From 50d5cecbda3b0d03344eed326287adc1f6c7f3ef Mon Sep 17 00:00:00 2001 From: Diego Devesa Date: Mon, 25 Nov 2024 22:05:39 +0100 Subject: [PATCH 01/30] 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 Date: Mon, 25 Nov 2024 16:56:24 -0500 Subject: [PATCH 02/30] 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 --- 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 +#else +#include +#endif + +#include +#include +#include +#include +#include +#include +#include +#include + +#include "llama-cpp.h" + +typedef std::unique_ptr 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 string_args; + std::unordered_map int_args; + std::vector 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(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 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 & 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 & 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(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 & 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 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 & 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 owned_content; + std::vector 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 + +#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_ptr; +typedef std::unique_ptr llama_context_ptr; +typedef std::unique_ptr llama_sampler_ptr; From 0eb4e12beebabae46d37b78742f4c5d4dbe52dc1 Mon Sep 17 00:00:00 2001 From: Junil Kim Date: Tue, 26 Nov 2024 10:47:20 +0900 Subject: [PATCH 03/30] 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 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 04/30] CANN: RoPE and CANCAT operator optimization (#10488) Co-authored-by: noemotiovon --- 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 #include @@ -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 05/30] CANN: Improve the Inferencing Performance for Ascend NPU Device (#10454) * improve inferencing performance for ascend npu. Co-authored-by: Frank Mai * some modification after review * some modifications after review * restore some modifications * restore some modifications --------- Co-authored-by: shanshan shen Co-authored-by: Frank Mai --- 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 #include #include +#include +#include #include #include #include @@ -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_backend_cann_context::new_pool_for_device( int device) { - // return std::unique_ptr(new ggml_cann_pool_leg(device)); return std::unique_ptr(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 Date: Tue, 26 Nov 2024 12:29:38 +0200 Subject: [PATCH 06/30] 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(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 Date: Tue, 26 Nov 2024 13:36:40 +0200 Subject: [PATCH 07/30] 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 Date: Tue, 26 Nov 2024 12:37:05 +0100 Subject: [PATCH 08/30] 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 ${CMAKE_REQUIRED_FLAGS_PREV}) + elseif (APPLE) + if (GGML_NATIVE) + set(USER_PROVIDED_MARCH FALSE) + foreach(flag_var IN ITEMS CMAKE_C_FLAGS CMAKE_CXX_FLAGS CMAKE_REQUIRED_FLAGS) + if ("${${flag_var}}" MATCHES "-march=[a-zA-Z0-9+._-]+") + set(USER_PROVIDED_MARCH TRUE) + break() + endif() + endforeach() + + if (NOT USER_PROVIDED_MARCH) + set(MARCH_FLAGS "-march=armv8.2a") + + check_cxx_source_compiles("#include \nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD) + if (GGML_COMPILER_SUPPORT_DOTPROD) + set(MARCH_FLAGS "${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} ${TEST_I8MM_FLAGS}") + check_cxx_source_compiles("#include \nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8) + if (GGML_COMPILER_SUPPORT_MATMUL_INT8) + set(MARCH_FLAGS "${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 "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "") From c6807b3f28cc3dbfda3ec390bcb87e69fb5634e2 Mon Sep 17 00:00:00 2001 From: Diego Devesa Date: Tue, 26 Nov 2024 13:05:07 +0100 Subject: [PATCH 09/30] ci : add ubuntu cuda build, build with one arch on windows (#10456) --- .github/labeler.yml | 15 +++----- .github/workflows/build.yml | 59 ++++++++++++++++++++++++++++++- .github/workflows/nix-ci.yml | 2 ++ .github/workflows/python-lint.yml | 9 ++++- 4 files changed, 73 insertions(+), 12 deletions(-) diff --git a/.github/labeler.yml b/.github/labeler.yml index 89436740d1ffb..1b47bc96885c4 100644 --- a/.github/labeler.yml +++ b/.github/labeler.yml @@ -3,19 +3,18 @@ Kompute: - changed-files: - any-glob-to-any-file: - ggml/include/ggml-kompute.h - - ggml/src/ggml-kompute.cpp + - ggml/src/ggml-kompute/** - README-kompute.md Apple Metal: - changed-files: - any-glob-to-any-file: - ggml/include/ggml-metal.h - - ggml/src/ggml-metal.cpp + - ggml/src/ggml-metal/** - README-metal.md SYCL: - changed-files: - any-glob-to-any-file: - ggml/include/ggml-sycl.h - - ggml/src/ggml-sycl.cpp - ggml/src/ggml-sycl/** - docs/backend/SYCL.md - examples/sycl/** @@ -27,8 +26,8 @@ Nvidia GPU: Vulkan: - changed-files: - any-glob-to-any-file: - - ggml/ggml_vk_generate_shaders.py - - ggml/src/ggml-vulkan* + - ggml/include/ggml-vulkan.h + - ggml/src/ggml-vulkan/** documentation: - changed-files: - any-glob-to-any-file: @@ -75,11 +74,7 @@ server: ggml: - changed-files: - any-glob-to-any-file: - - ggml/include/ggml*.h - - ggml/src/ggml*.c - - ggml/src/ggml*.cpp - - ggml/src/ggml*.h - - ggml-cuda/** + - ggml/** nix: - changed-files: - any-glob-to-any-file: diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index abaf2c504c8f4..6281663ecb347 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -871,8 +871,65 @@ jobs: path: llama-${{ 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 Date: Tue, 26 Nov 2024 13:05:20 +0100 Subject: [PATCH 10/30] 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: ${{ github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }} uses: docker/build-push-action@v6 with: context: . From ab96610b1e58684bc5e8b810130c4cf6d8252e21 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 26 Nov 2024 14:18:08 +0200 Subject: [PATCH 11/30] cmake : enable warnings in llama (#10474) * cmake : enable warnings in llama ggml-ci * cmake : add llama_get_flags and respect LLAMA_FATAL_WARNINGS * cmake : get_flags -> ggml_get_flags * speculative-simple : fix warnings * cmake : reuse ggml_get_flags ggml-ci * speculative-simple : fix compile warning ggml-ci --- CMakeLists.txt | 1 + cmake/common.cmake | 33 +++++++++++++++++++ common/CMakeLists.txt | 2 ++ examples/CMakeLists.txt | 4 +++ .../speculative-simple/speculative-simple.cpp | 6 ++-- ggml/src/CMakeLists.txt | 5 +-- ggml/src/ggml-cuda/CMakeLists.txt | 2 +- src/CMakeLists.txt | 2 ++ 8 files changed, 49 insertions(+), 6 deletions(-) create mode 100644 cmake/common.cmake diff --git a/CMakeLists.txt b/CMakeLists.txt index e7d91a5b5c9ac..0d389dccbfb42 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -82,6 +82,7 @@ option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF) # Required for relocatable CMake package include(${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} ${CMAKE_CXX_COMPILER_VERSION}) + + add_compile_options("$<$:${C_FLAGS};${GF_C_FLAGS}>" + "$<$:${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 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} ${CMAKE_CXX_COMPILER_VERSION}) add_compile_options("$<$:${C_FLAGS};${GF_C_FLAGS}>" "$<$:${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 Date: Tue, 26 Nov 2024 21:43:47 +0800 Subject: [PATCH 12/30] 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 Date: Tue, 26 Nov 2024 16:20:18 +0100 Subject: [PATCH 13/30] 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 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 tokens are predicted matching twind|Anna - And the completion is 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 \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 \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 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 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 max tokens to predict - And streaming is - Given concurrent OAI completions requests - Then the server is busy - Then the server is idle - Then all prompts are predicted with 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 max tokens to predict - And streaming is - Given concurrent OAI completions requests no v1 - Then the server is busy - Then the server is idle - Then all prompts are predicted with 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 max tokens to predict - And streaming is - Given concurrent OAI completions requests - Then the server is busy - Then the server is idle - Then all prompts are predicted with 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 from HF repo - And as batch size - And as number of junk - And server max tokens to predict - And 42 as seed - And 0.0 temperature - And KV cache size - And 1 slots - And group attention factor to extend context size through self-extend - And group attention width to extend context size through self-extend - # Can be override with N_GPU_LAYERS - And 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 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 Remember it. 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 challenge prompt with the passkey inserted every junk - And a completion request with no api error - Then tokens are predicted matching - - 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 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 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 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 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 slots - And KV cache size - And 1.0 temperature - And 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 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 - And 4 max tokens to predict - And a completion request with 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 - Given an OAI compatible chat completions request with 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 - 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 - Then CORS header is set to - - 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 - And max tokens to predict - And a completion request with no api error - Then tokens are predicted matching - And the completion is truncated - And prompt tokens are processed - And prometheus metrics are exposed - And metric llamacpp:tokens_predicted is - - 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 - And a system prompt - And a user prompt - And max tokens to predict - And streaming is - Given an OAI compatible chat completions request with no api error - Then tokens are predicted matching - And prompt tokens are processed - And the completion is 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 - And 10 max tokens to predict - Given an OAI compatible chat completions request with no api error - Then tokens are predicted matching - - 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 \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 \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 Date: Tue, 26 Nov 2024 09:45:05 -0600 Subject: [PATCH 14/30] 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(dst->op_params[0]); + enum ggml_op_pool op = static_cast(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> 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 Date: Wed, 27 Nov 2024 00:00:41 +0800 Subject: [PATCH 15/30] mtgpu: Add MUSA_DOCKER_ARCH in Dockerfiles && update cmake and make (#10516) Signed-off-by: Xiaodong Ye --- .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 ${COMPILE_FLAGS}) endforeach() ggml_add_backend_library(ggml-musa From be0e350c8b69632b27d5fb41fa064fa256dd7fbf Mon Sep 17 00:00:00 2001 From: Tristan Druyen Date: Tue, 26 Nov 2024 19:27:28 +0100 Subject: [PATCH 16/30] Fix HIP flag inconsistency & build docs (#10524) * Fix inconsistency of HIP flags in cmake & make * Fix docs regarding GGML_HIP --- Makefile | 4 ++-- docs/build.md | 4 ++-- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/Makefile b/Makefile index 9a079a2377205..25214ec053df1 100644 --- a/Makefile +++ b/Makefile @@ -752,7 +752,7 @@ vulkan-shaders-gen: ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp endif # GGML_VULKAN -ifdef GGML_HIPBLAS +ifdef GGML_HIP ifeq ($(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 Date: Tue, 26 Nov 2024 21:01:47 +0100 Subject: [PATCH 17/30] 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 Date: Tue, 26 Nov 2024 21:13:54 +0100 Subject: [PATCH 18/30] 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 Date: Tue, 26 Nov 2024 12:55:29 -0800 Subject: [PATCH 19/30] 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 Date: Tue, 26 Nov 2024 22:12:10 +0100 Subject: [PATCH 20/30] 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 Date: Wed, 27 Nov 2024 01:00:50 -0600 Subject: [PATCH 21/30] 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 Date: Wed, 27 Nov 2024 01:08:54 -0600 Subject: [PATCH 22/30] 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 Date: Wed, 27 Nov 2024 01:21:59 -0600 Subject: [PATCH 23/30] 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 Date: Wed, 27 Nov 2024 01:30:27 -0600 Subject: [PATCH 24/30] 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 l_warptile, m_warptile, s_warptile, l_warptile_mmq, m_warptile_mmq, s_warptile_mmq; - std::array l_wg_denoms, m_wg_denoms, s_wg_denoms; + std::vector l_warptile, m_warptile, s_warptile, + l_warptile_mmq, m_warptile_mmq, s_warptile_mmq; + std::array 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(); device->pipeline_matmul_f32_f16 = std::make_shared(); @@ -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 Date: Wed, 27 Nov 2024 01:32:54 -0600 Subject: [PATCH 25/30] 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 Date: Tue, 26 Nov 2024 15:50:26 +0200 Subject: [PATCH 26/30] 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 #endif // __ARM_FEATURE_SVE -#if defined(__ARM_NEON) +#if defined(__ARM_NEON) && !defined(__CUDACC__) // if YCM cannot find , 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 Date: Wed, 27 Nov 2024 11:10:42 +0200 Subject: [PATCH 27/30] 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 Date: Wed, 27 Nov 2024 11:22:14 +0200 Subject: [PATCH 28/30] 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 Date: Wed, 27 Nov 2024 11:03:25 +0100 Subject: [PATCH 29/30] 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 From 3ad5451f3b75809e3033e4e577b9f60bcaf6676a Mon Sep 17 00:00:00 2001 From: uvos Date: Wed, 27 Nov 2024 17:10:08 +0100 Subject: [PATCH 30/30] Add some minimal optimizations for CDNA (#10498) * Add some minimal optimizations for CDNA * ggml_cuda: set launch bounds also for GCN as it helps there too --- ggml/src/ggml-cuda/common.cuh | 17 ++++++++++++++--- ggml/src/ggml-cuda/ggml-cuda.cu | 11 ++++++++++- ggml/src/ggml-cuda/mmq.cu | 2 +- ggml/src/ggml-cuda/mmq.cuh | 4 ++-- ggml/src/ggml-cuda/mmvq.cu | 2 +- ggml/src/ggml-cuda/vendors/hip.h | 8 ++++++++ 6 files changed, 36 insertions(+), 8 deletions(-) diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index b0dd16066b4ba..535118d87928e 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -47,9 +47,20 @@ #define CC_TURING 750 #define CC_AMPERE 800 #define CC_OFFSET_AMD 1000000 -#define CC_RDNA1 (CC_OFFSET_AMD + 1010) -#define CC_RDNA2 (CC_OFFSET_AMD + 1030) -#define CC_RDNA3 (CC_OFFSET_AMD + 1100) + +// GCN/CNDA, wave size is 64 +#define CC_GCN4 (CC_OFFSET_AMD + 803) // Tonga, Fiji, Polaris, minimum for fast fp16 +#define CC_VEGA (CC_OFFSET_AMD + 900) // Vega56/64, minimum for fp16 dual issue +#define CC_VEGA20 (CC_OFFSET_AMD + 906) // MI50/Radeon VII, minimum for dp4a +#define CC_CDNA (CC_OFFSET_AMD + 908) // MI100, minimum for MFMA, acc registers +#define CC_CDNA2 (CC_OFFSET_AMD + 910) // MI210, minimum acc register renameing +#define CC_CDNA3 (CC_OFFSET_AMD + 942) // MI300 + +// RNDA removes MFMA, dp4a, xnack, acc registers, wave size is 32 +#define CC_RDNA1 (CC_OFFSET_AMD + 1010) // RX 5000 +#define CC_RDNA2 (CC_OFFSET_AMD + 1030) // RX 6000, minimum for dp4a +#define CC_RDNA3 (CC_OFFSET_AMD + 1100) // RX 7000, minimum for WMMA + #define CC_QY1 210 #define CC_QY2 220 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 2a78a4393d0f7..d6e4bfdd0d437 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -1107,6 +1107,11 @@ static void ggml_cuda_op_mul_mat_cublas( const half alpha_f16 = 1.0f; const half beta_f16 = 0.0f; + cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F; + if (ggml_cuda_info().devices[ctx.device].cc == CC_CDNA) { + cu_compute_type = CUBLAS_COMPUTE_32F; + } + CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream)); CUBLAS_CHECK( cublasGemmEx(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N, @@ -1114,7 +1119,7 @@ static void ggml_cuda_op_mul_mat_cublas( &alpha_f16, src0_ptr, CUDA_R_16F, ne00, src1_ptr, CUDA_R_16F, ne10, &beta_f16, dst_f16.get(), CUDA_R_16F, ldc, - CUBLAS_COMPUTE_16F, + cu_compute_type, CUBLAS_GEMM_DEFAULT_TENSOR_OP)); const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16); @@ -1607,6 +1612,10 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F; cudaDataType_t cu_data_type = CUDA_R_16F; + if (ggml_cuda_info().devices[ctx.device].cc == CC_CDNA) { + cu_compute_type = CUBLAS_COMPUTE_32F; + } + // dst strides size_t nbd2 = dst->nb[2]; size_t nbd3 = dst->nb[3]; diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu index ae5c68ab35129..7f7c8c90b6fe2 100644 --- a/ggml/src/ggml-cuda/mmq.cu +++ b/ggml/src/ggml-cuda/mmq.cu @@ -148,5 +148,5 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) { return cc < CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; } - return cc < CC_RDNA3 || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; + return (cc < CC_RDNA3 && cc != CC_CDNA && cc != CC_VEGA20) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; } diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh index 425acb20da311..8d8867121f321 100644 --- a/ggml/src/ggml-cuda/mmq.cuh +++ b/ggml/src/ggml-cuda/mmq.cuh @@ -2570,9 +2570,9 @@ static __device__ void mul_mat_q_process_tile( template #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) +#if defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN) __launch_bounds__(WARP_SIZE*nwarps, 2) -#endif // defined(RDNA3) || defined(RDNA2) +#endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN) #else #if __CUDA_ARCH__ >= CC_VOLTA __launch_bounds__(WARP_SIZE*nwarps, 1) diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu index 735975c160dd0..02d1509836c40 100644 --- a/ggml/src/ggml-cuda/mmvq.cu +++ b/ggml/src/ggml-cuda/mmvq.cu @@ -142,7 +142,7 @@ static void mul_mat_vec_q_cuda( int64_t nwarps = 1; int64_t rows_per_cuda_block = 1; - if (ggml_cuda_info().devices[id].cc < CC_RDNA2) { // NVIDIA and AMD older than RDNA2 + if (ggml_cuda_info().devices[id].cc < CC_CDNA || ggml_cuda_info().devices[id].cc == CC_RDNA1) { // NVIDIA and AMD older than RDNA2 but not CDNA switch(ncols_y) { case 1: nwarps = 4; diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h index 1f3c70c2e6934..3205534d66f10 100644 --- a/ggml/src/ggml-cuda/vendors/hip.h +++ b/ggml/src/ggml-cuda/vendors/hip.h @@ -95,6 +95,14 @@ #define __CUDA_ARCH__ 1300 +#if defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) +#define GCN +#endif + +#if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx942__) +#define CDNA +#endif + #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \ defined(__gfx1150__) || defined(__gfx1151__) #define RDNA3