA architecture to build for (defaults to all supported archs)
+ARG MUSA_DOCKER_ARCH=default
+
 RUN apt-get update && \
     apt-get install -y build-essential git cmake
 
@@ -15,7 +18,11 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+# Use the default MUSA archs if not specified
+RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
+        export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
+    fi && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release --target llama-cli -j$(nproc) && \
     mkdir -p /app/lib && \
     find build -name "*.so" -exec cp {} /app/lib \;
diff --git a/.devops/llama-server-musa.Dockerfile b/.devops/llama-server-musa.Dockerfile
index eb67201c18775..cebe51d42fa95 100644
--- a/.devops/llama-server-musa.Dockerfile
+++ b/.devops/llama-server-musa.Dockerfile
@@ -8,6 +8,9 @@ ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU
 
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 
+# MUSA architecture to build for (defaults to all supported archs)
+ARG MUSA_DOCKER_ARCH=default
+
 RUN apt-get update && \
     apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
@@ -15,7 +18,11 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+# Use the default MUSA archs if not specified
+RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
+        export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
+    fi && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release --target llama-server -j$(nproc) && \
     mkdir -p /app/lib && \
     find build -name "*.so" -exec cp {} /app/lib \;
diff --git a/Makefile b/Makefile
index cfc74c1dc71b2..9a079a2377205 100644
--- a/Makefile
+++ b/Makefile
@@ -815,7 +815,7 @@ ifdef GGML_MUSA
 	else
 		MUSA_PATH ?= /opt/musa
 	endif
-	MTGPU_TARGETS ?= mp_21 mp_22
+	MUSA_ARCHITECTURES ?= 21;22
 
 	MK_CPPFLAGS += -DGGML_USE_MUSA -DGGML_USE_CUDA
 	MK_LDFLAGS += -L$(MUSA_PATH)/lib -Wl,-rpath=$(MUSA_PATH)/lib
@@ -834,7 +834,8 @@ ifdef GGML_MUSA
 	CXX := $(MUSA_PATH)/bin/clang++
 	MCC := $(CCACHE) $(MUSA_PATH)/bin/mcc
 
-	MUSAFLAGS += $(addprefix --cuda-gpu-arch=, $(MTGPU_TARGETS))
+	MUSAFLAGS  = -x musa -mtgpu
+	MUSAFLAGS += $(foreach arch,$(subst ;, ,$(MUSA_ARCHITECTURES)),--cuda-gpu-arch=mp_$(arch))
 
 ifdef GGML_CUDA_FORCE_MMQ
 	MUSAFLAGS += -DGGML_CUDA_FORCE_MMQ
@@ -878,14 +879,14 @@ ggml/src/ggml-cuda/ggml-cuda.o: \
 	ggml/src/ggml-backend-impl.h \
 	ggml/src/ggml-common.h \
 	$(wildcard ggml/src/ggml-cuda/*.cuh)
-	$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -x musa -mtgpu -c -o $@ $<
+	$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -c -o $@ $<
 
 ggml/src/ggml-cuda/%.o: \
 	ggml/src/ggml-cuda/%.cu \
 	ggml/include/ggml.h \
 	ggml/src/ggml-common.h \
 	ggml/src/ggml-cuda/common.cuh
-	$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -x musa -mtgpu -c -o $@ $<
+	$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -c -o $@ $<
 endif # GGML_MUSA
 
 ifdef GGML_METAL
diff --git a/ggml/src/ggml-musa/CMakeLists.txt b/ggml/src/ggml-musa/CMakeLists.txt
index e1a69186e669f..415b2b2e09c1b 100644
--- a/ggml/src/ggml-musa/CMakeLists.txt
+++ b/ggml/src/ggml-musa/CMakeLists.txt
@@ -20,6 +20,11 @@ find_package(MUSAToolkit)
 if (MUSAToolkit_FOUND)
     message(STATUS "MUSA Toolkit found")
 
+    if (NOT DEFINED MUSA_ARCHITECTURES)
+        set(MUSA_ARCHITECTURES "21;22")
+    endif()
+    message(STATUS "Using MUSA architectures: ${MUSA_ARCHITECTURES}")
+
     file(GLOB   GGML_HEADERS_MUSA "../ggml-cuda/*.cuh")
     list(APPEND GGML_HEADERS_MUSA "../../include/ggml-cuda.h")
 
@@ -44,7 +49,11 @@ if (MUSAToolkit_FOUND)
 
     set_source_files_properties(${GGML_SOURCES_MUSA} PROPERTIES LANGUAGE CXX)
     foreach(SOURCE ${GGML_SOURCES_MUSA})
-        set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22")
+        set(COMPILE_FLAGS "-x musa -mtgpu")
+        foreach(ARCH ${MUSA_ARCHITECTURES})
+            set(COMPILE_FLAGS "${COMPILE_FLAGS} --cuda-gpu-arch=mp_${ARCH}")
+        endforeach()
+        set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS ${COMPILE_FLAGS})
     endforeach()
 
     ggml_add_backend_library(ggml-musa

From be0e350c8b69632b27d5fb41fa064fa256dd7fbf Mon Sep 17 00:00:00 2001
From: Tristan Druyen <tristan@vault81.mozmail.com>
Date: Tue, 26 Nov 2024 19:27:28 +0100
Subject: [PATCH 089/102] Fix HIP flag inconsistency & build docs (#10524)

* Fix inconsistency of HIP flags in cmake & make

* Fix docs regarding GGML_HIP
---
 Makefile      | 4 ++--
 docs/build.md | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/Makefile b/Makefile
index 9a079a2377205..25214ec053df1 100644
--- a/Makefile
+++ b/Makefile
@@ -752,7 +752,7 @@ vulkan-shaders-gen: ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
 
 endif # GGML_VULKAN
 
-ifdef GGML_HIPBLAS
+ifdef GGML_HIP
 	ifeq ($(wildcard /opt/rocm),)
 		ROCM_PATH      ?= /usr
 		AMDGPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
@@ -807,7 +807,7 @@ ggml/src/ggml-cuda/%.o: \
 	ggml/src/ggml-common.h \
 	ggml/src/ggml-cuda/common.cuh
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
-endif # GGML_HIPBLAS
+endif # GGML_HIP
 
 ifdef GGML_MUSA
 	ifeq ($(wildcard /opt/musa),)
diff --git a/docs/build.md b/docs/build.md
index 359952b30ff93..72b81043796dc 100644
--- a/docs/build.md
+++ b/docs/build.md
@@ -221,7 +221,7 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 
 - Using `make`:
   ```bash
-  make GGML_HIPBLAS=1
+  make GGML_HIP=1
   ```
 - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
   ```bash
@@ -249,7 +249,7 @@ You can download it from your Linux distro's package manager or from here: [ROCm
 
 - Using `make` (example for target gfx1030, build with 16 CPU threads):
   ```bash
-  make -j16 GGML_HIPBLAS=1 GGML_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
+  make -j16 GGML_HIP=1 GGML_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
   ```
 
 - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):

From 30ec39832165627dd6ed98938df63adfc6e6a21a Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Tue, 26 Nov 2024 21:01:47 +0100
Subject: [PATCH 090/102] llama : disable warnings for 3rd party sha1
 dependency (#10527)

---
 examples/gguf-hash/CMakeLists.txt | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/examples/gguf-hash/CMakeLists.txt b/examples/gguf-hash/CMakeLists.txt
index 633f4553594bb..7a494ce327ba5 100644
--- a/examples/gguf-hash/CMakeLists.txt
+++ b/examples/gguf-hash/CMakeLists.txt
@@ -4,10 +4,17 @@ install(TARGETS ${TARGET} RUNTIME)
 
 # clibs dependencies
 include_directories(deps/)
+
 add_library(xxhash OBJECT deps/xxhash/xxhash.c deps/xxhash/xxhash.h)
 target_link_libraries(${TARGET} PRIVATE xxhash)
+
 add_library(sha1 OBJECT deps/sha1/sha1.c deps/sha1/sha1.h)
 target_link_libraries(${TARGET} PRIVATE sha1)
+if (NOT MSVC)
+    # disable warnings in 3rd party code
+    target_compile_options(sha1 PRIVATE -w)
+endif()
+
 add_library(sha256 OBJECT deps/sha256/sha256.c deps/sha256/sha256.h)
 target_link_libraries(${TARGET} PRIVATE sha256)
 

From 5a349f2809dc825960dfcfdf8f76b19cd0345be7 Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Tue, 26 Nov 2024 21:13:54 +0100
Subject: [PATCH 091/102] ci : remove nix workflows (#10526)

---
 .github/workflows/nix-ci-aarch64.yml    | 72 ----------------------
 .github/workflows/nix-ci.yml            | 81 -------------------------
 .github/workflows/nix-flake-update.yml  | 22 -------
 .github/workflows/nix-publish-flake.yml | 36 -----------
 4 files changed, 211 deletions(-)
 delete mode 100644 .github/workflows/nix-ci-aarch64.yml
 delete mode 100644 .github/workflows/nix-ci.yml
 delete mode 100644 .github/workflows/nix-flake-update.yml
 delete mode 100644 .github/workflows/nix-publish-flake.yml

diff --git a/.github/workflows/nix-ci-aarch64.yml b/.github/workflows/nix-ci-aarch64.yml
deleted file mode 100644
index 0da6acdf1c81e..0000000000000
--- a/.github/workflows/nix-ci-aarch64.yml
+++ /dev/null
@@ -1,72 +0,0 @@
-name: Nix aarch64 builds
-
-on:
-  workflow_dispatch: # allows manual triggering
-  schedule:
-    # Rebuild daily rather than on every push because QEMU is expensive (e.g.
-    # 1.5h instead of minutes with the cold cache).
-    #
-    # randint(0, 59), randint(0, 23)
-    - cron: '26 12 * * *'
-  # But also rebuild if we touched any of the Nix expressions:
-  push:
-    branches:
-      - master
-    paths: ['**/*.nix', 'flake.lock']
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: ['**/*.nix', 'flake.lock']
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-# Fine-grant permission
-# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
-permissions:
-  # https://github.com/DeterminateSystems/nix-installer-action?tab=readme-ov-file#with-flakehub
-  id-token: write
-  contents: read
-
-jobs:
-  nix-build-aarch64:
-    runs-on: ubuntu-latest
-    steps:
-    - name: Checkout repository
-      uses: actions/checkout@v4
-    - name: Install QEMU
-      # Copy-paste from https://github.com/orgs/community/discussions/8305#discussioncomment-5888654
-      run: |
-        sudo apt-get update
-        sudo apt-get install -y qemu-user-static qemu-system-aarch64
-        sudo usermod -a -G kvm $USER
-    - name: Install Nix
-      uses: DeterminateSystems/nix-installer-action@v9
-      with:
-        github-token: ${{ secrets.GITHUB_TOKEN }}
-        extra-conf: |
-          extra-platforms = aarch64-linux
-          extra-system-features = nixos-test kvm
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
-    - uses: DeterminateSystems/magic-nix-cache-action@v2
-      with:
-        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
-    - name: Set-up cachix to push the results to
-      uses: cachix/cachix-action@v13
-      with:
-        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: llama-cpp
-    - name: Show all output paths
-      run: >
-          nix run github:nix-community/nix-eval-jobs
-          -- --gc-roots-dir gcroot
-          --flake
-          ".#packages.aarch64-linux"
-    - name: Build
-      run: >
-          nix run github:Mic92/nix-fast-build
-          -- --skip-cached --no-nom
-          --systems aarch64-linux
-          --flake
-          ".#checks.aarch64-linux"
diff --git a/.github/workflows/nix-ci.yml b/.github/workflows/nix-ci.yml
deleted file mode 100644
index 3fe94157644ad..0000000000000
--- a/.github/workflows/nix-ci.yml
+++ /dev/null
@@ -1,81 +0,0 @@
-name: Nix CI
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: ['.github/workflows/nix-ci.yml', '**/flake.nix', '**/flake.lock', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/nix-ci.yml', '**/flake.nix', '**/flake.lock', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-# Fine-grant permission
-# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
-permissions:
-  # https://github.com/DeterminateSystems/nix-installer-action?tab=readme-ov-file#with-flakehub
-  id-token: write
-  contents: read
-
-jobs:
-  nix-eval:
-    strategy:
-      fail-fast: false
-      matrix:
-        os: [ ubuntu-latest, macos-latest ]
-    runs-on: ${{ matrix.os }}
-    steps:
-    - name: Checkout repository
-      uses: actions/checkout@v4
-    - name: Install Nix
-      uses: DeterminateSystems/nix-installer-action@v9
-      with:
-        github-token: ${{ secrets.GITHUB_TOKEN }}
-        extra-conf: |
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
-    - uses: DeterminateSystems/magic-nix-cache-action@v2
-      with:
-        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
-    - name: List all flake outputs
-      run: nix flake show --all-systems
-    - name: Show all output paths
-      run: >
-          nix run github:nix-community/nix-eval-jobs
-          -- --gc-roots-dir gcroot
-          --flake
-          ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
-  nix-build:
-    strategy:
-      fail-fast: false
-      matrix:
-        os: [ ubuntu-latest, macos-latest ]
-    runs-on: ${{ matrix.os }}
-    steps:
-    - name: Checkout repository
-      uses: actions/checkout@v4
-    - name: Install Nix
-      uses: DeterminateSystems/nix-installer-action@v9
-      with:
-        github-token: ${{ secrets.GITHUB_TOKEN }}
-        extra-conf: |
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
-    - uses: DeterminateSystems/magic-nix-cache-action@v2
-      with:
-        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
-    - name: Set-up cachix to push the results to
-      uses: cachix/cachix-action@v13
-      with:
-        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: llama-cpp
-    - name: Build
-      run: >
-          nix run github:Mic92/nix-fast-build
-          -- --skip-cached --no-nom
-          --flake
-          ".#checks.$(nix eval --raw --impure --expr builtins.currentSystem)"
diff --git a/.github/workflows/nix-flake-update.yml b/.github/workflows/nix-flake-update.yml
deleted file mode 100644
index 3a6a96e263e59..0000000000000
--- a/.github/workflows/nix-flake-update.yml
+++ /dev/null
@@ -1,22 +0,0 @@
-name: update-flake-lock
-on:
-  workflow_dispatch:
-  schedule:
-    - cron: '0 0 * * 0' # runs weekly on Sunday at 00:00
-
-jobs:
-  lockfile:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-      - name: Install Nix
-        uses: DeterminateSystems/nix-installer-action@main
-      - name: Update flake.lock
-        uses: DeterminateSystems/update-flake-lock@main
-        with:
-          pr-title: "nix: update flake.lock"
-          pr-labels: |
-            nix
-          pr-reviewers: philiptaron,SomeoneSerge
-          token: ${{ secrets.FLAKE_TOKEN }}
diff --git a/.github/workflows/nix-publish-flake.yml b/.github/workflows/nix-publish-flake.yml
deleted file mode 100644
index 2c3c1ebdaeff1..0000000000000
--- a/.github/workflows/nix-publish-flake.yml
+++ /dev/null
@@ -1,36 +0,0 @@
-# Make the flake discoverable on https://flakestry.dev and https://flakehub.com/flakes
-name: "Publish a flake to flakestry & flakehub"
-on:
-    push:
-        tags:
-        - "*"
-    workflow_dispatch:
-        inputs:
-            tag:
-                description: "The existing tag to publish"
-                type: "string"
-                required: true
-jobs:
-    flakestry-publish:
-        runs-on: ubuntu-latest
-        permissions:
-            id-token: "write"
-            contents: "read"
-        steps:
-            - uses: flakestry/flakestry-publish@main
-              with:
-                version: "${{ inputs.tag || github.ref_name }}"
-    flakehub-publish:
-      runs-on: "ubuntu-latest"
-      permissions:
-        id-token: "write"
-        contents: "read"
-      steps:
-        - uses: "actions/checkout@v4"
-          with:
-            ref: "${{ (inputs.tag != null) && format('refs/tags/{0}', inputs.tag) || '' }}"
-        - uses: "DeterminateSystems/nix-installer-action@main"
-        - uses: "DeterminateSystems/flakehub-push@main"
-          with:
-            visibility: "public"
-            tag: "${{ inputs.tag }}"

From de5097351caffb3deaea3393633609df49ef41d0 Mon Sep 17 00:00:00 2001
From: Shane A <shanea@allenai.org>
Date: Tue, 26 Nov 2024 12:55:29 -0800
Subject: [PATCH 092/102] Add OLMo 2 model in docs (#10530)

* Add link to OLMo 2 model in docs

* Change link to landing page
---
 README.md | 1 +
 1 file changed, 1 insertion(+)

diff --git a/README.md b/README.md
index 5f7933c132dc3..414c5b1c0824a 100644
--- a/README.md
+++ b/README.md
@@ -79,6 +79,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
+- [x] [OLMo 2](https://allenai.org/olmo)
 - [x] [OLMoE](https://huggingface.co/allenai/OLMoE-1B-7B-0924)
 - [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
 - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)

From c9b00a70b080d5c0668608024afc3e0e2fed822f Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Tue, 26 Nov 2024 22:12:10 +0100
Subject: [PATCH 093/102] ci : fix cuda releases (#10532)

---
 .github/workflows/build.yml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index c6aecec6e94d5..399641c4af3f0 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -728,7 +728,7 @@ jobs:
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows-latest-cmake:
-    runs-on: windows-2019
+    runs-on: windows-latest
 
     env:
       OPENBLAS_VERSION: 0.3.23
@@ -929,7 +929,7 @@ jobs:
 
   windows-2019-cmake-cuda:
     runs-on: windows-2019
-    if: ${{ github.event == 'push' && github.ref == 'refs/heads/master' }}
+    if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
     strategy:
       matrix:

From 4a57d362e1948ada50af997a92c3cbff9711e78b Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Wed, 27 Nov 2024 01:00:50 -0600
Subject: [PATCH 094/102] vulkan: optimize Q2_K and Q3_K mul_mat_vec (#10459)

---
 .../vulkan-shaders/mul_mat_vec_q2_k.comp      | 74 +++++++++++++------
 .../vulkan-shaders/mul_mat_vec_q3_k.comp      | 46 +++++++++---
 2 files changed, 87 insertions(+), 33 deletions(-)

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
index e2625d32b902d..fcf02210ea6c7 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
@@ -1,4 +1,5 @@
 #version 450
+#extension GL_EXT_shader_explicit_arithmetic_types : require
 
 #include "mul_mat_vec_base.comp"
 
@@ -32,38 +33,67 @@ void main() {
     const uint s_offset = 8*v_im;
     const uint y_offset = 128*v_im + l0;
 
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
         const uint y_idx = i * QUANT_K + y_offset;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
+        f16vec2 d = data_a[ib0 + i].d;
+        const FLOAT_TYPE dall = d.x;
+        const FLOAT_TYPE dmin = d.y;
+
+        B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
+        B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
+        B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
+        B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
+        B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
+        B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
+        B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
+        B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
+
+        uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
+        uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
+
+        uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
+        uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
+        uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
+        uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;
+
+        uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
+        uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
+        uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
+        uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));
+
+        uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
+        uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
+        uvec2 qs0 =  uvec2(unpack8(qs0_u16));
+        uvec2 qs16 = uvec2(unpack8(qs16_u16));
 
         FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
         FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum1 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3), sum1))))))));
-            sum2 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF), sum2))))))));
+        [[unroll]] for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            sum1 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 0) & 3),
+                   fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
+                   fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 2) & 3),
+                   fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
+                   fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 4) & 3),
+                   fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
+                   fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 6) & 3),
+                   fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
+            sum2 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_hi4[0]),
+                   fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_hi4[1]),
+                   fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_hi4[2]),
+                   fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_hi4[3]),
+                   fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_hi4[0]),
+                   fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_hi4[1]),
+                   fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_hi4[2]),
+                   fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
         }
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(dall, sum1, fma(-dmin, sum2, tmp[tmp_idx]));
+        temp = fma(dall, sum1, fma(-dmin, sum2, temp));
     }
 
+    tmp[gl_LocalInvocationID.x] = temp;
+
     // sum up partial sums and write back result
     barrier();
     [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
index a28804533c320..723fadde0d52b 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
@@ -1,4 +1,5 @@
 #version 450
+#extension GL_EXT_shader_explicit_arithmetic_types : require
 
 #include "mul_mat_vec_base.comp"
 
@@ -33,7 +34,7 @@ void main() {
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 128*v_im + l0;
 
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
     const uint s_shift = 4 * v_im;
 
@@ -42,21 +43,44 @@ void main() {
 
         const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
 
+        B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
+        B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
+        B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
+        B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
+        B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
+        B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
+        B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
+        B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
+
+        uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
+        uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
+        uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
+        uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
+        uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
+        uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
+        u8vec2 s0 = unpack8(s0_16);
+        u8vec2 s2 = unpack8(s2_16);
+        u8vec2 s4 = unpack8(s4_16);
+        u8vec2 s6 = unpack8(s6_16);
+        u8vec2 s8 = unpack8(s8_16);
+        u8vec2 s10 = unpack8(s10_16);
+
         FLOAT_TYPE sum = FLOAT_TYPE(0.0);
         for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
+            sum = fma(FLOAT_TYPE(b0[l])   * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b32[l])  * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b64[l])  * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b96[l])  * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b16[l])  * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b48[l])  * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b80[l])  * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
         }
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(d, sum, tmp[tmp_idx]);
+        temp = fma(d, sum, temp);
     }
 
+    tmp[gl_LocalInvocationID.x] = temp;
+
     // sum up partial sums and write back result
     barrier();
     [[unroll]] for (uint s = 16; s > 0; s >>= 1) {

From 71a64989a5d2e25c13507efada145f12cf358914 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Wed, 27 Nov 2024 01:08:54 -0600
Subject: [PATCH 095/102] vulkan: skip integer div/mod in get_offsets for
 batch_idx==0 (#10506)

---
 .../vulkan-shaders/mul_mat_vec_base.comp            | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
index 8d0a5791374d3..2ec1af5c75542 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
@@ -52,13 +52,16 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #endif
 
 #ifndef MUL_MAT_ID
-    const uint i13 = batch_idx / p.ne12;
-    const uint i12 = batch_idx % p.ne12;
+    uint batch_idx_a = 0;
+    if (batch_idx != 0) {
+        const uint i13 = batch_idx / p.ne12;
+        const uint i12 = batch_idx % p.ne12;
 
-    const uint i03 = i13 / p.broadcast3;
-    const uint i02 = i12 / p.broadcast2;
+        const uint i03 = i13 / p.broadcast3;
+        const uint i02 = i12 / p.broadcast2;
 
-    const uint batch_idx_a = i03 * p.ne02 + i02;
+        batch_idx_a = i03 * p.ne02 + i02;
+    }
 #else
     const uint expert_id = data_ids[expert_idx];
 #endif

From 249a7902ec710c8d027b9cc0ed10219d2b4184f8 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Wed, 27 Nov 2024 01:21:59 -0600
Subject: [PATCH 096/102] vulkan: further optimize q5_k mul_mat_vec (#10479)

---
 .../vulkan-shaders/mul_mat_vec_q5_k.comp      | 52 ++++++++++---------
 1 file changed, 28 insertions(+), 24 deletions(-)

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
index 22a6bfae4fb61..b455cbd31ec91 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
@@ -34,9 +34,6 @@ void main() {
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 64*v_im + l0;
 
-    const uint8_t hm1 = uint8_t(1 << (2*v_im));
-    const uint8_t hm2 = uint8_t(hm1 << 4);
-
     FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += 2) {
@@ -71,6 +68,18 @@ void main() {
         uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
         uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
 
+        uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8]));
+
+        uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4;
+        uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3;
+        uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010) << 0;
+        uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1;
+
+        qs0_16_u32_lo4 += qs0_16_lo4_offset16;
+        qs0_16_u32_hi4 += qs0_16_hi4_offset16;
+        qs64_80_u32_lo4 += qs64_80_lo4_offset16;
+        qs64_80_u32_hi4 += qs64_80_hi4_offset16;
+
         uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
         uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
         uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
@@ -102,31 +111,26 @@ void main() {
         B_TYPE_VEC2 by232 = data_b_v2[(b_offset + y2_idx) / 2 + 16];
         B_TYPE_VEC2 by248 = data_b_v2[(b_offset + y2_idx) / 2 + 24];
 
-        uint32_t qh0 = data_a_packed16[ib0 + i].qh[l0 / 2];
-        uint32_t qh1 = qh0 >> 8;
-        uint32_t qh16 = data_a_packed16[ib0 + i].qh[l0 / 2 + 8];
-        uint32_t qh17 = qh16 >> 8;
-
         const FLOAT_TYPE sx =
-          fma(FLOAT_TYPE(by10.x), (q4_0 + (((qh0 & hm1) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(by10.y), (q4_1 + (((qh1 & hm1) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(by116.x), (q4_2 + (((qh16 & hm1) != 0) ? 16 : 0)),
-             FLOAT_TYPE(by116.y) * (q4_3 + (((qh17 & hm1) != 0) ? 16 : 0)))));
+          fma(FLOAT_TYPE(by10.x), q4_0,
+          fma(FLOAT_TYPE(by10.y), q4_1,
+          fma(FLOAT_TYPE(by116.x), q4_2,
+             FLOAT_TYPE(by116.y) * q4_3)));
         const FLOAT_TYPE sy =
-          fma(FLOAT_TYPE(by132.x), (q4_4 + (((qh0 & (hm1 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(by132.y), (q4_5 + (((qh1 & (hm1 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(by148.x), (q4_6 + (((qh16 & (hm1 << 1)) != 0) ? 16 : 0)),
-             FLOAT_TYPE(by148.y) * (q4_7 + (((qh17 & (hm1 << 1)) != 0) ? 16 : 0)))));
+          fma(FLOAT_TYPE(by132.x), q4_4,
+          fma(FLOAT_TYPE(by132.y), q4_5,
+          fma(FLOAT_TYPE(by148.x), q4_6,
+             FLOAT_TYPE(by148.y) * q4_7)));
         const FLOAT_TYPE sz =
-          fma(FLOAT_TYPE(by20.x), (q4_8  + (((qh0 & hm2) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(by20.y), (q4_9  + (((qh1 & hm2) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(by216.x), (q4_10 + (((qh16 & hm2) != 0) ? 16 : 0)),
-             FLOAT_TYPE(by216.y) * (q4_11 + (((qh17 & hm2) != 0) ? 16 : 0)))));
+          fma(FLOAT_TYPE(by20.x), q4_8,
+          fma(FLOAT_TYPE(by20.y), q4_9,
+          fma(FLOAT_TYPE(by216.x), q4_10,
+             FLOAT_TYPE(by216.y) * q4_11)));
         const FLOAT_TYPE sw =
-          fma(FLOAT_TYPE(by232.x), (q4_12 + (((qh0 & (hm2 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(by232.y), (q4_13 + (((qh1 & (hm2 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(by248.x), (q4_14 + (((qh16 & (hm2 << 1)) != 0) ? 16 : 0)),
-             FLOAT_TYPE(by248.y) * (q4_15 + (((qh17 & (hm2 << 1)) != 0) ? 16 : 0)))));
+          fma(FLOAT_TYPE(by232.x), q4_12,
+          fma(FLOAT_TYPE(by232.y), q4_13,
+          fma(FLOAT_TYPE(by248.x), q4_14,
+             FLOAT_TYPE(by248.y) * q4_15)));
         const FLOAT_TYPE smin =
           fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
           fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,

From 5b3466bedfa84aa29c6871c7254467550186ecc6 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Wed, 27 Nov 2024 01:30:27 -0600
Subject: [PATCH 097/102] vulkan: Handle GPUs with less shared memory (#10468)

There have been reports of failure to compile on systems with <= 32KB
of shared memory (e.g. #10037). This change makes the large tile size
fall back to a smaller size if necessary, and makes mul_mat_id fall
back to CPU if there's only 16KB of shared memory.
---
 ggml/src/ggml-vulkan/ggml-vulkan.cpp | 174 +++++++++++++++++----------
 1 file changed, 111 insertions(+), 63 deletions(-)

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index da1cfd24eac22..a833007fb3d02 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -1232,8 +1232,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
     std::cerr << "ggml_vulkan: Compiling shaders";
 
     // mulmat
-    std::vector<uint32_t> l_warptile, m_warptile, s_warptile, l_warptile_mmq, m_warptile_mmq, s_warptile_mmq;
-    std::array<uint32_t, 3> l_wg_denoms, m_wg_denoms, s_wg_denoms;
+    std::vector<uint32_t> l_warptile, m_warptile, s_warptile,
+                          l_warptile_mmq, m_warptile_mmq, s_warptile_mmq;
+    std::array<uint32_t, 3> l_wg_denoms, m_wg_denoms, s_wg_denoms,
+                            l_mmq_wg_denoms, m_mmq_wg_denoms, s_mmq_wg_denoms;
     uint32_t l_align, m_align, s_align;
 
     l_warptile = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
@@ -1244,14 +1246,48 @@ static void ggml_vk_load_shaders(vk_device& device) {
     m_warptile_mmq = { 128,  64,  64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
     s_warptile_mmq = { std::max(device->subgroup_size, 16u),  32,  32, 32, 32, 32, 2, 2, 2, device->subgroup_size };
 
-    l_wg_denoms = {128, 128, 1 };
-    m_wg_denoms = { 64,  64, 1 };
-    s_wg_denoms = { 32,  32, 1 };
+    l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
+    m_mmq_wg_denoms = m_wg_denoms = { 64,  64, 1 };
+    s_mmq_wg_denoms = s_wg_denoms = { 32,  32, 1 };
 
     l_align = 128;
     m_align =  64;
     s_align =  32;
 
+    // Fallback to smaller sizes if there's not enough shared memory. Given the current shaders
+    // and tile sizes, this should handle 16KB, 32KB, and 48KB+.
+    // This logic doesn't explicitly account for the 12KB row_ids in the mul_mat_mat_id shaders.
+    // But the numbers happen to work out for 32KB shared memory size that when using the medium
+    // size there's enough room for everything, and we assert for this.
+    uint32_t shmem_needed = (l_warptile[1] + l_warptile[2]) * (l_warptile[3] + 1) * sizeof(float);
+    if (shmem_needed > device->properties.limits.maxComputeSharedMemorySize) {
+        l_warptile = m_warptile;
+        l_wg_denoms = m_wg_denoms;
+        shmem_needed = (l_warptile[1] + l_warptile[2]) * (l_warptile[3] + 1) * sizeof(float);
+        GGML_ASSERT(shmem_needed <= device->properties.limits.maxComputeSharedMemorySize);
+    }
+    if (device->properties.limits.maxComputeSharedMemorySize >= 32768) {
+        // assert mul_mat_mat_id shaders will fit.
+        GGML_ASSERT(shmem_needed + 3072*4 <= device->properties.limits.maxComputeSharedMemorySize);
+    }
+
+    shmem_needed = (l_warptile_mmq[1] + l_warptile_mmq[2]) * (l_warptile_mmq[3] + 1) * sizeof(float);
+    if (shmem_needed > device->properties.limits.maxComputeSharedMemorySize) {
+        if (device->properties.limits.maxComputeSharedMemorySize == 32768) {
+            l_warptile_mmq = m_warptile_mmq;
+            l_mmq_wg_denoms = m_mmq_wg_denoms;
+        } else {
+            l_warptile_mmq = s_warptile_mmq;
+            l_mmq_wg_denoms = s_mmq_wg_denoms;
+        }
+        shmem_needed = (l_warptile_mmq[1] + l_warptile_mmq[2]) * (l_warptile_mmq[3] + 1) * sizeof(float);
+        GGML_ASSERT(shmem_needed <= device->properties.limits.maxComputeSharedMemorySize);
+    }
+    if (device->properties.limits.maxComputeSharedMemorySize >= 32768) {
+        // assert mul_mat_mat_id shaders will fit.
+        GGML_ASSERT(shmem_needed + 3072*4 <= device->properties.limits.maxComputeSharedMemorySize);
+    }
+
     device->pipeline_matmul_f32 = std::make_shared<vk_matmul_pipeline_struct>();
     device->pipeline_matmul_f32_f16 = std::make_shared<vk_matmul_pipeline_struct>();
 
@@ -1299,35 +1335,38 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
         CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
 
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-
-        CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
-        CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
-        CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
-
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+
+        // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
+        if (device->properties.limits.maxComputeSharedMemorySize >= 32768) {
+            CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+            CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+            CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        }
 #undef CREATE_MM
     } else {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
@@ -1344,35 +1383,38 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
         CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3);
 
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
-
-        CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
-        CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
-        CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
-
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3);
+
+        // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
+        if (device->properties.limits.maxComputeSharedMemorySize >= 32768) {
+            CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+            CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+            CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4);
+
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4);
+        }
 #undef CREATE_MM
     }
 
@@ -6541,6 +6583,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
+                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+                if (op->op == GGML_OP_MUL_MAT_ID &&
+                    ggml_vk_get_device(ctx->device)->properties.limits.maxComputeSharedMemorySize < 32768) {
+                    // If there's not enough shared memory for row_ids and the result tile, fallback to CPU
+                    return false;
+                }
                 switch (op->src[0]->type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:

From c31ed2abfce05c38a2a5189586bfae45a139a547 Mon Sep 17 00:00:00 2001
From: Jeff Bolz <jbolz@nvidia.com>
Date: Wed, 27 Nov 2024 01:32:54 -0600
Subject: [PATCH 098/102] vulkan: define all quant data structures in
 types.comp (#10440)

---
 .../src/ggml-vulkan/vulkan-shaders/types.comp | 147 +++++++++---------
 1 file changed, 76 insertions(+), 71 deletions(-)

diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
index bc28e0ab857aa..eecc47f3a9764 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
@@ -30,10 +30,8 @@
 #endif
 #endif
 
-#if defined(DATA_A_Q4_0)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 32
-#define QUANT_R 2
+#define QUANT_K_Q4_0 32
+#define QUANT_R_Q4_0 2
 
 struct block_q4_0
 {
@@ -46,14 +44,15 @@ struct block_q4_0_packed16
     uint16_t qs[16/2];
 };
 
+#if defined(DATA_A_Q4_0)
+#define QUANT_K QUANT_K_Q4_0
+#define QUANT_R QUANT_R_Q4_0
 #define A_TYPE block_q4_0
 #define A_TYPE_PACKED16 block_q4_0_packed16
 #endif
 
-#if defined(DATA_A_Q4_1)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 32
-#define QUANT_R 2
+#define QUANT_K_Q4_1 32
+#define QUANT_R_Q4_1 2
 
 struct block_q4_1
 {
@@ -69,15 +68,15 @@ struct block_q4_1_packed16
     uint16_t qs[16/2];
 };
 
+#if defined(DATA_A_Q4_1)
+#define QUANT_K QUANT_K_Q4_1
+#define QUANT_R QUANT_R_Q4_1
 #define A_TYPE block_q4_1
 #define A_TYPE_PACKED16 block_q4_1_packed16
 #endif
 
-#if defined(DATA_A_Q5_0)
-#extension GL_EXT_shader_16bit_storage : require
-#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
-#define QUANT_K 32
-#define QUANT_R 2
+#define QUANT_K_Q5_0 32
+#define QUANT_R_Q5_0 2
 
 struct block_q5_0
 {
@@ -93,15 +92,15 @@ struct block_q5_0_packed16
     uint16_t qs[16/2];
 };
 
+#if defined(DATA_A_Q5_0)
+#define QUANT_K QUANT_K_Q5_0
+#define QUANT_R QUANT_R_Q5_0
 #define A_TYPE block_q5_0
 #define A_TYPE_PACKED16 block_q5_0_packed16
 #endif
 
-#if defined(DATA_A_Q5_1)
-#extension GL_EXT_shader_16bit_storage : require
-#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
-#define QUANT_K 32
-#define QUANT_R 2
+#define QUANT_K_Q5_1 32
+#define QUANT_R_Q5_1 2
 
 struct block_q5_1
 {
@@ -119,14 +118,15 @@ struct block_q5_1_packed16
     uint16_t qs[16/2];
 };
 
+#if defined(DATA_A_Q5_1)
+#define QUANT_K QUANT_K_Q5_1
+#define QUANT_R QUANT_R_Q5_1
 #define A_TYPE block_q5_1
 #define A_TYPE_PACKED16 block_q5_1_packed16
 #endif
 
-#if defined(DATA_A_Q8_0)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 32
-#define QUANT_R 1
+#define QUANT_K_Q8_0 32
+#define QUANT_R_Q8_0 1
 
 struct block_q8_0
 {
@@ -139,164 +139,164 @@ struct block_q8_0_packed16
     uint16_t qs[32/2];
 };
 
+#if defined(DATA_A_Q8_0)
+#define QUANT_K QUANT_K_Q8_0
+#define QUANT_R QUANT_R_Q8_0
 #define A_TYPE block_q8_0
 #define A_TYPE_PACKED16 block_q8_0_packed16
 #endif
 
 // K-quants
-#if defined(DATA_A_Q2_K)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 256
+#define QUANT_K_Q2_K 256
 
 struct block_q2_K
 {
-    uint8_t scales[QUANT_K/16];
-    uint8_t qs[QUANT_K/4];
+    uint8_t scales[QUANT_K_Q2_K/16];
+    uint8_t qs[QUANT_K_Q2_K/4];
     f16vec2 d;
 };
 
 struct block_q2_K_packed16
 {
-    uint16_t scales[QUANT_K/16/2];
-    uint16_t qs[QUANT_K/4/2];
+    uint16_t scales[QUANT_K_Q2_K/16/2];
+    uint16_t qs[QUANT_K_Q2_K/4/2];
     f16vec2 d;
 };
 
 struct block_q2_K_packed32
 {
-    uint32_t scales[QUANT_K/16/4];
-    uint32_t qs[QUANT_K/4/4];
+    uint32_t scales[QUANT_K_Q2_K/16/4];
+    uint32_t qs[QUANT_K_Q2_K/4/4];
     f16vec2 d;
 };
 
+#if defined(DATA_A_Q2_K)
+#define QUANT_K QUANT_K_Q2_K
 #define A_TYPE block_q2_K
 #define A_TYPE_PACKED16 block_q2_K_packed16
 #define A_TYPE_PACKED32 block_q2_K_packed32
 #endif
 
-#if defined(DATA_A_Q3_K)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 256
+#define QUANT_K_Q3_K 256
 
 struct block_q3_K
 {
-    uint8_t hmask[QUANT_K/8];
-    uint8_t qs[QUANT_K/4];
+    uint8_t hmask[QUANT_K_Q3_K/8];
+    uint8_t qs[QUANT_K_Q3_K/4];
     uint8_t scales[12];
     float16_t d;
 };
 
 struct block_q3_K_packed16
 {
-    uint16_t hmask[QUANT_K/8/2];
-    uint16_t qs[QUANT_K/4/2];
+    uint16_t hmask[QUANT_K_Q3_K/8/2];
+    uint16_t qs[QUANT_K_Q3_K/4/2];
     uint16_t scales[12/2];
     float16_t d;
 };
 
+#if defined(DATA_A_Q3_K)
+#define QUANT_K QUANT_K_Q3_K
 #define A_TYPE block_q3_K
 #define A_TYPE_PACKED16 block_q3_K_packed16
 #endif
 
-#if defined(DATA_A_Q4_K)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 256
+#define QUANT_K_Q4_K 256
 
 struct block_q4_K
 {
     f16vec2 d;
-    uint8_t scales[3*QUANT_K/64];
-    uint8_t qs[QUANT_K/2];
+    uint8_t scales[3*QUANT_K_Q4_K/64];
+    uint8_t qs[QUANT_K_Q4_K/2];
 };
 
 struct block_q4_K_packed16
 {
     f16vec2 d;
-    uint16_t scales[3*QUANT_K/64/2];
-    uint16_t qs[QUANT_K/2/2];
+    uint16_t scales[3*QUANT_K_Q4_K/64/2];
+    uint16_t qs[QUANT_K_Q4_K/2/2];
 };
 
 struct block_q4_K_packed32
 {
     f16vec2 d;
-    uint32_t scales[3*QUANT_K/64/4];
-    uint32_t qs[QUANT_K/2/4];
+    uint32_t scales[3*QUANT_K_Q4_K/64/4];
+    uint32_t qs[QUANT_K_Q4_K/2/4];
 };
 
+#if defined(DATA_A_Q4_K)
+#define QUANT_K QUANT_K_Q4_K
 #define A_TYPE block_q4_K
 #define A_TYPE_PACKED16 block_q4_K_packed16
 #define A_TYPE_PACKED32 block_q4_K_packed32
 #endif
 
-#if defined(DATA_A_Q5_K)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 256
+#define QUANT_K_Q5_K 256
 
 struct block_q5_K
 {
     f16vec2 d;
     uint8_t scales[12];
-    uint8_t qh[QUANT_K/8];
-    uint8_t qs[QUANT_K/2];
+    uint8_t qh[QUANT_K_Q5_K/8];
+    uint8_t qs[QUANT_K_Q5_K/2];
 };
 
 struct block_q5_K_packed16
 {
     f16vec2 d;
     uint16_t scales[12/2];
-    uint16_t qh[QUANT_K/8/2];
-    uint16_t qs[QUANT_K/2/2];
+    uint16_t qh[QUANT_K_Q5_K/8/2];
+    uint16_t qs[QUANT_K_Q5_K/2/2];
 };
 
+#if defined(DATA_A_Q5_K)
+#define QUANT_K QUANT_K_Q5_K
 #define A_TYPE block_q5_K
 #define A_TYPE_PACKED16 block_q5_K_packed16
 #endif
 
-#if defined(DATA_A_Q6_K)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 256
+#define QUANT_K_Q6_K 256
 
 struct block_q6_K
 {
-    uint8_t ql[QUANT_K/2];
-    uint8_t qh[QUANT_K/4];
-    int8_t scales[QUANT_K/16];
+    uint8_t ql[QUANT_K_Q6_K/2];
+    uint8_t qh[QUANT_K_Q6_K/4];
+    int8_t scales[QUANT_K_Q6_K/16];
     float16_t d;
 };
 
 struct block_q6_K_packed16
 {
-    uint16_t ql[QUANT_K/2/2];
-    uint16_t qh[QUANT_K/4/2];
-    int8_t scales[QUANT_K/16];
+    uint16_t ql[QUANT_K_Q6_K/2/2];
+    uint16_t qh[QUANT_K_Q6_K/4/2];
+    int8_t scales[QUANT_K_Q6_K/16];
     float16_t d;
 };
 
+#if defined(DATA_A_Q6_K)
+#define QUANT_K QUANT_K_Q6_K
 #define A_TYPE block_q6_K
 #define A_TYPE_PACKED16 block_q6_K_packed16
 #endif
 
 // IQuants
 
-#if defined(DATA_A_IQ4_NL)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 32
-#define QUANT_R 2
+#define QUANT_K_IQ4_NL 32
+#define QUANT_R_IQ4_NL 2
 
 struct block_iq4_nl
 {
     float16_t d;
-    uint8_t qs[QUANT_K/2];
+    uint8_t qs[QUANT_K_IQ4_NL/2];
 };
 
 struct block_iq4_nl_packed16
 {
     float16_t d;
-    uint16_t qs[QUANT_K/2/2];
+    uint16_t qs[QUANT_K_IQ4_NL/2/2];
 };
 
-#define A_TYPE block_iq4_nl
-#define A_TYPE_PACKED16 block_iq4_nl_packed16
+#if defined(DATA_A_IQ4_NL)
 
 const int8_t kvalues_iq4nl_const[16] = {
     int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10),
@@ -313,6 +313,11 @@ void init_iq4nl_shmem()
     }
     barrier();
 }
+
+#define QUANT_K QUANT_K_IQ4_NL
+#define QUANT_R QUANT_R_IQ4_NL
+#define A_TYPE block_iq4_nl
+#define A_TYPE_PACKED16 block_iq4_nl_packed16
 #endif
 
 #endif // !defined(GGML_TYPES_COMP)

From 9150f8fef95327474d39ccd6c6e30787e85f3529 Mon Sep 17 00:00:00 2001
From: Frankie Robertson <frankier@users.noreply.github.com>
Date: Tue, 26 Nov 2024 15:50:26 +0200
Subject: [PATCH 099/102] Do not include arm_neon.h when compiling CUDA code
 (ggml/1028)

---
 ggml/src/ggml-impl.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
index 3965be78755a1..f39b7a88c8c4b 100644
--- a/ggml/src/ggml-impl.h
+++ b/ggml/src/ggml-impl.h
@@ -14,7 +14,7 @@
 #include <arm_sve.h>
 #endif // __ARM_FEATURE_SVE
 
-#if defined(__ARM_NEON)
+#if defined(__ARM_NEON) && !defined(__CUDACC__)
 // if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
 //
 //   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/

From fee824a1a1e35b5c49d482f654613addade61764 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Wed, 27 Nov 2024 11:10:42 +0200
Subject: [PATCH 100/102] sync : ggml

---
 scripts/sync-ggml.last | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last
index d101d2b57f8fa..fd71a0a46fa8e 100644
--- a/scripts/sync-ggml.last
+++ b/scripts/sync-ggml.last
@@ -1 +1 @@
-6fcbd60bc72ac3f7ad43f78c87e535f2e6206f58
+c598cbe30621251e80acbcf3b601589a37c17f4d

From 9e2301f4a4ef1690bd99360c11de43fe830b1c8d Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Wed, 27 Nov 2024 11:22:14 +0200
Subject: [PATCH 101/102] metal : fix group_norm support condition (#0)

---
 ggml/src/ggml-metal/ggml-metal.m | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index c47f07a9e0416..ae6b25edea95a 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -997,9 +997,10 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_SUM_ROWS:
         case GGML_OP_SOFT_MAX:
-        case GGML_OP_RMS_NORM:
         case GGML_OP_GROUP_NORM:
             return has_simdgroup_reduction;
+        case GGML_OP_RMS_NORM:
+            return has_simdgroup_reduction && (op->ne[0] % 4 == 0);
         case GGML_OP_NORM:
         case GGML_OP_ROPE:
             return true;
@@ -2672,7 +2673,6 @@ static void ggml_metal_encode_node(
             } break;
         case GGML_OP_GROUP_NORM:
             {
-                GGML_ASSERT(ne00 % 4 == 0);
                 GGML_ASSERT(ggml_is_contiguous(src0));
 
                 float eps;

From 46c69e0e752ff16206347bb12f96ed69f4a01abf Mon Sep 17 00:00:00 2001
From: Diego Devesa <slarengh@gmail.com>
Date: Wed, 27 Nov 2024 11:03:25 +0100
Subject: [PATCH 102/102] ci : faster CUDA toolkit installation method and use
 ccache (#10537)

* ci : faster CUDA toolkit installation method and use ccache

* remove fetch-depth

* only pack CUDA runtime on master
---
 .github/workflows/build.yml | 117 +++++++++++++++++++++---------------
 1 file changed, 70 insertions(+), 47 deletions(-)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 399641c4af3f0..e02b5c6200a82 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -892,12 +892,12 @@ jobs:
             cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=89-real -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined -DLLAMA_FATAL_WARNINGS=ON
             cmake --build build
 
-  windows-latest-cmake-cuda:
-    runs-on: windows-latest
+  windows-2019-cmake-cuda:
+    runs-on: windows-2019
 
     strategy:
       matrix:
-        cuda: ['12.6.2']
+        cuda: ['12.4', '11.7']
         build: ['cuda']
 
     steps:
@@ -905,13 +905,66 @@ jobs:
         id: checkout
         uses: actions/checkout@v4
 
-      - name: Install CUDA toolkit
-        id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.19
+      - name: Install Cuda Toolkit 11.7
+        if: ${{ matrix.cuda == '11.7' }}
+        run: |
+          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7"
+          choco install unzip -y
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-11.7.99-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-11.7.99-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-11.7.99-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-11.7.4.6-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-11.7.91-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-11.7.91-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-11.7.101-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-11.7.91-archive.zip"
+          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7"
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cudart-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvcc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvrtc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libcublas-windows-x86_64-11.7.4.6-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvtx-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\visual_studio_integration-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvprof-windows-x86_64-11.7.101-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cccl-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V11_7=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
+      - name: Install Cuda Toolkit 12.4
+        if: ${{ matrix.cuda == '12.4' }}
+        run: |
+          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
+          choco install unzip -y
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-12.4.127-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-12.4.131-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-12.4.127-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-12.4.5.8-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-12.4.127-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-12.4.127-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-12.4.127-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-12.4.127-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-12.4.127-archive.zip"
+          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4"
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cudart-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvcc-windows-x86_64-12.4.131-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvrtc-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libcublas-windows-x86_64-12.4.5.8-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvtx-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_profiler_api-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\visual_studio_integration-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvprof-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cccl-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V12_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+
+      - name: Install ccache
+        uses: hendrikmuhs/ccache-action@v1.2
         with:
-          cuda: ${{ matrix.cuda }}
-          method: 'network'
-          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
+          key: ${{ github.job }}-${{ matrix.cuda }}-${{ matrix.build }}
 
       - name: Install Ninja
         id: install_ninja
@@ -922,44 +975,12 @@ jobs:
         id: cmake_build
         shell: cmd
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
-          cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON -DCMAKE_CUDA_ARCHITECTURES=89-real
-          cmake --build build --config Release -t ggml-cuda
+          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          cmake -S . -B build -G "Ninja Multi-Config" -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON
+          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
+          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
           cmake --build build --config Release
 
-  windows-2019-cmake-cuda:
-    runs-on: windows-2019
-    if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-    strategy:
-      matrix:
-        cuda: ['12.2.0', '11.7.1']
-        build: ['cuda']
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: Install CUDA toolkit
-        id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.15
-        with:
-          cuda: ${{ matrix.cuda }}
-          method: 'network'
-          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON
-          cmake --build . --config Release -j $((${env:NUMBER_OF_PROCESSORS} - 1)) -t ggml
-          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
-
       - name: Determine tag name
         id: tag
         shell: bash
@@ -987,10 +1008,12 @@ jobs:
           name: llama-bin-win-cu${{ matrix.cuda }}-x64.zip
 
       - name: Copy and pack Cuda runtime
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
         run: |
-          echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
+          echo "Cuda install location: ${{ env.CUDA_PATH }}"
           $dst='.\build\bin\cudart\'
-          robocopy "${{steps.cuda-toolkit.outputs.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
           7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime