Merge pull request #18 from l3utterfly/merge

merged from upstream

.github/workflows/close-issue.yml

@@ -12,7 +12,7 @@ jobs:
     steps:
       - uses: actions/stale@v5
         with:
-          exempt-issue-labels: "refactor,help wanted,good first issue,research"
+          exempt-issue-labels: "refactor,help wanted,good first issue,research,bug"
           days-before-issue-stale: 30
           days-before-issue-close: 14
           stale-issue-label: "stale"

ci/run.sh

@@ -518,10 +518,10 @@ function gg_run_open_llama_7b_v2 {
 
     (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/save-load-state --model     -ngl 10 ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state --model -fa -ngl 10 ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state --model     -ngl 99 ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state --model -fa -ngl 99 ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"

common/common.h

@@ -135,7 +135,7 @@ struct gpt_params {
     bool   multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt
     size_t multiple_choice_tasks = 0;     // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed
 
-    bool   kl_divergence   = false; // compute KL-divergence
+    bool   kl_divergence   = false; // compute KL divergence
 
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs

common/log.h

@@ -234,7 +234,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // INTERNAL, DO NOT USE
 //  USE LOG() INSTEAD
 //
-#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER)
+#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__)
     #define LOG_IMPL(str, ...)                                                                                      \
     do {                                                                                                            \
         if (LOG_TARGET != nullptr)                                                                                  \
@@ -257,7 +257,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // INTERNAL, DO NOT USE
 //  USE LOG_TEE() INSTEAD
 //
-#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER)
+#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__)
     #define LOG_TEE_IMPL(str, ...)                                                                                                      \
     do {                                                                                                                                \
         if (LOG_TARGET != nullptr)                                                                                                      \

convert-hf-to-gguf-update.py

@@ -128,7 +128,7 @@ def download_file_with_auth(url, token, save_path):
     print(f"chkhsh: {chkhsh}")
 
     # print the "pre_tokenizer" content from the tokenizer.json
-    with open(f"models/tokenizers/{name}/tokenizer.json", "r") as f:
+    with open(f"models/tokenizers/{name}/tokenizer.json", "r", encoding="utf-8") as f:
         cfg = json.load(f)
         pre_tokenizer = cfg["pre_tokenizer"]
         print("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4))
@@ -156,15 +156,19 @@ def download_file_with_auth(url, token, save_path):
 src_func +=  "\n"
 src_func +=  "        res = None\n"
 src_func +=  "\n"
-src_func +=  "        # NOTE: if you get an error here, you need to add the model to the if-elif chain below\n"
-src_func +=  "        #       don't do this manually - use the convert-hf-to-gguf-update.py script!\n"
+src_func +=  "        # NOTE: if you get an error here, you need to update the convert-hf-to-gguf-update.py script\n"
+src_func +=  "        #       or pull the latest version of the model from Huggingface\n"
+src_func +=  "        #       don't edit the hashes manually!\n"
 src_func += f"{src_ifs}\n"
 src_func +=  "        if res is None:\n"
 src_func +=  "            print(\"\\n\")\n"
 src_func +=  "            print(\"**************************************************************************************\")\n"
 src_func +=  "            print(\"** WARNING: The BPE pre-tokenizer was not recognized!\")\n"
-src_func +=  "            print(\"**          This means that it was not added yet or you are using an older version.\")\n"
-src_func +=  "            print(\"**          Check convert-hf-to-gguf-update.py and update it accordingly.\")\n"
+src_func +=  "            print(\"**          There are 2 possible reasons for this:\")\n"
+src_func +=  "            print(\"**          - the model has not been added to convert-hf-to-gguf-update.py yet\")\n"
+src_func +=  "            print(\"**          - the pre-tokenization config has changed upstream\")\n"
+src_func +=  "            print(\"**          Check your model files and convert-hf-to-gguf-update.py and update them accordingly.\")\n"
+src_func +=  "            print(\"** ref:     https://github.com/ggerganov/llama.cpp/pull/6920\")\n"
 src_func +=  "            print(\"**\")\n"
 src_func +=  "            print(f\"** chkhsh:  {chkhsh}\")\n"
 src_func +=  "            print(\"**************************************************************************************\")\n"
@@ -249,7 +253,7 @@ def download_file_with_auth(url, token, save_path):
     from transformers import AutoTokenizer
     tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
 
-    with open(f"models/ggml-vocab-{name}.gguf.inp", "w") as f:
+    with open(f"models/ggml-vocab-{name}.gguf.inp", "w", encoding="utf-8") as f:
         for text in tests:
             f.write(f"{text}")
             f.write("\n__ggml_vocab_test__\n")

convert-hf-to-gguf.py

@@ -279,8 +279,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
 
         res = None
 
-        # NOTE: if you get an error here, you need to add the model to the if-elif chain below
-        #       don't do this manually - use the convert-hf-to-gguf-update.py script!
+        # NOTE: if you get an error here, you need to update the convert-hf-to-gguf-update.py script
+        #       or pull the latest version of the model from Huggingface
+        #       don't edit the hashes manually!
         if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
             # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
             res = "llama-bpe"
@@ -310,8 +311,11 @@ def get_vocab_base_pre(self, tokenizer) -> str:
             print("\n")
             print("**************************************************************************************")
             print("** WARNING: The BPE pre-tokenizer was not recognized!")
-            print("**          This means that it was not added yet or you are using an older version.")
-            print("**          Check convert-hf-to-gguf-update.py and update it accordingly.")
+            print("**          There are 2 possible reasons for this:")
+            print("**          - the model has not been added to convert-hf-to-gguf-update.py yet")
+            print("**          - the pre-tokenization config has changed upstream")
+            print("**          Check your model files and convert-hf-to-gguf-update.py and update them accordingly.")
+            print("** ref:     https://github.com/ggerganov/llama.cpp/pull/6920")
             print("**")
             print(f"** chkhsh:  {chkhsh}")
             print("**************************************************************************************")
@@ -1423,7 +1427,7 @@ def write_tensors(self):
         experts = dict()
         for name, data_torch in self.get_tensors():
             # we don't need these
-            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
+            if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
                 continue
 
             old_dtype = data_torch.dtype

examples/main/main.cpp

@@ -546,7 +546,7 @@ int main(int argc, char ** argv) {
                 // if we run out of context:
                 // - take the n_keep first tokens from the original prompt (via n_past)
                 // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
-                if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
+                if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) >= n_ctx) {
                     if (params.n_predict == -2) {
                         LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
                         break;