diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index f3326a5fb..886d33d2d 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -552,35 +552,44 @@ jobs: -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO -# TODO: tmp disabled. see for possible re-enable: -# https://github.com/ggerganov/llama.cpp/pull/10525 -# macOS-latest-swift: -# runs-on: macos-latest -# -# strategy: -# matrix: -# destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS'] -# -# steps: -# - name: Clone -# id: checkout -# uses: actions/checkout@v4 -# -# - name: Dependencies -# id: depends -# continue-on-error: true -# run: | -# brew update -# -# - name: xcodebuild for swift package -# id: xcodebuild -# run: | -# xcodebuild -scheme llama -destination "${{ matrix.destination }}" -# -# - name: Build Swift Example -# id: make_build_swift_example -# run: | -# make swift + macOS-latest-swift: + runs-on: macos-latest + + strategy: + matrix: + destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS'] + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v4 + + - name: Dependencies + id: depends + continue-on-error: true + run: | + brew update + + - name: Build llama.cpp with CMake + id: cmake_build + run: | + sysctl -a + mkdir build + cd build + cmake -G Xcode .. \ + -DGGML_METAL_USE_BF16=ON \ + -DGGML_METAL_EMBED_LIBRARY=ON \ + -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TESTS=OFF \ + -DLLAMA_BUILD_SERVER=OFF \ + -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" + cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) + sudo cmake --install . --config Release + + - name: xcodebuild for swift package + id: xcodebuild + run: | + xcodebuild -scheme llama-Package -destination "${{ matrix.destination }}" windows-msys2: runs-on: windows-latest @@ -1104,6 +1113,29 @@ jobs: - name: Checkout code uses: actions/checkout@v4 + - name: Build + id: cmake_build + run: | + sysctl -a + mkdir build + cd build + cmake -G Xcode .. \ + -DGGML_METAL_USE_BF16=ON \ + -DGGML_METAL_EMBED_LIBRARY=ON \ + -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TESTS=OFF \ + -DLLAMA_BUILD_SERVER=OFF \ + -DCMAKE_SYSTEM_NAME=iOS \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \ + -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml + cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO + sudo cmake --install . --config Release + + - name: xcodebuild for swift package + id: xcodebuild + run: | + xcodebuild -scheme llama-Package -destination 'generic/platform=iOS' + - name: Build Xcode project run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build @@ -1131,23 +1163,6 @@ jobs: ./gradlew build --no-daemon -# freeBSD-latest: -# runs-on: macos-12 -# steps: -# - name: Clone -# uses: actions/checkout@v4 -# -# - name: Build -# uses: cross-platform-actions/action@v0.19.0 -# with: -# operating_system: freebsd -# version: '13.2' -# hypervisor: 'qemu' -# run: | -# sudo pkg update -# sudo pkg install -y gmake automake autoconf pkgconf llvm15 openblas -# gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j `sysctl -n hw.ncpu` - release: if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} diff --git a/CMakeLists.txt b/CMakeLists.txt index f84fff9e6..a717a508f 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -46,11 +46,9 @@ if (WIN32) add_compile_definitions(_CRT_SECURE_NO_WARNINGS) endif() -if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC") - add_compile_options("$<$:/source-charset:utf-8>") - add_compile_options("$<$:/source-charset:utf-8>") - add_compile_options("$<$:/execution-charset:utf-8>") - add_compile_options("$<$:/execution-charset:utf-8>") +if (MSVC) + add_compile_options("$<$:/utf-8>") + add_compile_options("$<$:/utf-8>") endif() # diff --git a/CMakePresets.json b/CMakePresets.json index 436448967..13bdd7907 100644 --- a/CMakePresets.json +++ b/CMakePresets.json @@ -31,6 +31,13 @@ { "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } }, { "name": "vulkan", "hidden": true, "cacheVariables": { "GGML_VULKAN": "ON" } }, + { + "name": "x64-windows-llvm", "hidden": true, + "cacheVariables": { + "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/x64-windows-llvm.cmake" + } + }, + { "name": "arm64-windows-msvc", "hidden": true, "architecture": { "value": "arm64", "strategy": "external" }, @@ -70,6 +77,11 @@ { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg" ] }, { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg", "static" ] }, + { "name": "x64-windows-llvm-debug", "inherits": [ "base", "x64-windows-llvm", "debug" ] }, + { "name": "x64-windows-llvm-release", "inherits": [ "base", "x64-windows-llvm", "release" ] }, + { "name": "x64-windows-llvm-reldbg", "inherits": [ "base", "x64-windows-llvm", "reldbg" ] }, + { "name": "x64-windows-llvm+static-release", "inherits": [ "base", "x64-windows-llvm", "reldbg", "static" ] }, + { "name": "x64-windows-msvc-debug", "inherits": [ "base", "debug" ] }, { "name": "x64-windows-msvc-release", "inherits": [ "base", "reldbg" ] }, { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] }, diff --git a/Makefile b/Makefile index 8c6ef2440..e4efb242b 100644 --- a/Makefile +++ b/Makefile @@ -448,6 +448,10 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64)) MK_CFLAGS += -march=native -mtune=native HOST_CXXFLAGS += -march=native -mtune=native + # Usage AMX build test + #MK_CFLAGS += -march=graniterapids -mtune=graniterapids + #HOST_CXXFLAGS += -march=graniterapids -mtune=graniterapids + # Usage AVX-only #MK_CFLAGS += -mfma -mf16c -mavx #MK_CXXFLAGS += -mfma -mf16c -mavx @@ -951,7 +955,6 @@ DIR_COMMON = common OBJ_GGML = \ $(DIR_GGML)/src/ggml.o \ - $(DIR_GGML)/src/ggml-aarch64.o \ $(DIR_GGML)/src/ggml-alloc.o \ $(DIR_GGML)/src/ggml-backend.o \ $(DIR_GGML)/src/ggml-backend-reg.o \ @@ -959,9 +962,11 @@ OBJ_GGML = \ $(DIR_GGML)/src/ggml-quants.o \ $(DIR_GGML)/src/ggml-threading.o \ $(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \ - $(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o \ + $(DIR_GGML)/src/ggml-cpu/ggml-cpu_cpp.o \ $(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \ + $(DIR_GGML)/src/ggml-cpu/ggml-cpu-hbm.o \ $(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \ + $(DIR_GGML)/src/ggml-cpu/ggml-cpu-traits.o \ $(OBJ_GGML_EXT) OBJ_LLAMA = \ @@ -1102,17 +1107,10 @@ DEP_FILES = $(OBJ_GGML:.o=.d) $(OBJ_LLAMA:.o=.d) $(OBJ_COMMON:.o=.d) # Default target all: $(BUILD_TARGETS) +# force c++ build for source file that have same name as c file # Note: need this exception because `ggml-cpu.c` and `ggml-cpu.cpp` both produce the same obj/dep files -# g++ -M -I ./ggml/include/ -I ./ggml/src ggml/src/ggml-cpu/ggml-cpu.cpp | grep ggml -$(DIR_GGML)/src/ggml-cpu/ggml-cpu-cpp.o: \ - ggml/src/ggml-cpu/ggml-cpu.cpp \ - ggml/include/ggml-backend.h \ - ggml/include/ggml.h \ - ggml/include/ggml-alloc.h \ - ggml/src/ggml-backend-impl.h \ - ggml/include/ggml-cpu.h \ - ggml/src/ggml-impl.h - $(CXX) $(CXXFLAGS) -c $< -o $@ +$(DIR_GGML)/%_cpp.o: $(DIR_GGML)/%.cpp + $(CXX) $(CXXFLAGS) -MMD -c $< -o $@ # Rules for building object files $(DIR_GGML)/%.o: $(DIR_GGML)/%.c diff --git a/Package.swift b/Package.swift index d32b74a63..01c996d24 100644 --- a/Package.swift +++ b/Package.swift @@ -2,59 +2,6 @@ import PackageDescription -var sources = [ - "src/llama.cpp", - "src/llama-vocab.cpp", - "src/llama-grammar.cpp", - "src/llama-sampling.cpp", - "src/unicode.cpp", - "src/unicode-data.cpp", - "ggml/src/ggml.c", - "ggml/src/ggml-aarch64.c", - "ggml/src/ggml-alloc.c", - "ggml/src/ggml-backend.cpp", - "ggml/src/ggml-backend-reg.cpp", - "ggml/src/ggml-cpu/ggml-cpu.c", - "ggml/src/ggml-cpu/ggml-cpu.cpp", - "ggml/src/ggml-cpu/ggml-cpu-aarch64.c", - "ggml/src/ggml-cpu/ggml-cpu-quants.c", - "ggml/src/ggml-threading.cpp", - "ggml/src/ggml-quants.c", -] - -var resources: [Resource] = [] -var linkerSettings: [LinkerSetting] = [] -var cSettings: [CSetting] = [ - .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]), - .unsafeFlags(["-fno-objc-arc"]), - .headerSearchPath("ggml/src"), - .headerSearchPath("ggml/src/ggml-cpu"), - // NOTE: NEW_LAPACK will required iOS version 16.4+ - // We should consider add this in the future when we drop support for iOS 14 - // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc) - // .define("ACCELERATE_NEW_LAPACK"), - // .define("ACCELERATE_LAPACK_ILP64") - .define("GGML_USE_CPU"), -] - - -#if canImport(Darwin) -sources.append("ggml/src/ggml-common.h") -sources.append("ggml/src/ggml-metal/ggml-metal.m") -resources.append(.process("ggml/src/ggml-metal/ggml-metal.metal")) -linkerSettings.append(.linkedFramework("Accelerate")) -cSettings.append( - contentsOf: [ - .define("GGML_USE_ACCELERATE"), - .define("GGML_USE_METAL"), - ] -) -#endif - -#if os(Linux) - cSettings.append(.define("_GNU_SOURCE")) -#endif - let package = Package( name: "llama", platforms: [ @@ -67,26 +14,6 @@ let package = Package( .library(name: "llama", targets: ["llama"]), ], targets: [ - .target( - name: "llama", - path: ".", - exclude: [ - "build", - "cmake", - "examples", - "scripts", - "models", - "tests", - "CMakeLists.txt", - "Makefile", - "ggml/src/ggml-metal-embed.metal" - ], - sources: sources, - resources: resources, - publicHeadersPath: "spm-headers", - cSettings: cSettings, - linkerSettings: linkerSettings - ) - ], - cxxLanguageStandard: .cxx17 + .systemLibrary(name: "llama", pkgConfig: "llama"), + ] ) diff --git a/Sources/llama/llama.h b/Sources/llama/llama.h new file mode 100644 index 000000000..41725880e --- /dev/null +++ b/Sources/llama/llama.h @@ -0,0 +1,4 @@ +#pragma once + +#include + diff --git a/Sources/llama/module.modulemap b/Sources/llama/module.modulemap new file mode 100644 index 000000000..d010555b1 --- /dev/null +++ b/Sources/llama/module.modulemap @@ -0,0 +1,5 @@ +module llama [system] { + header "llama.h" + link "llama" + export * +} diff --git a/cmake/llama.pc.in b/cmake/llama.pc.in index 326acbb61..0b2b6bcfa 100644 --- a/cmake/llama.pc.in +++ b/cmake/llama.pc.in @@ -6,5 +6,5 @@ includedir=${prefix}/include Name: llama Description: Port of Facebook's LLaMA model in C/C++ Version: @PROJECT_VERSION@ -Libs: -L${libdir} -lllama +Libs: -L${libdir} -lggml -lggml-base -lllama Cflags: -I${includedir} diff --git a/cmake/x64-windows-llvm.cmake b/cmake/x64-windows-llvm.cmake new file mode 100644 index 000000000..0603d738f --- /dev/null +++ b/cmake/x64-windows-llvm.cmake @@ -0,0 +1,11 @@ +set( CMAKE_SYSTEM_NAME Windows ) +set( CMAKE_SYSTEM_PROCESSOR x86_64 ) + +set( CMAKE_C_COMPILER clang ) +set( CMAKE_CXX_COMPILER clang++ ) + +set( arch_c_flags "-march=native" ) + +set( CMAKE_C_FLAGS_INIT "${arch_c_flags}" ) +set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags}" ) + diff --git a/common/common.h b/common/common.h index 1e5890404..92cd4e9c1 100644 --- a/common/common.h +++ b/common/common.h @@ -222,7 +222,7 @@ struct common_params { struct common_params_speculative speculative; std::string model = ""; // model path // NOLINT - std::string model_alias = "unknown"; // model alias // NOLINT + std::string model_alias = ""; // model alias // NOLINT std::string model_url = ""; // model url to download // NOLINT std::string hf_token = ""; // HF token // NOLINT std::string hf_repo = ""; // HF repo // NOLINT diff --git a/common/speculative.cpp b/common/speculative.cpp index e559675c4..3fcbb0020 100644 --- a/common/speculative.cpp +++ b/common/speculative.cpp @@ -62,6 +62,10 @@ struct common_speculative * common_speculative_init( } void common_speculative_free(struct common_speculative * spec) { + if (spec == nullptr) { + return; + } + common_sampler_free(spec->smpl); llama_batch_free(spec->batch); diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 9f1419e29..c63d929c1 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -661,6 +661,9 @@ class Model: if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35": # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0 res = "minerva-7b" + if chkhsh == "8b5a93ed704057481f240da0be7e7dca721d7f8f4755263b6807227a2cbeae65": + # ref: https://huggingface.co/sentence-transformers/stsb-roberta-base + res = "roberta-bpe" if res is None: logger.warning("\n") @@ -1989,6 +1992,14 @@ class Qwen2Model(Model): except FileNotFoundError: self._set_vocab_gpt2() + def set_gguf_parameters(self): + super().set_gguf_parameters() + if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: + if self.hparams["rope_scaling"].get("type") == "yarn": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) + self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"]) + @Model.register("Qwen2MoeForCausalLM") class Qwen2MoeModel(Model): @@ -2533,7 +2544,7 @@ class InternLM2Model(Model): return [(self.map_tensor_name(name), data_torch)] -@Model.register("BertModel", "CamembertModel") +@Model.register("BertModel", "CamembertModel", "RobertaModel") class BertModel(Model): model_arch = gguf.MODEL_ARCH.BERT @@ -2574,7 +2585,8 @@ class BertModel(Model): # we need this to validate the size of the token_type embeddings # though currently we are passing all zeros to the token_type embeddings - self.gguf_writer.add_token_type_count(2) # "Sequence A" or "Sequence B" + # "Sequence A" or "Sequence B" + self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1)) # convert to phantom space vocab def phantom(tok): diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index ce3c571df..aa10e5db7 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -103,6 +103,7 @@ models = [ {"name": "phi-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", }, {"name": "chameleon", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/facebook/chameleon-7b", }, {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", }, + {"name": "roberta-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sentence-transformers/stsb-roberta-base"}, ] diff --git a/docs/build.md b/docs/build.md index a4964cbd1..84019b204 100644 --- a/docs/build.md +++ b/docs/build.md @@ -55,7 +55,14 @@ cmake --build build --config Release cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF cmake --build build-arm64-windows-llvm-release ``` - Building for arm64 can also be done with the MSVC compiler with the build-arm64-windows-MSVC preset, or the standard CMake build instructions. However, note that the MSVC compiler does not support inline ARM assembly code, used e.g. for the accelerated Q4_0_4_8 CPU kernels. + Building for arm64 can also be done with the MSVC compiler with the build-arm64-windows-MSVC preset, or the standard CMake build instructions. However, note that the MSVC compiler does not support inline ARM assembly code, used e.g. for the accelerated Q4_0_N_M CPU kernels. + + For building with ninja generator and clang compiler as default: + -set path:set LIB=C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\um\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\lib\x64\uwp;C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\ucrt\x64 + ```bash + cmake --preset x64-windows-llvm-release + cmake --build build-x64-windows-llvm-release + ``` ## BLAS Build diff --git a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift index 65cd4eb51..998c673d5 100644 --- a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift +++ b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift @@ -210,20 +210,20 @@ actor LlamaContext { llama_kv_cache_clear(context) - let t_pp_start = ggml_time_us() + let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000; if llama_decode(context, batch) != 0 { print("llama_decode() failed during prompt") } llama_synchronize(context) - let t_pp_end = ggml_time_us() + let t_pp_end = DispatchTime.now().uptimeNanoseconds / 1000; // bench text generation llama_kv_cache_clear(context) - let t_tg_start = ggml_time_us() + let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000; for i in 0.. QUANT_OPTIONS = { { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 5.33G, +0.0569 ppl @ Llama-3-8B", }, { "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 6.14G, +0.0217 ppl @ Llama-3-8B", }, { "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 7.96G, +0.0026 ppl @ Llama-3-8B", }, - { "Q4_0_4_4", LLAMA_FTYPE_MOSTLY_Q4_0_4_4, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, - { "Q4_0_4_8", LLAMA_FTYPE_MOSTLY_Q4_0_4_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, - { "Q4_0_8_8", LLAMA_FTYPE_MOSTLY_Q4_0_8_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, { "F16", LLAMA_FTYPE_MOSTLY_F16, "14.00G, +0.0020 ppl @ Mistral-7B", }, { "BF16", LLAMA_FTYPE_MOSTLY_BF16, "14.00G, -0.0050 ppl @ Mistral-7B", }, { "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", }, diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt index 0035859a6..63fca1d59 100644 --- a/examples/server/CMakeLists.txt +++ b/examples/server/CMakeLists.txt @@ -34,14 +34,6 @@ endforeach() add_executable(${TARGET} ${TARGET_SRCS}) install(TARGETS ${TARGET} RUNTIME) -# clean up generated files in pre-build step -foreach(asset ${PUBLIC_ASSETS}) - set(output "${CMAKE_CURRENT_BINARY_DIR}/${asset}.hpp") - add_custom_command(TARGET ${TARGET} PRE_BUILD - COMMAND "${CMAKE_COMMAND}" -E remove -f "${output}" - ) -endforeach() - target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT}) if (LLAMA_SERVER_SSL) diff --git a/examples/server/README.md b/examples/server/README.md index 4d1da8da0..738153981 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -473,9 +473,11 @@ Notice that each `probs` is an array of length `n_probs`. - `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`. These options may differ from the original ones in some way (e.g. bad values filtered out, strings converted to tokens, etc.). - `model`: The path to the model loaded with `-m` - `prompt`: The provided `prompt` -- `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token -- `stopped_limit`: Indicating whether the completion stopped because `n_predict` tokens were generated before stop words or EOS was encountered -- `stopped_word`: Indicating whether the completion stopped due to encountering a stopping word from `stop` JSON array provided +- `stop_type`: Indicating whether the completion has stopped. Possible values are: + - `none`: Generating (not stopped) + - `eos`: Stopped because it encountered the EOS token + - `limit`: Stopped because `n_predict` tokens were generated before stop words or EOS was encountered + - `word`: Stopped due to encountering a stopping word from `stop` JSON array provided - `stopping_word`: The stopping word encountered which stopped the generation (or "" if not stopped due to a stopping word) - `timings`: Hash of timing information about the completion such as the number of tokens `predicted_per_second` - `tokens_cached`: Number of tokens from the prompt which could be re-used from previous completion (`n_past`) @@ -616,14 +618,83 @@ This endpoint is public (no API key check). By default, it is read-only. To make ```json { - "default_generation_settings": { ... }, + "default_generation_settings": { + "id": 0, + "id_task": -1, + "n_ctx": 1024, + "speculative": false, + "is_processing": false, + "params": { + "n_predict": -1, + "seed": 4294967295, + "temperature": 0.800000011920929, + "dynatemp_range": 0.0, + "dynatemp_exponent": 1.0, + "top_k": 40, + "top_p": 0.949999988079071, + "min_p": 0.05000000074505806, + "xtc_probability": 0.0, + "xtc_threshold": 0.10000000149011612, + "typical_p": 1.0, + "repeat_last_n": 64, + "repeat_penalty": 1.0, + "presence_penalty": 0.0, + "frequency_penalty": 0.0, + "dry_multiplier": 0.0, + "dry_base": 1.75, + "dry_allowed_length": 2, + "dry_penalty_last_n": -1, + "dry_sequence_breakers": [ + "\n", + ":", + "\"", + "*" + ], + "mirostat": 0, + "mirostat_tau": 5.0, + "mirostat_eta": 0.10000000149011612, + "penalize_nl": false, + "stop": [], + "max_tokens": -1, + "n_keep": 0, + "n_discard": 0, + "ignore_eos": false, + "stream": true, + "n_probs": 0, + "min_keep": 0, + "grammar": "", + "samplers": [ + "dry", + "top_k", + "typ_p", + "top_p", + "min_p", + "xtc", + "temperature" + ], + "speculative.n_max": 16, + "speculative.n_min": 5, + "speculative.p_min": 0.8999999761581421, + "timings_per_token": false + }, + "prompt": "", + "next_token": { + "has_next_token": true, + "has_new_line": false, + "n_remain": -1, + "n_decoded": 0, + "stopping_word": "" + } + }, "total_slots": 1, - "chat_template": "" + "model_path": "../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf", + "chat_template": "..." } ``` - `default_generation_settings` - the default generation settings for the `/completion` endpoint, which has the same fields as the `generation_settings` response object from the `/completion` endpoint. - `total_slots` - the total number of slots for process requests (defined by `--parallel` option) +- `model_path` - the path to model file (same with `-m` argument) - `chat_template` - the model's original Jinja2 prompt template ### POST `/props`: Change server global properties. @@ -816,56 +887,74 @@ Example: ```json [ - { - "dynatemp_exponent": 1.0, - "dynatemp_range": 0.0, - "frequency_penalty": 0.0, - "grammar": "", - "id": 0, - "ignore_eos": false, - "is_processing": false, - "logit_bias": [], - "min_p": 0.05000000074505806, - "mirostat": 0, - "mirostat_eta": 0.10000000149011612, - "mirostat_tau": 5.0, - "model": "llama-2-7b-32k-instruct.Q2_K.gguf", - "n_ctx": 2048, - "n_keep": 0, - "n_predict": 100000, - "n_probs": 0, - "next_token": { - "has_next_token": true, - "n_remain": -1, - "n_decoded": 0, - "stopped_eos": false, - "stopped_limit": false, - "stopped_word": false, - "stopping_word": "" - }, - "penalize_nl": true, - "presence_penalty": 0.0, - "prompt": "Say hello to llama.cpp", - "repeat_last_n": 64, - "repeat_penalty": 1.100000023841858, - "samplers": [ - "top_k", - "typical_p", - "top_p", - "min_p", - "temperature" - ], - "seed": 42, - "stop": [ - "\n" - ], - "stream": false, - "task_id": 0, - "temperature": 0.0, - "top_k": 40, - "top_p": 0.949999988079071, - "typical_p": 1.0 + { + "id": 0, + "id_task": -1, + "n_ctx": 1024, + "speculative": false, + "is_processing": false, + "params": { + "n_predict": -1, + "seed": 4294967295, + "temperature": 0.800000011920929, + "dynatemp_range": 0.0, + "dynatemp_exponent": 1.0, + "top_k": 40, + "top_p": 0.949999988079071, + "min_p": 0.05000000074505806, + "xtc_probability": 0.0, + "xtc_threshold": 0.10000000149011612, + "typical_p": 1.0, + "repeat_last_n": 64, + "repeat_penalty": 1.0, + "presence_penalty": 0.0, + "frequency_penalty": 0.0, + "dry_multiplier": 0.0, + "dry_base": 1.75, + "dry_allowed_length": 2, + "dry_penalty_last_n": -1, + "dry_sequence_breakers": [ + "\n", + ":", + "\"", + "*" + ], + "mirostat": 0, + "mirostat_tau": 5.0, + "mirostat_eta": 0.10000000149011612, + "penalize_nl": false, + "stop": [], + "max_tokens": -1, + "n_keep": 0, + "n_discard": 0, + "ignore_eos": false, + "stream": true, + "n_probs": 0, + "min_keep": 0, + "grammar": "", + "samplers": [ + "dry", + "top_k", + "typ_p", + "top_p", + "min_p", + "xtc", + "temperature" + ], + "speculative.n_max": 16, + "speculative.n_min": 5, + "speculative.p_min": 0.8999999761581421, + "timings_per_token": false + }, + "prompt": "", + "next_token": { + "has_next_token": true, + "has_new_line": false, + "n_remain": -1, + "n_decoded": 0, + "stopping_word": "" } + } ] ``` diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 007c6ec82..10cf5262f 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -33,8 +33,10 @@ using json = nlohmann::ordered_json; enum stop_type { - STOP_TYPE_FULL, - STOP_TYPE_PARTIAL, + STOP_TYPE_NONE, + STOP_TYPE_EOS, + STOP_TYPE_WORD, + STOP_TYPE_LIMIT, }; // state diagram: https://github.com/ggerganov/llama.cpp/pull/9283 @@ -52,7 +54,10 @@ enum server_state { }; enum server_task_type { - SERVER_TASK_TYPE_INFERENCE, + SERVER_TASK_TYPE_COMPLETION, + SERVER_TASK_TYPE_EMBEDDING, + SERVER_TASK_TYPE_RERANK, + SERVER_TASK_TYPE_INFILL, SERVER_TASK_TYPE_CANCEL, SERVER_TASK_TYPE_NEXT_RESPONSE, SERVER_TASK_TYPE_METRICS, @@ -62,40 +67,15 @@ enum server_task_type { SERVER_TASK_TYPE_SET_LORA, }; -enum server_task_inf_type { - SERVER_TASK_INF_TYPE_COMPLETION, - SERVER_TASK_INF_TYPE_EMBEDDING, - SERVER_TASK_INF_TYPE_RERANK, - SERVER_TASK_INF_TYPE_INFILL, -}; - -struct server_task { - int id = -1; // to be filled by server_queue - int id_target = -1; // used by SERVER_TASK_TYPE_CANCEL - - llama_tokens prompt_tokens; - server_task_type type; - json data; - - server_task_inf_type inf_type = SERVER_TASK_INF_TYPE_COMPLETION; - - // utility function - static std::unordered_set get_list_id(const std::vector & tasks) { - std::unordered_set ids(tasks.size()); - for (size_t i = 0; i < tasks.size(); i++) { - ids.insert(tasks[i].id); - } - return ids; - } -}; - -struct server_task_result { - int id = -1; - - json data; - - bool stop; - bool error; +// https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11 +enum error_type { + ERROR_TYPE_INVALID_REQUEST, + ERROR_TYPE_AUTHENTICATION, + ERROR_TYPE_SERVER, + ERROR_TYPE_NOT_FOUND, + ERROR_TYPE_PERMISSION, + ERROR_TYPE_UNAVAILABLE, // custom error + ERROR_TYPE_NOT_SUPPORTED, // custom error }; struct slot_params { @@ -113,17 +93,906 @@ struct slot_params { json input_prefix; json input_suffix; std::vector antiprompt; + std::vector grammar_triggers; + bool timings_per_token = false; + bool ignore_eos = false; struct common_params_sampling sampling; struct common_params_speculative speculative; + + // OAI-compat fields + bool verbose = false; + bool oaicompat = false; + bool oaicompat_chat = true; + std::string oaicompat_model; + std::string oaicompat_cmpl_id; + json oaicompat_tools; + llama_tool_call_style oaicompat_tool_call_style = llama_tool_call_style::None; + + json to_json() const { + std::vector samplers; + samplers.reserve(sampling.samplers.size()); + for (const auto & sampler : sampling.samplers) { + samplers.emplace_back(common_sampler_type_to_str(sampler)); + } + + return json { + {"n_predict", n_predict}, // Server configured n_predict + {"seed", sampling.seed}, + {"temperature", sampling.temp}, + {"dynatemp_range", sampling.dynatemp_range}, + {"dynatemp_exponent", sampling.dynatemp_exponent}, + {"top_k", sampling.top_k}, + {"top_p", sampling.top_p}, + {"min_p", sampling.min_p}, + {"xtc_probability", sampling.xtc_probability}, + {"xtc_threshold", sampling.xtc_threshold}, + {"typical_p", sampling.typ_p}, + {"repeat_last_n", sampling.penalty_last_n}, + {"repeat_penalty", sampling.penalty_repeat}, + {"presence_penalty", sampling.penalty_present}, + {"frequency_penalty", sampling.penalty_freq}, + {"dry_multiplier", sampling.dry_multiplier}, + {"dry_base", sampling.dry_base}, + {"dry_allowed_length", sampling.dry_allowed_length}, + {"dry_penalty_last_n", sampling.dry_penalty_last_n}, + {"dry_sequence_breakers", sampling.dry_sequence_breakers}, + {"mirostat", sampling.mirostat}, + {"mirostat_tau", sampling.mirostat_tau}, + {"mirostat_eta", sampling.mirostat_eta}, + {"penalize_nl", sampling.penalize_nl}, + {"stop", antiprompt}, + {"max_tokens", n_predict}, // User configured n_predict + {"n_keep", n_keep}, + {"n_discard", n_discard}, + {"ignore_eos", sampling.ignore_eos}, + {"stream", stream}, + {"logit_bias", format_logit_bias(sampling.logit_bias)}, + {"n_probs", sampling.n_probs}, + {"min_keep", sampling.min_keep}, + {"grammar", sampling.grammar}, + {"samplers", samplers}, + {"speculative.n_max", speculative.n_max}, + {"speculative.n_min", speculative.n_min}, + {"speculative.p_min", speculative.p_min}, + {"timings_per_token", timings_per_token}, + }; + } +}; + +struct server_task { + int id = -1; // to be filled by server_queue + int index = -1; // used when there are multiple prompts (batch request) + + server_task_type type; + + // used by SERVER_TASK_TYPE_CANCEL + int id_target = -1; + + // used by SERVER_TASK_TYPE_INFERENCE + slot_params params; + llama_tokens prompt_tokens; + int id_selected_slot = -1; + + // used by SERVER_TASK_TYPE_SLOT_SAVE, SERVER_TASK_TYPE_SLOT_RESTORE, SERVER_TASK_TYPE_SLOT_ERASE + struct slot_action { + int slot_id; + std::string filename; + std::string filepath; + }; + slot_action slot_action; + + // used by SERVER_TASK_TYPE_METRICS + bool metrics_reset_bucket = false; + + server_task(server_task_type type) : type(type) {} + + static slot_params params_from_json_cmpl( + const llama_model * model, + const common_params & params_base, + const json & data) { + slot_params params; + + // Sampling parameter defaults are loaded from the global server context (but individual requests can still override them) + slot_params defaults; + defaults.sampling = params_base.sampling; + defaults.speculative = params_base.speculative; + + // enabling this will output extra debug information in the HTTP responses from the server + params.verbose = params_base.verbosity > 9; + params.timings_per_token = json_value(data, "timings_per_token", false); + + params.stream = json_value(data, "stream", false); + params.cache_prompt = json_value(data, "cache_prompt", true); + params.n_predict = json_value(data, "n_predict", json_value(data, "max_tokens", defaults.n_predict)); + params.n_indent = json_value(data, "n_indent", defaults.n_indent); + params.n_keep = json_value(data, "n_keep", defaults.n_keep); + params.n_discard = json_value(data, "n_discard", defaults.n_discard); + //params.t_max_prompt_ms = json_value(data, "t_max_prompt_ms", defaults.t_max_prompt_ms); // TODO: implement + params.t_max_predict_ms = json_value(data, "t_max_predict_ms", defaults.t_max_predict_ms); + + params.sampling.top_k = json_value(data, "top_k", defaults.sampling.top_k); + params.sampling.top_p = json_value(data, "top_p", defaults.sampling.top_p); + params.sampling.min_p = json_value(data, "min_p", defaults.sampling.min_p); + params.sampling.xtc_probability = json_value(data, "xtc_probability", defaults.sampling.xtc_probability); + params.sampling.xtc_threshold = json_value(data, "xtc_threshold", defaults.sampling.xtc_threshold); + params.sampling.typ_p = json_value(data, "typical_p", defaults.sampling.typ_p); + params.sampling.temp = json_value(data, "temperature", defaults.sampling.temp); + params.sampling.dynatemp_range = json_value(data, "dynatemp_range", defaults.sampling.dynatemp_range); + params.sampling.dynatemp_exponent = json_value(data, "dynatemp_exponent", defaults.sampling.dynatemp_exponent); + params.sampling.penalty_last_n = json_value(data, "repeat_last_n", defaults.sampling.penalty_last_n); + params.sampling.penalty_repeat = json_value(data, "repeat_penalty", defaults.sampling.penalty_repeat); + params.sampling.penalty_freq = json_value(data, "frequency_penalty", defaults.sampling.penalty_freq); + params.sampling.penalty_present = json_value(data, "presence_penalty", defaults.sampling.penalty_present); + params.sampling.dry_multiplier = json_value(data, "dry_multiplier", defaults.sampling.dry_multiplier); + params.sampling.dry_base = json_value(data, "dry_base", defaults.sampling.dry_base); + params.sampling.dry_allowed_length = json_value(data, "dry_allowed_length", defaults.sampling.dry_allowed_length); + params.sampling.dry_penalty_last_n = json_value(data, "dry_penalty_last_n", defaults.sampling.dry_penalty_last_n); + params.sampling.mirostat = json_value(data, "mirostat", defaults.sampling.mirostat); + params.sampling.mirostat_tau = json_value(data, "mirostat_tau", defaults.sampling.mirostat_tau); + params.sampling.mirostat_eta = json_value(data, "mirostat_eta", defaults.sampling.mirostat_eta); + params.sampling.penalize_nl = json_value(data, "penalize_nl", defaults.sampling.penalize_nl); + params.sampling.seed = json_value(data, "seed", defaults.sampling.seed); + params.sampling.n_probs = json_value(data, "n_probs", defaults.sampling.n_probs); + params.sampling.min_keep = json_value(data, "min_keep", defaults.sampling.min_keep); + + params.speculative.n_min = json_value(data, "speculative.n_min", defaults.speculative.n_min); + params.speculative.n_max = json_value(data, "speculative.n_max", defaults.speculative.n_max); + params.speculative.p_min = json_value(data, "speculative.p_min", defaults.speculative.p_min); + + params.speculative.n_min = std::min(params.speculative.n_max, params.speculative.n_min); + params.speculative.n_min = std::max(params.speculative.n_min, 2); + params.speculative.n_max = std::max(params.speculative.n_max, 0); + + if (params.sampling.dry_base < 1.0f) { + params.sampling.dry_base = defaults.sampling.dry_base; + } + + // sequence breakers for DRY + { + // Currently, this is not compatible with TextGen WebUI, Koboldcpp and SillyTavern format + // Ref: https://github.com/oobabooga/text-generation-webui/blob/d1af7a41ade7bd3c3a463bfa640725edb818ebaf/extensions/openai/typing.py#L39 + + if (data.contains("dry_sequence_breakers")) { + params.sampling.dry_sequence_breakers = json_value(data, "dry_sequence_breakers", std::vector()); + if (params.sampling.dry_sequence_breakers.empty()) { + throw std::runtime_error("Error: dry_sequence_breakers must be a non-empty array of strings"); + } + } + } + + // process "json_schema" and "grammar" + if (data.contains("json_schema") && !data.at("json_schema").is_null() && data.contains("grammar") && !data.at("grammar").is_null()) { + throw std::runtime_error("Either \"json_schema\" or \"grammar\" can be specified, but not both"); + } + if (data.contains("json_schema") && !data.contains("grammar")) { + try { + auto schema = json_value(data, "json_schema", json::object()); + params.sampling.grammar = json_schema_to_grammar(schema); + } catch (const std::exception & e) { + throw std::runtime_error(std::string("\"json_schema\": ") + e.what()); + } + } else { + params.sampling.grammar = json_value(data, "grammar", defaults.sampling.grammar); + } + + if (data.contains("tools")) { + params.oaicompat_tools = data.at("tools"); + } + if (data.contains("tool_call_style")) { + params.oaicompat_tool_call_style = data.at("tool_call_style"); + } + + { + params.sampling.logit_bias.clear(); + params.ignore_eos = json_value(data, "ignore_eos", false); + + const auto & logit_bias = data.find("logit_bias"); + if (logit_bias != data.end() && logit_bias->is_array()) { + const int n_vocab = llama_n_vocab(model); + for (const auto & el : *logit_bias) { + // TODO: we may want to throw errors here, in case "el" is incorrect + if (el.is_array() && el.size() == 2) { + float bias; + if (el[1].is_number()) { + bias = el[1].get(); + } else if (el[1].is_boolean() && !el[1].get()) { + bias = -INFINITY; + } else { + continue; + } + + if (el[0].is_number_integer()) { + llama_token tok = el[0].get(); + if (tok >= 0 && tok < n_vocab) { + params.sampling.logit_bias.push_back({tok, bias}); + } + } else if (el[0].is_string()) { + auto toks = common_tokenize(model, el[0].get(), false); + for (auto tok : toks) { + params.sampling.logit_bias.push_back({tok, bias}); + } + } + } + } + } + } + + if (data.contains("grammar_triggers")) { + const auto & triggers = data.at("grammar_triggers"); + if (triggers.is_array()) { + for (const auto & trigger : triggers) { + if (trigger.is_string()) { + params.grammar_triggers.push_back(trigger); + } + } + } + } + + { + params.antiprompt.clear(); + + const auto & stop = data.find("stop"); + if (stop != data.end() && stop->is_array()) { + for (const auto & word : *stop) { + if (!word.empty()) { + params.antiprompt.push_back(word); + } + } + } + } + + { + const auto & samplers = data.find("samplers"); + if (samplers != data.end()) { + if (samplers->is_array()) { + std::vector sampler_names; + for (const auto & name : *samplers) { + if (name.is_string()) { + sampler_names.emplace_back(name); + } + } + params.sampling.samplers = common_sampler_types_from_names(sampler_names, false); + } else if (samplers->is_string()){ + std::string sampler_string; + for (const auto & name : *samplers) { + sampler_string += name; + } + params.sampling.samplers = common_sampler_types_from_chars(sampler_string); + } + } else { + params.sampling.samplers = defaults.sampling.samplers; + } + } + + std::string model_name = params_base.model_alias.empty() ? DEFAULT_OAICOMPAT_MODEL : params_base.model_alias; + params.oaicompat_model = json_value(data, "model", model_name); + + return params; + } + + // utility function + static std::unordered_set get_list_id(const std::vector & tasks) { + std::unordered_set ids(tasks.size()); + for (size_t i = 0; i < tasks.size(); i++) { + ids.insert(tasks[i].id); + } + return ids; + } +}; + +struct result_timings { + int32_t prompt_n = -1; + double prompt_ms; + double prompt_per_token_ms; + double prompt_per_second; + + int32_t predicted_n = -1; + double predicted_ms; + double predicted_per_token_ms; + double predicted_per_second; + + json to_json() const { + return { + {"prompt_n", prompt_n}, + {"prompt_ms", prompt_ms}, + {"prompt_per_token_ms", prompt_per_token_ms}, + {"prompt_per_second", prompt_per_second}, + + {"predicted_n", predicted_n}, + {"predicted_ms", predicted_ms}, + {"predicted_per_token_ms", predicted_per_token_ms}, + {"predicted_per_second", predicted_per_second}, + }; + } +}; + +struct server_task_result { + int id = -1; + int id_slot = -1; + virtual bool is_error() { + // only used by server_task_result_error + return false; + } + virtual bool is_stop() { + // only used by server_task_result_cmpl_* + return false; + } + virtual int get_index() { + return -1; + } + virtual json to_json() = 0; + virtual ~server_task_result() = default; +}; + +// using shared_ptr for polymorphism of server_task_result +using server_task_result_ptr = std::unique_ptr; + +inline std::string stop_type_to_str(stop_type type) { + switch (type) { + case STOP_TYPE_EOS: return "eos"; + case STOP_TYPE_WORD: return "word"; + case STOP_TYPE_LIMIT: return "limit"; + default: return "none"; + } +} + +struct completion_token_output { + llama_token tok; + std::string text_to_send; + struct token_prob { + llama_token tok; + std::string tok_str; + float prob; + }; + std::vector probs; + + json to_json() const { + json probs_for_token = json::array(); + for (const auto & p : probs) { + probs_for_token.push_back(json { + {"tok_str", p.tok_str}, + {"prob", p.prob}, + }); + } + return probs_for_token; + } + + static json probs_vector_to_json(const std::vector & probs) { + json out = json::array(); + for (const auto & prob : probs) { + const std::string tok_str = prob.text_to_send; + out.push_back(json { + {"content", tok_str}, + {"probs", prob.to_json()}, + }); + } + return out; + } +}; + +struct server_task_result_cmpl_final : server_task_result { + int index = 0; + std::string content; + bool stream; + result_timings timings; + std::string prompt; + + bool truncated; + int32_t n_decoded; + int32_t n_prompt_tokens; + int32_t n_tokens_cached; + int32_t has_new_line; + std::string stopping_word; + stop_type stop = STOP_TYPE_NONE; + + std::vector probs_output; + + slot_params generation_params; + + // OAI-compat fields + bool verbose = false; + bool oaicompat = false; + bool oaicompat_chat = true; // TODO: support oaicompat for non-chat + std::string oaicompat_model; + std::string oaicompat_cmpl_id; + json oaicompat_tools; + llama_tool_call_style oaicompat_tool_call_style = llama_tool_call_style::None; + + virtual int get_index() override { + return index; + } + + virtual bool is_stop() override { + return true; // in stream mode, final responses are considered stop + } + + virtual json to_json() override { + return oaicompat + ? (stream ? to_json_oaicompat_chat_stream() : to_json_oaicompat_chat()) + : to_json_non_oaicompat(); + } + + json to_json_non_oaicompat() { + json res = json { + {"index", index}, + {"content", stream ? "" : content}, // in stream mode, content is already in last partial chunk + {"id_slot", id_slot}, + {"stop", true}, + {"model", oaicompat_model}, + {"tokens_predicted", n_decoded}, + {"tokens_evaluated", n_prompt_tokens}, + {"generation_settings", generation_params.to_json()}, + {"prompt", prompt}, + {"has_new_line", has_new_line}, + {"truncated", truncated}, + {"stop_type", stop_type_to_str(stop)}, + {"stopping_word", stopping_word}, + {"tokens_cached", n_tokens_cached}, + {"timings", timings.to_json()}, + }; + if (!probs_output.empty()) { + res["completion_probabilities"] = completion_token_output::probs_vector_to_json(probs_output); + } + return res; + } + + json to_json_oaicompat_chat() { + std::string finish_reason = "length"; + if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) { + finish_reason = "stop"; + } + + llama_tool_calls parsed_tool_calls; + json tool_calls; + json message_content; + if (!oaicompat_tools.is_null()) { + parsed_tool_calls = parse_tool_calls(oaicompat_tool_call_style, oaicompat_tools, content); + if (!parsed_tool_calls.tool_calls.empty()) { + finish_reason = "tool_calls"; + message_content = parsed_tool_calls.content; + tool_calls = json::array(); + for (const auto & tc : parsed_tool_calls.tool_calls) { + tool_calls.push_back({ + {"type", "function"}, + {"function", { + {"name", tc.name}, + {"arguments", tc.arguments}, + }}, + {"id", tc.id.empty() ? json() : json(tc.id)}, + }); + } + } else { + message_content = parsed_tool_calls.content; + } + } else { + message_content = content; + } + + json choices = json::array({json{ + {"finish_reason", finish_reason}, + {"index", 0}, + {"message", json{{"content", message_content}, + {"tool_calls", tool_calls}, + {"role", "assistant"} + } + }}}); + + std::time_t t = std::time(0); + + json res = json { + {"choices", choices}, + {"created", t}, + {"model", oaicompat_model}, + {"object", "chat.completion"}, + {"usage", json { + {"completion_tokens", n_decoded}, + {"prompt_tokens", n_prompt_tokens}, + {"total_tokens", n_decoded + n_prompt_tokens} + }}, + {"id", oaicompat_cmpl_id} + }; + + // extra fields for debugging purposes + if (verbose) { + res["__verbose"] = to_json_non_oaicompat(); + } + if (timings.prompt_n >= 0) { + res.push_back({"timings", timings.to_json()}); + } + + return res; + } + + json to_json_oaicompat_chat_stream() { + std::time_t t = std::time(0); + std::string finish_reason = "length"; + if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) { + finish_reason = "stop"; + } + + json choices = json::array({json{{"finish_reason", finish_reason}, + {"index", 0}, + {"delta", json::object()}}}); + + json ret = json { + {"choices", choices}, + {"created", t}, + {"id", oaicompat_cmpl_id}, + {"model", oaicompat_model}, + {"object", "chat.completion.chunk"}, + {"usage", json { + {"completion_tokens", n_decoded}, + {"prompt_tokens", n_prompt_tokens}, + {"total_tokens", n_decoded + n_prompt_tokens}, + }}, + }; + + if (timings.prompt_n >= 0) { + ret.push_back({"timings", timings.to_json()}); + } + + return ret; + } +}; + +struct server_task_result_cmpl_partial : server_task_result { + int index = 0; + std::string content; + + int32_t n_decoded; + int32_t n_prompt_tokens; + + std::vector probs_output; + result_timings timings; + + // OAI-compat fields + bool verbose = false; + bool oaicompat = false; + bool oaicompat_chat = true; // TODO: support oaicompat for non-chat + std::string oaicompat_model; + std::string oaicompat_cmpl_id; + json oaicompat_tools; + llama_tool_call_style oaicompat_tool_call_style = llama_tool_call_style::None; + + virtual int get_index() override { + return index; + } + + virtual bool is_stop() override { + return false; // in stream mode, partial responses are not considered stop + } + + virtual json to_json() override { + return oaicompat ? to_json_oaicompat() : to_json_non_oaicompat(); + } + + json to_json_non_oaicompat() { + // non-OAI-compat JSON + json res = json { + {"index", index}, + {"content", content}, + {"stop", false}, + {"id_slot", id_slot}, + {"tokens_predicted", n_decoded}, + {"tokens_evaluated", n_prompt_tokens}, + }; + // populate the timings object when needed (usually for the last response or with timings_per_token enabled) + if (timings.prompt_n > 0) { + res.push_back({"timings", timings.to_json()}); + } + if (!probs_output.empty()) { + res["completion_probabilities"] = completion_token_output::probs_vector_to_json(probs_output); + } + return res; + } + + json to_json_oaicompat() { + bool first = n_decoded == 0; + std::time_t t = std::time(0); + json choices; + + + // auto chat_template = json_value(request, "chat_template", std::string()); + // llama_tool_calls parsed_tool_calls; + // auto tools = json_value(request, "tools", json::array()); + // json tool_calls; + // json message_content; + // if (json_value(request, "parse_tool_calls", false)) { + // parsed_tool_calls = parse_tool_calls(tool_call_style, tools, content); + // if (!parsed_tool_calls.tool_calls.empty()) { + // finish_reason = "tool_calls"; + // message_content = parsed_tool_calls.content; + // tool_calls = json::array(); + // for (const auto & tc : parsed_tool_calls.tool_calls) { + // tool_calls.push_back({ + // {"type", "function"}, + // {"function", { + // {"name", tc.name}, + // {"arguments", tc.arguments}, + // }}, + // {"id", tc.id.empty() ? json() : json(tc.id)}, + // }); + // } + // } else { + // message_content = parsed_tool_calls.content; + // } + // } else { + // message_content = content; + // } + + // json choices = + // streaming ? json::array({json{{"finish_reason", finish_reason}, + // {"index", 0}, + // {"delta", json::object()}}}) + // : json::array({json{{"finish_reason", finish_reason}, + // {"index", 0}, + // {"message", json{{"content", message_content}, + // {"tool_calls", tool_calls}, + // {"role", "assistant"}}}}}); + + + if (first) { + if (content.empty()) { + choices = json::array({json{{"finish_reason", nullptr}, + {"index", 0}, + {"delta", json{{"role", "assistant"}}}}}); + } else { + // We have to send this as two updates to conform to openai behavior + json initial_ret = json{{"choices", json::array({json{ + {"finish_reason", nullptr}, + {"index", 0}, + {"delta", json{ + {"role", "assistant"} + }}}})}, + {"created", t}, + {"id", oaicompat_cmpl_id}, + {"model", oaicompat_model}, + {"object", "chat.completion.chunk"}}; + + json second_ret = json{ + {"choices", json::array({json{{"finish_reason", nullptr}, + {"index", 0}, + {"delta", json{ + {"content", content}}} + }})}, + {"created", t}, + {"id", oaicompat_cmpl_id}, + {"model", oaicompat_model}, + {"object", "chat.completion.chunk"}}; + + return std::vector({initial_ret, second_ret}); + } + } else { + choices = json::array({json{ + {"finish_reason", nullptr}, + {"index", 0}, + {"delta", + json{ + {"content", content}, + }}, + }}); + } + + json ret = json { + {"choices", choices}, + {"created", t}, + {"id", oaicompat_cmpl_id}, + {"model", oaicompat_model}, + {"object", "chat.completion.chunk"} + }; + + if (timings.prompt_n >= 0) { + ret.push_back({"timings", timings.to_json()}); + } + + return std::vector({ret}); + } +}; + +struct server_task_result_embd : server_task_result { + int index = 0; + std::vector embedding; + + virtual int get_index() override { + return index; + } + + virtual json to_json() override { + return json { + {"index", index}, + {"embedding", embedding}, + }; + } +}; + +struct server_task_result_rerank : server_task_result { + int index = 0; + float score = -1e6; + + virtual int get_index() override { + return index; + } + + virtual json to_json() override { + return json { + {"index", index}, + {"score", score}, + }; + } +}; + +// this function maybe used outside of server_task_result_error +static json format_error_response(const std::string & message, const enum error_type type) { + std::string type_str; + int code = 500; + switch (type) { + case ERROR_TYPE_INVALID_REQUEST: + type_str = "invalid_request_error"; + code = 400; + break; + case ERROR_TYPE_AUTHENTICATION: + type_str = "authentication_error"; + code = 401; + break; + case ERROR_TYPE_NOT_FOUND: + type_str = "not_found_error"; + code = 404; + break; + case ERROR_TYPE_SERVER: + type_str = "server_error"; + code = 500; + break; + case ERROR_TYPE_PERMISSION: + type_str = "permission_error"; + code = 403; + break; + case ERROR_TYPE_NOT_SUPPORTED: + type_str = "not_supported_error"; + code = 501; + break; + case ERROR_TYPE_UNAVAILABLE: + type_str = "unavailable_error"; + code = 503; + break; + } + return json { + {"code", code}, + {"message", message}, + {"type", type_str}, + }; +} + +struct server_task_result_error : server_task_result { + int index = 0; + error_type err_type = ERROR_TYPE_SERVER; + std::string err_msg; + + virtual bool is_error() override { + return true; + } + + virtual json to_json() override { + return format_error_response(err_msg, err_type); + } +}; + +struct server_task_result_metrics : server_task_result { + int n_idle_slots; + int n_processing_slots; + int n_tasks_deferred; + int64_t t_start; + + int32_t kv_cache_tokens_count; + int32_t kv_cache_used_cells; + + // TODO: somehow reuse server_metrics in the future, instead of duplicating the fields + uint64_t n_prompt_tokens_processed_total = 0; + uint64_t t_prompt_processing_total = 0; + uint64_t n_tokens_predicted_total = 0; + uint64_t t_tokens_generation_total = 0; + + uint64_t n_prompt_tokens_processed = 0; + uint64_t t_prompt_processing = 0; + + uint64_t n_tokens_predicted = 0; + uint64_t t_tokens_generation = 0; + + uint64_t n_decode_total = 0; + uint64_t n_busy_slots_total = 0; + + // while we can also use std::vector this requires copying the slot object which can be quite messy + // therefore, we use json to temporarily store the slot.to_json() result + json slots_data = json::array(); + + virtual json to_json() override { + return json { + { "idle", n_idle_slots }, + { "processing", n_processing_slots }, + { "deferred", n_tasks_deferred }, + { "t_start", t_start }, + + { "n_prompt_tokens_processed_total", n_prompt_tokens_processed_total }, + { "t_tokens_generation_total", t_tokens_generation_total }, + { "n_tokens_predicted_total", n_tokens_predicted_total }, + { "t_prompt_processing_total", t_prompt_processing_total }, + + { "n_prompt_tokens_processed", n_prompt_tokens_processed }, + { "t_prompt_processing", t_prompt_processing }, + { "n_tokens_predicted", n_tokens_predicted }, + { "t_tokens_generation", t_tokens_generation }, + + { "n_decode_total", n_decode_total }, + { "n_busy_slots_total", n_busy_slots_total }, + + { "kv_cache_tokens_count", kv_cache_tokens_count }, + { "kv_cache_used_cells", kv_cache_used_cells }, + + { "slots", slots_data }, + }; + } +}; + +struct server_task_result_slot_save_load : server_task_result { + std::string filename; + bool is_save; // true = save, false = load + + size_t n_tokens; + size_t n_bytes; + double t_ms; + + virtual json to_json() override { + if (is_save) { + return json { + { "id_slot", id_slot }, + { "filename", filename }, + { "n_saved", n_tokens }, + { "n_written", n_bytes }, + { "timings", { + { "save_ms", t_ms } + }}, + }; + } else { + return json { + { "id_slot", id_slot }, + { "filename", filename }, + { "n_restored", n_tokens }, + { "n_read", n_bytes }, + { "timings", { + { "restore_ms", t_ms } + }}, + }; + } + } +}; + +struct server_task_result_slot_erase : server_task_result { + size_t n_erased; + + virtual json to_json() override { + return json { + { "id_slot", id_slot }, + { "n_erased", n_erased }, + }; + } +}; + +struct server_task_result_apply_lora : server_task_result { + virtual json to_json() override { + return json {{ "success", true }}; + } }; struct server_slot { int id; int id_task = -1; - llama_batch batch_spec; + // only used for completion/embedding/infill/rerank + server_task_type task_type = SERVER_TASK_TYPE_COMPLETION; + llama_batch batch_spec = {}; + + llama_context * ctx = nullptr; llama_context * ctx_dft = nullptr; common_speculative * spec = nullptr; @@ -159,20 +1028,11 @@ struct server_slot { llama_tokens cache_tokens; std::vector generated_token_probs; - server_task_inf_type inf_type = SERVER_TASK_INF_TYPE_COMPLETION; - bool has_next_token = true; bool has_new_line = false; bool truncated = false; - bool stopped_eos = false; - bool stopped_word = false; - bool stopped_limit = false; + stop_type stop; - bool timings_per_token = false; - - bool oaicompat = false; - - std::string oaicompat_model; std::string stopping_word; llama_antiprompts antiprompts; @@ -204,18 +1064,20 @@ struct server_slot { generated_text = ""; has_new_line = false; truncated = false; - stopped_eos = false; - stopped_word = false; - stopped_limit = false; + stop = STOP_TYPE_NONE; stopping_word = ""; n_past = 0; n_sent_text = 0; n_sent_token_probs = 0; - inf_type = SERVER_TASK_INF_TYPE_COMPLETION; + task_type = SERVER_TASK_TYPE_COMPLETION; generated_token_probs.clear(); } + bool is_non_causal() const { + return task_type == SERVER_TASK_TYPE_EMBEDDING || task_type == SERVER_TASK_TYPE_RERANK; + } + bool has_budget(const common_params & global_params) { if (params.n_predict == -1 && global_params.n_predict == -1) { return true; // limitless @@ -259,18 +1121,19 @@ struct server_slot { } } - json get_formated_timings() const { - return json { - {"prompt_n", n_prompt_tokens_processed}, - {"prompt_ms", t_prompt_processing}, - {"prompt_per_token_ms", t_prompt_processing / n_prompt_tokens_processed}, - {"prompt_per_second", 1e3 / t_prompt_processing * n_prompt_tokens_processed}, + result_timings get_timings() const { + result_timings timings; + timings.prompt_n = n_prompt_tokens_processed; + timings.prompt_ms = t_prompt_processing; + timings.prompt_per_token_ms = t_prompt_processing / n_prompt_tokens_processed; + timings.prompt_per_second = 1e3 / t_prompt_processing * n_prompt_tokens_processed; - {"predicted_n", n_decoded}, - {"predicted_ms", t_token_generation}, - {"predicted_per_token_ms", t_token_generation / n_decoded}, - {"predicted_per_second", 1e3 / t_token_generation * n_decoded}, - }; + timings.predicted_n = n_decoded; + timings.predicted_ms = t_token_generation; + timings.predicted_per_token_ms = t_token_generation / n_decoded; + timings.predicted_per_second = 1e3 / t_token_generation * n_decoded; + + return timings; } void print_timings() const { @@ -289,6 +1152,28 @@ struct server_slot { t_token_generation, n_decoded, t_gen, n_gen_second, t_prompt_processing + t_token_generation, n_prompt_tokens_processed + n_decoded); } + + json to_json() const { + return json { + {"id", id}, + {"id_task", id_task}, + {"n_ctx", n_ctx}, + {"speculative", can_speculate()}, + {"is_processing", is_processing()}, + {"non_causal", is_non_causal()}, + {"params", params.to_json()}, + {"prompt", common_detokenize(ctx, prompt_tokens)}, + {"next_token", + { + {"has_next_token", has_next_token}, + {"has_new_line", has_new_line}, + {"n_remain", n_remaining}, + {"n_decoded", n_decoded}, + {"stopping_word", stopping_word}, + } + }, + }; + } }; struct server_metrics { @@ -361,9 +1246,7 @@ struct server_queue { // Add a new task to the end of the queue int post(server_task task, bool front = false) { std::unique_lock lock(mutex_tasks); - if (task.id == -1) { - task.id = id++; - } + GGML_ASSERT(task.id != -1); QUE_DBG("new task, id = %d, front = %d\n", task.id, front); if (front) { queue_tasks.push_front(std::move(task)); @@ -487,8 +1370,8 @@ struct server_response { // for keeping track of all tasks waiting for the result std::unordered_set waiting_task_ids; - // the main result queue - std::vector queue_results; + // the main result queue (using ptr for polymorphism) + std::vector queue_results; std::mutex mutex_results; std::condition_variable condition_results; @@ -528,7 +1411,7 @@ struct server_response { } // This function blocks the thread until there is a response for one of the id_tasks - server_task_result recv(const std::unordered_set & id_tasks) { + server_task_result_ptr recv(const std::unordered_set & id_tasks) { while (true) { std::unique_lock lock(mutex_results); condition_results.wait(lock, [&]{ @@ -536,8 +1419,8 @@ struct server_response { }); for (int i = 0; i < (int) queue_results.size(); i++) { - if (id_tasks.find(queue_results[i].id) != id_tasks.end()) { - server_task_result res = queue_results[i]; + if (id_tasks.find(queue_results[i]->id) != id_tasks.end()) { + server_task_result_ptr res = std::move(queue_results[i]); queue_results.erase(queue_results.begin() + i); return res; } @@ -548,21 +1431,21 @@ struct server_response { } // single-task version of recv() - server_task_result recv(int id_task) { + server_task_result_ptr recv(int id_task) { std::unordered_set id_tasks = {id_task}; return recv(id_tasks); } // Send a new result to a waiting id_task - void send(server_task_result & result) { - SRV_DBG("sending result for task id = %d\n", result.id); + void send(server_task_result_ptr && result) { + SRV_DBG("sending result for task id = %d\n", result->id); std::unique_lock lock(mutex_results); for (const auto & id_task : waiting_task_ids) { - if (result.id == id_task) { - SRV_DBG("task id = %d moved to result queue\n", result.id); + if (result->id == id_task) { + SRV_DBG("task id = %d pushed to result queue\n", result->id); - queue_results.push_back(std::move(result)); + queue_results.emplace_back(std::move(result)); condition_results.notify_all(); return; } @@ -735,6 +1618,7 @@ struct server_context { server_slot slot; slot.id = i; + slot.ctx = ctx; slot.n_ctx = n_ctx_slot; slot.n_predict = params_base.n_predict; @@ -767,8 +1651,7 @@ struct server_context { slots.push_back(slot); } - default_generation_settings_for_props = get_formated_generation(slots.front()); - default_generation_settings_for_props["seed"] = -1; + default_generation_settings_for_props = slots[0].to_json(); // the update_slots() logic will always submit a maximum of n_batch or n_parallel tokens // note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used) @@ -855,99 +1738,14 @@ struct server_context { } bool launch_slot_with_task(server_slot & slot, const server_task & task) { - // Sampling parameter defaults are loaded from the global server context (but individual requests can still override them) - slot_params defaults; - defaults.sampling = params_base.sampling; - defaults.speculative = params_base.speculative; + slot.reset(); + slot.id_task = task.id; + slot.index = task.index; + slot.task_type = task.type; + slot.params = std::move(task.params); + slot.prompt_tokens = std::move(task.prompt_tokens); - const auto & data = task.data; - - if (data.count("__oaicompat") != 0) { - slot.oaicompat = true; - slot.oaicompat_model = json_value(data, "model", std::string(DEFAULT_OAICOMPAT_MODEL)); - } else { - slot.oaicompat = false; - slot.oaicompat_model = ""; - } - - slot.timings_per_token = json_value(data, "timings_per_token", false); - - slot.params.stream = json_value(data, "stream", false); - slot.params.cache_prompt = json_value(data, "cache_prompt", true); - slot.params.n_predict = json_value(data, "n_predict", json_value(data, "max_tokens", defaults.n_predict)); - slot.params.n_indent = json_value(data, "n_indent", defaults.n_indent); - slot.params.n_keep = json_value(data, "n_keep", defaults.n_keep); - slot.params.n_discard = json_value(data, "n_discard", defaults.n_discard); - //slot.params.t_max_prompt_ms = json_value(data, "t_max_prompt_ms", defaults.t_max_prompt_ms); // TODO: implement - slot.params.t_max_predict_ms = json_value(data, "t_max_predict_ms", defaults.t_max_predict_ms); - - slot.params.sampling.top_k = json_value(data, "top_k", defaults.sampling.top_k); - slot.params.sampling.top_p = json_value(data, "top_p", defaults.sampling.top_p); - slot.params.sampling.min_p = json_value(data, "min_p", defaults.sampling.min_p); - slot.params.sampling.xtc_probability = json_value(data, "xtc_probability", defaults.sampling.xtc_probability); - slot.params.sampling.xtc_threshold = json_value(data, "xtc_threshold", defaults.sampling.xtc_threshold); - slot.params.sampling.typ_p = json_value(data, "typical_p", defaults.sampling.typ_p); - slot.params.sampling.temp = json_value(data, "temperature", defaults.sampling.temp); - slot.params.sampling.dynatemp_range = json_value(data, "dynatemp_range", defaults.sampling.dynatemp_range); - slot.params.sampling.dynatemp_exponent = json_value(data, "dynatemp_exponent", defaults.sampling.dynatemp_exponent); - slot.params.sampling.penalty_last_n = json_value(data, "repeat_last_n", defaults.sampling.penalty_last_n); - slot.params.sampling.penalty_repeat = json_value(data, "repeat_penalty", defaults.sampling.penalty_repeat); - slot.params.sampling.penalty_freq = json_value(data, "frequency_penalty", defaults.sampling.penalty_freq); - slot.params.sampling.penalty_present = json_value(data, "presence_penalty", defaults.sampling.penalty_present); - slot.params.sampling.dry_multiplier = json_value(data, "dry_multiplier", defaults.sampling.dry_multiplier); - slot.params.sampling.dry_base = json_value(data, "dry_base", defaults.sampling.dry_base); - slot.params.sampling.dry_allowed_length = json_value(data, "dry_allowed_length", defaults.sampling.dry_allowed_length); - slot.params.sampling.dry_penalty_last_n = json_value(data, "dry_penalty_last_n", defaults.sampling.dry_penalty_last_n); - slot.params.sampling.mirostat = json_value(data, "mirostat", defaults.sampling.mirostat); - slot.params.sampling.mirostat_tau = json_value(data, "mirostat_tau", defaults.sampling.mirostat_tau); - slot.params.sampling.mirostat_eta = json_value(data, "mirostat_eta", defaults.sampling.mirostat_eta); - slot.params.sampling.penalize_nl = json_value(data, "penalize_nl", defaults.sampling.penalize_nl); - slot.params.sampling.seed = json_value(data, "seed", defaults.sampling.seed); - slot.params.sampling.n_probs = json_value(data, "n_probs", defaults.sampling.n_probs); - slot.params.sampling.min_keep = json_value(data, "min_keep", defaults.sampling.min_keep); - - slot.params.speculative.n_min = json_value(data, "speculative.n_min", defaults.speculative.n_min); - slot.params.speculative.n_max = json_value(data, "speculative.n_max", defaults.speculative.n_max); - slot.params.speculative.p_min = json_value(data, "speculative.p_min", defaults.speculative.p_min); - - slot.params.speculative.n_min = std::min(slot.params.speculative.n_max, slot.params.speculative.n_min); - slot.params.speculative.n_min = std::max(slot.params.speculative.n_min, 2); - slot.params.speculative.n_max = std::max(slot.params.speculative.n_max, 0); - - if (slot.params.sampling.dry_base < 1.0f) { - slot.params.sampling.dry_base = defaults.sampling.dry_base; - } - - // sequence breakers for DRY - { - // Currently, this is not compatible with TextGen WebUI, Koboldcpp and SillyTavern format - // Ref: https://github.com/oobabooga/text-generation-webui/blob/d1af7a41ade7bd3c3a463bfa640725edb818ebaf/extensions/openai/typing.py#L39 - - if (data.contains("dry_sequence_breakers")) { - slot.params.sampling.dry_sequence_breakers = json_value(data, "dry_sequence_breakers", std::vector()); - if (slot.params.sampling.dry_sequence_breakers.empty()) { - send_error(task, "Error: dry_sequence_breakers must be a non-empty array of strings", ERROR_TYPE_INVALID_REQUEST); - return false; - } - } - } - - // process "json_schema" and "grammar" - if (data.contains("json_schema") && !data.at("json_schema").is_null() && data.contains("grammar") && !data.at("grammar").is_null()) { - send_error(task, "Either \"json_schema\" or \"grammar\" can be specified, but not both", ERROR_TYPE_INVALID_REQUEST); - return false; - } - if (data.contains("json_schema") && !data.contains("grammar")) { - try { - auto schema = json_value(data, "json_schema", json::object()); - slot.params.sampling.grammar = json_schema_to_grammar(schema); - } catch (const std::exception & e) { - send_error(task, std::string("\"json_schema\": ") + e.what(), ERROR_TYPE_INVALID_REQUEST); - return false; - } - } else { - slot.params.sampling.grammar = json_value(data, "grammar", defaults.sampling.grammar); - } + SLT_DBG(slot, "launching slot : %s\n", safe_json_to_str(slot.to_json()).c_str()); if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) { // Might be better to reject the request with a 400 ? @@ -955,89 +1753,13 @@ struct server_context { SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d", slot.n_predict, slot.n_predict); } - { - slot.params.sampling.logit_bias.clear(); - - if (json_value(data, "ignore_eos", false) && has_eos_token) { - slot.params.sampling.logit_bias.push_back({llama_token_eos(model), -INFINITY}); - } - - const auto & logit_bias = data.find("logit_bias"); - if (logit_bias != data.end() && logit_bias->is_array()) { - const int n_vocab = llama_n_vocab(model); - for (const auto & el : *logit_bias) { - // TODO: we may want to throw errors here, in case "el" is incorrect - if (el.is_array() && el.size() == 2) { - float bias; - if (el[1].is_number()) { - bias = el[1].get(); - } else if (el[1].is_boolean() && !el[1].get()) { - bias = -INFINITY; - } else { - continue; - } - - if (el[0].is_number_integer()) { - llama_token tok = el[0].get(); - if (tok >= 0 && tok < n_vocab) { - slot.params.sampling.logit_bias.push_back({tok, bias}); - } - } else if (el[0].is_string()) { - auto toks = common_tokenize(model, el[0].get(), false); - for (auto tok : toks) { - slot.params.sampling.logit_bias.push_back({tok, bias}); - } - } - } - } - } + if (slot.params.ignore_eos && has_eos_token) { + slot.params.sampling.logit_bias.push_back({llama_token_eos(model), -INFINITY}); } { slot.antiprompts.clear(); - slot.params.sampling.grammar_trigger_words.clear(); - - auto copy_string_array = [&](const json & data, const std::string & key, std::vector & vec) { - const auto & arr = data.find(key); - if (arr != data.end() && arr->is_array()) { - for (const auto & word : *arr) { - if (word.is_string()) { - vec.push_back(word); - } - } - } - }; - - std::vector stop_words; - - copy_string_array(data, "stop", stop_words); - copy_string_array(data, "grammar_trigger_words", slot.params.sampling.grammar_trigger_words); - - slot.antiprompts.build(ctx, stop_words, slot.params.sampling.grammar_trigger_words); - - } - - { - const auto & samplers = data.find("samplers"); - if (samplers != data.end()) { - if (samplers->is_array()) { - std::vector sampler_names; - for (const auto & name : *samplers) { - if (name.is_string()) { - sampler_names.emplace_back(name); - } - } - slot.params.sampling.samplers = common_sampler_types_from_names(sampler_names, false); - } else if (samplers->is_string()){ - std::string sampler_string; - for (const auto & name : *samplers) { - sampler_string += name; - } - slot.params.sampling.samplers = common_sampler_types_from_chars(sampler_string); - } - } else { - slot.params.sampling.samplers = defaults.sampling.samplers; - } + slot.antiprompts.build(ctx, slot.params.antiprompt, slot.params.grammar_triggers); } { @@ -1085,7 +1807,7 @@ struct server_context { if (match.is_grammar_trigger) { common_sampler_trigger_grammar(model, slot.smpl, common_token_to_piece(ctx, result.tok, params_base.special)); } else { - slot.stopped_word = true; + // slot.stopped_word = true; slot.stopping_word = match.pattern; slot.has_next_token = false; return false; @@ -1132,7 +1854,7 @@ struct server_context { is_grammar_trigger = true; length = match.pos + match.matchLength; } else if (!match.is_grammar_trigger && match.pos != std::string::npos && !match.is_partial) { - slot.stopped_word = true; + // slot.stopped_word = true; slot.stopping_word = match.pattern; slot.has_next_token = false; @@ -1166,7 +1888,7 @@ struct server_context { // check the limits if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params_base)) { - slot.stopped_limit = true; + slot.stop = STOP_TYPE_LIMIT; slot.has_next_token = false; SLT_DBG(slot, "stopped by limit, n_decoded = %d, n_predict = %d\n", slot.n_decoded, slot.params.n_predict); @@ -1175,7 +1897,7 @@ struct server_context { if (slot.has_new_line) { // if we have already seen a new line, we stop after a certain time limit if (slot.params.t_max_predict_ms > 0 && (ggml_time_us() - slot.t_start_generation > 1000.0f*slot.params.t_max_predict_ms)) { - slot.stopped_limit = true; + slot.stop = STOP_TYPE_LIMIT; slot.has_next_token = false; SLT_DBG(slot, "stopped by time limit, n_decoded = %d, t_max_predict_ms = %d ms\n", slot.n_decoded, (int) slot.params.t_max_predict_ms); @@ -1195,7 +1917,7 @@ struct server_context { } if (pos < slot.generated_text.size() && n_indent < slot.params.n_indent) { - slot.stopped_limit = true; + slot.stop = STOP_TYPE_LIMIT; slot.has_next_token = false; // cut the last line @@ -1224,7 +1946,7 @@ struct server_context { // if context shift is disabled, we stop when it reaches the context limit if (slot.n_past >= slot.n_ctx) { slot.truncated = true; - slot.stopped_limit = true; + slot.stop = STOP_TYPE_LIMIT; slot.has_next_token = false; SLT_DBG(slot, "stopped due to running out of context capacity, n_past = %d, n_prompt_tokens = %d, n_decoded = %d, n_ctx = %d\n", @@ -1232,7 +1954,7 @@ struct server_context { } if (llama_token_is_eog(model, result.tok)) { - slot.stopped_eos = true; + slot.stop = STOP_TYPE_EOS; slot.has_next_token = false; SLT_DBG(slot, "%s", "stopped by EOS\n"); @@ -1242,7 +1964,7 @@ struct server_context { if (slot.params.n_predict < 1 && slot.n_predict < 1 && slot.n_prompt_tokens + slot.n_decoded >= n_ctx_train) { slot.truncated = true; - slot.stopped_limit = true; + slot.stop = STOP_TYPE_LIMIT; slot.has_next_token = false; // stop prediction SLT_WRN(slot, @@ -1292,7 +2014,7 @@ struct server_context { {"mirostat_eta", slot.params.sampling.mirostat_eta}, {"penalize_nl", slot.params.sampling.penalize_nl}, {"stop", slot.params.antiprompt}, - {"grammar_trigger", slot.antiprompts.grammar_trigger_words}, + {"grammar_trigger", slot.params.grammar_triggers}, {"max_tokens", slot.params.n_predict}, // User configured n_predict {"n_keep", slot.params.n_keep}, {"n_discard", slot.params.n_discard}, @@ -1307,7 +2029,6 @@ struct server_context { {"speculative.n_max", slot.params.speculative.n_max}, {"speculative.n_min", slot.params.speculative.n_min}, {"speculative.p_min", slot.params.speculative.p_min}, - {"timings_per_token", slot.timings_per_token}, }; } @@ -1322,112 +2043,107 @@ struct server_context { void send_error(const int id_task, const std::string & error, const enum error_type type = ERROR_TYPE_SERVER) { SRV_ERR("task id = %d, error: %s\n", id_task, error.c_str()); - server_task_result res; - res.id = id_task; - res.stop = false; - res.error = true; - res.data = format_error_response(error, type); + auto res = std::make_unique(); + res->id = id_task; + res->err_type = type; + res->err_msg = error; - queue_results.send(res); + queue_results.send(std::move(res)); } - void send_partial_response(server_slot & slot, completion_token_output tkn) { - server_task_result res; - res.id = slot.id_task; - res.error = false; - res.stop = false; - res.data = json { - {"content", tkn.text_to_send}, - {"stop", false}, - {"id_slot", slot.id}, - {"multimodal", false}, - {"index", slot.index}, - }; + void send_partial_response(server_slot & slot, const completion_token_output & tkn) { + auto res = std::make_unique(); + res->id = slot.id_task; + res->index = slot.index; + res->content = tkn.text_to_send; + + res->n_decoded = slot.n_decoded; + res->n_prompt_tokens = slot.n_prompt_tokens; + + res->verbose = slot.params.verbose; + res->oaicompat = slot.params.oaicompat; + res->oaicompat_chat = slot.params.oaicompat_chat; + res->oaicompat_model = slot.params.oaicompat_model; + res->oaicompat_cmpl_id = slot.params.oaicompat_cmpl_id; + // res->oaicompat_tools = slot.params.oaicompat_tools; + // res->oaicompat_tool_call_style = slot.params.oaicompat_tool_call_style; + + // populate res.probs_output if (slot.params.sampling.n_probs > 0) { const llama_tokens to_send_toks = common_tokenize(ctx, tkn.text_to_send, false); + const size_t probs_pos = std::min(slot.n_sent_token_probs, slot.generated_token_probs.size()); const size_t probs_stop_pos = std::min(slot.n_sent_token_probs + to_send_toks.size(), slot.generated_token_probs.size()); std::vector probs_output; if (probs_pos < probs_stop_pos) { - probs_output = std::vector( + res->probs_output = std::vector( slot.generated_token_probs.begin() + probs_pos, slot.generated_token_probs.begin() + probs_stop_pos); } - slot.n_sent_token_probs = probs_stop_pos; - - res.data["completion_probabilities"] = probs_vector_to_json(ctx, probs_output); } - if (slot.oaicompat) { - res.data["oaicompat_token_ctr"] = slot.n_decoded; - res.data["model"] = slot.oaicompat_model; + // populate timings if this is final response or timings_per_token is enabled + if (slot.stop != STOP_TYPE_NONE || slot.params.timings_per_token) { + res->timings = slot.get_timings(); } - if (slot.timings_per_token) { - res.data["timings"] = slot.get_formated_timings(); - } - - queue_results.send(res); + queue_results.send(std::move(res)); } - void send_final_response(const server_slot & slot) { - server_task_result res; - res.id = slot.id_task; - res.error = false; - res.stop = true; - res.data = json { - {"content", !slot.params.stream ? slot.generated_text : ""}, - {"id_slot", slot.id}, - {"stop", true}, - {"model", params_base.model_alias}, - {"tokens_predicted", slot.n_decoded}, - {"tokens_evaluated", slot.n_prompt_tokens}, - {"generation_settings", get_formated_generation(slot)}, - {"prompt", common_detokenize(ctx, slot.prompt_tokens)}, - {"has_new_line", slot.has_new_line}, - {"truncated", slot.truncated}, - {"stopped_eos", slot.stopped_eos}, - {"stopped_word", slot.stopped_word}, - {"stopped_limit", slot.stopped_limit}, - {"stopping_word", slot.stopping_word}, - {"tokens_cached", slot.n_past}, - {"timings", slot.get_formated_timings()}, - {"index", slot.index}, - }; + void send_final_response(server_slot & slot) { + auto res = std::make_unique(); + res->id = slot.id_task; + res->id_slot = slot.id; + res->index = slot.index; + res->content = slot.generated_text; + res->timings = slot.get_timings(); + res->prompt = common_detokenize(ctx, slot.prompt_tokens, true); + + res->truncated = slot.truncated; + res->n_decoded = slot.n_decoded; + res->n_prompt_tokens = slot.n_prompt_tokens; + res->n_tokens_cached = slot.n_past; + res->has_new_line = slot.has_new_line; + res->stopping_word = slot.stopping_word; + res->stop = slot.stop; + + res->verbose = slot.params.verbose; + res->stream = slot.params.stream; + res->oaicompat = slot.params.oaicompat; + res->oaicompat_chat = slot.params.oaicompat_chat; + res->oaicompat_model = slot.params.oaicompat_model; + res->oaicompat_cmpl_id = slot.params.oaicompat_cmpl_id; + res->oaicompat_tools = slot.params.oaicompat_tools; + res->oaicompat_tool_call_style = slot.params.oaicompat_tool_call_style; + + // populate res.probs_output if (slot.params.sampling.n_probs > 0) { - std::vector probs; - if (!slot.params.stream && slot.stopped_word) { + if (!slot.params.stream && slot.stop == STOP_TYPE_WORD) { const llama_tokens stop_word_toks = common_tokenize(ctx, slot.stopping_word, false); size_t safe_offset = std::min(slot.generated_token_probs.size(), stop_word_toks.size()); - probs = std::vector( + res->probs_output = std::vector( slot.generated_token_probs.begin(), slot.generated_token_probs.end() - safe_offset); } else { - probs = std::vector( + res->probs_output = std::vector( slot.generated_token_probs.begin(), slot.generated_token_probs.end()); } - - res.data["completion_probabilities"] = probs_vector_to_json(ctx, probs); } - if (slot.oaicompat) { - res.data["oaicompat_token_ctr"] = slot.n_decoded; - res.data["model"] = slot.oaicompat_model; - } + res->generation_params = slot.params; // copy the parameters - queue_results.send(res); + queue_results.send(std::move(res)); } void send_embedding(const server_slot & slot, const llama_batch & batch) { - server_task_result res; - res.id = slot.id_task; - res.error = false; - res.stop = true; + auto res = std::make_unique(); + res->id = slot.id_task; + res->index = slot.index; const int n_embd = llama_n_embd(model); @@ -1446,32 +2162,23 @@ struct server_context { if (embd == NULL) { SLT_ERR(slot, "failed to get embeddings, token = %d, seq_id = %d\n", batch.token[i], batch.seq_id[i][0]); - res.data = json { - {"embedding", std::vector(n_embd, 0.0f)}, - {"index", slot.index}, - }; - + res->embedding = std::vector(n_embd, 0.0f); continue; } common_embd_normalize(embd, embd_res.data(), n_embd); - - res.data = json { - {"embedding", embd_res}, - {"index", slot.index}, - }; + res->embedding = embd_res; } SLT_DBG(slot, "%s", "sending embeddings\n"); - queue_results.send(res); + queue_results.send(std::move(res)); } void send_rerank(const server_slot & slot, const llama_batch & batch) { - server_task_result res; - res.id = slot.id_task; - res.error = false; - res.stop = true; + auto res = std::make_unique(); + res->id = slot.id_task; + res->index = slot.index; for (int i = 0; i < batch.n_tokens; ++i) { if (!batch.logits[i] || batch.seq_id[i][0] != slot.id) { @@ -1486,104 +2193,29 @@ struct server_context { if (embd == NULL) { SLT_ERR(slot, "failed to get embeddings, token = %d, seq_id = %d\n", batch.token[i], batch.seq_id[i][0]); - res.data = json { - {"index", slot.index}, - {"score", -1e6}, - }; - + res->score = -1e6; continue; } - res.data = json { - {"index", slot.index}, - {"score", embd[0]}, - }; + res->score = embd[0]; } - SLT_DBG(slot, "sending rerank result, res = '%s'\n", res.data.dump().c_str()); + SLT_DBG(slot, "sending rerank result, res.score = %f\n", res->score); - queue_results.send(res); + queue_results.send(std::move(res)); } // // Functions to create new task(s) and receive result(s) // - // break the input "prompt" into multiple tasks if needed, then format and tokenize the input prompt(s) - std::vector create_tasks_inference(json data, server_task_inf_type inf_type) { - std::vector tasks; - auto create_task = [&](json & task_data, llama_tokens & prompt_tokens) { - SRV_DBG("create task, n_tokens = %d\n", (int) prompt_tokens.size()); - server_task task; - task.id = queue_tasks.get_new_id(); - task.inf_type = inf_type; - task.type = SERVER_TASK_TYPE_INFERENCE; - task.data = task_data; - task.prompt_tokens = std::move(prompt_tokens); - tasks.push_back(std::move(task)); - }; - - static constexpr const char * error_msg = "\"prompt\" must be a string, an array of token ids or an array of prompts"; - if (!data.contains("prompt")) { - throw std::runtime_error(error_msg); - } - - // because llama_tokenize api is thread-safe, we can tokenize the prompt from HTTP thread - bool add_special = inf_type != SERVER_TASK_INF_TYPE_RERANK && inf_type != SERVER_TASK_INF_TYPE_INFILL; - std::vector tokenized_prompts = tokenize_input_prompts(ctx, data.at("prompt"), add_special, true); - switch (inf_type) { - case SERVER_TASK_INF_TYPE_RERANK: - { - // prompts[0] is the question - // the rest are the answers/documents - GGML_ASSERT(tokenized_prompts.size() > 1); - SRV_DBG("creating rerank tasks, n_prompts = %d\n", (int) tokenized_prompts.size() - 1); - for (size_t i = 1; i < tokenized_prompts.size(); i++) { - data["index"] = i - 1; - auto tokens = format_rerank(model, tokenized_prompts[0], tokenized_prompts[i]); - create_task(data, tokens); - } - } break; - case SERVER_TASK_INF_TYPE_INFILL: - { - SRV_DBG("creating infill tasks, n_prompts = %d\n", (int) tokenized_prompts.size()); - for (size_t i = 0; i < tokenized_prompts.size(); i++) { - data["index"] = i; - auto tokens = format_infill( - ctx, - data.at("input_prefix"), - data.at("input_suffix"), - data.at("input_extra"), - params_base.n_batch, - params_base.n_predict, - slots[0].n_ctx, // TODO: there should be a better way - params_base.spm_infill, - tokenized_prompts[i] - ); - create_task(data, tokens); - } - } break; - default: - { - SRV_DBG("creating multi-prompt tasks, n_prompts = %d\n", (int) tokenized_prompts.size()); - for (size_t i = 0; i < tokenized_prompts.size(); i++) { - data["index"] = i; - create_task(data, tokenized_prompts[i]); - } - } - } - - return tasks; - } - void cancel_tasks(const std::unordered_set & id_tasks) { std::vector cancel_tasks; cancel_tasks.reserve(id_tasks.size()); for (const auto & id_task : id_tasks) { SRV_WRN("cancel task, id_task = %d\n", id_task); - server_task task; - task.type = SERVER_TASK_TYPE_CANCEL; + server_task task(SERVER_TASK_TYPE_CANCEL); task.id_target = id_task; cancel_tasks.push_back(task); queue_results.remove_waiting_task_id(id_task); @@ -1592,50 +2224,58 @@ struct server_context { queue_tasks.post(cancel_tasks, true); } - // receive the results from task(s) created by create_tasks_inference - void receive_cmpl_results( + // receive the results from task(s) + void receive_multi_results( const std::unordered_set & id_tasks, - const std::function&)> & result_handler, + const std::function&)> & result_handler, const std::function & error_handler) { - // TODO: currently, there is no way to detect the client has cancelled the request - std::vector results(id_tasks.size()); + std::vector results(id_tasks.size()); for (size_t i = 0; i < id_tasks.size(); i++) { - server_task_result result = queue_results.recv(id_tasks); + server_task_result_ptr result = queue_results.recv(id_tasks); - if (result.error) { - error_handler(result.data); + if (result->is_error()) { + error_handler(result->to_json()); cancel_tasks(id_tasks); return; } - const size_t idx = result.data["index"]; + GGML_ASSERT( + dynamic_cast(result.get()) != nullptr + || dynamic_cast(result.get()) != nullptr + || dynamic_cast(result.get()) != nullptr + ); + const size_t idx = result->get_index(); GGML_ASSERT(idx < results.size() && "index out of range"); - - results[idx] = result; + results[idx] = std::move(result); } result_handler(results); } - // receive the results from task(s) created by create_tasks_inference, in stream mode + // receive the results from task(s), in stream mode void receive_cmpl_results_stream( - const std::unordered_set & id_tasks, const - std::function & result_handler, const - std::function & error_handler) { + const std::unordered_set & id_tasks, + const std::function & result_handler, + const std::function & error_handler) { size_t n_finished = 0; while (true) { - server_task_result result = queue_results.recv(id_tasks); + server_task_result_ptr result = queue_results.recv(id_tasks); + + if (result->is_error()) { + error_handler(result->to_json()); + cancel_tasks(id_tasks); + return; + } + + GGML_ASSERT( + dynamic_cast(result.get()) != nullptr + || dynamic_cast(result.get()) != nullptr + ); if (!result_handler(result)) { cancel_tasks(id_tasks); break; } - if (result.error) { - error_handler(result.data); - cancel_tasks(id_tasks); - break; - } - - if (result.stop) { + if (result->is_stop()) { if (++n_finished == id_tasks.size()) { break; } @@ -1649,9 +2289,12 @@ struct server_context { void process_single_task(server_task task) { switch (task.type) { - case SERVER_TASK_TYPE_INFERENCE: + case SERVER_TASK_TYPE_COMPLETION: + case SERVER_TASK_TYPE_INFILL: + case SERVER_TASK_TYPE_EMBEDDING: + case SERVER_TASK_TYPE_RERANK: { - const int id_slot = json_value(task.data, "id_slot", -1); + const int id_slot = task.id_selected_slot; server_slot * slot = id_slot != -1 ? get_slot_by_id(id_slot) : get_available_slot(task); @@ -1668,13 +2311,6 @@ struct server_context { break; } - slot->reset(); - - slot->id_task = task.id; - slot->inf_type = task.inf_type; - slot->index = json_value(task.data, "index", 0); - slot->prompt_tokens = std::move(task.prompt_tokens); - if (!launch_slot_with_task(*slot, task)) { SRV_ERR("failed to launch slot with task, id_task = %d\n", task.id); break; @@ -1702,21 +2338,7 @@ struct server_context { int n_processing_slots = 0; for (server_slot & slot : slots) { - json slot_data = get_formated_generation(slot); - slot_data["id"] = slot.id; - slot_data["id_task"] = slot.id_task; - slot_data["is_processing"] = slot.is_processing(); - slot_data["prompt"] = common_detokenize(ctx, slot.prompt_tokens); - slot_data["next_token"] = { - {"has_next_token", slot.has_next_token}, - {"has_new_line", slot.has_new_line}, - {"n_remain", slot.n_remaining}, - {"n_decoded", slot.n_decoded}, - {"stopped_eos", slot.stopped_eos}, - {"stopped_word", slot.stopped_word}, - {"stopped_limit", slot.stopped_limit}, - {"stopping_word", slot.stopping_word}, - }; + json slot_data = slot.to_json(); if (slot.is_processing()) { n_processing_slots++; @@ -1728,43 +2350,38 @@ struct server_context { } SRV_DBG("n_idle_slots = %d, n_processing_slots = %d\n", n_idle_slots, n_processing_slots); - server_task_result res; - res.id = task.id; - res.stop = true; - res.error = false; - res.data = { - { "idle", n_idle_slots }, - { "processing", n_processing_slots }, - { "deferred", queue_tasks.queue_tasks_deferred.size() }, - { "t_start", metrics.t_start}, + auto res = std::make_unique(); + res->id = task.id; + res->slots_data = std::move(slots_data); + res->n_idle_slots = n_idle_slots; + res->n_processing_slots = n_processing_slots; + res->n_tasks_deferred = queue_tasks.queue_tasks_deferred.size(); + res->t_start = metrics.t_start; - { "n_prompt_tokens_processed_total", metrics.n_prompt_tokens_processed_total}, - { "t_tokens_generation_total", metrics.t_tokens_generation_total}, - { "n_tokens_predicted_total", metrics.n_tokens_predicted_total}, - { "t_prompt_processing_total", metrics.t_prompt_processing_total}, + res->kv_cache_tokens_count = llama_get_kv_cache_token_count(ctx); + res->kv_cache_used_cells = llama_get_kv_cache_used_cells(ctx); - { "n_prompt_tokens_processed", metrics.n_prompt_tokens_processed}, - { "t_prompt_processing", metrics.t_prompt_processing}, - { "n_tokens_predicted", metrics.n_tokens_predicted}, - { "t_tokens_generation", metrics.t_tokens_generation}, + res->n_prompt_tokens_processed_total = metrics.n_prompt_tokens_processed_total; + res->t_prompt_processing_total = metrics.t_prompt_processing_total; + res->n_tokens_predicted_total = metrics.n_tokens_predicted_total; + res->t_tokens_generation_total = metrics.t_tokens_generation_total; - { "n_decode_total", metrics.n_decode_total}, - { "n_busy_slots_total", metrics.n_busy_slots_total}, + res->n_prompt_tokens_processed = metrics.n_prompt_tokens_processed; + res->t_prompt_processing = metrics.t_prompt_processing; + res->n_tokens_predicted = metrics.n_tokens_predicted; + res->t_tokens_generation = metrics.t_tokens_generation; - { "kv_cache_tokens_count", llama_get_kv_cache_token_count(ctx)}, - { "kv_cache_used_cells", llama_get_kv_cache_used_cells(ctx)}, + res->n_decode_total = metrics.n_decode_total; + res->n_busy_slots_total = metrics.n_busy_slots_total; - { "slots", slots_data }, - }; - - if (json_value(task.data, "reset_bucket", false)) { + if (task.metrics_reset_bucket) { metrics.reset_bucket(); } - queue_results.send(res); + queue_results.send(std::move(res)); } break; case SERVER_TASK_TYPE_SLOT_SAVE: { - int id_slot = task.data.at("id_slot"); + int id_slot = task.slot_action.slot_id; server_slot * slot = get_slot_by_id(id_slot); if (slot == nullptr) { send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST); @@ -1780,32 +2397,27 @@ struct server_context { const size_t token_count = slot->cache_tokens.size(); const int64_t t_start = ggml_time_us(); - std::string filename = task.data.at("filename"); - std::string filepath = task.data.at("filepath"); + std::string filename = task.slot_action.filename; + std::string filepath = task.slot_action.filepath; const size_t nwrite = llama_state_seq_save_file(ctx, filepath.c_str(), slot->id, slot->cache_tokens.data(), token_count); const int64_t t_end = ggml_time_us(); const double t_save_ms = (t_end - t_start) / 1000.0; - server_task_result result; - result.id = task.id; - result.stop = true; - result.error = false; - result.data = json { - { "id_slot", id_slot }, - { "filename", filename }, - { "n_saved", token_count }, // tokens saved - { "n_written", nwrite }, // bytes written - { "timings", { - { "save_ms", t_save_ms } - } } - }; - queue_results.send(result); + auto res = std::make_unique(); + res->id = task.id; + res->id_slot = id_slot; + res->filename = filename; + res->is_save = true; + res->n_tokens = token_count; + res->n_bytes = nwrite; + res->t_ms = t_save_ms; + queue_results.send(std::move(res)); } break; case SERVER_TASK_TYPE_SLOT_RESTORE: { - int id_slot = task.data.at("id_slot"); + int id_slot = task.slot_action.slot_id; server_slot * slot = get_slot_by_id(id_slot); if (slot == nullptr) { send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST); @@ -1820,8 +2432,8 @@ struct server_context { const int64_t t_start = ggml_time_us(); - std::string filename = task.data.at("filename"); - std::string filepath = task.data.at("filepath"); + std::string filename = task.slot_action.filename; + std::string filepath = task.slot_action.filepath; slot->cache_tokens.resize(slot->n_ctx); size_t token_count = 0; @@ -1836,24 +2448,19 @@ struct server_context { const int64_t t_end = ggml_time_us(); const double t_restore_ms = (t_end - t_start) / 1000.0; - server_task_result result; - result.id = task.id; - result.stop = true; - result.error = false; - result.data = json { - { "id_slot", id_slot }, - { "filename", filename }, - { "n_restored", token_count }, // tokens restored - { "n_read", nread }, // bytes read - { "timings", { - { "restore_ms", t_restore_ms } - } } - }; - queue_results.send(result); + auto res = std::make_unique(); + res->id = task.id; + res->id_slot = id_slot; + res->filename = filename; + res->is_save = false; + res->n_tokens = token_count; + res->n_bytes = nread; + res->t_ms = t_restore_ms; + queue_results.send(std::move(res)); } break; case SERVER_TASK_TYPE_SLOT_ERASE: { - int id_slot = task.data.at("id_slot"); + int id_slot = task.slot_action.slot_id; server_slot * slot = get_slot_by_id(id_slot); if (slot == nullptr) { send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST); @@ -1871,25 +2478,18 @@ struct server_context { llama_kv_cache_seq_rm(ctx, slot->id, -1, -1); slot->cache_tokens.clear(); - server_task_result result; - result.id = task.id; - result.stop = true; - result.error = false; - result.data = json { - { "id_slot", id_slot }, - { "n_erased", n_erased } - }; - queue_results.send(result); + auto res = std::make_unique(); + res->id = task.id; + res->id_slot = id_slot; + res->n_erased = n_erased; + queue_results.send(std::move(res)); } break; case SERVER_TASK_TYPE_SET_LORA: { common_lora_adapters_apply(ctx, loras); - server_task_result result; - result.id = task.id; - result.stop = true; - result.error = false; - result.data = json{{ "success", true }}; - queue_results.send(result); + auto res = std::make_unique(); + res->id = task.id; + queue_results.send(std::move(res)); } break; } } @@ -1919,10 +2519,8 @@ struct server_context { { SRV_DBG("%s", "posting NEXT_RESPONSE\n"); - server_task task; - task.type = SERVER_TASK_TYPE_NEXT_RESPONSE; - task.id_target = -1; - + server_task task(SERVER_TASK_TYPE_NEXT_RESPONSE); + task.id = queue_tasks.get_new_id(); queue_tasks.post(task); } @@ -2036,7 +2634,7 @@ struct server_context { continue; } - if (slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING || slot.inf_type == SERVER_TASK_INF_TYPE_RERANK) { + if (slot.is_non_causal()) { if (slot.n_prompt_tokens > n_ubatch) { slot.release(); send_error(slot, "input is too large to process. increase the physical batch size", ERROR_TYPE_SERVER); @@ -2151,7 +2749,7 @@ struct server_context { } // non-causal tasks require to fit the entire prompt in the physical batch - if (slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING || slot.inf_type == SERVER_TASK_INF_TYPE_RERANK) { + if (slot.is_non_causal()) { // cannot fit the prompt in the current batch - will try next iter if (batch.n_tokens + slot.n_prompt_tokens > n_batch) { continue; @@ -2159,10 +2757,7 @@ struct server_context { } // check that we are in the right batch_type, if not defer the slot - const bool slot_type = - slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING || - slot.inf_type == SERVER_TASK_INF_TYPE_RERANK ? 1 : 0; - + int slot_type = slot.is_non_causal(); if (batch_type == -1) { batch_type = slot_type; } else if (batch_type != slot_type) { @@ -2279,7 +2874,7 @@ struct server_context { } if (slot.state == SLOT_STATE_DONE_PROMPT) { - if (slot.inf_type == SERVER_TASK_INF_TYPE_EMBEDDING) { + if (slot.task_type == SERVER_TASK_TYPE_EMBEDDING) { // prompt evaluated for embedding send_embedding(slot, batch_view); slot.release(); @@ -2287,7 +2882,7 @@ struct server_context { continue; // continue loop of slots } - if (slot.inf_type == SERVER_TASK_INF_TYPE_RERANK) { + if (slot.task_type == SERVER_TASK_TYPE_RERANK) { send_rerank(slot, batch_view); slot.release(); slot.i_batch = -1; @@ -2324,9 +2919,11 @@ struct server_context { const auto * cur_p = common_sampler_get_candidates(slot.smpl); for (size_t i = 0; i < (size_t) slot.params.sampling.n_probs; ++i) { + auto tok_id = cur_p->data[i].id; result.probs.push_back({ - cur_p->data[i].id, - i >= cur_p->size ? 0.0f : cur_p->data[i].p, + tok_id, + tokens_to_output_formatted_string(ctx, tok_id), + i >= cur_p->size ? 0.0f : cur_p->data[i].p, }); } @@ -2478,17 +3075,9 @@ int main(int argc, char ** argv) { common_init(); - // enabling this will output extra debug information in the HTTP responses from the server - // see format_final_response_oaicompat() - const bool verbose = params.verbosity > 9; - // struct that contains llama context and inference server_context ctx_server; - if (params.model_alias == "unknown") { - params.model_alias = params.model; - } - llama_backend_init(); llama_numa_init(params.numa); @@ -2523,20 +3112,20 @@ int main(int argc, char ** argv) { auto res_error = [](httplib::Response & res, const json & error_data) { json final_response {{"error", error_data}}; - res.set_content(final_response.dump(-1, ' ', false, json::error_handler_t::replace), MIMETYPE_JSON); + res.set_content(safe_json_to_str(final_response), MIMETYPE_JSON); res.status = json_value(error_data, "code", 500); }; auto res_ok = [](httplib::Response & res, const json & data) { - res.set_content(data.dump(-1, ' ', false, json::error_handler_t::replace), MIMETYPE_JSON); + res.set_content(safe_json_to_str(data), MIMETYPE_JSON); res.status = 200; }; - svr->set_exception_handler([&res_error](const httplib::Request &, httplib::Response & res, std::exception_ptr ep) { + svr->set_exception_handler([&res_error](const httplib::Request &, httplib::Response & res, const std::exception_ptr & ep) { std::string message; try { std::rethrow_exception(ep); - } catch (std::exception & e) { + } catch (const std::exception & e) { message = e.what(); } catch (...) { message = "Unknown Exception"; @@ -2665,27 +3254,33 @@ int main(int argc, char ** argv) { } // request slots data using task queue - server_task task; + server_task task(SERVER_TASK_TYPE_METRICS); task.id = ctx_server.queue_tasks.get_new_id(); - task.type = SERVER_TASK_TYPE_METRICS; - ctx_server.queue_results.add_waiting_task_id(task.id); ctx_server.queue_tasks.post(task, true); // high-priority task // get the result - server_task_result result = ctx_server.queue_results.recv(task.id); + server_task_result_ptr result = ctx_server.queue_results.recv(task.id); ctx_server.queue_results.remove_waiting_task_id(task.id); + if (result->is_error()) { + res_error(res, result->to_json()); + return; + } + + // TODO: get rid of this dynamic_cast + auto res_metrics = dynamic_cast(result.get()); + GGML_ASSERT(res_metrics != nullptr); + // optionally return "fail_on_no_slot" error - const int n_idle_slots = result.data.at("idle"); if (req.has_param("fail_on_no_slot")) { - if (n_idle_slots == 0) { + if (res_metrics->n_idle_slots == 0) { res_error(res, format_error_response("no slot available", ERROR_TYPE_UNAVAILABLE)); return; } } - res_ok(res, result.data.at("slots")); + res_ok(res, res_metrics->slots_data); }; const auto handle_metrics = [&](const httplib::Request &, httplib::Response & res) { @@ -2695,83 +3290,77 @@ int main(int argc, char ** argv) { } // request slots data using task queue - server_task task; + server_task task(SERVER_TASK_TYPE_METRICS); task.id = ctx_server.queue_tasks.get_new_id(); - task.id_target = -1; - task.type = SERVER_TASK_TYPE_METRICS; - task.data.push_back({{"reset_bucket", true}}); + task.metrics_reset_bucket = true; ctx_server.queue_results.add_waiting_task_id(task.id); ctx_server.queue_tasks.post(task, true); // high-priority task // get the result - server_task_result result = ctx_server.queue_results.recv(task.id); + server_task_result_ptr result = ctx_server.queue_results.recv(task.id); ctx_server.queue_results.remove_waiting_task_id(task.id); - json data = result.data; + if (result->is_error()) { + res_error(res, result->to_json()); + return; + } - const uint64_t n_prompt_tokens_processed = data.at("n_prompt_tokens_processed"); - const uint64_t t_prompt_processing = data.at("t_prompt_processing"); - - const uint64_t n_tokens_predicted = data.at("n_tokens_predicted"); - const uint64_t t_tokens_generation = data.at("t_tokens_generation"); - - const uint64_t n_decode_total = data.at("n_decode_total"); - const uint64_t n_busy_slots_total = data.at("n_busy_slots_total"); - - const int32_t kv_cache_used_cells = data.at("kv_cache_used_cells"); + // TODO: get rid of this dynamic_cast + auto res_metrics = dynamic_cast(result.get()); + GGML_ASSERT(res_metrics != nullptr); // metrics definition: https://prometheus.io/docs/practices/naming/#metric-names json all_metrics_def = json { {"counter", {{ {"name", "prompt_tokens_total"}, {"help", "Number of prompt tokens processed."}, - {"value", (uint64_t) data.at("n_prompt_tokens_processed_total")} + {"value", (uint64_t) res_metrics->n_prompt_tokens_processed_total} }, { {"name", "prompt_seconds_total"}, {"help", "Prompt process time"}, - {"value", (uint64_t) data.at("t_prompt_processing_total") / 1.e3} + {"value", (uint64_t) res_metrics->t_prompt_processing_total / 1.e3} }, { {"name", "tokens_predicted_total"}, {"help", "Number of generation tokens processed."}, - {"value", (uint64_t) data.at("n_tokens_predicted_total")} + {"value", (uint64_t) res_metrics->n_tokens_predicted_total} }, { {"name", "tokens_predicted_seconds_total"}, {"help", "Predict process time"}, - {"value", (uint64_t) data.at("t_tokens_generation_total") / 1.e3} + {"value", (uint64_t) res_metrics->t_tokens_generation_total / 1.e3} }, { {"name", "n_decode_total"}, {"help", "Total number of llama_decode() calls"}, - {"value", n_decode_total} + {"value", res_metrics->n_decode_total} }, { {"name", "n_busy_slots_per_decode"}, {"help", "Average number of busy slots per llama_decode() call"}, - {"value", (float) n_busy_slots_total / (float) n_decode_total} + {"value", (float) res_metrics->n_busy_slots_total / (float) res_metrics->n_decode_total} }}}, {"gauge", {{ {"name", "prompt_tokens_seconds"}, {"help", "Average prompt throughput in tokens/s."}, - {"value", n_prompt_tokens_processed ? 1.e3 / t_prompt_processing * n_prompt_tokens_processed : 0.} + {"value", res_metrics->n_prompt_tokens_processed ? 1.e3 / res_metrics->t_prompt_processing * res_metrics->n_prompt_tokens_processed : 0.} },{ {"name", "predicted_tokens_seconds"}, {"help", "Average generation throughput in tokens/s."}, - {"value", n_tokens_predicted ? 1.e3 / t_tokens_generation * n_tokens_predicted : 0.} + {"value", res_metrics->n_tokens_predicted ? 1.e3 / res_metrics->t_tokens_generation * res_metrics->n_tokens_predicted : 0.} },{ {"name", "kv_cache_usage_ratio"}, {"help", "KV-cache usage. 1 means 100 percent usage."}, - {"value", 1. * kv_cache_used_cells / params.n_ctx} + {"value", 1. * res_metrics->kv_cache_used_cells / params.n_ctx} },{ {"name", "kv_cache_tokens"}, {"help", "KV-cache tokens."}, - {"value", (uint64_t) data.at("kv_cache_tokens_count")} + {"value", (uint64_t) res_metrics->kv_cache_tokens_count} },{ {"name", "requests_processing"}, {"help", "Number of request processing."}, - {"value", (uint64_t) data.at("processing")} + {"value", (uint64_t) res_metrics->n_processing_slots} },{ {"name", "requests_deferred"}, {"help", "Number of request deferred."}, - {"value", (uint64_t) data.at("deferred")} + {"value", (uint64_t) res_metrics->n_tasks_deferred} }}} }; @@ -2792,8 +3381,7 @@ int main(int argc, char ** argv) { } } - const int64_t t_start = data.at("t_start"); - res.set_header("Process-Start-Time-Unix", std::to_string(t_start)); + res.set_header("Process-Start-Time-Unix", std::to_string(res_metrics->t_start)); res.set_content(prometheus.str(), "text/plain; version=0.0.4"); res.status = 200; // HTTP OK @@ -2808,25 +3396,24 @@ int main(int argc, char ** argv) { } std::string filepath = params.slot_save_path + filename; - server_task task; - task.type = SERVER_TASK_TYPE_SLOT_SAVE; - task.data = { - { "id_slot", id_slot }, - { "filename", filename }, - { "filepath", filepath }, - }; + server_task task(SERVER_TASK_TYPE_SLOT_SAVE); + task.id = ctx_server.queue_tasks.get_new_id(); + task.slot_action.slot_id = id_slot; + task.slot_action.filename = filename; + task.slot_action.filepath = filepath; - const int id_task = ctx_server.queue_tasks.post(task); - ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); - server_task_result result = ctx_server.queue_results.recv(id_task); - ctx_server.queue_results.remove_waiting_task_id(id_task); + server_task_result_ptr result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); - if (result.error) { - res_error(res, result.data); - } else { - res_ok(res, result.data); + if (result->is_error()) { + res_error(res, result->to_json()); + return; } + + res_ok(res, result->to_json()); }; const auto handle_slots_restore = [&ctx_server, &res_error, &res_ok, ¶ms](const httplib::Request & req, httplib::Response & res, int id_slot) { @@ -2838,45 +3425,45 @@ int main(int argc, char ** argv) { } std::string filepath = params.slot_save_path + filename; - server_task task; - task.type = SERVER_TASK_TYPE_SLOT_RESTORE; - task.data = { - { "id_slot", id_slot }, - { "filename", filename }, - { "filepath", filepath }, - }; + server_task task(SERVER_TASK_TYPE_SLOT_RESTORE); + task.id = ctx_server.queue_tasks.get_new_id(); + task.slot_action.slot_id = id_slot; + task.slot_action.filename = filename; + task.slot_action.filepath = filepath; - const int id_task = ctx_server.queue_tasks.post(task); - ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); - server_task_result result = ctx_server.queue_results.recv(id_task); - ctx_server.queue_results.remove_waiting_task_id(id_task); + server_task_result_ptr result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); - if (result.error) { - res_error(res, result.data); - } else { - res_ok(res, result.data); + if (result->is_error()) { + res_error(res, result->to_json()); + return; } + + GGML_ASSERT(dynamic_cast(result.get()) != nullptr); + res_ok(res, result->to_json()); }; const auto handle_slots_erase = [&ctx_server, &res_error, &res_ok](const httplib::Request & /* req */, httplib::Response & res, int id_slot) { - server_task task; - task.type = SERVER_TASK_TYPE_SLOT_ERASE; - task.data = { - { "id_slot", id_slot }, - }; + server_task task(SERVER_TASK_TYPE_SLOT_ERASE); + task.id = ctx_server.queue_tasks.get_new_id(); + task.slot_action.slot_id = id_slot; - const int id_task = ctx_server.queue_tasks.post(task); - ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); - server_task_result result = ctx_server.queue_results.recv(id_task); - ctx_server.queue_results.remove_waiting_task_id(id_task); + server_task_result_ptr result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); - if (result.error) { - res_error(res, result.data); - } else { - res_ok(res, result.data); + if (result->is_error()) { + res_error(res, result->to_json()); + return; } + + GGML_ASSERT(dynamic_cast(result.get()) != nullptr); + res_ok(res, result->to_json()); }; const auto handle_slots_action = [¶ms, &res_error, &handle_slots_save, &handle_slots_restore, &handle_slots_erase](const httplib::Request & req, httplib::Response & res) { @@ -2913,6 +3500,7 @@ int main(int argc, char ** argv) { json data = { { "default_generation_settings", ctx_server.default_generation_settings_for_props }, { "total_slots", ctx_server.params_base.n_parallel }, + { "model_path", ctx_server.params_base.model }, { "bos_token", common_token_to_piece(ctx_server.ctx, llama_token_bos(ctx_server.model), true) }, { "eos_token", common_token_to_piece(ctx_server.ctx, llama_token_eos(ctx_server.model), true) }, { "chat_template", chat_template.source()}, @@ -2940,13 +3528,54 @@ int main(int argc, char ** argv) { res_ok(res, {{ "success", true }}); }; - const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok](server_task_inf_type inf_type, json & data, httplib::Response & res) { + // handle completion-like requests (completion, chat, infill) + // we can optionally provide a custom format for partial results and final results + const auto handle_completions_generic = [&ctx_server, &res_error, &res_ok]( + server_task_type type, + json & data, + httplib::Response & res, + bool oaicompat = false, + bool oaicompat_chat = false, + llama_tool_call_style tool_call_style = llama_tool_call_style::None) { + GGML_ASSERT(type == SERVER_TASK_TYPE_COMPLETION || type == SERVER_TASK_TYPE_INFILL); + if (ctx_server.params_base.embedding) { res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED)); return; } - std::vector tasks = ctx_server.create_tasks_inference(data, inf_type); + auto completion_id = gen_chatcmplid(); + std::vector tasks; + + try { + std::vector tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, data.at("prompt"), true, true); + tasks.reserve(tokenized_prompts.size()); + for (size_t i = 0; i < tokenized_prompts.size(); i++) { + server_task task = server_task(type); + + task.id = ctx_server.queue_tasks.get_new_id(); + task.index = i; + + task.prompt_tokens = std::move(tokenized_prompts[i]); + task.params = server_task::params_from_json_cmpl(ctx_server.model, ctx_server.params_base, data); + task.id_selected_slot = json_value(data, "id_slot", -1); + + // OAI-compat + task.params.oaicompat = oaicompat; + task.params.oaicompat_chat = oaicompat_chat; + task.params.oaicompat_cmpl_id = completion_id; + task.params.oaicompat_tools = json_value(data, "tools", json::array()); + task.params.oaicompat_tool_call_style = tool_call_style; + + // oaicompat_model is already populated by params_from_json_cmpl + + tasks.push_back(task); + } + } catch (const std::exception & e) { + res_error(res, format_error_response(e.what(), ERROR_TYPE_INVALID_REQUEST)); + return; + } + ctx_server.queue_results.add_waiting_tasks(tasks); ctx_server.queue_tasks.post(tasks); @@ -2954,15 +3583,15 @@ int main(int argc, char ** argv) { const auto task_ids = server_task::get_list_id(tasks); if (!stream) { - ctx_server.receive_cmpl_results(task_ids, [&](std::vector & results) { + ctx_server.receive_multi_results(task_ids, [&](std::vector & results) { if (results.size() == 1) { // single result - res_ok(res, results[0].data); + res_ok(res, results[0]->to_json()); } else { // multiple results (multitask) json arr = json::array(); - for (const auto & res : results) { - arr.push_back(res.data); + for (auto & res : results) { + arr.push_back(res->to_json()); } res_ok(res, arr); } @@ -2972,12 +3601,26 @@ int main(int argc, char ** argv) { ctx_server.queue_results.remove_waiting_task_ids(task_ids); } else { - const auto chunked_content_provider = [task_ids, &ctx_server](size_t, httplib::DataSink & sink) { - ctx_server.receive_cmpl_results_stream(task_ids, [&](const server_task_result & result) -> bool { - return server_sent_event(sink, "data", result.data); + const auto chunked_content_provider = [task_ids, &ctx_server, oaicompat](size_t, httplib::DataSink & sink) { + ctx_server.receive_cmpl_results_stream(task_ids, [&](server_task_result_ptr & result) -> bool { + json res_json = result->to_json(); + if (res_json.is_array()) { + for (const auto & res : res_json) { + if (!server_sent_event(sink, "data", res)) { + return false; + } + } + return true; + } else { + return server_sent_event(sink, "data", res_json); + } }, [&](const json & error_data) { server_sent_event(sink, "error", error_data); }); + if (oaicompat) { + static const std::string ev_done = "data: [DONE]\n\n"; + sink.write(ev_done.data(), ev_done.size()); + } sink.done(); return false; }; @@ -2992,7 +3635,12 @@ int main(int argc, char ** argv) { const auto handle_completions = [&handle_completions_generic](const httplib::Request & req, httplib::Response & res) { json data = json::parse(req.body); - return handle_completions_generic(SERVER_TASK_INF_TYPE_COMPLETION, data, res); + return handle_completions_generic( + SERVER_TASK_TYPE_COMPLETION, + data, + res, + /* oaicompat */ false, + /* oaicompat_chat */ false); }; const auto handle_infill = [&ctx_server, &res_error, &handle_completions_generic](const httplib::Request & req, httplib::Response & res) { @@ -3015,6 +3663,11 @@ int main(int argc, char ** argv) { json data = json::parse(req.body); // validate input + if (data.contains("prompt") && !data.at("prompt").is_string()) { + // prompt is optional + res_error(res, format_error_response("\"prompt\" must be a string", ERROR_TYPE_INVALID_REQUEST)); + } + if (!data.contains("input_prefix")) { res_error(res, format_error_response("\"input_prefix\" is required", ERROR_TYPE_INVALID_REQUEST)); } @@ -3024,9 +3677,11 @@ int main(int argc, char ** argv) { } if (data.contains("input_extra") && !data.at("input_extra").is_array()) { + // input_extra is optional res_error(res, format_error_response("\"input_extra\" must be an array of {\"filename\": string, \"text\": string}", ERROR_TYPE_INVALID_REQUEST)); return; } + json input_extra = json_value(data, "input_extra", json::array()); for (const auto & chunk : input_extra) { // { "text": string, "filename": string } @@ -3042,11 +3697,25 @@ int main(int argc, char ** argv) { } data["input_extra"] = input_extra; // default to empty array if it's not exist - return handle_completions_generic(SERVER_TASK_INF_TYPE_INFILL, data, res); + std::string prompt = json_value(data, "prompt", std::string()); + std::vector tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, prompt, true, true); + SRV_DBG("creating infill tasks, n_prompts = %d\n", (int) tokenized_prompts.size()); + data["prompt"] = format_infill( + ctx_server.ctx, + data.at("input_prefix"), + data.at("input_suffix"), + data.at("input_extra"), + ctx_server.params_base.n_batch, + ctx_server.params_base.n_predict, + ctx_server.slots[0].n_ctx, // TODO: there should be a better way + ctx_server.params_base.spm_infill, + tokenized_prompts[0] + ); + + return handle_completions_generic(SERVER_TASK_TYPE_INFILL, data, res); }; - // TODO: maybe merge this function with "handle_completions_generic" - const auto handle_chat_completions = [&ctx_server, ¶ms, &res_error, &res_ok, verbose](const httplib::Request & req, httplib::Response & res) { + const auto handle_chat_completions = [&ctx_server, ¶ms, &res_error, &handle_completions_generic](const httplib::Request & req, httplib::Response & res) { if (ctx_server.params_base.embedding) { res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED)); return; @@ -3057,67 +3726,17 @@ int main(int argc, char ** argv) { auto tool_call_style = llama_tool_call_style_detect(chat_template); LOG_INF("Tool call style: %s\n", llama_tool_call_style_name(tool_call_style).c_str()); - json data; - try { - data = oaicompat_completion_params_parse(ctx_server.model, body, chat_template, tool_call_style, params.use_jinja); - } catch (const std::exception & e) { - res_error(res, format_error_response(e.what(), ERROR_TYPE_NOT_SUPPORTED)); - return; - } - - std::vector tasks = ctx_server.create_tasks_inference(data, SERVER_TASK_INF_TYPE_COMPLETION); - ctx_server.queue_results.add_waiting_tasks(tasks); - ctx_server.queue_tasks.post(tasks); - - bool stream = json_value(data, "stream", false); - const auto task_ids = server_task::get_list_id(tasks); - const auto completion_id = gen_chatcmplid(); - - if (!stream) { - ctx_server.receive_cmpl_results(task_ids, [&](const std::vector & results) { - // multitask is never support in chat completion, there is only one result - try { - json result_oai = format_final_response_oaicompat(data, results[0].data, completion_id, tool_call_style, /*.streaming =*/ false, verbose); - res_ok(res, result_oai); - } catch (const std::runtime_error & e) { - res_error(res, format_error_response(e.what(), ERROR_TYPE_SERVER)); - } - }, [&](const json & error_data) { - res_error(res, error_data); - }); - - ctx_server.queue_results.remove_waiting_task_ids(task_ids); - } else { - const auto chunked_content_provider = [task_ids, &ctx_server, completion_id](size_t, httplib::DataSink & sink) { - ctx_server.receive_cmpl_results_stream(task_ids, [&](const server_task_result & result) -> bool { - std::vector result_array = format_partial_response_oaicompat(result.data, completion_id); - for (auto & event_data : result_array) { - if (event_data.empty()) { - continue; // skip the stop token - } - if (!server_sent_event(sink, "data", event_data)) { - return false; // connection is closed - } - } - return true; // ok - }, [&](const json & error_data) { - server_sent_event(sink, "error", error_data); - }); - static const std::string ev_done = "data: [DONE]\n\n"; - sink.write(ev_done.data(), ev_done.size()); - sink.done(); - return true; - }; - - auto on_complete = [task_ids, &ctx_server] (bool) { - ctx_server.queue_results.remove_waiting_task_ids(task_ids); - }; - - res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); - } + json data = oaicompat_completion_params_parse(ctx_server.model, json::parse(req.body), chat_template, tool_call_style, params.use_jinja); + return handle_completions_generic( + SERVER_TASK_TYPE_COMPLETION, + data, + res, + /* oaicompat */ true, + /* oaicompat_chat */ true, + tool_call_style); }; - const auto handle_models = [¶ms, &ctx_server](const httplib::Request &, httplib::Response & res) { + const auto handle_models = [¶ms, &ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) { json models = { {"object", "list"}, {"data", { @@ -3131,7 +3750,7 @@ int main(int argc, char ** argv) { }} }; - res.set_content(models.dump(), MIMETYPE_JSON); + res_ok(res, models); }; const auto handle_tokenize = [&ctx_server, &res_ok](const httplib::Request & req, httplib::Response & res) { @@ -3189,12 +3808,12 @@ int main(int argc, char ** argv) { const auto handle_embeddings = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) { const json body = json::parse(req.body); - bool is_openai = false; + bool oaicompat = false; // an input prompt can be a string or a list of tokens (integer) json prompt; if (body.count("input") != 0) { - is_openai = true; + oaicompat = true; prompt = body.at("input"); } else if (body.count("content") != 0) { // with "content", we only support single prompt @@ -3208,16 +3827,26 @@ int main(int argc, char ** argv) { json responses = json::array(); bool error = false; { - std::vector tasks = ctx_server.create_tasks_inference({{"prompt", prompt}}, SERVER_TASK_INF_TYPE_EMBEDDING); + std::vector tasks; + std::vector tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, prompt, /* add_special */ false, true); + for (size_t i = 0; i < tokenized_prompts.size(); i++) { + server_task task = server_task(SERVER_TASK_TYPE_EMBEDDING); + task.id = ctx_server.queue_tasks.get_new_id(); + task.index = i; + task.prompt_tokens = std::move(tokenized_prompts[i]); + tasks.push_back(task); + } + ctx_server.queue_results.add_waiting_tasks(tasks); ctx_server.queue_tasks.post(tasks); // get the result std::unordered_set task_ids = server_task::get_list_id(tasks); - ctx_server.receive_cmpl_results(task_ids, [&](std::vector & results) { - for (const auto & res : results) { - responses.push_back(res.data); + ctx_server.receive_multi_results(task_ids, [&](std::vector & results) { + for (auto & res : results) { + GGML_ASSERT(dynamic_cast(res.get()) != nullptr); + responses.push_back(res->to_json()); } }, [&](const json & error_data) { res_error(res, error_data); @@ -3232,9 +3861,9 @@ int main(int argc, char ** argv) { } // write JSON response - json root = is_openai + json root = oaicompat ? format_embeddings_response_oaicompat(body, responses) - : responses[0]; + : responses.size() == 1 ? responses[0] : json(responses); res_ok(res, root); }; @@ -3273,29 +3902,33 @@ int main(int argc, char ** argv) { return; } - // construct prompt object: array of ["query", "doc0", "doc1", ...] - json prompt; - prompt.push_back(query); - for (const auto & doc : documents) { - prompt.push_back(doc); - } - - LOG_DBG("rerank prompt: %s\n", prompt.dump().c_str()); + llama_tokens tokenized_query = tokenize_input_prompts(ctx_server.ctx, query, /* add_special */ false, true)[0]; // create and queue the task json responses = json::array(); bool error = false; { - std::vector tasks = ctx_server.create_tasks_inference({{"prompt", prompt}}, SERVER_TASK_INF_TYPE_RERANK); + std::vector tasks; + std::vector tokenized_docs = tokenize_input_prompts(ctx_server.ctx, documents, /* add_special */ false, true); + tasks.reserve(tokenized_docs.size()); + for (size_t i = 0; i < tokenized_docs.size(); i++) { + server_task task = server_task(SERVER_TASK_TYPE_RERANK); + task.id = ctx_server.queue_tasks.get_new_id(); + task.index = i; + task.prompt_tokens = format_rerank(ctx_server.model, tokenized_query, tokenized_docs[i]); + tasks.push_back(task); + } + ctx_server.queue_results.add_waiting_tasks(tasks); ctx_server.queue_tasks.post(tasks); // get the result std::unordered_set task_ids = server_task::get_list_id(tasks); - ctx_server.receive_cmpl_results(task_ids, [&](std::vector & results) { - for (const auto & res : results) { - responses.push_back(res.data); + ctx_server.receive_multi_results(task_ids, [&](std::vector & results) { + for (auto & res : results) { + GGML_ASSERT(dynamic_cast(res.get()) != nullptr); + responses.push_back(res->to_json()); } }, [&](const json & error_data) { res_error(res, error_data); @@ -3346,16 +3979,21 @@ int main(int argc, char ** argv) { } } - server_task task; - task.type = SERVER_TASK_TYPE_SET_LORA; - const int id_task = ctx_server.queue_tasks.post(task); - ctx_server.queue_results.add_waiting_task_id(id_task); + server_task task(SERVER_TASK_TYPE_SET_LORA); + task.id = ctx_server.queue_tasks.get_new_id(); + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); - server_task_result result = ctx_server.queue_results.recv(id_task); - ctx_server.queue_results.remove_waiting_task_id(id_task); + server_task_result_ptr result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); - res_ok(res, result.data); - res.status = 200; // HTTP OK + if (result->is_error()) { + res_error(res, result->to_json()); + return; + } + + GGML_ASSERT(dynamic_cast(result.get()) != nullptr); + res_ok(res, result->to_json()); }; // diff --git a/examples/server/tests/README.md b/examples/server/tests/README.md index eb7c54d8e..fa3d0a2f5 100644 --- a/examples/server/tests/README.md +++ b/examples/server/tests/README.md @@ -44,11 +44,10 @@ To run with stdout/stderr display in real time (verbose output, but useful for d DEBUG=1 ./tests.sh -s -v -x ``` -Some tests (especially `@slow` ones) require model downloads. Since this can time out the tests, you can pre-download them in the cache ahead of time with: +Hint: You can compile and run test in single command, useful for local developement: ```shell -pip install -r examples/server/tests/requirements.txt -python scripts/fetch_server_test_models.py +cmake --build build -j --target llama-server && ./examples/server/tests/tests.sh ``` 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/tests.sh b/examples/server/tests/tests.sh index f57a9b40f..87526c3b4 100755 --- a/examples/server/tests/tests.sh +++ b/examples/server/tests/tests.sh @@ -1,5 +1,9 @@ #!/bin/bash +# make sure we are in the right directory +SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd ) +cd $SCRIPT_DIR + set -eu if [ $# -lt 1 ] diff --git a/examples/server/tests/unit/test_basic.py b/examples/server/tests/unit/test_basic.py index d82d54a5a..1d5124016 100644 --- a/examples/server/tests/unit/test_basic.py +++ b/examples/server/tests/unit/test_basic.py @@ -22,7 +22,12 @@ def test_server_props(): server.start() res = server.make_request("GET", "/props") assert res.status_code == 200 + assert ".gguf" in res.body["model_path"] assert res.body["total_slots"] == server.n_slots + default_val = res.body["default_generation_settings"] + assert server.n_ctx is not None and server.n_slots is not None + assert default_val["n_ctx"] == server.n_ctx / server.n_slots + assert default_val["params"]["seed"] == server.seed def test_server_models(): @@ -33,6 +38,31 @@ def test_server_models(): assert len(res.body["data"]) == 1 assert res.body["data"][0]["id"] == server.model_alias + +def test_server_slots(): + global server + + # without slots endpoint enabled, this should return error + server.server_slots = False + server.start() + res = server.make_request("GET", "/slots") + assert res.status_code == 501 # ERROR_TYPE_NOT_SUPPORTED + assert "error" in res.body + server.stop() + + # with slots endpoint enabled, this should return slots info + server.server_slots = True + server.n_slots = 2 + server.start() + res = server.make_request("GET", "/slots") + assert res.status_code == 200 + assert len(res.body) == server.n_slots + assert server.n_ctx is not None and server.n_slots is not None + assert res.body[0]["n_ctx"] == server.n_ctx / server.n_slots + assert "params" in res.body[0] + assert res.body[0]["params"]["seed"] == server.seed + + def test_load_split_model(): global server server.model_hf_repo = "ggml-org/models" diff --git a/examples/server/tests/unit/test_chat_completion.py b/examples/server/tests/unit/test_chat_completion.py index d2dab04ca..3b1f25f97 100644 --- a/examples/server/tests/unit/test_chat_completion.py +++ b/examples/server/tests/unit/test_chat_completion.py @@ -12,13 +12,13 @@ def create_server(): @pytest.mark.parametrize( - "model,system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,truncated", + "model,system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,finish_reason", [ - ("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), + (None, "Book", "What is the best book", 8, "(Suddenly)+", 77, 8, "length"), + ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length"), ] ) -def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, truncated): +def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, finish_reason): global server server.start() res = server.make_request("POST", "/chat/completions", data={ @@ -30,29 +30,28 @@ def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_conte ], }) assert res.status_code == 200 + assert "cmpl" in res.body["id"] # make sure the completion id has the expected format + assert res.body["model"] == model if model is not None else server.model_alias 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" + assert choice["finish_reason"] == finish_reason @pytest.mark.parametrize( - "model,system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,truncated", + "system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,finish_reason", [ - ("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), + ("Book", "What is the best book", 8, "(Suddenly)+", 77, 8, "length"), + ("You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length"), ] ) -def test_chat_completion_stream(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, truncated): +def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, finish_reason): global server + server.model_alias = None # try using DEFAULT_OAICOMPAT_MODEL server.start() res = server.make_stream_request("POST", "/chat/completions", data={ - "model": model, "max_tokens": max_tokens, "messages": [ {"role": "system", "content": system_prompt}, @@ -61,18 +60,19 @@ def test_chat_completion_stream(model, system_prompt, user_prompt, max_tokens, r "stream": True, }) content = "" + last_cmpl_id = None for data in res: choice = data["choices"][0] + assert "gpt-3.5" in data["model"] # DEFAULT_OAICOMPAT_MODEL, maybe changed in the future + if last_cmpl_id is None: + last_cmpl_id = data["id"] + assert last_cmpl_id == data["id"] # make sure the completion id is the same for all events in the stream 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" + assert choice["finish_reason"] == finish_reason else: assert choice["finish_reason"] is None content += choice["delta"]["content"] @@ -93,7 +93,7 @@ def test_chat_completion_with_openai_library(): temperature=0.8, ) print(res) - assert res.choices[0].finish_reason == "stop" + assert res.choices[0].finish_reason == "length" assert res.choices[0].message.content is not None assert match_regex("(Suddenly)+", res.choices[0].message.content) diff --git a/examples/server/tests/unit/test_completion.py b/examples/server/tests/unit/test_completion.py index 2fa30dd03..7f4f9cd03 100644 --- a/examples/server/tests/unit/test_completion.py +++ b/examples/server/tests/unit/test_completion.py @@ -42,15 +42,39 @@ def test_completion_stream(prompt: str, n_predict: int, re_content: str, n_promp }) content = "" for data in res: + assert "stop" in data and type(data["stop"]) == bool if data["stop"]: assert data["timings"]["prompt_n"] == n_prompt assert data["timings"]["predicted_n"] == n_predicted assert data["truncated"] == truncated + assert data["stop_type"] == "limit" + assert "generation_settings" in data + assert server.n_predict is not None + assert data["generation_settings"]["n_predict"] == min(n_predict, server.n_predict) + assert data["generation_settings"]["seed"] == server.seed assert match_regex(re_content, content) else: content += data["content"] +def test_completion_stream_vs_non_stream(): + global server + server.start() + res_stream = server.make_stream_request("POST", "/completion", data={ + "n_predict": 8, + "prompt": "I believe the meaning of life is", + "stream": True, + }) + res_non_stream = server.make_request("POST", "/completion", data={ + "n_predict": 8, + "prompt": "I believe the meaning of life is", + }) + content_stream = "" + for data in res_stream: + content_stream += data["content"] + assert content_stream == res_non_stream.body["content"] + + @pytest.mark.parametrize("n_slots", [1, 2]) def test_consistent_result_same_seed(n_slots: int): global server @@ -221,3 +245,24 @@ def test_completion_parallel_slots(n_slots: int, n_requests: int): 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"]) + + +def test_n_probs(): + global server + server.start() + res = server.make_request("POST", "/completion", data={ + "prompt": "I believe the meaning of life is", + "n_probs": 10, + "temperature": 0.0, + "n_predict": 5, + }) + assert res.status_code == 200 + assert "completion_probabilities" in res.body + assert len(res.body["completion_probabilities"]) == 5 + for tok in res.body["completion_probabilities"]: + assert "probs" in tok + assert len(tok["probs"]) == 10 + for prob in tok["probs"]: + assert "prob" in prob + assert "tok_str" in prob + assert 0.0 <= prob["prob"] <= 1.0 diff --git a/examples/server/tests/unit/test_infill.py b/examples/server/tests/unit/test_infill.py index 6a6d40a1c..ad4b8192a 100644 --- a/examples/server/tests/unit/test_infill.py +++ b/examples/server/tests/unit/test_infill.py @@ -13,28 +13,28 @@ 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_prefix": "#include \n#include \"llama.h\"\n\nint main() {\n", + "prompt": " 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"]) + assert match_regex("(Ann|small|shiny)+", 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_prefix": "#include \n#include \"llama.h\"\n\nint main() {\n", + "prompt": " 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"]) + assert match_regex("(Dad|excited|park)+", res.body["content"]) @pytest.mark.parametrize("input_extra", [ @@ -48,10 +48,30 @@ def test_invalid_input_extra_req(input_extra): global server server.start() res = server.make_request("POST", "/infill", data={ - "prompt": "Complete this", "input_extra": [input_extra], - "input_prefix": "#include \n#include \"llama.h\"\n\nint main() {\n int n_threads = llama_", + "input_prefix": "#include \n#include \"llama.h\"\n\nint main() {\n", + "prompt": " int n_threads = llama_", "input_suffix": "}\n", }) assert res.status_code == 400 assert "error" in res.body + + +@pytest.mark.skipif(not is_slow_test_allowed(), reason="skipping slow test") +def test_with_qwen_model(): + global server + server.model_file = None + server.model_hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-IQ3_XXS-GGUF" + server.model_hf_file = "qwen2.5-coder-1.5b-iq3_xxs-imat.gguf" + server.start(timeout_seconds=600) + res = server.make_request("POST", "/infill", data={ + "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", + "prompt": " int n_threads = llama_", + "input_suffix": "}\n", + }) + assert res.status_code == 200 + assert res.body["content"] == "n_threads();\n printf(\"Number of threads: %d\\n\", n_threads);\n return 0;\n" diff --git a/examples/server/tests/utils.py b/examples/server/tests/utils.py index 65080402a..8837c6c7d 100644 --- a/examples/server/tests/utils.py +++ b/examples/server/tests/utils.py @@ -64,6 +64,7 @@ class ServerProcess: server_embeddings: bool | None = False server_reranking: bool | None = False server_metrics: bool | None = False + server_slots: bool | None = False draft: int | None = None api_key: str | None = None response_format: str | None = None @@ -93,7 +94,6 @@ class ServerProcess: else: server_path = "../../../build/bin/llama-server" server_args = [ - "--slots", # requires to get slot status via /slots endpoint "--host", self.server_host, "--port", @@ -131,6 +131,8 @@ class ServerProcess: server_args.append("--reranking") if self.server_metrics: server_args.append("--metrics") + if self.server_slots: + server_args.append("--slots") if self.model_alias: server_args.extend(["--alias", self.model_alias]) if self.n_ctx: @@ -187,7 +189,7 @@ class ServerProcess: start_time = time.time() while time.time() - start_time < timeout_seconds: try: - response = self.make_request("GET", "/slots", headers={ + response = self.make_request("GET", "/health", headers={ "Authorization": f"Bearer {self.api_key}" if self.api_key else None }) if response.status_code == 200: @@ -230,7 +232,7 @@ class ServerProcess: 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) + print("Response from server", json.dumps(result.body, indent=2)) return result def make_stream_request( @@ -251,7 +253,7 @@ class ServerProcess: break elif line.startswith('data: '): data = json.loads(line[6:]) - print("Partial response from server", data) + print("Partial response from server", json.dumps(data, indent=2)) yield data @@ -375,3 +377,6 @@ def match_regex(regex: str, text: str) -> bool: ).search(text) is not None ) + +def is_slow_test_allowed(): + return os.environ.get("SLOW_TESTS") == "1" or os.environ.get("SLOW_TESTS") == "ON" diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp index ac0843695..c73a5f042 100644 --- a/examples/server/utils.hpp +++ b/examples/server/utils.hpp @@ -22,6 +22,7 @@ #include #include #include +#include #define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613" @@ -42,17 +43,6 @@ using json = nlohmann::ordered_json; #define QUE_ERR(fmt, ...) LOG_ERR("que %12.*s: " fmt, 12, __func__, __VA_ARGS__) #define QUE_DBG(fmt, ...) LOG_DBG("que %12.*s: " fmt, 12, __func__, __VA_ARGS__) -// https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11 -enum error_type { - ERROR_TYPE_INVALID_REQUEST, - ERROR_TYPE_AUTHENTICATION, - ERROR_TYPE_SERVER, - ERROR_TYPE_NOT_FOUND, - ERROR_TYPE_PERMISSION, - ERROR_TYPE_UNAVAILABLE, // custom error - ERROR_TYPE_NOT_SUPPORTED, // custom error -}; - template static T json_value(const json & body, const std::string & key, const T & default_value) { // Fallback null to default value @@ -176,6 +166,9 @@ static std::vector tokenize_input_prompts(llama_context * ctx, con } else { throw std::runtime_error("\"prompt\" must be a string, an list of tokens, a list of mixed strings & tokens, or a list of prompts"); } + if (result.empty()) { + throw std::runtime_error("\"prompt\" must not be empty"); + } return result; } @@ -474,48 +467,11 @@ static std::string tokens_to_output_formatted_string(const llama_context * ctx, return out; } -struct completion_token_output { - llama_token tok; - std::string text_to_send; - - struct token_prob { - llama_token tok; - float prob; - }; - - std::vector probs; -}; - -// convert a vector of completion_token_output to json -static json probs_vector_to_json(const llama_context * ctx, const std::vector & probs) { - json out = json::array(); - - for (const auto & prob : probs) { - json probs_for_token = json::array(); - - for (const auto & p : prob.probs) { - const std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok); - probs_for_token.push_back(json { - {"tok_str", tok_str}, - {"prob", p.prob}, - }); - } - - const std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok); - out.push_back(json { - {"content", tok_str}, - {"probs", probs_for_token}, - }); - } - - return out; -} - static bool server_sent_event(httplib::DataSink & sink, const char * event, const json & data) { const std::string str = std::string(event) + ": " + data.dump(-1, ' ', false, json::error_handler_t::replace) + - "\n\n"; // note: these newlines are important (not sure why though, if you know, add a comment to explain) + "\n\n"; // required by RFC 8895 - A message is terminated by a blank line (two line terminators in a row). LOG_DBG("data stream, to_send: %s", str.c_str()); @@ -535,11 +491,14 @@ static json oaicompat_completion_params_parse( { json llama_params; - llama_params["__oaicompat"] = true; - auto tools = json_value(body, "tools", json()); auto has_tools = tools.is_array() && !tools.empty(); + auto stream = json_value(body, "stream", json()); + if (stream && has_tools) { + throw std::runtime_error("Cannot use tools with stream"); + } + // Apply chat template to the list of messages llama_params["chat_template"] = tmpl.source(); @@ -589,22 +548,24 @@ static json oaicompat_completion_params_parse( if (use_jinja) { bool allow_content = tool_choice != "required"; if (tool_choice != "none" && has_tools) { + llama_params["tools"] = tools; + llama_params["tool_call_style"] = tool_call_style; + auto parallel_tool_calls = body.contains("parallel_tool_calls") ? body.at("parallel_tool_calls") : json(); - llama_params["parse_tool_calls"] = true; llama_params["parallel_tool_calls"] = parallel_tool_calls; auto handler = llama_tool_call_handler_init(tool_call_style, tmpl, allow_content, parallel_tool_calls, body.at("messages"), tools, llama_params["json_schema"]); llama_params["prompt"] = handler.prompt; - for (const auto & stop : handler.additional_stop_words) { + for (const auto & stop : handler.additional_stops) { llama_params["stop"].push_back(stop); } - if (!handler.grammar_trigger_words.empty()) { + if (!handler.grammar_triggers.empty()) { auto triggers = json::array(); - for (const auto & word : handler.grammar_trigger_words) { + for (const auto & word : handler.grammar_triggers) { triggers.push_back(word); } - llama_params["grammar_trigger_words"] = triggers; + llama_params["grammar_triggers"] = triggers; } if (!handler.grammar.empty()) { if (llama_params.contains("grammar")) { @@ -656,192 +617,6 @@ static json oaicompat_completion_params_parse( return llama_params; } -static json format_final_response_oaicompat(const json & request, const json & result, const std::string & completion_id, llama_tool_call_style tool_call_style, bool streaming = false, bool verbose = false) { - bool stopped_word = result.count("stopped_word") != 0; - bool stopped_eos = json_value(result, "stopped_eos", false); - int num_tokens_predicted = json_value(result, "tokens_predicted", 0); - int num_prompt_tokens = json_value(result, "tokens_evaluated", 0); - std::string content = json_value(result, "content", std::string("")); - - std::string finish_reason = "length"; - if (stopped_word || stopped_eos) { - finish_reason = "stop"; - } - auto chat_template = json_value(request, "chat_template", std::string()); - llama_tool_calls parsed_tool_calls; - auto tools = json_value(request, "tools", json::array()); - json tool_calls; - json message_content; - if (json_value(request, "parse_tool_calls", false)) { - parsed_tool_calls = parse_tool_calls(tool_call_style, tools, content); - if (!parsed_tool_calls.tool_calls.empty()) { - finish_reason = "tool_calls"; - message_content = parsed_tool_calls.content; - tool_calls = json::array(); - for (const auto & tc : parsed_tool_calls.tool_calls) { - tool_calls.push_back({ - {"type", "function"}, - {"function", { - {"name", tc.name}, - {"arguments", tc.arguments}, - }}, - {"id", tc.id.empty() ? json() : json(tc.id)}, - }); - } - } else { - message_content = parsed_tool_calls.content; - } - } else { - message_content = content; - } - - json choices = - streaming ? json::array({json{{"finish_reason", finish_reason}, - {"index", 0}, - {"delta", json::object()}}}) - : json::array({json{{"finish_reason", finish_reason}, - {"index", 0}, - {"message", json{{"content", message_content}, - {"tool_calls", tool_calls}, - {"role", "assistant"}}}}}); - - std::time_t t = std::time(0); - - json res = json { - {"choices", choices}, - {"created", t}, - {"model", - json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))}, - {"object", streaming ? "chat.completion.chunk" : "chat.completion"}, - {"usage", json { - {"completion_tokens", num_tokens_predicted}, - {"prompt_tokens", num_prompt_tokens}, - {"total_tokens", num_tokens_predicted + num_prompt_tokens} - }}, - {"id", completion_id} - }; - - // extra fields for debugging purposes - if (verbose) { - res["__verbose"] = result; - } - - if (result.contains("completion_probabilities")) { - res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array()); - } - - if (result.contains("timings")) { - res.push_back({"timings", json_value(result, "timings", json::object())}); - } - - return res; -} - -// return value is vector as there is one case where we might need to generate two responses -static std::vector format_partial_response_oaicompat(const json & result, const std::string & completion_id) { - if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) { - return std::vector({result}); - } - - bool first = json_value(result, "oaicompat_token_ctr", 0) == 0; - std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL)); - - bool stopped_word = json_value(result, "stopped_word", false); - bool stopped_eos = json_value(result, "stopped_eos", false); - bool stopped_limit = json_value(result, "stopped_limit", false); - std::string content = json_value(result, "content", std::string("")); - - std::string finish_reason; - if (stopped_word || stopped_eos) { - finish_reason = "stop"; - } - if (stopped_limit) { - finish_reason = "length"; - } - - std::time_t t = std::time(0); - - json choices; - - if (!finish_reason.empty()) { - choices = json::array({json{{"finish_reason", finish_reason}, - {"index", 0}, - {"delta", json::object()}}}); - } else { - if (first) { - if (content.empty()) { - choices = json::array({json{{"finish_reason", nullptr}, - {"index", 0}, - {"delta", json{{"role", "assistant"}}}}}); - } else { - // We have to send this as two updates to conform to openai behavior - json initial_ret = json{{"choices", json::array({json{ - {"finish_reason", nullptr}, - {"index", 0}, - {"delta", json{ - {"role", "assistant"} - }}}})}, - {"created", t}, - {"id", completion_id}, - {"model", modelname}, - {"object", "chat.completion.chunk"}}; - - json second_ret = json{ - {"choices", json::array({json{{"finish_reason", nullptr}, - {"index", 0}, - {"delta", json{ - {"content", content}}} - }})}, - {"created", t}, - {"id", completion_id}, - {"model", modelname}, - {"object", "chat.completion.chunk"}}; - - return std::vector({initial_ret, second_ret}); - } - } else { - // Some idiosyncrasy in task processing logic makes several trailing calls - // with empty content, we ignore these at the calee site. - if (content.empty()) { - return std::vector({json::object()}); - } - - choices = json::array({json{ - {"finish_reason", nullptr}, - {"index", 0}, - {"delta", - json{ - {"content", content}, - }}, - }}); - } - } - - json ret = json { - {"choices", choices}, - {"created", t}, - {"id", completion_id}, - {"model", modelname}, - {"object", "chat.completion.chunk"} - }; - - if (result.contains("timings")) { - ret.push_back({"timings", json_value(result, "timings", json::object())}); - } - - if (!finish_reason.empty()) { - int num_tokens_predicted = json_value(result, "tokens_predicted", 0); - int num_prompt_tokens = json_value(result, "tokens_evaluated", 0); - ret.push_back({"usage", json { - {"completion_tokens", num_tokens_predicted}, - {"prompt_tokens", num_prompt_tokens}, - {"total_tokens", num_tokens_predicted + num_prompt_tokens} - }}); - } - - return std::vector({ret}); -} - static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) { json data = json::array(); int i = 0; @@ -934,42 +709,17 @@ static json format_detokenized_response(const std::string & content) { }; } -static json format_error_response(const std::string & message, const enum error_type type) { - std::string type_str; - int code = 500; - switch (type) { - case ERROR_TYPE_INVALID_REQUEST: - type_str = "invalid_request_error"; - code = 400; - break; - case ERROR_TYPE_AUTHENTICATION: - type_str = "authentication_error"; - code = 401; - break; - case ERROR_TYPE_NOT_FOUND: - type_str = "not_found_error"; - code = 404; - break; - case ERROR_TYPE_SERVER: - type_str = "server_error"; - code = 500; - break; - case ERROR_TYPE_PERMISSION: - type_str = "permission_error"; - code = 403; - break; - case ERROR_TYPE_NOT_SUPPORTED: - type_str = "not_supported_error"; - code = 501; - break; - case ERROR_TYPE_UNAVAILABLE: - type_str = "unavailable_error"; - code = 503; - break; +static json format_logit_bias(const std::vector & logit_bias) { + json data = json::array(); + for (const auto & lb : logit_bias) { + data.push_back(json{ + {"bias", lb.bias}, + {"token", lb.token}, + }); } - return json { - {"code", code}, - {"message", message}, - {"type", type_str}, - }; + return data; +} + +static std::string safe_json_to_str(json data) { + return data.dump(-1, ' ', false, json::error_handler_t::replace); } diff --git a/ggml/include/ggml-cpu.h b/ggml/include/ggml-cpu.h index e14ea9ea5..3aa71badb 100644 --- a/ggml/include/ggml-cpu.h +++ b/ggml/include/ggml-cpu.h @@ -103,24 +103,14 @@ extern "C" { // Internal types and functions exposed for tests and benchmarks - typedef void (*ggml_from_float_to_mat_t) - (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs); typedef void (*ggml_vec_dot_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx, const void * GGML_RESTRICT y, size_t by, int nrc); - typedef void (*ggml_gemv_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, - const void * GGML_RESTRICT y, int nr, int nc); - typedef void (*ggml_gemm_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, - const void * GGML_RESTRICT y, int nr, int nc); struct ggml_type_traits_cpu { ggml_from_float_t from_float; - ggml_from_float_to_mat_t from_float_to_mat; ggml_vec_dot_t vec_dot; enum ggml_type vec_dot_type; int64_t nrows; // number of rows to process simultaneously - int64_t ncols; // number of columns to process simultaneously - ggml_gemv_t gemv; - ggml_gemm_t gemm; }; GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type); @@ -140,13 +130,6 @@ extern "C" { GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void); -#ifdef GGML_USE_CPU_HBM - GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void); -#endif - - GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void); - GGML_BACKEND_API bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft); - #ifdef __cplusplus } #endif diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 1c8cc11b6..386d5a15d 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -384,15 +384,15 @@ extern "C" { GGML_TYPE_F64 = 28, GGML_TYPE_IQ1_M = 29, GGML_TYPE_BF16 = 30, - GGML_TYPE_Q4_0_4_4 = 31, - GGML_TYPE_Q4_0_4_8 = 32, - GGML_TYPE_Q4_0_8_8 = 33, + // GGML_TYPE_Q4_0_4_4 = 31, support has been removed from gguf files + // GGML_TYPE_Q4_0_4_8 = 32, + // GGML_TYPE_Q4_0_8_8 = 33, GGML_TYPE_TQ1_0 = 34, GGML_TYPE_TQ2_0 = 35, - GGML_TYPE_IQ4_NL_4_4 = 36, + // GGML_TYPE_IQ4_NL_4_4 = 36, // GGML_TYPE_IQ4_NL_4_8 = 37, // GGML_TYPE_IQ4_NL_8_8 = 38, - GGML_TYPE_COUNT, + GGML_TYPE_COUNT = 39, }; // precision @@ -433,9 +433,6 @@ extern "C" { GGML_FTYPE_MOSTLY_IQ4_XS = 22, // except 1d tensors GGML_FTYPE_MOSTLY_IQ1_M = 23, // except 1d tensors GGML_FTYPE_MOSTLY_BF16 = 24, // except 1d tensors - GGML_FTYPE_MOSTLY_Q4_0_4_4 = 25, // except 1d tensors - GGML_FTYPE_MOSTLY_Q4_0_4_8 = 26, // except 1d tensors - GGML_FTYPE_MOSTLY_Q4_0_8_8 = 27, // except 1d tensors }; // available tensor operations: @@ -2205,11 +2202,19 @@ extern "C" { GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx); GGML_API void gguf_get_meta_data(const struct gguf_context * ctx, void * data); -#ifdef __cplusplus -// restrict not standard in C++ -#define GGML_RESTRICT +#ifdef __cplusplus + // restrict not standard in C++ +# if defined(__GNUC__) +# define GGML_RESTRICT __restrict__ +# elif defined(__clang__) +# define GGML_RESTRICT __restrict +# elif defined(_MSC_VER) +# define GGML_RESTRICT __restrict +# else +# define GGML_RESTRICT +# endif #else -#define GGML_RESTRICT restrict +# define GGML_RESTRICT restrict #endif typedef void (*ggml_to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index f07533fdb..a267a8b59 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -220,9 +220,7 @@ add_library(ggml-base ggml-threading.cpp ggml-threading.h ggml-quants.c - ggml-quants.h - ggml-aarch64.c - ggml-aarch64.h) + ggml-quants.h) target_include_directories(ggml-base PRIVATE .) diff --git a/ggml/src/ggml-aarch64.c b/ggml/src/ggml-aarch64.c deleted file mode 100644 index 013912051..000000000 --- a/ggml/src/ggml-aarch64.c +++ /dev/null @@ -1,129 +0,0 @@ -#define GGML_COMMON_DECL_C -#include "ggml-common.h" - -#include "ggml-aarch64.h" -#include "ggml-impl.h" -#include "ggml-quants.h" -#include - -#define UNUSED GGML_UNUSED - -static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) { - block_q4_0x4 out; - - for (int i = 0; i < 4; i++) { - out.d[i] = in[i].d; - } - - const int end = QK4_0 * 2 / blck_size_interleave; - - if (blck_size_interleave == 8) { - const uint64_t xor_mask = 0x8888888888888888ULL; - for (int i = 0; i < end; ++i) { - int src_id = i % 4; - int src_offset = (i / 4) * blck_size_interleave; - int dst_offset = i * blck_size_interleave; - - uint64_t elems; - // Using memcpy to avoid unaligned memory accesses - memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t)); - elems ^= xor_mask; - memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t)); - } - } else if (blck_size_interleave == 4) { - const uint32_t xor_mask = 0x88888888; - for (int i = 0; i < end; ++i) { - int src_id = i % 4; - int src_offset = (i / 4) * blck_size_interleave; - int dst_offset = i * blck_size_interleave; - - uint32_t elems; - memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t)); - elems ^= xor_mask; - memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t)); - } - } else { - GGML_ASSERT(false); - } - - return out; -} - -// interleave 8 block_q4_0s in blocks of blck_size_interleave -// returns an interleaved block_q4_0x8 -// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks -// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave -static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) { - block_q4_0x8 out; - - for (int i = 0; i < 8; i++) { - out.d[i] = in[i].d; - } - - const int end = QK4_0 * 4 / blck_size_interleave; - const uint64_t xor_mask = 0x8888888888888888ULL; - - for (int i = 0; i < end; ++i) { - int src_id = i % 8; - int src_offset = (i / 8) * blck_size_interleave; - int dst_offset = i * blck_size_interleave; - - uint64_t elems; - memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t)); - elems ^= xor_mask; - memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t)); - } - - return out; -} - -static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int blck_size_interleave) { - assert(n_per_row % QK4_0 == 0); - const int nb = n_per_row / QK4_0; - - void * out_ptr = NULL; - if (nrows_interleaved == 8) { - out_ptr = (block_q4_0x8 *) dst; - } - else if (nrows_interleaved == 4) { - out_ptr = (block_q4_0x4 *) dst; - } - assert(nrows_interleaved <= 8); - block_q4_0 dst_tmp[8]; - - for (int b = 0; b < (nrow * n_per_row); b += nrows_interleaved * n_per_row) { - - for (int64_t x = 0; x < nb; x++) { - - for (int i = 0; i < nrows_interleaved; i++ ) { - quantize_row_q4_0_ref(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0); - } - - if (nrows_interleaved == 8) { - *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave); - out_ptr = (block_q4_0x8 *) out_ptr + 1; - } - else if (nrows_interleaved == 4) { - *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave); - out_ptr = (block_q4_0x4 *) out_ptr + 1; - } - } - } - - return ((nrow * n_per_row) / QK4_0 * sizeof(block_q4_0)); -} - -size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { - UNUSED(quant_weights); - return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4); -} - -size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { - UNUSED(quant_weights); - return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8); -} - -size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { - UNUSED(quant_weights); - return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8); -} diff --git a/ggml/src/ggml-aarch64.h b/ggml/src/ggml-aarch64.h deleted file mode 100644 index a57868591..000000000 --- a/ggml/src/ggml-aarch64.h +++ /dev/null @@ -1,19 +0,0 @@ -#pragma once - -#include "ggml.h" - -// GGML internal header - -#ifdef __cplusplus -extern "C" { -#endif - -// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization") -size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); -size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); -size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); - -#ifdef __cplusplus -} -#endif - diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp index 04e25b8ab..fa04ab84f 100644 --- a/ggml/src/ggml-cann/ggml-cann.cpp +++ b/ggml/src/ggml-cann/ggml-cann.cpp @@ -2089,7 +2089,7 @@ static void * ggml_backend_cann_reg_get_proc_address(ggml_backend_reg_t reg, con static const ggml_backend_reg_i ggml_backend_cann_reg_interface = { /* .get_name = */ ggml_backend_cann_reg_get_name, /* .get_device_count = */ ggml_backend_cann_reg_get_device_count, - /* .get_device_get = */ ggml_backend_cann_reg_get_device, + /* .get_device = */ ggml_backend_cann_reg_get_device, /* .get_proc_address = */ ggml_backend_cann_reg_get_proc_address, }; diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index 27253a6c2..7fd2aadec 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -6,7 +6,20 @@ typedef uint16_t ggml_half; typedef uint32_t ggml_half2; -#define GGML_COMMON_AGGR +#define GGML_COMMON_AGGR_U +#define GGML_COMMON_AGGR_S + +#define GGML_COMMON_DECL +#elif defined(GGML_COMMON_DECL_CPP) +#include + +typedef uint16_t ggml_half; +typedef uint32_t ggml_half2; + +// std-c++ allow anonymous unions but some compiler warn on it +#define GGML_COMMON_AGGR_U data +// std-c++ do not allow it. +#define GGML_COMMON_AGGR_S data #define GGML_COMMON_DECL #elif defined(GGML_COMMON_DECL_METAL) @@ -15,7 +28,8 @@ typedef uint32_t ggml_half2; typedef half ggml_half; typedef half2 ggml_half2; -#define GGML_COMMON_AGGR +#define GGML_COMMON_AGGR_U +#define GGML_COMMON_AGGR_S #define GGML_COMMON_DECL #elif defined(GGML_COMMON_DECL_CUDA) @@ -29,7 +43,8 @@ typedef half2 ggml_half2; typedef half ggml_half; typedef half2 ggml_half2; -#define GGML_COMMON_AGGR data +#define GGML_COMMON_AGGR_U +#define GGML_COMMON_AGGR_S data #define GGML_COMMON_DECL #elif defined(GGML_COMMON_DECL_HIP) @@ -39,7 +54,8 @@ typedef half2 ggml_half2; typedef half ggml_half; typedef half2 ggml_half2; -#define GGML_COMMON_AGGR data +#define GGML_COMMON_AGGR_U +#define GGML_COMMON_AGGR_S data #define GGML_COMMON_DECL #elif defined(GGML_COMMON_DECL_SYCL) @@ -49,7 +65,8 @@ typedef half2 ggml_half2; typedef sycl::half ggml_half; typedef sycl::half2 ggml_half2; -#define GGML_COMMON_AGGR data +#define GGML_COMMON_AGGR_U +#define GGML_COMMON_AGGR_S data #define GGML_COMMON_DECL #endif @@ -154,9 +171,9 @@ typedef struct { struct { ggml_half d; // delta ggml_half m; // min - } GGML_COMMON_AGGR; + } GGML_COMMON_AGGR_S; ggml_half2 dm; - }; + } GGML_COMMON_AGGR_U; uint8_t qs[QK4_1 / 2]; // nibbles / quants } block_q4_1; static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding"); @@ -175,9 +192,9 @@ typedef struct { struct { ggml_half d; // delta ggml_half m; // min - } GGML_COMMON_AGGR; + } GGML_COMMON_AGGR_S; ggml_half2 dm; - }; + } GGML_COMMON_AGGR_U; uint8_t qh[4]; // 5-th bit of quants uint8_t qs[QK5_1 / 2]; // nibbles / quants } block_q5_1; @@ -196,37 +213,13 @@ typedef struct { struct { ggml_half d; // delta ggml_half s; // d * sum(qs[i]) - } GGML_COMMON_AGGR; + } GGML_COMMON_AGGR_S; ggml_half2 ds; - }; + } GGML_COMMON_AGGR_U; int8_t qs[QK8_1]; // quants } block_q8_1; static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding"); -typedef struct { - ggml_half d[4]; // deltas for 4 q4_0 blocks - uint8_t qs[QK4_0 * 2]; // nibbles / quants for 4 q4_0 blocks -} block_q4_0x4; -static_assert(sizeof(block_q4_0x4) == 4 * sizeof(ggml_half) + QK4_0 * 2, "wrong q4_0x4 block size/padding"); - -typedef struct { - ggml_half d[8]; // deltas for 8 q4_0 blocks - uint8_t qs[QK4_0 * 4]; // nibbles / quants for 8 q4_0 blocks -} block_q4_0x8; -static_assert(sizeof(block_q4_0x8) == 8 * sizeof(ggml_half) + QK4_0 * 4, "wrong q4_0x8 block size/padding"); - -typedef struct { - ggml_half d[4]; // deltas for 4 q8_0 blocks - int8_t qs[QK8_0 * 4]; // quants for 4 q8_0 blocks -} block_q8_0x4; -static_assert(sizeof(block_q8_0x4) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong q8_0x4 block size/padding"); - -typedef struct { - ggml_half d[8]; // deltas for 8 q8_0 blocks - int8_t qs[QK8_0 * 8]; // quants for 8 q8_0 blocks -} block_q8_0x8; -static_assert(sizeof(block_q8_0x8) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong q8_0x8 block size/padding"); - // // Ternary quantization // @@ -261,9 +254,9 @@ typedef struct { struct { ggml_half d; // super-block scale for quantized scales ggml_half dmin; // super-block scale for quantized mins - } GGML_COMMON_AGGR; + } GGML_COMMON_AGGR_S; ggml_half2 dm; - }; + } GGML_COMMON_AGGR_U; } block_q2_K; static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding"); @@ -288,9 +281,9 @@ typedef struct { struct { ggml_half d; // super-block scale for quantized scales ggml_half dmin; // super-block scale for quantized mins - } GGML_COMMON_AGGR; + } GGML_COMMON_AGGR_S; ggml_half2 dm; - }; + } GGML_COMMON_AGGR_U; uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits uint8_t qs[QK_K/2]; // 4--bit quants } block_q4_K; @@ -305,9 +298,9 @@ typedef struct { struct { ggml_half d; // super-block scale for quantized scales ggml_half dmin; // super-block scale for quantized mins - } GGML_COMMON_AGGR; + } GGML_COMMON_AGGR_S; ggml_half2 dm; - }; + } GGML_COMMON_AGGR_U; uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits uint8_t qh[QK_K/8]; // quants, high bit uint8_t qs[QK_K/2]; // quants, low 4 bits @@ -418,12 +411,6 @@ typedef struct { } block_iq4_xs; static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding"); -typedef struct { - ggml_half d[4]; // deltas for 4 iq4_nl blocks - uint8_t qs[QK4_NL * 2];// nibbles / quants for 4 iq4_nl blocks -} block_iq4_nlx4; -static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding"); - #endif // GGML_COMMON_DECL #endif // GGML_COMMON_DECL @@ -437,6 +424,13 @@ static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wro #define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = { #define GGML_TABLE_END() }; +#define GGML_COMMON_IMPL +#elif defined(GGML_COMMON_IMPL_CPP) +#include + +#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = { +#define GGML_TABLE_END() }; + #define GGML_COMMON_IMPL #elif defined(GGML_COMMON_IMPL_METAL) #include diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index bc326c059..0e0556703 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -10,10 +10,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name) list (APPEND GGML_CPU_SOURCES ggml-cpu/ggml-cpu.c ggml-cpu/ggml-cpu.cpp - ggml-cpu/ggml-cpu-aarch64.c + ggml-cpu/ggml-cpu-aarch64.cpp ggml-cpu/ggml-cpu-aarch64.h + ggml-cpu/ggml-cpu-hbm.cpp + ggml-cpu/ggml-cpu-hbm.h ggml-cpu/ggml-cpu-quants.c ggml-cpu/ggml-cpu-quants.h + ggml-cpu/ggml-cpu-traits.cpp + ggml-cpu/ggml-cpu-traits.h ggml-cpu/amx/amx.cpp ggml-cpu/amx/amx.h ggml-cpu/amx/mmq.cpp diff --git a/ggml/src/ggml-cpu/amx/amx.cpp b/ggml/src/ggml-cpu/amx/amx.cpp index 09c0df0f5..b9074cb3a 100644 --- a/ggml/src/ggml-cpu/amx/amx.cpp +++ b/ggml/src/ggml-cpu/amx/amx.cpp @@ -5,6 +5,7 @@ #include "ggml-backend.h" #include "ggml-impl.h" #include "ggml-cpu.h" +#include "ggml-cpu-traits.h" #if defined(__gnu_linux__) #include @@ -17,31 +18,65 @@ #if defined(__AMX_INT8__) && defined(__AVX512VNNI__) +// AMX type_trais +namespace ggml::cpu::amx { +class tensor_traits : public ggml::cpu::tensor_traits { + bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override { + size = ggml_backend_amx_desired_wsize(op); + return true; + } + + bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override { + if (op->op == GGML_OP_MUL_MAT) { + ggml_backend_amx_mul_mat(params, op); + return true; + } + return false; + } +}; + +static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struct ggml_tensor *) { + static tensor_traits traits; + return &traits; +} +} // namespace ggml::cpu::amx + // AMX buffer interface static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) { free(buffer->context); } static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) { - return (void *)(buffer->context); + return (void *) (buffer->context); } -static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { - memset((char *)tensor->data + offset, value, size); +static void ggml_backend_amx_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { + tensor->extra = (void *) ggml::cpu::amx::get_tensor_traits(buffer, tensor); GGML_UNUSED(buffer); } -static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { +static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, + uint8_t value, size_t offset, size_t size) { + memset((char *) tensor->data + offset, value, size); + + GGML_UNUSED(buffer); +} + +static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, + const void * data, size_t offset, size_t size) { if (qtype_has_amx_kernels(tensor->type)) { + GGML_LOG_DEBUG("%s: amx repack tensor %s of type %s\n", __func__, tensor->name, ggml_type_name(tensor->type)); ggml_backend_amx_convert_weight(tensor, data, offset, size); } else { - memcpy((char *)tensor->data + offset, data, size); + memcpy((char *) tensor->data + offset, data, size); } GGML_UNUSED(buffer); } +/* +// need to figure what we need to do with buffer->extra. static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { GGML_ASSERT(!qtype_has_amx_kernels(tensor->type)); memcpy(data, (const char *)tensor->data + offset, size); @@ -62,6 +97,7 @@ static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con GGML_UNUSED(buffer); } +*/ static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { memset(buffer->context, value, buffer->size); @@ -70,13 +106,13 @@ static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = { /* .free_buffer = */ ggml_backend_amx_buffer_free_buffer, /* .get_base = */ ggml_backend_amx_buffer_get_base, - /* .init_tensor = */ NULL, // no initialization required + /* .init_tensor = */ ggml_backend_amx_buffer_init_tensor, /* .memset_tensor = */ ggml_backend_amx_buffer_memset_tensor, /* .set_tensor = */ ggml_backend_amx_buffer_set_tensor, - /* .get_tensor = */ ggml_backend_amx_buffer_get_tensor, - /* .cpy_tensor = */ ggml_backend_amx_buffer_cpy_tensor, + /* .get_tensor = */ nullptr, + /* .cpy_tensor = */ nullptr, /* .clear = */ ggml_backend_amx_buffer_clear, - /* .reset = */ NULL, + /* .reset = */ nullptr, }; static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) { @@ -101,18 +137,48 @@ static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_typ GGML_UNUSED(buft); } -static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) { +namespace ggml::cpu::amx { +class extra_buffer_type : ggml::cpu::extra_buffer_type { + bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override { + // handle only 2d gemm for now + auto is_contiguous_2d = [](const struct ggml_tensor * t) { + return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1; + }; + + if (op->op == GGML_OP_MUL_MAT && is_contiguous_2d(op->src[0]) && // src0 must be contiguous + is_contiguous_2d(op->src[1]) && // src1 must be contiguous + op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_amx_buffer_type() && + op->ne[0] % (TILE_N * 2) == 0 && // out_features is 32x + (qtype_has_amx_kernels(op->src[0]->type) || (op->src[0]->type == GGML_TYPE_F16))) { + // src1 must be host buffer + if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) { + return false; + } + // src1 must be float32 + if (op->src[1]->type == GGML_TYPE_F32) { + return true; + } + } + return false; + } + + ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override { + if (op->op == GGML_OP_MUL_MAT && op->src[0]->buffer && + op->src[0]->buffer->buft == ggml_backend_amx_buffer_type()) { + return (ggml::cpu::tensor_traits *) op->src[0]->extra; + } + + return nullptr; + } +}; +} // namespace ggml::cpu::amx + +static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { return ggml_backend_amx_get_alloc_size(tensor); GGML_UNUSED(buft); } -static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) { - return false; - - GGML_UNUSED(buft); -} - #define ARCH_GET_XCOMP_PERM 0x1022 #define ARCH_REQ_XCOMP_PERM 0x1023 #define XFEATURE_XTILECFG 17 @@ -129,68 +195,26 @@ static bool ggml_amx_init() { return true; #endif } + ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() { static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = { /* .iface = */ { - /* .get_name = */ ggml_backend_amx_buffer_type_get_name, - /* .alloc_buffer = */ ggml_backend_amx_buffer_type_alloc_buffer, - /* .get_alignment = */ ggml_backend_amx_buffer_type_get_alignment, - /* .get_max_size = */ NULL, // defaults to SIZE_MAX - /* .get_alloc_size = */ ggml_backend_amx_buffer_type_get_alloc_size, - /* .is_host = */ ggml_backend_amx_buffer_type_is_host, - }, + /* .get_name = */ ggml_backend_amx_buffer_type_get_name, + /* .alloc_buffer = */ ggml_backend_amx_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_amx_buffer_type_get_alignment, + /* .get_max_size = */ nullptr, // defaults to SIZE_MAX + /* .get_alloc_size = */ ggml_backend_amx_buffer_type_get_alloc_size, + /* .is_host = */ nullptr, + }, /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0), - /* .context = */ NULL, + /* .context = */ new ggml::cpu::amx::extra_buffer_type(), }; if (!ggml_amx_init()) { - return NULL; + return nullptr; } return &ggml_backend_buffer_type_amx; } -bool ggml_backend_amx_buft_is_amx(ggml_backend_buffer_type_t buft) { - return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name; -} - -bool ggml_backend_amx_device_supports_op(const struct ggml_tensor * op) { - // handle only 2d gemm for now - auto is_contiguous_2d = [](const struct ggml_tensor * t) { - return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1; - }; - - switch (op->op) { - case GGML_OP_NONE: - case GGML_OP_RESHAPE: - case GGML_OP_VIEW: - case GGML_OP_PERMUTE: - case GGML_OP_TRANSPOSE: - return true; - - case GGML_OP_MUL_MAT: { - const struct ggml_tensor * src0 = op->src[0]; - const struct ggml_tensor * src1 = op->src[1]; - - const enum ggml_type type = src0->type; - const int64_t ne0 = op->ne[0]; - - // amx kernels enables for Q4_0, Q4_1, Q8_0, F16 - // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256 - bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16); - - bool can_use_amx = - is_contiguous_2d(src0) && // src0 must be contiguous - is_contiguous_2d(src1) && // src1 must be contiguous - src1->type == GGML_TYPE_F32 && // src1 must be float32 - has_amx_kernels && // with amx kernel impls - ne0 % (TILE_N * 2) == 0; // out_features is 32x - - return can_use_amx; - } - default: - return false; - } -} - -#endif // defined(__AMX_INT8__) && defined(__AVX512VNNI__) +#endif // defined(__AMX_INT8__) && defined(__AVX512VNNI__) diff --git a/ggml/src/ggml-cpu/amx/amx.h b/ggml/src/ggml-cpu/amx/amx.h index c43546273..5b65d76bd 100644 --- a/ggml/src/ggml-cpu/amx/amx.h +++ b/ggml/src/ggml-cpu/amx/amx.h @@ -1,20 +1,8 @@ #include "ggml-backend.h" #include "ggml-cpu-impl.h" -#ifdef __cplusplus -extern "C" { -#endif +// GGML internal header #if defined(__AMX_INT8__) && defined(__AVX512VNNI__) - ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void); -bool ggml_backend_amx_buft_is_amx(ggml_backend_buffer_type_t buft); -bool ggml_backend_amx_device_supports_op(const struct ggml_tensor * op); -void ggml_backend_amx_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst); -size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst); - -#endif - -#ifdef __cplusplus -} #endif diff --git a/ggml/src/ggml-cpu/amx/common.h b/ggml/src/ggml-cpu/amx/common.h index 40074c3fc..f392e8985 100644 --- a/ggml/src/ggml-cpu/amx/common.h +++ b/ggml/src/ggml-cpu/amx/common.h @@ -7,7 +7,7 @@ #include #include -#if defined(_OPENMP) +#if defined(GGML_USE_OPENMP) #include #endif @@ -56,11 +56,11 @@ inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) { } template -inline void parallel_for(int nth, int n, const func_t& f) { -#if defined(_OPENMP) -#pragma omp parallel num_threads(nth) +inline void parallel_for(int n, const func_t& f) { +#if defined(GGML_USE_OPENMP) +#pragma omp parallel { - //int nth = omp_get_num_threads(); + int nth = omp_get_num_threads(); int ith = omp_get_thread_num(); int tbegin, tend; balance211(n, nth, ith, tbegin, tend); @@ -68,8 +68,6 @@ inline void parallel_for(int nth, int n, const func_t& f) { } #else f(0, n); - - GGML_UNUSED(nth); #endif } @@ -91,10 +89,3 @@ inline bool qtype_has_amx_kernels(const enum ggml_type type) { (type == GGML_TYPE_Q6_K) || (type == GGML_TYPE_IQ4_XS); } - -// ggml backend context -struct ggml_backend_amx_context { - int n_threads = GGML_DEFAULT_N_THREADS; - std::unique_ptr work_data; - size_t work_size = 0; -}; diff --git a/ggml/src/ggml-cpu/amx/mmq.cpp b/ggml/src/ggml-cpu/amx/mmq.cpp index 0ec3aa86d..0ea91596b 100644 --- a/ggml/src/ggml-cpu/amx/mmq.cpp +++ b/ggml/src/ggml-cpu/amx/mmq.cpp @@ -18,10 +18,6 @@ #include #endif -#if defined(_OPENMP) -#include -#endif - #if (defined(_WIN32) || defined(_WIN64)) #define RESTRICT __restrict #else @@ -1382,13 +1378,13 @@ struct tinygemm_kernel_avx #define PACKED_INDEX(n, k, KB, tile_size) (n * KB + k) * tile_size template -void convert_B_packed_format(void * RESTRICT packed_B, const TB * RESTRICT B, int N, int K, int n_threads) { +void convert_B_packed_format(void * RESTRICT packed_B, const TB * RESTRICT B, int N, int K) { const int NB = N / TILE_N; const int KB = K / BLOCK_K; const int TILE_SIZE = get_tile_size(); // parallel on NB should be enough - parallel_for(n_threads, NB, [&](int begin, int end) { + parallel_for(NB, [&](int begin, int end) { for (int n = begin; n < end; ++n) { for (int k = 0; k < KB; ++k) { int n0 = n * TILE_N; @@ -2334,15 +2330,8 @@ void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * d const int K = tensor->ne[0]; // ne0: in_features const int N = tensor->ne[1]; // ne1: out_features -#if defined(_OPENMP) - // the buffer ctx is not initialized when .set_tensor is called - int n_threads = omp_get_num_threads(); -#else - int n_threads = 1; -#endif - GGML_DISPATCH_QTYPES(TYPE, [&] { - convert_B_packed_format((void *)((char *)tensor->data + offset), (const type *)data, N, K, n_threads); + convert_B_packed_format((void *)((char *)tensor->data + offset), (const type *)data, N, K); }); } diff --git a/ggml/src/ggml-cpu/amx/mmq.h b/ggml/src/ggml-cpu/amx/mmq.h index f37366093..baf768477 100644 --- a/ggml/src/ggml-cpu/amx/mmq.h +++ b/ggml/src/ggml-cpu/amx/mmq.h @@ -1,16 +1,10 @@ #pragma once #include "common.h" -#ifdef __cplusplus -extern "C" { -#endif +size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst); size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor); void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); void ggml_backend_amx_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst); - -#ifdef __cplusplus -} -#endif diff --git a/ggml/src/ggml-cpu/ggml-cpu-aarch64.c b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp similarity index 85% rename from ggml/src/ggml-cpu/ggml-cpu-aarch64.c rename to ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp index 11152385e..a51d1a6c5 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-aarch64.c +++ b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp @@ -1,20 +1,57 @@ -#define GGML_COMMON_IMPL_C +#define GGML_COMMON_IMPL_CPP +#define GGML_COMMON_DECL_CPP #include "ggml-common.h" +#include "ggml-backend-impl.h" #include "ggml-quants.h" #include "ggml-impl.h" #include "ggml-cpu.h" -#include "ggml-cpu/ggml-cpu-impl.h" +#include "ggml-cpu-impl.h" +#include "ggml-cpu-traits.h" -#include -#include -#include -#include -#include // for qsort -#include // for GGML_ASSERT +#include +#include +#include +#include +#include // for qsort +#include // for GGML_ASSERT #include "ggml-cpu-aarch64.h" +// TODO: move to include file? +template constexpr int QK_0() { + if constexpr (K == 4) { + return QK4_0; + } + if constexpr (K == 8) { + return QK8_0; + } + return -1; +} + +template struct block { + ggml_half d[N]; // deltas for N qK_0 blocks + int8_t qs[(QK_0() * N * K) / 8]; // quants for N qK_0 blocks +}; + +// control size +static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding"); +static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding"); +static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding"); +static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding"); + +using block_q4_0x4 = block<4, 4>; +using block_q4_0x8 = block<4, 8>; +using block_q8_0x4 = block<8, 4>; +using block_q8_0x8 = block<8, 8>; + +struct block_iq4_nlx4 { + ggml_half d[4]; // deltas for 4 iq4_nl blocks + uint8_t qs[QK4_NL * 2]; // nibbles / quants for 4 iq4_nl blocks +}; + +static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding"); + #if defined(__GNUC__) #pragma GCC diagnostic ignored "-Woverlength-strings" #elif defined(_MSC_VER) @@ -185,12 +222,12 @@ static inline __m256i mul_sum_i8_pairs_int32x8(const __m256i x, const __m256i y) static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; -static void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k) { +static void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(QK8_0 == 32); assert(k % QK8_0 == 0); const int nb = k / QK8_0; - block_q8_0x4 * restrict y = (block_q8_0x4 *) vy; + block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy; #if defined(__ARM_NEON) float32x4_t srcv[4][8]; @@ -279,12 +316,12 @@ static void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int6 #endif } -static void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k) { +static void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(QK8_0 == 32); assert(k % QK8_0 == 0); const int nb = k / QK8_0; - block_q8_0x4 * restrict y = (block_q8_0x4 *) vy; + block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy; #if defined(__ARM_NEON) float32x4_t srcv[4][8]; @@ -494,7 +531,7 @@ static void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int6 #endif } -void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) { +static void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) { assert(nrow == 4); UNUSED(nrow); if (blck_size_interleave == 4) { @@ -506,7 +543,7 @@ void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nro } } -void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { +static void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; const int ncols_interleaved = 4; @@ -591,7 +628,7 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * } } -void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { +static void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; const int ncols_interleaved = 4; @@ -701,7 +738,7 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * } } -void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { +static void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; const int ncols_interleaved = 8; @@ -974,7 +1011,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * } } -void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { +static void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; const int ncols_interleaved = 4; @@ -1070,7 +1107,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * restrict s, size_t bs, const void } } -void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { +static void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; const int ncols_interleaved = 4; @@ -1586,7 +1623,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * } } -void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { +static void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; const int ncols_interleaved = 4; @@ -2040,7 +2077,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * } } -void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { +static void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; const int ncols_interleaved = 8; @@ -2560,31 +2597,31 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31) // Shuffle pattern one - right side input - const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3) - const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3) + const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3) + const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3) - const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11) - const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11) + const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11) + const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11) - const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19) - const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19) + const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19) + const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19) - const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27) - const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27) + const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27) + const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27) // Shuffle pattern two - right side input - const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7) - const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7) + const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7) + const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7) - const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15) - const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15) + const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15) + const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15) - const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23) - const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23) + const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23) + const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23) - const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31) - const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31) + const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31) + const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31) // Scale values - Load the weight scale values of two block_q4_0x8 const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); @@ -2618,31 +2655,31 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * // Shuffle pattern one - left side input - const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) - const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) + const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) + const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) - const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) - const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) + const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) + const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) - const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) - const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) + const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) + const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) - const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) - const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) + const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) + const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) // Shuffle pattern two - left side input - const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) - const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) + const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) + const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) - const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) - const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) + const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) + const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) - const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) - const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) + const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) + const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) - const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) - const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) + const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) + const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane // Resembles MMLAs into 2x2 matrices in ARM Version @@ -2671,10 +2708,10 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * // Straighten out to make 4 row vectors - __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, 78)); - __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01); - __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, 78)); - __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11); + __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78)); + __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01); + __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78)); + __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11); // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68); @@ -2753,31 +2790,31 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31) // Shuffle pattern one - right side input - const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3) - const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3) + const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3) + const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3) - const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11) - const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11) + const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11) + const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11) - const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19) - const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19) + const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19) + const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19) - const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27) - const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27) + const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27) + const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27) // Shuffle pattern two - right side input - const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7) - const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7) + const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7) + const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7) - const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15) - const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15) + const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15) + const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15) - const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23) - const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23) + const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23) + const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23) - const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31) - const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31) + const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31) + const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31) // Scale values - Load the weight scale values of two block_q4_0x8 @@ -2809,31 +2846,31 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * // Shuffle pattern one - left side input - const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) - const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) + const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) + const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) - const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) - const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) + const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) + const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) - const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) - const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) + const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) + const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) - const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) - const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) + const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) + const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) // Shuffle pattern two - left side input - const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) - const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) + const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) + const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) - const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) - const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) + const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) + const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) - const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) - const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) + const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) + const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) - const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) - const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) + const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) + const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane // Resembles MMLAs into 2x2 matrices in ARM Version @@ -2862,10 +2899,10 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * // Straighten out to make 4 row vectors - __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, 78)); - __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01); - __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, 78)); - __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11); + __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78)); + __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01); + __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78)); + __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11); // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68); @@ -3460,7 +3497,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * } } -void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { +static void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; const int ncols_interleaved = 4; @@ -3571,7 +3608,6 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * restrict s, size_t bs, const void } } -// FIXME: this code is duplicated from ggml-aarch64.c static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) { block_q4_0x4 out; @@ -3641,20 +3677,20 @@ static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_in return out; } -static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * restrict data, size_t data_size) { +static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { GGML_ASSERT(t->type == GGML_TYPE_Q4_0); GGML_ASSERT(interleave_block == 4 || interleave_block == 8); + constexpr int nrows_interleaved = 4; block_q4_0x4 * dst = (block_q4_0x4 *)t->data; const block_q4_0 * src = (const block_q4_0 *)data; block_q4_0 dst_tmp[4]; - int nrow = t->ne[1]; // Number of rows - int nrows_interleaved = 4; + int nrow = ggml_nrows(t); int nblocks = t->ne[0] / QK4_0; GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0)); - if (nrow % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { + if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { return -1; } @@ -3672,20 +3708,20 @@ static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block GGML_UNUSED(data_size); } -static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor *t, int interleave_block, const void * restrict data, size_t data_size) { +static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { GGML_ASSERT(t->type == GGML_TYPE_Q4_0); GGML_ASSERT(interleave_block == 8); + constexpr int nrows_interleaved = 8; block_q4_0x8 * dst = (block_q4_0x8*)t->data; const block_q4_0 * src = (const block_q4_0*) data; block_q4_0 dst_tmp[8]; - int nrow = t->ne[1]; // Number of rows - int nrows_interleaved = 8; + int nrow = ggml_nrows(t); int nblocks = t->ne[0] / QK4_0; GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0)); - if (nrow % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { + if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { return -1; } @@ -3712,16 +3748,18 @@ static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_s const int end = QK4_NL * 2 / blck_size_interleave; - if (blck_size_interleave == 8) { - for (int i = 0; i < end; ++i) { - int src_id = i % 4; - int src_offset = (i / 4) * blck_size_interleave; - int dst_offset = i * blck_size_interleave; + // TODO: this branch seems wrong + //if (blck_size_interleave == 8) { + // for (int i = 0; i < end; ++i) { + // int src_id = i % 4; + // int src_offset = (i / 4) * blck_size_interleave; + // int dst_offset = i * blck_size_interleave; - // Using memcpy to avoid unaligned memory accesses - memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t)); - } - } else if (blck_size_interleave == 4) { + // // Using memcpy to avoid unaligned memory accesses + // memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t)); + // } + //} else + if (blck_size_interleave == 4) { for (int i = 0; i < end; ++i) { int src_id = i % 4; int src_offset = (i / 4) * blck_size_interleave; @@ -3736,20 +3774,21 @@ static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_s return out; } -static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * restrict data, size_t data_size) { +static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL); - GGML_ASSERT(interleave_block == 4 || interleave_block == 8); + //GGML_ASSERT(interleave_block == 4 || interleave_block == 8); + GGML_ASSERT(interleave_block == 4); block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data; const block_iq4_nl * src = (const block_iq4_nl *)data; block_iq4_nl dst_tmp[4]; - int nrow = t->ne[1]; // Number of rows + int nrow = ggml_nrows(t); int nrows_interleaved = 4; int nblocks = t->ne[0] / QK4_0; GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl)); - if (nrow % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { + if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { return -1; } @@ -3767,57 +3806,457 @@ static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_b GGML_UNUSED(data_size); } -// Prepare for optimized kernels if applicable -void ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * restrict data, size_t data_size) { - if (cur->type == repack_type) { - memcpy(cur->data, data, data_size); - return; - } +namespace ggml::cpu::aarch64 { +// repack +template +int repack(struct ggml_tensor *, const void *, size_t); - if (cur->type == GGML_TYPE_Q4_0) { - switch (repack_type) { - case GGML_TYPE_Q4_0_8_8: - repack_q4_0_to_q4_0_8_bl(cur, 8, data, data_size); - break; - case GGML_TYPE_Q4_0_4_8: - repack_q4_0_to_q4_0_4_bl(cur, 8, data, data_size); - break; - case GGML_TYPE_Q4_0_4_4: - repack_q4_0_to_q4_0_4_bl(cur, 4, data, data_size); - break; - default: - GGML_ABORT("Unsupported type"); - } - } else if (cur->type == GGML_TYPE_IQ4_NL) { - switch (repack_type) { - case GGML_TYPE_IQ4_NL_4_4: - repack_iq4_nl_to_iq4_nl_4_bl(cur, 4, data, data_size); - break; - default: - GGML_ABORT("Unsupported type"); - } - } else { - GGML_ABORT("Unsupported type"); - } +// TODO: generalise. +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_q4_0_to_q4_0_4_bl(t, 4, data, data_size); } -enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur) { +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_q4_0_to_q4_0_4_bl(t, 8, data, data_size); +} + +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_q4_0_to_q4_0_8_bl(t, 8, data, data_size); +} + +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size); +} + +// TODO: needs to be revisited +//template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { +// return repack_iq4_nl_to_iq4_nl_4_bl(t, 8, data, data_size); +//} + +// gemv +template +void gemv(int, float *, size_t, const void *, const void *, int, int); + +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_q4_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> +void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); +} + +// gemm +template +void gemm(int, float *, size_t, const void *, const void *, int, int); + +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_q4_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc); +} + +template <> +void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); +} + +class tensor_traits_base : public ggml::cpu::tensor_traits { + public: + virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0; +}; + +template class tensor_traits : public tensor_traits_base { + + bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override { + // not realy a GGML_TYPE_Q8_0 but same size. + switch (op->op) { + case GGML_OP_MUL_MAT: + size = ggml_row_size(GGML_TYPE_Q8_0, ggml_nelements(op->src[1])); + return true; + case GGML_OP_MUL_MAT_ID: + size = ggml_row_size(GGML_TYPE_Q8_0, ggml_nelements(op->src[1])); + size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc. + size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2]; + return true; + default: + // GGML_ABORT("fatal error"); + break; + } + return false; + } + + bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override { + switch (op->op) { + case GGML_OP_MUL_MAT: + forward_mul_mat(params, op); + return true; + case GGML_OP_MUL_MAT_ID: + forward_mul_mat_id(params, op); + return true; + default: + // GGML_ABORT("fatal error"); + break; + } + return false; + } + + void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) { + const ggml_tensor * src0 = op->src[0]; + const ggml_tensor * src1 = op->src[1]; + ggml_tensor * dst = op; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + GGML_ASSERT(ne0 == ne01); + GGML_ASSERT(ne1 == ne11); + GGML_ASSERT(ne2 == ne12); + GGML_ASSERT(ne3 == ne13); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + GGML_ASSERT(ggml_n_dims(op->src[0]) == 2); + // GGML_ASSERT(ggml_n_dims(op->src[1]) == 2); + + char * wdata = static_cast(params->wdata); + const size_t nbw1 = ggml_row_size(GGML_TYPE_Q8_0, ne10); + + assert(params->wsize >= nbw1 * ne11); + + const ggml_from_float_t from_float = ggml_get_type_traits_cpu(GGML_TYPE_Q8_0)->from_float; + + int64_t i11_processed = 0; + for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) { + quantize_mat_q8_0((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), 4, ne10, + INTER_SIZE); + } + i11_processed = ne11 - ne11 % 4; + for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) { + from_float((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), ne10); + } + + ggml_barrier(params->threadpool); + + const void * src1_wdata = params->wdata; + const size_t src1_col_stride = ggml_row_size(GGML_TYPE_Q8_0, ne10); + int64_t src0_start = (ith * ne01) / nth; + int64_t src0_end = ((ith + 1) * ne01) / nth; + src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start; + src0_end = (src0_end % NB_COLS) ? src0_end + NB_COLS - (src0_end % NB_COLS) : src0_end; + if (src0_start >= src0_end) { + return; + } + + // If there are more than three rows in src1, use gemm; otherwise, use gemv. + if (ne11 > 3) { + gemm(ne00, (float *) ((char *) dst->data) + src0_start, ne01, + (const char *) src0->data + src0_start * nb01, + (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start); + } + for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) { + gemv(ne00, (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01, + (const char *) src0->data + src0_start * nb01, + (const char *) src1_wdata + (src1_col_stride * iter), 1, + src0_end - src0_start); + } + } + + void forward_mul_mat_id(ggml_compute_params * params, ggml_tensor * op) { + const ggml_tensor * src0 = op->src[0]; + const ggml_tensor * src1 = op->src[1]; + const ggml_tensor * ids = op->src[2]; + ggml_tensor * dst = op; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + const ggml_from_float_t from_float = ggml_get_type_traits_cpu(GGML_TYPE_Q8_0)->from_float; + + // we don't support permuted src0 or src1 + GGML_ASSERT(nb00 == ggml_type_size(src0->type)); + GGML_ASSERT(nb10 == ggml_type_size(src1->type)); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + GGML_ASSERT(ne03 == 1); + GGML_ASSERT(ne13 == 1); + GGML_ASSERT(ne3 == 1); + + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + // row groups + const int n_ids = ids->ne[0]; // n_expert_used + const int n_as = ne02; // n_expert + + const size_t nbw1 = ggml_row_size(GGML_TYPE_Q8_0, ne10); + const size_t nbw2 = nbw1*ne11; + const size_t nbw3 = nbw2*ne12; + + struct mmid_row_mapping { + int32_t i1; + int32_t i2; + }; + + GGML_ASSERT(params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as * sizeof(int64_t) + + n_as * ne12 * sizeof(mmid_row_mapping))); + + auto wdata = (char *) params->wdata; + auto wdata_src1_end = (char *) wdata + GGML_PAD(nbw3, sizeof(int64_t)); + int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as] + struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12] + + // src1: float32 => block_q8_0 + for (int64_t i12 = 0; i12 < ne12; ++i12) { + for (int64_t i11 = ith; i11 < ne11; i11 += nth) { + from_float((float *)((char *) src1->data + i12 * nb12 + i11 * nb11), + (void *) (wdata + i12 * nbw2 + i11 * nbw1), + ne10); + } + } + +#define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id) * ne12 + (i1)] + + if (ith == 0) { + // initialize matrix_row_counts + memset(matrix_row_counts, 0, n_as * sizeof(int64_t)); + + // group rows by src0 matrix + for (int32_t iid1 = 0; iid1 < ids->ne[1]; ++iid1) { + for (int32_t id = 0; id < n_ids; ++id) { + const int32_t i02 = + *(const int32_t *) ((const char *) ids->data + iid1 * ids->nb[1] + id * ids->nb[0]); + + GGML_ASSERT(i02 >= 0 && i02 < n_as); + + MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = { id, iid1 }; + matrix_row_counts[i02] += 1; + } + } + } + + ggml_barrier(params->threadpool); + + // compute each matrix multiplication in sequence + for (int cur_a = 0; cur_a < n_as; ++cur_a) { + const int64_t cne1 = matrix_row_counts[cur_a]; + + if (cne1 == 0) { + continue; + } + + auto src0_cur = (const char *) src0->data + cur_a*nb02; + + //const int64_t nr0 = ne01; // src0 rows + const int64_t nr1 = cne1; // src1 rows + + int64_t src0_cur_start = (ith * ne01) / nth; + int64_t src0_cur_end = ((ith + 1) * ne01) / nth; + src0_cur_start = + (src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start; + src0_cur_end = (src0_cur_end % NB_COLS) ? src0_cur_end + NB_COLS - (src0_cur_end % NB_COLS) : src0_cur_end; + + if (src0_cur_start >= src0_cur_end) return; + + for (int ir1 = 0; ir1 < nr1; ir1++) { + struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1); + const int id = row_mapping.i1; // selected expert index + + const int64_t i11 = id % ne11; + const int64_t i12 = row_mapping.i2; // row index in src1 + + const int64_t i1 = id; // selected expert index + const int64_t i2 = i12; // row + + auto src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2); + + gemv( + ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, + ne01, src0_cur + src0_cur_start * nb01, + src1_col, 1, src0_cur_end - src0_cur_start); + } + } +#undef MMID_MATRIX_ROW + } + + int repack(struct ggml_tensor * t, const void * data, size_t data_size) override { + GGML_LOG_DEBUG("%s: repack tensor %s with %s_%dx%d\n", __func__, t->name, ggml_type_name(t->type), + (int) NB_COLS, (int) INTER_SIZE); + return ggml::cpu::aarch64::repack(t, data, data_size); + } +}; + +// instance for Q4 +static const tensor_traits q4_0_4x4_q8_0; +static const tensor_traits q4_0_4x8_q8_0; +static const tensor_traits q4_0_8x8_q8_0; + +// instance for IQ4 +static const tensor_traits iq4_nl_4x4_q8_0; + +} // namespace ggml::cpu::aarch64 + +static const ggml::cpu::tensor_traits * ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur) { if (cur->type == GGML_TYPE_Q4_0) { - // TODO: enable for AVX2 - currently disabled due to bad gemv performance - if (/* ggml_cpu_has_avx2() || */ (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) { - return GGML_TYPE_Q4_0_8_8; + if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) { + if (cur->ne[1] % 8 == 0) { + return &ggml::cpu::aarch64::q4_0_8x8_q8_0; + } } if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { - return GGML_TYPE_Q4_0_4_8; + if (cur->ne[1] % 4 == 0) { + return &ggml::cpu::aarch64::q4_0_4x8_q8_0; + } } if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { - return GGML_TYPE_Q4_0_4_4; + if (cur->ne[1] % 4 == 0) { + return &ggml::cpu::aarch64::q4_0_4x4_q8_0; + } } } else if (cur->type == GGML_TYPE_IQ4_NL) { if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { - return GGML_TYPE_IQ4_NL_4_4; + if (cur->ne[1] % 4 == 0) { + return &ggml::cpu::aarch64::iq4_nl_4x4_q8_0; + } } } - return cur->type; + return nullptr; +} + +static void ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { + tensor->extra = (void *) const_cast(ggml_aarch64_get_optimal_repack_type(tensor)); + + GGML_UNUSED(buffer); +} + +static void ggml_backend_cpu_aarch64_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, + const void * data, size_t offset, size_t size) { + GGML_ASSERT(offset == 0); + GGML_ASSERT(size == ggml_nbytes(tensor)); + + auto tensor_traits = (ggml::cpu::aarch64::tensor_traits_base *) tensor->extra; + auto OK = tensor_traits->repack(tensor, data, size); + + GGML_ASSERT(OK == 0); + GGML_UNUSED(buffer); +} + +static const char * ggml_backend_cpu_aarch64_buffer_type_get_name(ggml_backend_buffer_type_t buft) { + return "CPU_AARCH64"; + + GGML_UNUSED(buft); +} + +static ggml_backend_buffer_t ggml_backend_cpu_aarch64_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size); + + if (buffer == nullptr) { + return nullptr; + } + + buffer->buft = buft; + buffer->iface.init_tensor = ggml_backend_cpu_aarch64_buffer_init_tensor; + buffer->iface.set_tensor = ggml_backend_cpu_aarch64_buffer_set_tensor; + return buffer; +} + +static size_t ggml_backend_cpu_aarch64_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { + return TENSOR_ALIGNMENT; + + GGML_UNUSED(buft); +} + +namespace ggml::cpu::aarch64 { +class extra_buffer_type : ggml::cpu::extra_buffer_type { + bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override { + if ( op->op == GGML_OP_MUL_MAT && + op->src[0]->buffer && + (ggml_n_dims(op->src[0]) == 2) && + op->src[0]->buffer->buft == ggml_backend_cpu_aarch64_buffer_type() && + ggml_aarch64_get_optimal_repack_type(op->src[0]) + ) { + if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) { + return false; + } + if (op->src[1]->type == GGML_TYPE_F32) { + return true; + } + //if (op->src[1]->type == GGML_TYPE_Q8_0) { + // return true; + //} + // may be possible if Q8_0 packed... + } else if (op->op == GGML_OP_MUL_MAT_ID + && op->src[0]->buffer + && (ggml_n_dims(op->src[0]) == 3) + && op->src[0]->buffer->buft == ggml_backend_cpu_aarch64_buffer_type() + && ggml_aarch64_get_optimal_repack_type(op->src[0]) + ) { + if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) { + return false; + } + if (op->src[1]->type == GGML_TYPE_F32) { + return true; + } + //if (op->src[1]->type == GGML_TYPE_Q8_0) { + // return true; + //} + } + return false; + } + + ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override { + if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_MUL_MAT_ID) { + if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_aarch64_buffer_type()) { + return (ggml::cpu::tensor_traits *) op->src[0]->extra; + } + } + return nullptr; + } +}; +} // namespace ggml::cpu::aarch64 + +ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void) { + static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_aarch64 = { + /* .iface = */ { + /* .get_name = */ ggml_backend_cpu_aarch64_buffer_type_get_name, + /* .alloc_buffer = */ ggml_backend_cpu_aarch64_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_cpu_aarch64_buffer_type_get_alignment, + /* .get_max_size = */ nullptr, // defaults to SIZE_MAX + /* .get_alloc_size = */ nullptr, // defaults to ggml_nbytes + /* .is_host = */ nullptr, + }, + /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0), + /* .context = */ new ggml::cpu::aarch64::extra_buffer_type(), + }; + + return &ggml_backend_cpu_buffer_type_aarch64; } diff --git a/ggml/src/ggml-cpu/ggml-cpu-aarch64.h b/ggml/src/ggml-cpu/ggml-cpu-aarch64.h index 3d9db6a19..6e84c826b 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-aarch64.h +++ b/ggml/src/ggml-cpu/ggml-cpu-aarch64.h @@ -1,32 +1,8 @@ #pragma once +#include "ggml-cpu-traits.h" #include "ggml.h" // GGML internal header -#ifdef __cplusplus -extern "C" { -#endif - -// Quantization -void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave); - -// GEMV -void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); -void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); -void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); -void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); - -// GEMM -void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); -void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); -void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); -void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); - -void ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * data, size_t data_size); -enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur); - -#ifdef __cplusplus -} -#endif - +ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void); diff --git a/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp b/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp new file mode 100644 index 000000000..fa8dea2af --- /dev/null +++ b/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp @@ -0,0 +1,55 @@ +#ifdef GGML_USE_CPU_HBM + +#include "ggml-backend.h" +#include "ggml-backend-impl.h" +#include "ggml-cpu.h" +#include "ggml-impl.h" + +#include "ggml-cpu-hbm.h" + +// buffer type HBM + +#include + +static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) { + return "CPU_HBM"; + + GGML_UNUSED(buft); +} + +static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) { + hbw_free(buffer->context); +} + +static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, + size_t size) { + void * ptr; + int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size); + if (result != 0) { + GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size); + return NULL; + } + + ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size); + buffer->buft = buft; + buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer; + + return buffer; +} + +ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) { + static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = { + /* .iface = */ { + /* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name, + /* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment, + /* .get_max_size = */ nullptr, // defaults to SIZE_MAX + /* .get_alloc_size = */ nullptr, // defaults to ggml_nbytes + /* .is_host = */ ggml_backend_cpu_buffer_type_is_host, + }, + /* .context = */ nullptr, + }; + + return &ggml_backend_cpu_buffer_type_hbm; +} +#endif diff --git a/ggml/src/ggml-cpu/ggml-cpu-hbm.h b/ggml/src/ggml-cpu/ggml-cpu-hbm.h new file mode 100644 index 000000000..09a1f09d7 --- /dev/null +++ b/ggml/src/ggml-cpu/ggml-cpu-hbm.h @@ -0,0 +1,8 @@ +#pragma once + +#include "ggml-backend.h" +#include "ggml.h" + +// GGML CPU internal header + +ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void); diff --git a/ggml/src/ggml-cpu/ggml-cpu-traits.cpp b/ggml/src/ggml-cpu/ggml-cpu-traits.cpp new file mode 100644 index 000000000..62a0712da --- /dev/null +++ b/ggml/src/ggml-cpu/ggml-cpu-traits.cpp @@ -0,0 +1,36 @@ +#include "ggml-cpu-traits.h" + +#include "ggml-backend-impl.h" +#include "ggml-backend.h" + +namespace ggml::cpu { +tensor_traits::~tensor_traits() {} + +extra_buffer_type::~extra_buffer_type() {} +} // namespace ggml::cpu + +bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) { + for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) { + if (extra && extra->context) { + auto buf_extra = (ggml::cpu::extra_buffer_type *) extra->context; + auto tensor_traits = buf_extra->get_tensor_traits(op); + if (tensor_traits && tensor_traits->compute_forward(params, op)) { + return true; + } + } + } + return false; +} + +bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) { + for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) { + if (extra && extra->context) { + auto buf_extra = (ggml::cpu::extra_buffer_type *) extra->context; + auto tensor_traits = buf_extra->get_tensor_traits(op); + if (tensor_traits && tensor_traits->work_size(n_threads, op, *size)) { + return true; + } + } + } + return false; +} diff --git a/ggml/src/ggml-cpu/ggml-cpu-traits.h b/ggml/src/ggml-cpu/ggml-cpu-traits.h new file mode 100644 index 000000000..99a6186b1 --- /dev/null +++ b/ggml/src/ggml-cpu/ggml-cpu-traits.h @@ -0,0 +1,38 @@ +#pragma once +#include "ggml-backend-impl.h" +#include "ggml-cpu-impl.h" +#include "ggml.h" + +#ifdef __cplusplus +# include +extern "C" { +#endif + +// return true if op part of extra "accelerator" +bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op); +bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size); + +#ifdef __cplusplus +} + +namespace ggml::cpu { +// register in tensor->extra +class tensor_traits { + public: + virtual ~tensor_traits(); + virtual bool work_size(int n_threads, const struct ggml_tensor * op, size_t & size) = 0; + virtual bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) = 0; +}; + +class extra_buffer_type { + public: + virtual ~extra_buffer_type(); + virtual bool supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) = 0; + virtual tensor_traits * get_tensor_traits(const struct ggml_tensor * op) = 0; +}; +} // namespace ggml::cpu + +// implemented in ggml-cpu.cpp. +std::vector & ggml_backend_cpu_get_extra_buffers_type(); + +#endif diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index f12b62e3b..ea17d6077 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -3,7 +3,7 @@ #include "ggml-backend-impl.h" #include "ggml-backend.h" -#include "ggml-cpu-aarch64.h" +#include "ggml-cpu-traits.h" #include "ggml-cpu-impl.h" #include "ggml-cpu.h" #include "ggml-impl.h" @@ -224,10 +224,6 @@ typedef void * thread_ret_t; typedef pthread_t ggml_thread_t; -#ifdef GGML_USE_CPU_HBM -#include -#endif - #if defined(__APPLE__) #include #include @@ -301,7 +297,6 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = { }, [GGML_TYPE_Q8_0] = { .from_float = quantize_row_q8_0, - .from_float_to_mat = quantize_mat_q8_0, .vec_dot = ggml_vec_dot_q8_0_q8_0, .vec_dot_type = GGML_TYPE_Q8_0, #if defined (__ARM_FEATURE_MATMUL_INT8) @@ -409,33 +404,6 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = { .vec_dot_type = GGML_TYPE_BF16, .nrows = 1, }, - [GGML_TYPE_Q4_0_4_4] = { - .from_float = NULL, - .vec_dot = NULL, - .vec_dot_type = GGML_TYPE_Q8_0, - .nrows = 1, - .ncols = 4, - .gemv = ggml_gemv_q4_0_4x4_q8_0, - .gemm = ggml_gemm_q4_0_4x4_q8_0, - }, - [GGML_TYPE_Q4_0_4_8] = { - .from_float = NULL, - .vec_dot = NULL, - .vec_dot_type = GGML_TYPE_Q8_0, - .nrows = 1, - .ncols = 4, - .gemv = ggml_gemv_q4_0_4x8_q8_0, - .gemm = ggml_gemm_q4_0_4x8_q8_0, - }, - [GGML_TYPE_Q4_0_8_8] = { - .from_float = NULL, - .vec_dot = NULL, - .vec_dot_type = GGML_TYPE_Q8_0, - .nrows = 1, - .ncols = 8, - .gemv = ggml_gemv_q4_0_8x8_q8_0, - .gemm = ggml_gemm_q4_0_8x8_q8_0, - }, [GGML_TYPE_TQ1_0] = { .from_float = quantize_row_tq1_0, .vec_dot = ggml_vec_dot_tq1_0_q8_K, @@ -448,15 +416,6 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = { .vec_dot_type = GGML_TYPE_Q8_K, .nrows = 1, }, - [GGML_TYPE_IQ4_NL_4_4] = { - .from_float = NULL, - .vec_dot = NULL, - .vec_dot_type = GGML_TYPE_Q8_0, - .nrows = 1, - .ncols = 4, - .gemv = ggml_gemv_iq4_nl_4x4_q8_0, - .gemm = ggml_gemm_iq4_nl_4x4_q8_0, - }, }; const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type) { @@ -4509,9 +4468,6 @@ static void ggml_compute_forward_add( case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: - case GGML_TYPE_Q4_0_4_4: - case GGML_TYPE_Q4_0_4_8: - case GGML_TYPE_Q4_0_8_8: { ggml_compute_forward_add_q_f32(params, dst); } break; @@ -4889,9 +4845,6 @@ static void ggml_compute_forward_add1( case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: - case GGML_TYPE_Q4_0_4_4: - case GGML_TYPE_Q4_0_4_8: - case GGML_TYPE_Q4_0_8_8: { ggml_compute_forward_add1_q_f32(params, dst); } break; @@ -5019,9 +4972,6 @@ static void ggml_compute_forward_acc( case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: - case GGML_TYPE_Q4_0_4_4: - case GGML_TYPE_Q4_0_4_8: - case GGML_TYPE_Q4_0_8_8: default: { GGML_ABORT("fatal error"); @@ -7437,27 +7387,9 @@ static void ggml_compute_forward_mul_mat( const int ith = params->ith; const int nth = params->nth; - enum ggml_type type = src0->type; - - if (src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) { - type = (enum ggml_type)(intptr_t)src0->extra; - } - -#if defined(__AMX_INT8__) && defined(__AVX512VNNI__) - if (src0->buffer && ggml_backend_amx_buft_is_amx(src0->buffer->buft)) { - ggml_backend_amx_mul_mat(params, dst); - return; - } -#endif - - enum ggml_type const vec_dot_type = type_traits_cpu[type].vec_dot_type; + enum ggml_type const vec_dot_type = type_traits_cpu[src0->type].vec_dot_type; ggml_from_float_t const from_float = type_traits_cpu[vec_dot_type].from_float; - ggml_from_float_to_mat_t const from_float_to_mat = type_traits_cpu[vec_dot_type].from_float_to_mat; - int64_t const vec_dot_num_rows = type_traits_cpu[type].nrows; - int64_t const matmul_num_cols = type_traits_cpu[type].ncols; - int64_t const blck_size_interleave = ggml_get_type_traits(type)->blck_size_interleave; - ggml_gemv_t const gemv = type_traits_cpu[type].gemv; - ggml_gemm_t const gemm = type_traits_cpu[type].gemm; + int64_t const vec_dot_num_rows = type_traits_cpu[src0->type].nrows; GGML_ASSERT(ne0 == ne01); GGML_ASSERT(ne1 == ne11); @@ -7465,7 +7397,7 @@ static void ggml_compute_forward_mul_mat( GGML_ASSERT(ne3 == ne13); // we don't support permuted src0 or src1 - GGML_ASSERT(nb00 == ggml_type_size(type)); + GGML_ASSERT(nb00 == ggml_type_size(src0->type)); GGML_ASSERT(nb10 == ggml_type_size(src1->type)); // dst cannot be transposed or permuted @@ -7477,6 +7409,7 @@ static void ggml_compute_forward_mul_mat( // nb01 >= nb00 - src0 is not transposed // compute by src0 rows + // TODO: extract to "extra_op" #if GGML_USE_LLAMAFILE // broadcast factors const int64_t r2 = ne12 / ne02; @@ -7487,15 +7420,15 @@ static void ggml_compute_forward_mul_mat( if (src1_cont) { for (int64_t i13 = 0; i13 < ne13; i13++) for (int64_t i12 = 0; i12 < ne12; i12++) - if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type), + if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type), (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03, - nb01/ggml_type_size(type), + nb01/ggml_type_size(src0->type), (const char *)src1->data + i12*nb12 + i13*nb13, nb11/ggml_type_size(src1->type), (char *)dst->data + i12*nb2 + i13*nb3, nb1/ggml_type_size(dst->type), ith, nth, - type, + src0->type, src1->type, dst->type)) goto UseGgmlGemm1; @@ -7516,19 +7449,10 @@ UseGgmlGemm1:; for (int64_t i13 = 0; i13 < ne13; ++i13) { for (int64_t i12 = 0; i12 < ne12; ++i12) { - int64_t i11_processed = 0; - if ((ggml_n_dims(src1) == 2) && from_float_to_mat && gemm) { - for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) { - from_float_to_mat((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), - (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1), - 4, ne10, blck_size_interleave); - } - i11_processed = ne11 - ne11 % 4; - } - for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) { + for (int64_t i11 = ith; i11 < ne11; i11 += nth) { from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), - (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1), - ne10); + (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1), + ne10); } } } @@ -7548,15 +7472,15 @@ UseGgmlGemm1:; for (int64_t i13 = 0; i13 < ne13; i13++) for (int64_t i12 = 0; i12 < ne12; i12++) - if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type), + if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type), (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03, - nb01/ggml_type_size(type), + nb01/ggml_type_size(src0->type), (const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size, row_size/ggml_type_size(vec_dot_type), (char *)dst->data + i12*nb2 + i13*nb3, nb1/ggml_type_size(dst->type), ith, nth, - type, + src0->type, vec_dot_type, dst->type)) goto UseGgmlGemm2; @@ -7598,28 +7522,6 @@ UseGgmlGemm2:; const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0; const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1; - if ((ggml_n_dims(src0) == 2) && gemv) { - const void * src1_wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; - const size_t src1_col_stride = ggml_is_contiguous(src1) || src1->type != vec_dot_type ? ggml_row_size(vec_dot_type, ne10) : nb11; - int64_t src0_start = (ith * ne01) / nth; - int64_t src0_end = ((ith + 1) * ne01) / nth; - src0_start = (src0_start % matmul_num_cols) ? src0_start + matmul_num_cols - (src0_start % matmul_num_cols): src0_start; - src0_end = (src0_end % matmul_num_cols) ? src0_end + matmul_num_cols - (src0_end % matmul_num_cols): src0_end; - if (src0_start >= src0_end) return; - - // If there are more than three rows in src1, use gemm; otherwise, use gemv. - if (gemm && (ne11 > 3)) { - gemm(ne00, (float *)((char *) dst->data) + src0_start, ne01, (const char *) src0->data + src0_start * nb01, - (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start); - } - for (int iter = gemm ? ne11 - ne11 % 4 : 0; iter < ne11; iter++) { - gemv(ne00, (float *)((char *) dst->data + (iter * nb1)) + src0_start, ne01, - (const char *) src0->data + src0_start * nb01, (const char *) src1_wdata + (src1_col_stride * iter), 1, - src0_end - src0_start); - } - return; - } - // The first chunk comes from our thread_id, the rest will get auto-assigned. int current_chunk = ith; @@ -7642,7 +7544,7 @@ UseGgmlGemm2:; num_rows_per_vec_dot = 1; } - ggml_compute_forward_mul_mat_one_chunk(params, dst, type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end); + ggml_compute_forward_mul_mat_one_chunk(params, dst, src0->type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end); if (nth >= nchunk0 * nchunk1) { break; @@ -7674,8 +7576,6 @@ static void ggml_compute_forward_mul_mat_id( ggml_vec_dot_t const vec_dot = type_traits_cpu[type].vec_dot; enum ggml_type const vec_dot_type = type_traits_cpu[type].vec_dot_type; ggml_from_float_t const from_float = type_traits_cpu[vec_dot_type].from_float; - int64_t const matmul_num_cols = type_traits_cpu[type].ncols; - ggml_gemv_t const gemv = type_traits_cpu[type].gemv; // we don't support permuted src0 or src1 GGML_ASSERT(nb00 == ggml_type_size(type)); @@ -7761,34 +7661,6 @@ static void ggml_compute_forward_mul_mat_id( const int64_t nr0 = ne01; // src0 rows const int64_t nr1 = cne1; // src1 rows - if (((ggml_n_dims(src0) - 1) == 2) && gemv) { - int64_t src0_cur_start = (ith * ne01) / nth; - int64_t src0_cur_end = ((ith + 1) * ne01) / nth; - src0_cur_start = (src0_cur_start % matmul_num_cols) ? src0_cur_start + matmul_num_cols - (src0_cur_start % matmul_num_cols): src0_cur_start; - src0_cur_end = (src0_cur_end % matmul_num_cols) ? src0_cur_end + matmul_num_cols - (src0_cur_end % matmul_num_cols): src0_cur_end; - if (src0_cur_start >= src0_cur_end) return; - - for (int ir1 = 0; ir1 < nr1; ir1++) { - struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1); - const int id = row_mapping.i1; // selected expert index - - const int64_t i11 = id % ne11; - const int64_t i12 = row_mapping.i2; // row index in src1 - - const int64_t i1 = id; // selected expert index - const int64_t i2 = i12; // row - - const char * src1_col = (const char *) wdata + - (src1_cont || src1->type != vec_dot_type - ? (i11 + i12 * ne11) * row_size - : (i11 * nb11 + i12 * nb12)); - - gemv(ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01, - (const char *) src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start); - } - continue; - } - // distribute the thread work across the inner or outer loop based on which one is larger const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows @@ -8096,9 +7968,6 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: - case GGML_TYPE_Q4_0_4_4: - case GGML_TYPE_Q4_0_4_8: - case GGML_TYPE_Q4_0_8_8: { ggml_compute_forward_out_prod_q_f32(params, dst); } break; @@ -8361,9 +8230,6 @@ static void ggml_compute_forward_set( case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: - case GGML_TYPE_Q4_0_4_4: - case GGML_TYPE_Q4_0_4_8: - case GGML_TYPE_Q4_0_8_8: default: { GGML_ABORT("fatal error"); @@ -8625,9 +8491,6 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: - case GGML_TYPE_Q4_0_4_4: - case GGML_TYPE_Q4_0_4_8: - case GGML_TYPE_Q4_0_8_8: { ggml_compute_forward_get_rows_q(params, dst); } break; @@ -9217,10 +9080,6 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_IQ3_S: case GGML_TYPE_IQ2_S: case GGML_TYPE_Q8_K: - case GGML_TYPE_Q4_0_4_4: - case GGML_TYPE_Q4_0_4_8: - case GGML_TYPE_Q4_0_8_8: - case GGML_TYPE_IQ4_NL_4_4: case GGML_TYPE_I8: case GGML_TYPE_I16: case GGML_TYPE_I32: @@ -12426,6 +12285,9 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm return; } + // extra_buffer op? + if (ggml_cpu_extra_compute_forward(params, tensor)) return; + switch (tensor->op) { case GGML_OP_DUP: { @@ -13373,146 +13235,142 @@ struct ggml_cplan ggml_graph_plan( size_t cur = 0; - switch (node->op) { - case GGML_OP_CPY: - case GGML_OP_DUP: - { - if (ggml_is_quantized(node->type) || - // F16 -> BF16 and BF16 -> F16 copies go through intermediate F32 - (node->src[0]->type == GGML_TYPE_F16 && node->src[1] && node->src[1]->type == GGML_TYPE_BF16) || - (node->src[0]->type == GGML_TYPE_BF16 && node->src[1] && node->src[1]->type == GGML_TYPE_F16)) { + if (!ggml_cpu_extra_work_size(n_threads, node, &cur)) { + + switch (node->op) { + case GGML_OP_CPY: + case GGML_OP_DUP: + { + if (ggml_is_quantized(node->type) || + // F16 -> BF16 and BF16 -> F16 copies go through intermediate F32 + (node->src[0]->type == GGML_TYPE_F16 && node->src[1] && node->src[1]->type == GGML_TYPE_BF16) || + (node->src[0]->type == GGML_TYPE_BF16 && node->src[1] && node->src[1]->type == GGML_TYPE_F16)) { + cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks; + } + } break; + case GGML_OP_ADD: + case GGML_OP_ADD1: + { + if (ggml_is_quantized(node->src[0]->type)) { + cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks; + } + } break; + case GGML_OP_ACC: + { + if (ggml_is_quantized(node->src[0]->type)) { + cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks; + } + } break; + case GGML_OP_COUNT_EQUAL: + { + cur = ggml_type_size(node->type)*n_tasks; + } break; + case GGML_OP_MUL_MAT: + { + const enum ggml_type vec_dot_type = type_traits_cpu[node->src[0]->type].vec_dot_type; + + if (node->src[1]->type != vec_dot_type) { + cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1])); + } + } break; + case GGML_OP_MUL_MAT_ID: + { + cur = 0; + const struct ggml_tensor * src0 = node->src[0]; + const struct ggml_tensor * src1 = node->src[1]; + const enum ggml_type vec_dot_type = type_traits_cpu[src0->type].vec_dot_type; + if (src1->type != vec_dot_type) { + cur += ggml_row_size(vec_dot_type, ggml_nelements(src1)); + } + const int n_as = src0->ne[2]; + cur += GGML_PAD(cur, sizeof(int64_t)); // align + cur += n_as * sizeof(int64_t); // matrix_row_counts + cur += n_as * src1->ne[2] * sizeof(int64_t); // matrix_rows + } break; + case GGML_OP_OUT_PROD: + { + if (ggml_is_quantized(node->src[0]->type)) { + cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks; + } + } break; + case GGML_OP_SOFT_MAX: + case GGML_OP_ROPE: + { cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks; - } - } break; - case GGML_OP_ADD: - case GGML_OP_ADD1: - { - if (ggml_is_quantized(node->src[0]->type)) { - cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks; - } - } break; - case GGML_OP_ACC: - { - if (ggml_is_quantized(node->src[0]->type)) { - cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks; - } - } break; - case GGML_OP_COUNT_EQUAL: - { - cur = ggml_type_size(node->type)*n_tasks; - } break; - case GGML_OP_MUL_MAT: - { -#if defined(__AMX_INT8__) && defined(__AVX512VNNI__) - if (node->src[0]->buffer && ggml_backend_amx_buft_is_amx(node->src[0]->buffer->buft)) { - cur = ggml_backend_amx_desired_wsize(node); - } -#endif - const enum ggml_type vec_dot_type = type_traits_cpu[node->src[0]->type].vec_dot_type; + } break; + case GGML_OP_CONV_TRANSPOSE_1D: + { + GGML_ASSERT(node->src[0]->ne[3] == 1); + GGML_ASSERT(node->src[1]->ne[2] == 1); + GGML_ASSERT(node->src[1]->ne[3] == 1); - if (node->src[1]->type != vec_dot_type) { - size_t cur2 = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1])); - cur = MAX(cur, cur2); - } - } break; - case GGML_OP_MUL_MAT_ID: - { - cur = 0; - const struct ggml_tensor * src0 = node->src[0]; - const struct ggml_tensor * src1 = node->src[1]; - const enum ggml_type vec_dot_type = type_traits_cpu[src0->type].vec_dot_type; - if (src1->type != vec_dot_type) { - cur += ggml_row_size(vec_dot_type, ggml_nelements(src1)); - } - const int n_as = src0->ne[2]; - cur += GGML_PAD(cur, sizeof(int64_t)); // align - cur += n_as * sizeof(int64_t); // matrix_row_counts - cur += n_as * src1->ne[2] * sizeof(int64_t); // matrix_rows - } break; - case GGML_OP_OUT_PROD: - { - if (ggml_is_quantized(node->src[0]->type)) { - cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks; - } - } break; - case GGML_OP_SOFT_MAX: - case GGML_OP_ROPE: - { - cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks; - } break; - case GGML_OP_CONV_TRANSPOSE_1D: - { - GGML_ASSERT(node->src[0]->ne[3] == 1); - GGML_ASSERT(node->src[1]->ne[2] == 1); - GGML_ASSERT(node->src[1]->ne[3] == 1); + const int64_t ne00 = node->src[0]->ne[0]; // K + const int64_t ne01 = node->src[0]->ne[1]; // Cout + const int64_t ne02 = node->src[0]->ne[2]; // Cin + const int64_t ne10 = node->src[1]->ne[0]; // L + const int64_t ne11 = node->src[1]->ne[1]; // Cin - const int64_t ne00 = node->src[0]->ne[0]; // K - const int64_t ne01 = node->src[0]->ne[1]; // Cout - const int64_t ne02 = node->src[0]->ne[2]; // Cin + if ((node->src[0]->type == GGML_TYPE_F16 || + node->src[0]->type == GGML_TYPE_BF16) && + node->src[1]->type == GGML_TYPE_F32) { + cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02; + cur += sizeof(ggml_fp16_t)*ne10*ne11; + } else if (node->src[0]->type == GGML_TYPE_F32 && + node->src[1]->type == GGML_TYPE_F32) { + cur += sizeof(float)*ne00*ne01*ne02; + cur += sizeof(float)*ne10*ne11; + } else { + GGML_ABORT("fatal error"); + } + } break; + case GGML_OP_CONV_TRANSPOSE_2D: + { + const int64_t ne00 = node->src[0]->ne[0]; // W + const int64_t ne01 = node->src[0]->ne[1]; // H + const int64_t ne02 = node->src[0]->ne[2]; // Channels Out + const int64_t ne03 = node->src[0]->ne[3]; // Channels In - const int64_t ne10 = node->src[1]->ne[0]; // L - const int64_t ne11 = node->src[1]->ne[1]; // Cin + const int64_t ne10 = node->src[1]->ne[0]; // W + const int64_t ne11 = node->src[1]->ne[1]; // H + const int64_t ne12 = node->src[1]->ne[2]; // Channels In - if ((node->src[0]->type == GGML_TYPE_F16 || - node->src[0]->type == GGML_TYPE_BF16) && - node->src[1]->type == GGML_TYPE_F32) { - cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02; - cur += sizeof(ggml_fp16_t)*ne10*ne11; - } else if (node->src[0]->type == GGML_TYPE_F32 && - node->src[1]->type == GGML_TYPE_F32) { - cur += sizeof(float)*ne00*ne01*ne02; - cur += sizeof(float)*ne10*ne11; - } else { + cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03; + cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12; + } break; + case GGML_OP_FLASH_ATTN_EXT: + { + const int64_t ne00 = node->src[0]->ne[0]; // D + + cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread + } break; + case GGML_OP_FLASH_ATTN_BACK: + { + const int64_t D = node->src[0]->ne[0]; + const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL); + const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back + if (node->src[1]->type == GGML_TYPE_F32) { + cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1) + cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2 + } else if (node->src[1]->type == GGML_TYPE_F16) { + cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1) + cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2 + } else if (node->src[1]->type == GGML_TYPE_BF16) { + cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1) + cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2 + } + } break; + + case GGML_OP_CROSS_ENTROPY_LOSS: + { + cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks); + } break; + case GGML_OP_COUNT: + { GGML_ABORT("fatal error"); } - } break; - case GGML_OP_CONV_TRANSPOSE_2D: - { - const int64_t ne00 = node->src[0]->ne[0]; // W - const int64_t ne01 = node->src[0]->ne[1]; // H - const int64_t ne02 = node->src[0]->ne[2]; // Channels Out - const int64_t ne03 = node->src[0]->ne[3]; // Channels In - - const int64_t ne10 = node->src[1]->ne[0]; // W - const int64_t ne11 = node->src[1]->ne[1]; // H - const int64_t ne12 = node->src[1]->ne[2]; // Channels In - - cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03; - cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12; - } break; - case GGML_OP_FLASH_ATTN_EXT: - { - const int64_t ne00 = node->src[0]->ne[0]; // D - - cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread - } break; - case GGML_OP_FLASH_ATTN_BACK: - { - const int64_t D = node->src[0]->ne[0]; - const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL); - const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back - if (node->src[1]->type == GGML_TYPE_F32) { - cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1) - cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2 - } else if (node->src[1]->type == GGML_TYPE_F16) { - cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1) - cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2 - } else if (node->src[1]->type == GGML_TYPE_BF16) { - cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1) - cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2 - } - } break; - - case GGML_OP_CROSS_ENTROPY_LOSS: - { - cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks); - } break; - case GGML_OP_COUNT: - { - GGML_ABORT("fatal error"); - } - default: - break; + default: + break; + } } work_size = MAX(work_size, cur); diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp index d3b4bdb96..c390957af 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.cpp +++ b/ggml/src/ggml-cpu/ggml-cpu.cpp @@ -2,12 +2,18 @@ #include "ggml-backend-impl.h" #include "ggml-cpu.h" #include "ggml-cpu-aarch64.h" +#include "ggml-cpu-traits.h" #include "ggml-impl.h" #include "amx/amx.h" + #include #include #include +#ifdef GGML_USE_CPU_HBM +#include "ggml-cpu-hbm.h" +#endif + #if defined(__APPLE__) #include #include @@ -23,115 +29,7 @@ // ggml-backend interface -#ifdef GGML_USE_CPU_HBM - -// buffer type HBM - -#include - -static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) { - return "CPU_HBM"; - - GGML_UNUSED(buft); -} - -static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) { - hbw_free(buffer->context); -} - -static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { - void * ptr; - int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size); - if (result != 0) { - GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size); - return NULL; - } - - ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size); - buffer->buft = buft; - buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer; - - return buffer; -} - -ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) { - static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = { - /* .iface = */ { - /* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name, - /* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer, - /* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment, - /* .get_max_size = */ NULL, // defaults to SIZE_MAX - /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes - /* .is_host = */ ggml_backend_cpu_buffer_type_is_host, - }, - /* .context = */ NULL, - }; - - return &ggml_backend_cpu_buffer_type_hbm; -} -#endif - -// buffer type AARCH64 - -static void ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { - tensor->extra = (void *)ggml_aarch64_get_optimal_repack_type(tensor); // NOLINT - - GGML_UNUSED(buffer); -} - -static void ggml_backend_cpu_aarch64_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { - GGML_ASSERT(offset == 0); - GGML_ASSERT(size == ggml_nbytes(tensor)); - - enum ggml_type repack_type = (enum ggml_type)(intptr_t)tensor->extra; - - ggml_aarch64_repack_tensor(tensor, repack_type, data, size); - - GGML_UNUSED(buffer); -} - -static const char * ggml_backend_cpu_aarch64_buffer_type_get_name(ggml_backend_buffer_type_t buft) { - return "CPU_AARCH64"; - - GGML_UNUSED(buft); -} - -static ggml_backend_buffer_t ggml_backend_cpu_aarch64_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { - auto * buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size); - - if (buffer == NULL) { - return NULL; - } - - buffer->buft = buft; - buffer->iface.init_tensor = ggml_backend_cpu_aarch64_buffer_init_tensor; - buffer->iface.set_tensor = ggml_backend_cpu_aarch64_buffer_set_tensor; - - return buffer; -} - -ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void) { - static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_aarch64 = { - /* .iface = */ { - /* .get_name = */ ggml_backend_cpu_aarch64_buffer_type_get_name, - /* .alloc_buffer = */ ggml_backend_cpu_aarch64_buffer_type_alloc_buffer, - /* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment, - /* .get_max_size = */ NULL, // defaults to SIZE_MAX - /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes - /* .is_host = */ NULL, - }, - /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0), - /* .context = */ NULL, - }; - - return &ggml_backend_cpu_buffer_type_aarch64; -} - -bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft) { - return buft == ggml_backend_cpu_aarch64_buffer_type(); -} - -static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) { +std::vector& ggml_backend_cpu_get_extra_buffers_type() { static std::vector bufts = []() { std::vector bufts; @@ -152,11 +50,22 @@ static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backen return bufts; }(); - return bufts.data(); + return bufts; +} + +static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_type(ggml_backend_dev_t device) { + return ggml_backend_cpu_get_extra_buffers_type().data(); GGML_UNUSED(device); } +static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) { + for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) { + if (extra && extra == buft) return true; + } + return false; +} + // CPU backend - backend (stream) struct ggml_backend_cpu_context { @@ -465,25 +374,19 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st return true; } - if (src0 && src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) { - if (op->op != GGML_OP_MUL_MAT || src0->type == ggml_aarch64_get_optimal_repack_type(src0)) { - return false; + // extra_buffer_op? + for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) { + if (extra) { + auto buf_extra = (ggml::cpu::extra_buffer_type*) extra->context; + if (buf_extra && buf_extra->supports_op(dev, op)) { + return true; + } } } -#if defined(__AMX_INT8__) && defined(__AVX512VNNI__) - if (src0 && src0->buffer && ggml_backend_amx_buft_is_amx(src0->buffer->buft)) { - return ggml_backend_amx_device_supports_op(op); - } - for (int i = 1; i < GGML_MAX_SRC; i++) { - if (op->src[i] && op->src[i]->buffer && ggml_backend_amx_buft_is_amx(op->src[i]->buffer->buft)) { - return false; - } - } -#endif - - for (int i = 1; i < GGML_MAX_SRC; i++) { - if (op->src[i] && op->src[i]->buffer && ggml_backend_cpu_buft_is_aarch64(op->src[i]->buffer->buft)) { + // the other case need host buffer. + for (int i = 0; i < GGML_MAX_SRC; i++) { + if (op->src[i] && op->src[i]->buffer && !ggml_backend_buft_is_host(op->src[i]->buffer->buft)) { return false; } } @@ -506,19 +409,10 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st default: return true; } - - GGML_UNUSED(dev); } static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) { - bool supported = ggml_backend_buft_is_host(buft) || ggml_backend_cpu_buft_is_aarch64(buft); - -#if defined(__AMX_INT8__) && defined(__AVX512VNNI__) - supported = supported || ggml_backend_amx_buft_is_amx(buft); -#endif - - return supported; - + return ggml_backend_buft_is_host(buft) || ggml_backend_cpu_is_extra_buffer_type(buft); GGML_UNUSED(dev); } @@ -666,10 +560,12 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) { if (strcmp(name, "ggml_backend_set_n_threads") == 0) { - return (void *)ggml_backend_cpu_set_n_threads; + ggml_backend_set_n_threads_t fct = ggml_backend_cpu_set_n_threads; + return (void *)fct; } if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) { - return (void *)ggml_backend_cpu_get_extra_bufts; + ggml_backend_dev_get_extra_bufts_t fct = ggml_backend_cpu_device_get_extra_buffers_type; + return (void *)fct; } if (strcmp(name, "ggml_backend_get_features") == 0) { return (void *)ggml_backend_cpu_get_features; diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index d6e4bfdd0..15fcb2a65 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -3210,7 +3210,7 @@ static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, con static const ggml_backend_reg_i ggml_backend_cuda_reg_interface = { /* .get_name = */ ggml_backend_cuda_reg_get_name, /* .get_device_count = */ ggml_backend_cuda_reg_get_device_count, - /* .get_device_get = */ ggml_backend_cuda_reg_get_device, + /* .get_device = */ ggml_backend_cuda_reg_get_device, /* .get_proc_address = */ ggml_backend_cuda_reg_get_proc_address, }; diff --git a/ggml/src/ggml-cuda/mmv.cu b/ggml/src/ggml-cuda/mmv.cu index cfe91f428..a4b4f6bc1 100644 --- a/ggml/src/ggml-cuda/mmv.cu +++ b/ggml/src/ggml-cuda/mmv.cu @@ -57,7 +57,7 @@ static __global__ void mul_mat_vec( if (block_size > WARP_SIZE) { buf_iw[tid/WARP_SIZE] = sumf; __syncthreads(); - if (tid > WARP_SIZE) { + if (tid >= WARP_SIZE) { return; } sumf = buf_iw[tid]; diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m index f80fda7a4..34fe5778e 100644 --- a/ggml/src/ggml-metal/ggml-metal.m +++ b/ggml/src/ggml-metal/ggml-metal.m @@ -510,6 +510,35 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de #endif NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"]; + if (path_lib == nil) { + // Try to find the resource in the directory where the current binary located. + NSString * current_binary = [[NSProcessInfo processInfo] arguments][0]; + NSString * bin_dir = [current_binary stringByDeletingLastPathComponent]; + NSString * default_metallib_path = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]]; + if ([[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) { + GGML_LOG_INFO("%s: found '%s'\n", __func__, [default_metallib_path UTF8String]); + NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:default_metallib_path error:&error]; + if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) { + // Optionally, if this is a symlink, try to resolve it. + default_metallib_path = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:default_metallib_path error:&error]; + if (default_metallib_path && [default_metallib_path length] > 0 && ![[default_metallib_path substringToIndex:1] isEqualToString:@"/"]) { + // It is a relative path, adding the binary directory as directory prefix. + default_metallib_path = [NSString pathWithComponents:@[bin_dir, default_metallib_path]]; + } + if (!default_metallib_path || ![[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) { + // Link to the resource could not be resolved. + default_metallib_path = nil; + } else { + GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [default_metallib_path UTF8String]); + } + } + } else { + // The resource couldn't be found in the binary's directory. + default_metallib_path = nil; + } + path_lib = default_metallib_path; + } + if (try_metallib && path_lib != nil) { // pre-compiled library found NSURL * libURL = [NSURL fileURLWithPath:path_lib]; diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c index 7301a9c6c..7918388ae 100644 --- a/ggml/src/ggml-quants.c +++ b/ggml/src/ggml-quants.c @@ -5220,15 +5220,6 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte { VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_nl, data, nb); } break; - case GGML_TYPE_Q4_0_4_4: - case GGML_TYPE_Q4_0_4_8: - { - VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x4, data, nbytes / sizeof(block_q4_0x4), 4); - } break; - case GGML_TYPE_Q4_0_8_8: - { - VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x8, data, nbytes / sizeof(block_q4_0x8), 8); - } break; case GGML_TYPE_I8: case GGML_TYPE_I16: diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index 135efb521..ae3baedc7 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -4630,7 +4630,7 @@ static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, cons static const ggml_backend_reg_i ggml_backend_sycl_reg_interface = { /* .get_name = */ ggml_backend_sycl_reg_get_name, /* .get_device_count = */ ggml_backend_sycl_reg_get_device_count, - /* .get_device_get = */ ggml_backend_sycl_reg_get_device, + /* .get_device = */ ggml_backend_sycl_reg_get_device, /* .get_proc_address = */ ggml_backend_sycl_reg_get_proc_address, }; diff --git a/ggml/src/ggml-vulkan/CMakeLists.txt b/ggml/src/ggml-vulkan/CMakeLists.txt index ae0485e04..546c853b6 100644 --- a/ggml/src/ggml-vulkan/CMakeLists.txt +++ b/ggml/src/ggml-vulkan/CMakeLists.txt @@ -8,6 +8,20 @@ if (Vulkan_FOUND) ../../include/ggml-vulkan.h ) + # Compile a test shader to determine whether GL_NV_cooperative_matrix2 is supported. + # If it's not, there will be an error to stderr. + # If it's supported, set a define to indicate that we should compile those shaders + execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat2_support.comp" + OUTPUT_VARIABLE glslc_output + ERROR_VARIABLE glslc_error) + + if (${glslc_error} MATCHES ".*extension not supported: GL_NV_cooperative_matrix2.*") + message(STATUS "GL_NV_cooperative_matrix2 not supported by glslc") + else() + message(STATUS "GL_NV_cooperative_matrix2 supported by glslc") + add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + endif() + target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan) target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR}) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index c7ac0e8f7..5d9eba983 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -1,7 +1,8 @@ #include "ggml-vulkan.h" #include -#if defined(GGML_VULKAN_RUN_TESTS) || defined(GGML_VULKAN_PERF) +#if defined(GGML_VULKAN_RUN_TESTS) || defined(GGML_VULKAN_PERF) || defined(GGML_VULKAN_CHECK_RESULTS) #include +#include "ggml-cpu.h" #endif #include @@ -169,8 +170,22 @@ struct vk_device_struct { bool uma; bool coopmat2; + bool coopmat_support; + bool coopmat_acc_f32_support; + bool coopmat_acc_f16_support; + uint32_t coopmat_m; + uint32_t coopmat_n; + uint32_t coopmat_k; + size_t idx; + bool mul_mat_l; + bool mul_mat_m; + bool mul_mat_s; + bool mul_mat_id_l; + bool mul_mat_id_m; + bool mul_mat_id_s; + vk_matmul_pipeline pipeline_matmul_f32; vk_matmul_pipeline pipeline_matmul_f32_f16; vk_matmul_pipeline2 pipeline_matmul_f16; @@ -181,10 +196,10 @@ struct vk_device_struct { vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT]; vk_matmul_pipeline pipeline_matmul_id_f32; - vk_matmul_pipeline pipeline_matmul_id_f16; - vk_matmul_pipeline pipeline_matmul_id_f16_f32; + vk_matmul_pipeline2 pipeline_matmul_id_f16; + vk_matmul_pipeline2 pipeline_matmul_id_f16_f32; - vk_matmul_pipeline pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT]; + vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT]; vk_pipeline pipeline_dequant[GGML_TYPE_COUNT]; vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT]; @@ -412,7 +427,7 @@ static_assert(sizeof(vk_op_unary_push_constants) <= 128, "sizeof(vk_op_unary_pus // and a shift: // // n/d = (mulhi(n, mp) + n) >> L; -void init_fastdiv_values(uint32_t d, uint32_t &mp, uint32_t &L) +static void init_fastdiv_values(uint32_t d, uint32_t &mp, uint32_t &L) { // compute L = ceil(log2(d)); L = 0; @@ -424,6 +439,7 @@ void init_fastdiv_values(uint32_t d, uint32_t &mp, uint32_t &L) } template void init_pushconst_fastdiv(T &p) { + GGML_UNUSED(p); static_assert(!std::is_const::value, "unexpected type"); } @@ -1325,6 +1341,18 @@ static std::array fa_rows_cols(uint32_t D, uint32_t clamp, ggml_typ return {64, 64}; }; +static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vector& warptile, bool mul_mat_id) { + // Needs to be kept up to date on shader changes + const uint32_t bank_conflict_offset = device->coopmat_support ? 8 : 1; + const uint32_t type_size = device->fp16 ? sizeof(ggml_fp16_t) : sizeof(float); + const uint32_t warps = warptile[0] / device->subgroup_size; + + const uint32_t load_bufs = (warptile[1] + warptile[2]) * (warptile[3] + bank_conflict_offset) * type_size; + const uint32_t mmid_row_ids = mul_mat_id ? 3072 * sizeof(uint32_t) : 0; + const uint32_t coopmat_stage = device->coopmat_support ? warptile[7] * warptile[8] / warps * sizeof(float) : 0; + + return (load_bufs + mmid_row_ids + coopmat_stage) <= device->properties.limits.maxComputeSharedMemorySize; +} static void ggml_vk_load_shaders(vk_device& device) { VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")"); @@ -1382,12 +1410,25 @@ static void ggml_vk_load_shaders(vk_device& device) { m_align = 64; s_align = 32; } else { - l_warptile = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size }; - m_warptile = { 128, 64, 64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size }; - s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, 2, 2, device->subgroup_size }; - l_warptile_mmq = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size }; - m_warptile_mmq = { 128, 64, 64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size }; - s_warptile_mmq = { subgroup_size_16, 32, 32, 32, 32, 32, 2, 2, 2, device->subgroup_size }; + // Matrix cores require different warp group sizes + const uint32_t tm_l = device->coopmat_support ? device->coopmat_m : 4; + const uint32_t tm_m = device->coopmat_support ? device->coopmat_m : 4; + const uint32_t tm_s = device->coopmat_support ? device->coopmat_m : 2; + const uint32_t tn_l = device->coopmat_support ? device->coopmat_n : 4; + const uint32_t tn_m = device->coopmat_support ? device->coopmat_n : 2; + const uint32_t tn_s = device->coopmat_support ? device->coopmat_n : 2; + const uint32_t tk_l = device->coopmat_support ? device->coopmat_k : 1; + const uint32_t tk_m = device->coopmat_support ? device->coopmat_k : 1; + const uint32_t tk_s = device->coopmat_support ? device->coopmat_k : 1; + + l_warptile = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, tm_l, tn_l, tk_l, device->subgroup_size }; + m_warptile = { 128, 64, 64, 16, device->subgroup_size, 32, 2, tm_m, tn_m, tk_m, device->subgroup_size }; + s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, tm_s, tn_s, tk_s, device->subgroup_size }; + + l_warptile_mmq = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, tm_l, tn_l, tk_l, device->subgroup_size }; + m_warptile_mmq = { 128, 64, 64, 32, device->subgroup_size, 32, 2, tm_m, tn_m, tk_m, device->subgroup_size }; + s_warptile_mmq = { subgroup_size_16, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, device->subgroup_size }; + 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 }; @@ -1428,25 +1469,36 @@ static void ggml_vk_load_shaders(vk_device& device) { // assert mul_mat_mat_id shaders will fit. GGML_ASSERT(shmem_needed + 3072*4 <= device->properties.limits.maxComputeSharedMemorySize); } + // Disable medium and large matrix multiplication if not enough shared memory is available + // Check mmq warptiles as the largest configuration + // Throw an error if not enough for any matrix multiplication is available + if (!ggml_vk_matmul_shmem_support(device, s_warptile_mmq, false)) { + std::cerr << "ggml_vulkan: Error: Shared memory size too small for matrix multiplication." << std::endl; + throw std::runtime_error("Shared memory size too small for matrix multiplication."); + } else if (!ggml_vk_matmul_shmem_support(device, m_warptile_mmq, false)) { + device->mul_mat_m = false; + device->mul_mat_l = false; + } else if (!ggml_vk_matmul_shmem_support(device, l_warptile_mmq, false)) { + device->mul_mat_l = false; + } + + // Disable mul_mat_id if not enough shared memory is available + if (!ggml_vk_matmul_shmem_support(device, s_warptile_mmq, true)) { + device->mul_mat_id_s = false; + device->mul_mat_id_m = false; + device->mul_mat_id_l = false; + } else if (!ggml_vk_matmul_shmem_support(device, m_warptile_mmq, true)) { + device->mul_mat_id_m = false; + device->mul_mat_id_l = false; + } else if (!ggml_vk_matmul_shmem_support(device, l_warptile_mmq, true)) { + device->mul_mat_id_l = false; + } } device->pipeline_matmul_f32 = std::make_shared(); device->pipeline_matmul_f32_f16 = std::make_shared(); device->pipeline_matmul_id_f32 = std::make_shared(); - device->pipeline_matmul_id_f16_f32 = std::make_shared(); - device->pipeline_matmul_id_f16 = std::make_shared(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0] = std::make_shared(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1] = std::make_shared(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0] = std::make_shared(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1] = std::make_shared(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0] = std::make_shared(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K] = std::make_shared(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K] = std::make_shared(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K] = std::make_shared(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K] = std::make_shared(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K] = std::make_shared(); - device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared(); std::vector> compiles; auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array wg_denoms, const std::vector& specialization_constants, uint32_t align, bool disable_robustness = false) { @@ -1462,7 +1514,7 @@ static void ggml_vk_load_shaders(vk_device& device) { compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align, disable_robustness)); }; -#if defined(VK_NV_cooperative_matrix2) +#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) if (device->coopmat2) { auto const &fa_wg_denoms = [&](uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array { @@ -1543,119 +1595,191 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f16, _f16acc, mmq_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_id_push_constants, 4) - CREATE_MM(pipeline_matmul_id_f16, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4) - CREATE_MM(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4) + CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4) + CREATE_MM2(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4) - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL], matmul_id_iq4_nl_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) #undef CREATE_MM #undef CREATE_MM2 } else -#endif - if (device->fp16) { +#endif // defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + if (device->coopmat_support) { // Create 6 variants, {s,m,l}x{unaligned,aligned} -#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \ +#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + if (device->mul_mat ## ID ## _l) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _m) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _s) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _l) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \ + if (device->mul_mat ## ID ## _m) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \ + if (device->mul_mat ## ID ## _s) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \ - CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3); - 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 2 variants, {f16,f32} accumulator +#define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + CREATE_MM(PIPELINE_NAME . f16acc, NAMELC, _f16acc, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ - 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_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, 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); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, 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); + if (device->mul_mat_id_s || device->mul_mat_id_m || device->mul_mat_id_l) { + CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM2(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); - 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_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - 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); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + } +#undef CREATE_MM + } else if (device->fp16) { + // Create 6 variants, {s,m,l}x{unaligned,aligned} +#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + if (device->mul_mat ## ID ## _l) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _m) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _s) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _l) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \ + if (device->mul_mat ## ID ## _m) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \ + if (device->mul_mat ## ID ## _s) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \ + + CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM2(pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, 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->mul_mat_id_s || device->mul_mat_id_m || device->mul_mat_id_l) { + CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM2(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); } #undef CREATE_MM } else { // Create 6 variants, {s,m,l}x{unaligned,aligned} -#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \ - ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \ +#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + if (device->mul_mat ## ID ## _l) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _m) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _s) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _l) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align); \ + if (device->mul_mat ## ID ## _m) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align); \ + if (device->mul_mat ## ID ## _s) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \ - CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3); - CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3); - 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_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + 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, , 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_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); + 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); + if (device->mul_mat_id_s || device->mul_mat_id_m || device->mul_mat_id_l) { + CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM(pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM(pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); - 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_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - 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); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); } +#undef CREATE_MM2 #undef CREATE_MM } @@ -1851,8 +1975,10 @@ static vk_device ggml_vk_get_device(size_t idx) { bool fp16_compute = false; bool maintenance4_support = false; bool sm_builtins = false; + bool amd_shader_core_properties2 = false; bool pipeline_robustness = false; bool coopmat2_support = false; + device->coopmat_support = false; // Check if maintenance4 is supported for (const auto& properties : ext_props) { @@ -1864,10 +1990,18 @@ static vk_device ggml_vk_get_device(size_t idx) { fp16_compute = true; } else if (strcmp("VK_NV_shader_sm_builtins", properties.extensionName) == 0) { sm_builtins = true; + } else if (strcmp("VK_AMD_shader_core_properties2", properties.extensionName) == 0) { + amd_shader_core_properties2 = true; } else if (strcmp("VK_EXT_pipeline_robustness", properties.extensionName) == 0) { pipeline_robustness = true; + } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_COOPMAT")) { + device->coopmat_support = true; + device->coopmat_m = 0; + device->coopmat_n = 0; + device->coopmat_k = 0; } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 && - !getenv("GGML_VULKAN_DISABLE_COOPMAT2")) { + !getenv("GGML_VK_DISABLE_COOPMAT2")) { coopmat2_support = true; } } @@ -1876,11 +2010,14 @@ static vk_device ggml_vk_get_device(size_t idx) { vk::PhysicalDeviceMaintenance3Properties props3; vk::PhysicalDeviceMaintenance4Properties props4; vk::PhysicalDeviceSubgroupProperties subgroup_props; + vk::PhysicalDeviceDriverProperties driver_props; vk::PhysicalDeviceShaderSMBuiltinsPropertiesNV sm_props; + vk::PhysicalDeviceShaderCoreProperties2AMD amd_shader_core_properties2_props; props2.pNext = &props3; props3.pNext = &subgroup_props; + subgroup_props.pNext = &driver_props; - VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&subgroup_props; + VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&driver_props; if (maintenance4_support) { last_struct->pNext = (VkBaseOutStructure *)&props4; @@ -1890,6 +2027,10 @@ static vk_device ggml_vk_get_device(size_t idx) { last_struct->pNext = (VkBaseOutStructure *)&sm_props; last_struct = (VkBaseOutStructure *)&sm_props; } + if (amd_shader_core_properties2) { + last_struct->pNext = (VkBaseOutStructure *)&amd_shader_core_properties2_props; + last_struct = (VkBaseOutStructure *)&amd_shader_core_properties2_props; + } #if defined(VK_NV_cooperative_matrix2) vk::PhysicalDeviceCooperativeMatrix2PropertiesNV coopmat2_props; @@ -1905,7 +2046,7 @@ static vk_device ggml_vk_get_device(size_t idx) { const char* GGML_VK_FORCE_MAX_ALLOCATION_SIZE = getenv("GGML_VK_FORCE_MAX_ALLOCATION_SIZE"); if (GGML_VK_FORCE_MAX_ALLOCATION_SIZE != nullptr) { - device->max_memory_allocation_size = std::stoi(GGML_VK_FORCE_MAX_ALLOCATION_SIZE); + device->max_memory_allocation_size = std::stoul(GGML_VK_FORCE_MAX_ALLOCATION_SIZE); } else if (maintenance4_support) { device->max_memory_allocation_size = std::min(props3.maxMemoryAllocationSize, props4.maxBufferSize); } else { @@ -1917,15 +2058,22 @@ static vk_device ggml_vk_get_device(size_t idx) { device->uma = device->properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu; if (sm_builtins) { device->shader_core_count = sm_props.shaderSMCount; + } else if (amd_shader_core_properties2) { + device->shader_core_count = amd_shader_core_properties2_props.activeComputeUnitCount; } else { device->shader_core_count = 0; } - const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16"); - const bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr; + const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr; device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute; + if (device->vendor_id == VK_VENDOR_ID_INTEL || (props2.properties.vendorID == VK_VENDOR_ID_AMD && driver_props.driverID == vk::DriverId::eAmdProprietary)) { + // Intel drivers don't support coopmat properly yet + // Only RADV supports coopmat properly on AMD + device->coopmat_support = false; + } + std::vector queue_family_props = device->physical_device.getQueueFamilyProperties(); // Try to find a non-graphics compute queue and transfer-focused queues @@ -1976,6 +2124,16 @@ static vk_device ggml_vk_get_device(size_t idx) { device_extensions.push_back("VK_EXT_pipeline_robustness"); } + VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features; + coopmat_features.pNext = nullptr; + coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR; + coopmat_features.cooperativeMatrix = VK_FALSE; + + if (device->coopmat_support) { + last_struct->pNext = (VkBaseOutStructure *)&coopmat_features; + last_struct = (VkBaseOutStructure *)&coopmat_features; + } + #if defined(VK_NV_cooperative_matrix2) VkPhysicalDeviceCooperativeMatrix2FeaturesNV coopmat2_features {}; coopmat2_features.pNext = nullptr; @@ -1993,8 +2151,10 @@ static vk_device ggml_vk_get_device(size_t idx) { device->pipeline_robustness = pl_robustness_features.pipelineRobustness; + device->coopmat_support = device->coopmat_support && coopmat_features.cooperativeMatrix; + if (coopmat2_support) { -#if defined(VK_NV_cooperative_matrix2) +#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) if (coopmat2_features.cooperativeMatrixWorkgroupScope && coopmat2_features.cooperativeMatrixFlexibleDimensions && coopmat2_features.cooperativeMatrixReductions && @@ -2083,6 +2243,74 @@ static vk_device ggml_vk_get_device(size_t idx) { if (device->fp16) { device_extensions.push_back("VK_KHR_shader_float16_int8"); } + + if (device->coopmat_support) { + // Query supported shapes + std::vector cm_props; + + PFN_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR pfn_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR = + (PFN_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR)vkGetInstanceProcAddr(vk_instance.instance, "vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR"); + + uint32_t cm_props_num; + + pfn_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR(device->physical_device, &cm_props_num, nullptr); + + cm_props.resize(cm_props_num); + + for (auto& prop : cm_props) { + prop.sType = VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_PROPERTIES_KHR; + } + + pfn_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR(device->physical_device, &cm_props_num, cm_props.data()); + + VK_LOG_DEBUG("ggml_vulkan: Cooperative Matrix Shapes: " << cm_props.size()); + + for (auto& prop : cm_props) { + VK_LOG_DEBUG("ggml_vulkan: M: " << prop.MSize << " N: " << prop.NSize << " K: " << prop.KSize << " A: " << vk::to_string((vk::ComponentTypeKHR)prop.AType) << " B: " << vk::to_string((vk::ComponentTypeKHR)prop.BType) << " C: " << vk::to_string((vk::ComponentTypeKHR)prop.CType) << " Result: " << vk::to_string((vk::ComponentTypeKHR)prop.ResultType) << " saturatingAccumulation: " << prop.saturatingAccumulation << " scope: " << vk::to_string((vk::ScopeKHR)prop.scope)); + + if ((vk::ComponentTypeKHR)prop.AType == vk::ComponentTypeKHR::eFloat16 && + (vk::ComponentTypeKHR)prop.BType == vk::ComponentTypeKHR::eFloat16 && + (vk::ScopeKHR)prop.scope == vk::ScopeKHR::eSubgroup + ) { + if ((vk::ComponentTypeKHR)prop.CType == vk::ComponentTypeKHR::eFloat32 && + (vk::ComponentTypeKHR)prop.ResultType == vk::ComponentTypeKHR::eFloat32) { + // coopmat sizes not set yet + if (device->coopmat_m == 0) { + device->coopmat_acc_f32_support = true; + device->coopmat_m = prop.MSize; + device->coopmat_n = prop.NSize; + device->coopmat_k = prop.KSize; + } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) { + // Only enable if shape is identical + device->coopmat_acc_f32_support = true; + } + } else if ((vk::ComponentTypeKHR)prop.CType == vk::ComponentTypeKHR::eFloat16 && + (vk::ComponentTypeKHR)prop.ResultType == vk::ComponentTypeKHR::eFloat16) { + // coopmat sizes not set yet + if (device->coopmat_m == 0) { + device->coopmat_acc_f16_support = true; + device->coopmat_m = prop.MSize; + device->coopmat_n = prop.NSize; + device->coopmat_k = prop.KSize; + } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) { + // Only enable if shape is identical + device->coopmat_acc_f16_support = true; + } + } + } + } + + if (device->coopmat_m == 0) { + // No suitable matmul mode found + GGML_LOG_DEBUG("ggml_vulkan: WARNING: No suitable matrix core mode found. Disabling matrix cores.\n"); + device->coopmat_support = false; + } + } + + if (device->coopmat_support) { + device_extensions.push_back("VK_KHR_cooperative_matrix"); + } + device->name = GGML_VK_NAME + std::to_string(idx); device_create_info = { @@ -2098,6 +2326,37 @@ static vk_device ggml_vk_get_device(size_t idx) { ggml_vk_create_queue(device, device->compute_queue, compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer }, false); // Shaders + // Disable matmul tile sizes early if performance low or not supported + switch (device->vendor_id) { +#ifndef GGML_VULKAN_RUN_TESTS + case VK_VENDOR_ID_AMD: + case VK_VENDOR_ID_INTEL: + device->mul_mat_l = false; + device->mul_mat_m = true; + device->mul_mat_s = true; + device->mul_mat_id_l = false; + device->mul_mat_id_m = true; + device->mul_mat_id_s = true; + break; + case VK_VENDOR_ID_APPLE: + device->mul_mat_l = false; + device->mul_mat_m = true; + device->mul_mat_s = false; + device->mul_mat_id_l = false; + device->mul_mat_id_m = true; + device->mul_mat_id_s = false; + break; +#endif + default: + device->mul_mat_l = true; + device->mul_mat_m = true; + device->mul_mat_s = true; + device->mul_mat_id_l = true; + device->mul_mat_id_m = true; + device->mul_mat_id_s = true; + break; + } + ggml_vk_load_shaders(device); if (!device->single_queue) { @@ -2155,15 +2414,29 @@ static void ggml_vk_print_gpu_info(size_t idx) { bool fp16_storage = false; bool fp16_compute = false; + bool coopmat_support = false; + bool coopmat2_support = false; for (auto properties : ext_props) { if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) { fp16_storage = true; } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) { fp16_compute = true; + } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_COOPMAT")) { + coopmat_support = true; + } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_COOPMAT2")) { + coopmat2_support = true; } } + if (props2.properties.vendorID == VK_VENDOR_ID_INTEL || (props2.properties.vendorID == VK_VENDOR_ID_AMD && driver_props.driverID == vk::DriverId::eAmdProprietary)) { + // Intel drivers don't support coopmat properly yet + // Only RADV supports coopmat properly on AMD + coopmat_support = false; + } + const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16"); bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr; @@ -2186,13 +2459,30 @@ static void ggml_vk_print_gpu_info(size_t idx) { vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES; vk11_features.pNext = &vk12_features; + // Pointer to the last chain element + VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&vk12_features; + + VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features; + coopmat_features.pNext = nullptr; + coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR; + coopmat_features.cooperativeMatrix = VK_FALSE; + + if (coopmat_support) { + last_struct->pNext = (VkBaseOutStructure *)&coopmat_features; + last_struct = (VkBaseOutStructure *)&coopmat_features; + } + vkGetPhysicalDeviceFeatures2(physical_device, &device_features2); fp16 = fp16 && vk12_features.shaderFloat16; + coopmat_support = coopmat_support && coopmat_features.cooperativeMatrix; + + std::string matrix_cores = coopmat2_support ? "NV_coopmat2" : coopmat_support ? "KHR_coopmat" : "none"; + std::string device_name = props2.properties.deviceName.data(); - GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu\n", - idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size); + GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu | matrix cores: %s\n", + idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size, matrix_cores.c_str()); if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) { GGML_LOG_DEBUG("ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want.\n"); @@ -2428,7 +2718,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) { return ctx->device->pipeline_matmul_f32_f16; } - if (prec == GGML_PREC_DEFAULT && ctx->device->coopmat2) { + if (prec == GGML_PREC_DEFAULT && ctx->device->fp16) { if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { return ctx->device->pipeline_matmul_f16_f32.f16acc; } @@ -2469,7 +2759,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte assert(src1_type == GGML_TYPE_F16); return ctx->device->pipeline_dequant_mul_mat_mat_f16[src0_type].f16acc; } - return ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc; + return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc; } static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) { @@ -2498,16 +2788,25 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type]; } -static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type) { +static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type, ggml_prec prec) { VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_id_pipeline()"); if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) { return ctx->device->pipeline_matmul_id_f32; } - if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { - return ctx->device->pipeline_matmul_id_f16_f32; - } - if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { - return ctx->device->pipeline_matmul_id_f16; + if (prec == GGML_PREC_DEFAULT && ctx->device->fp16) { + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { + return ctx->device->pipeline_matmul_id_f16_f32.f16acc; + } + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { + return ctx->device->pipeline_matmul_id_f16.f16acc; + } + } else { + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { + return ctx->device->pipeline_matmul_id_f16_f32.f32acc; + } + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { + return ctx->device->pipeline_matmul_id_f16.f32acc; + } } GGML_ASSERT(src1_type == GGML_TYPE_F32); @@ -2529,7 +2828,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co return nullptr; } - return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type]; + return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc; } static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) { @@ -3119,54 +3418,23 @@ static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, int m, int return split_k; } -static vk_pipeline ggml_vk_guess_matmul_pipeline_amd(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) { - if (m <= 32 || n <= 32) { - return aligned ? mmp->a_s : mmp->s; - } - return aligned ? mmp->a_m : mmp->m; - - GGML_UNUSED(ctx); -} - -static vk_pipeline ggml_vk_guess_matmul_pipeline_apple(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, bool aligned) { - return aligned ? mmp->a_m : mmp->m; - - GGML_UNUSED(ctx); -} - -static vk_pipeline ggml_vk_guess_matmul_pipeline_intel(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, bool aligned) { - return aligned ? mmp->a_s : mmp->s; - - GGML_UNUSED(ctx); -} - static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) { VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ")"); - switch (ctx->device->vendor_id) { - case VK_VENDOR_ID_AMD: - return ggml_vk_guess_matmul_pipeline_amd(ctx, mmp, m, n, aligned); - case VK_VENDOR_ID_APPLE: - return ggml_vk_guess_matmul_pipeline_apple(ctx, mmp, aligned); - case VK_VENDOR_ID_INTEL: - return ggml_vk_guess_matmul_pipeline_intel(ctx, mmp, aligned); - default: - break; - } if (ctx->device->coopmat2) { - if ((m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) { + if ((ctx->device->mul_mat_l && (m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) || (!ctx->device->mul_mat_m && !ctx->device->mul_mat_s)) { return aligned ? mmp->a_l : mmp->l; } - if ((m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) { + if ((ctx->device->mul_mat_m && (m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) || !ctx->device->mul_mat_s) { return aligned ? mmp->a_m : mmp->m; } return aligned ? mmp->a_s : mmp->s; } - if (m <= 32 || n <= 32) { + if ((ctx->device->mul_mat_s && (m <= 32 || n <= 32)) || (!ctx->device->mul_mat_m && !ctx->device->mul_mat_l)) { return aligned ? mmp->a_s : mmp->s; } - if (m <= 64 || n <= 64) { + if ((ctx->device->mul_mat_m && (m <= 64 || n <= 64)) || !ctx->device->mul_mat_l) { return aligned ? mmp->a_m : mmp->m; } return aligned ? mmp->a_l : mmp->l; @@ -3201,6 +3469,33 @@ static void ggml_vk_matmul( ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2.size() * sizeof(uint32_t), pc2.data(), { m * n * batch, 1, 1 }); } +static vk_pipeline ggml_vk_guess_matmul_id_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) { + VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ")"); + + if (ctx->device->coopmat2) { + if ((ctx->device->mul_mat_id_l && (m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) || (!ctx->device->mul_mat_id_m && !ctx->device->mul_mat_id_s)) { + return aligned ? mmp->a_l : mmp->l; + } + if ((ctx->device->mul_mat_id_m && (m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) || !ctx->device->mul_mat_id_s) { + return aligned ? mmp->a_m : mmp->m; + } + return aligned ? mmp->a_s : mmp->s; + } + + if ((ctx->device->mul_mat_id_s && (m <= 32 || n <= 32)) || (!ctx->device->mul_mat_id_m && !ctx->device->mul_mat_id_l)) { + return aligned ? mmp->a_s : mmp->s; + } + if ((ctx->device->mul_mat_id_m && (m <= 64 || n <= 64)) || !ctx->device->mul_mat_id_l) { + return aligned ? mmp->a_m : mmp->m; + } + return aligned ? mmp->a_l : mmp->l; +} + +static uint32_t ggml_vk_guess_matmul_id_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n) { + VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ")"); + return ggml_vk_guess_matmul_id_pipeline(ctx, mmp, m, n, true)->align; +} + static void ggml_vk_matmul_id( ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline& pipeline, vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids, @@ -3277,6 +3572,7 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context& (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], 1 , (uint32_t)tensor->ne[0] , (uint32_t)(tensor->ne[0] * tensor->ne[1]) , (uint32_t)(tensor->ne[0] * tensor->ne[1] * tensor->ne[2]), 0, 0.0f, 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }; init_pushconst_fastdiv(pc); ggml_vk_sync_buffers(subctx); @@ -3904,7 +4200,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig; - vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type); + vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]); const bool qx_needs_dequant = mmp == nullptr || x_non_contig; const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig; @@ -3920,10 +4216,10 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& const uint64_t y_ne = ne11 * ne10; const uint64_t d_ne = ne21 * ne20; - const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, nei1)); + const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1)); const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8; - vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, nei1, aligned); + vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned); const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type); const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); @@ -5057,7 +5353,8 @@ static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, con (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, - op_params[0], 0.0f + op_params[0], 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, dryrun); } @@ -5071,6 +5368,7 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, dryrun); } @@ -5084,6 +5382,7 @@ static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, dryrun); } @@ -5097,6 +5396,7 @@ static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, dryrun); } @@ -5111,6 +5411,7 @@ static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, con (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, op_params[0], op_params[1], + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, dryrun); } @@ -5124,6 +5425,7 @@ static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, dryrun); } @@ -5137,6 +5439,7 @@ static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, co (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, 0, 0.0f, 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, dryrun); } @@ -5151,6 +5454,7 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, d_offset, 0.0f, 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, }, dryrun); } @@ -5504,19 +5808,27 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t for (size_t i = 0; i < x_ne; i++) { if (std::is_same()) { x[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; + // x[i] = 1.0f; + // x[i] = i + 1; + // x[i] = (i % k == i / k) ? 1.0f : 0.0f; } else if (std::is_same()) { x[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f); + // x[i] = ggml_fp32_to_fp16(1.0f); + // x[i] = ggml_fp32_to_fp16(i + 1); + // x[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f); } else { GGML_ABORT("fatal error"); } } for (size_t i = 0; i < y_ne; i++) { if (std::is_same()) { - // y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; - y[i] = (i % k == i / k) ? 1.0f : 0.0f; + y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; + // y[i] = (i % k == i / k) ? 1.0f : 0.0f; + // y[i] = i + 1; } else if (std::is_same()) { - // y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f); - y[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f); + y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f); + // y[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f); + // y[i] = ggml_fp32_to_fp16(i + 1); } else { GGML_ABORT("fatal error"); } @@ -5600,7 +5912,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t double err = std::fabs(d[i] - d_chk[i]); avg_err += err; - if (err > 0.05f && first_err_n == -1) { + if ((err > 0.05f || std::isnan(err)) && first_err_n == -1) { first_err_b = i / (m * n); first_err_n = (i % (m * n)) / m; first_err_m = (i % (m * n)) % m; @@ -5613,12 +5925,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t std::cerr << "TEST " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl; - if (avg_err > 0.1) { + if (avg_err > 0.1 || std::isnan(avg_err)) { std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl; std::cerr << "Actual result: " << std::endl << std::endl; ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); - std::cerr << std::endl; - ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n + 15, first_err_b); std::cerr << "Expected result: " << std::endl << std::endl; ggml_vk_print_matrix_area(d_chk, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); @@ -5801,13 +6111,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, vk_pipeline p; std::string shname; if (shader_size == 0) { - p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_s; + p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->a_s : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_s; shname = std::string(ggml_type_name(quant)) + "_ALIGNED_S"; } else if (shader_size == 1) { - p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_m; + p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->a_m : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_m; shname = std::string(ggml_type_name(quant)) + "_ALIGNED_M"; } else if (shader_size == 2) { - p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_l; + p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->a_l : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_l; shname = std::string(ggml_type_name(quant)) + "_ALIGNED_L"; } else { GGML_ASSERT(0); @@ -5817,13 +6127,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, if (k != kpad) { if (shader_size == 0) { - p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->s; + p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->s : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->s; shname = std::string(ggml_type_name(quant)) + "_S"; } else if (shader_size == 1) { - p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->m; + p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->m : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->m; shname = std::string(ggml_type_name(quant)) + "_M"; } else if (shader_size == 2) { - p = ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->l; + p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->l : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->l; shname = std::string(ggml_type_name(quant)) + "_L"; } else { GGML_ASSERT(0); @@ -5982,105 +6292,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { #if defined(GGML_VULKAN_RUN_TESTS) - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_F32); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_0); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_1); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_0); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_1); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q8_0); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q2_K); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q3_K); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_K); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_K); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q6_K); - ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_IQ4_NL); - - ggml_vk_test_matmul(ctx, 512, 512, 100, 32, 100, 1, 2); - - ggml_vk_test_matmul(ctx, 128, 512, 512, 2, 100, 1, 0); - ggml_vk_test_matmul(ctx, 128, 512, 512, 2, 100, 1, 1); - ggml_vk_test_matmul(ctx, 128, 512, 512, 2, 100, 1, 2); - // ggml_vk_test_matmul(ctx, 128, 512, 512, 2, 100, 4, 0); - // ggml_vk_test_matmul(ctx, 128, 512, 512, 2, 100, 4, 1); - // ggml_vk_test_matmul(ctx, 128, 512, 512, 2, 100, 4, 2); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_0); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_0); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_0); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_1); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_1); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_1); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_1); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_1); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_1); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_0); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_0); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_0); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_1); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_1); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_1); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_1); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_1); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_1); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q8_0); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q8_0); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q8_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q8_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q8_0); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q8_0); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q2_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q2_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q2_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q2_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q2_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q2_K); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q3_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q3_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q3_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q3_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q3_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q3_K); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_K); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_K); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q6_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q6_K); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q6_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q6_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q6_K); - // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q6_K); - - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_IQ4_NL); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_IQ4_NL); - ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_IQ4_NL); - - std::cerr << std::endl; - const std::vector vals { + 512, 512, 128, + 128, 512, 512, + 4096, 512, 4096, + 11008, 512, 4096, + 4096, 512, 11008, + 32000, 512, 4096, 8, 8, 8, 100, 46, 576, 623, 111, 128, @@ -6093,15 +6311,6 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { 49, 49, 128, 128, 49, 49, 4096, 49, 4096, - 11008, 49, 4096, - 4096, 49, 11008, - 32000, 49, 4096, - 512, 512, 128, - 128, 512, 512, - 4096, 512, 4096, - 11008, 512, 4096, - 4096, 512, 11008, - 32000, 512, 4096, }; const size_t num_it = 100; @@ -6109,10 +6318,45 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0); ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1); ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2); - // ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0); - // ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1); - // ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2); - std::cerr << std::endl; + std::cerr << '\n'; + ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 0); + ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 1); + ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 2); + std::cerr << '\n'; + ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0); + ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1); + ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2); + std::cerr << '\n' << std::endl; + + if (vals[i + 2] % 32 == 0) { + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2, GGML_TYPE_Q4_0); + std::cerr << '\n'; + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 0, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 1, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 2, GGML_TYPE_Q4_0); + std::cerr << '\n'; + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2, GGML_TYPE_Q4_0); + std::cerr << '\n' << std::endl; + } + + if (vals[i + 2] % 256 == 0) { + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0, GGML_TYPE_Q4_K); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1, GGML_TYPE_Q4_K); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2, GGML_TYPE_Q4_K); + std::cerr << '\n'; + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 0, GGML_TYPE_Q4_K); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 1, GGML_TYPE_Q4_K); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 2, GGML_TYPE_Q4_K); + std::cerr << '\n'; + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0, GGML_TYPE_Q4_K); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1, GGML_TYPE_Q4_K); + ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2, GGML_TYPE_Q4_K); + std::cerr << '\n' << std::endl; + } } GGML_ABORT("fatal error"); @@ -7200,8 +7444,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm 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) { + const vk_device& device = ggml_vk_get_device(ctx->device); + if (op->op == GGML_OP_MUL_MAT_ID && !device->mul_mat_id_s && !device->mul_mat_id_m && !device->mul_mat_id_l) { // If there's not enough shared memory for row_ids and the result tile, fallback to CPU return false; } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp index 2ff5c4305..48122cbef 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp @@ -7,6 +7,12 @@ #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require #endif +#ifdef COOPMAT +#extension GL_KHR_cooperative_matrix : enable +#extension GL_KHR_memory_scope_semantics : enable +#extension GL_KHR_shader_subgroup_basic : enable +#endif + #ifdef MUL_MAT_ID #extension GL_EXT_shader_explicit_arithmetic_types_int16 : require #endif @@ -57,6 +63,7 @@ layout (push_constant) uniform parameter #endif } p; +layout (constant_id = 0) const uint BLOCK_SIZE = 64; layout (constant_id = 1) const uint BM = 64; layout (constant_id = 2) const uint BN = 64; layout (constant_id = 3) const uint BK = 16; // Assumed to be 32 if working with a quant @@ -65,13 +72,26 @@ layout (constant_id = 5) const uint WN = 32; layout (constant_id = 6) const uint WMITER = 2; layout (constant_id = 7) const uint TM = 4; layout (constant_id = 8) const uint TN = 2; -layout (constant_id = 9) const uint WARP = 32; +layout (constant_id = 9) const uint TK = 1; // Only needed for coopmat +layout (constant_id = 10) const uint WARP = 32; -shared FLOAT_TYPE buf_a[BM * (BK+1)]; -shared FLOAT_TYPE buf_b[BN * (BK+1)]; +#ifdef COOPMAT +#define SHMEM_STRIDE (BK + 8) +#else +#define SHMEM_STRIDE (BK + 1) +#endif + +shared FLOAT_TYPE buf_a[BM * SHMEM_STRIDE]; +shared FLOAT_TYPE buf_b[BN * SHMEM_STRIDE]; #ifdef MUL_MAT_ID shared u16vec2 row_ids[3072]; +#endif // MUL_MAT_ID + +#define NUM_WARPS (BLOCK_SIZE / WARP) + +#ifdef COOPMAT +shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS]; #endif void main() { @@ -98,17 +118,32 @@ void main() { const uint ik = gl_WorkGroupID.x / blocks_m; const uint ic = gl_WorkGroupID.y; - const uint warp_i = gl_LocalInvocationID.x / WARP; - const uint warp_r = warp_i % (BM / WM); - const uint warp_c = warp_i / (BM / WM); - const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER); const uint WSUBM = WM / WMITER; const uint WSUBN = WN / WNITER; +#ifdef COOPMAT + const uint warp_i = gl_SubgroupID; + + const uint tiw = gl_SubgroupInvocationID; + + const uint cms_per_row = WM / TM; + const uint cms_per_col = WN / TN; + + const uint storestride = WARP / TM; + const uint store_r = tiw % TM; + const uint store_c = tiw / TM; +#else + const uint warp_i = gl_LocalInvocationID.x / WARP; + const uint tiw = gl_LocalInvocationID.x % WARP; + const uint tiwr = tiw % (WSUBM / TM); const uint tiwc = tiw / (WSUBM / TM); +#endif + + const uint warp_r = warp_i % (BM / WM); + const uint warp_c = warp_i / (BM / WM); const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A); const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A); @@ -156,21 +191,31 @@ void main() { uint pos_b = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / LOAD_VEC_B; #endif - float sums[WMITER * TM * WNITER * TN]; +#ifdef COOPMAT + coopmat cache_a; + coopmat cache_b; + coopmat sums[cms_per_row * cms_per_col]; + + [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) { + sums[i] = coopmat(0.0f); + } +#else + ACC_TYPE sums[WMITER * TM * WNITER * TN]; FLOAT_TYPE cache_a[WMITER * TM]; FLOAT_TYPE cache_b[WNITER * TN]; [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) { - sums[i] = 0.0f; + sums[i] = ACC_TYPE(0.0f); } +#endif - [[unroll]] for (uint block = start_k; block < end_k; block += BK) { + for (uint block = start_k; block < end_k; block += BK) { [[unroll]] for (uint l = 0; l < BM; l += loadstride_a) { #if defined(DATA_A_F32) || defined(DATA_A_F16) #if LOAD_VEC_A == 8 const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; buf_a[buf_idx ] = FLOAT_TYPE(data_a[idx][0].x); buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx][0].y); buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx][0].z); @@ -181,21 +226,21 @@ void main() { buf_a[buf_idx + 7] = FLOAT_TYPE(data_a[idx][1].w); #elif LOAD_VEC_A == 4 const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; buf_a[buf_idx ] = FLOAT_TYPE(data_a[idx].x); buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx].y); buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx].z); buf_a[buf_idx + 3] = FLOAT_TYPE(data_a[idx].w); #else if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) { - buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]); + buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]); } else { - buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(0.0f); + buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = FLOAT_TYPE(0.0f); } #endif #elif defined(DATA_A_Q4_0) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a; const uint ib = idx / 16; const uint iqs = idx & 0xF; @@ -208,7 +253,7 @@ void main() { buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); #elif defined(DATA_A_Q4_1) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a; const uint ib = idx / 16; const uint iqs = idx & 0xF; @@ -222,7 +267,7 @@ void main() { buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); #elif defined(DATA_A_Q5_0) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a; const uint ib = idx / 16; const uint iqs = idx & 0xF; @@ -237,7 +282,7 @@ void main() { buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); #elif defined(DATA_A_Q5_1) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a; const uint ib = idx / 16; const uint iqs = idx & 0xF; @@ -253,7 +298,7 @@ void main() { buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); #elif defined(DATA_A_Q8_0) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; const uint ib = idx / 16; const uint iqs = (idx & 0xF) * 2; @@ -265,7 +310,7 @@ void main() { buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); #elif defined(DATA_A_Q2_K) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; const uint ib = idx / 128; // 2 values per idx const uint iqs = idx % 128; // 0..127 @@ -284,7 +329,7 @@ void main() { buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); #elif defined(DATA_A_Q3_K) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; const uint ib = idx / 128; // 2 values per idx const uint iqs = idx % 128; // 0..127 @@ -308,7 +353,7 @@ void main() { buf_a[buf_idx + 1] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4))); #elif defined(DATA_A_Q4_K) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; const uint ib = idx / 128; // 2 values per idx const uint iqs = idx % 128; // 0..127 @@ -320,15 +365,20 @@ void main() { const vec2 loadd = vec2(data_a[ib].d); - uint8_t sc; - uint8_t mbyte; - if (is < 4) { - sc = uint8_t(data_a[ib].scales[is ] & 63); - mbyte = uint8_t(data_a[ib].scales[is + 4] & 63); - } else { - sc = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4)); - mbyte = uint8_t((data_a[ib].scales[is + 4] >> 4) | ((data_a[ib].scales[is ] >> 6) << 4)); - } + const uint scidx0 = (is < 4) ? is : (is + 4); + const uint scidx1 = (is < 4) ? is : (is - 4); + const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0; + const uint scidxshift1 = (is < 4) ? 0 : 2; + const uint mbidx0 = is + 4; + const uint mbidx1 = (is < 4) ? is + 4 : is; + const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0; + const uint mbidxshift0 = (is < 4) ? 0 : 4; + const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0; + const uint mbidxshift1 = (is < 4) ? 0 : 2; + + const uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1)); + const uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1)); + const float d = loadd.x * sc; const float m = -loadd.y * mbyte; @@ -336,7 +386,7 @@ void main() { buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF), m)); #elif defined(DATA_A_Q5_K) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; const uint ib = idx / 128; // 2 values per idx const uint iqs = idx % 128; // 0..127 @@ -351,15 +401,20 @@ void main() { const vec2 loadd = vec2(data_a[ib].d); - uint8_t sc; - uint8_t mbyte; - if (is < 4) { - sc = uint8_t(data_a[ib].scales[is ] & 63); - mbyte = uint8_t(data_a[ib].scales[is + 4] & 63); - } else { - sc = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4)); - mbyte = uint8_t((data_a[ib].scales[is + 4] >> 4) | ((data_a[ib].scales[is ] >> 6) << 4)); - } + const uint scidx0 = (is < 4) ? is : (is + 4); + const uint scidx1 = (is < 4) ? is : (is - 4); + const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0; + const uint scidxshift1 = (is < 4) ? 0 : 2; + const uint mbidx0 = is + 4; + const uint mbidx1 = (is < 4) ? is + 4 : is; + const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0; + const uint mbidxshift0 = (is < 4) ? 0 : 4; + const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0; + const uint mbidxshift1 = (is < 4) ? 0 : 2; + + const uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1)); + const uint8_t mbyte = uint8_t(((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1)); + const float d = loadd.x * sc; const float m = -loadd.y * mbyte; @@ -367,7 +422,7 @@ void main() { buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m)); #elif defined(DATA_A_Q6_K) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; const uint ib = idx / 128; // 2 values per idx const uint iqs = idx % 128; // 0..127 @@ -386,7 +441,7 @@ void main() { buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32)); #elif defined(DATA_A_IQ4_NL) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a; const uint ib = idx / 16; const uint iqs = idx & 0xF; @@ -407,7 +462,7 @@ void main() { #else const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b; #endif - const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B; + const uint buf_idx = (loadc_b + l) * SHMEM_STRIDE + loadr_b * LOAD_VEC_B; buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx][0].x); buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx][0].y); buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx][0].z); @@ -423,24 +478,24 @@ void main() { #else const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b; #endif - const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B; + const uint buf_idx = (loadc_b + l) * SHMEM_STRIDE + loadr_b * LOAD_VEC_B; buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx].x); buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx].y); buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx].z); buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx].w); #elif !MUL_MAT_ID if (ic * BN + loadc_b + l < p.N && block + loadr_b < end_k) { - buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]); + buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]); } else { - buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f); + buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f); } #else const uint row_i = ic * BN + loadc_b + l; if (row_i < _ne1) { const u16vec2 row_idx = row_ids[row_i]; - buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]); + buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]); } else { - buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f); + buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f); } #endif } @@ -450,16 +505,30 @@ void main() { pos_a += BK / LOAD_VEC_A; pos_b += BK / LOAD_VEC_B; - for (uint i = 0; i < BK; i++) { +#ifdef COOPMAT + [[unroll]] for (uint i = 0; i < BK; i += TK) { + [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { + // Load from shared into cache + coopMatLoad(cache_a, buf_a, (warp_r * WM + cm_row * TM) * SHMEM_STRIDE + i, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor); + + [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { + coopMatLoad(cache_b, buf_b, (warp_c * WN + cm_col * TN) * SHMEM_STRIDE + i, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor); + + sums[cm_col * cms_per_row + cm_row] = coopMatMulAdd(cache_a, cache_b, sums[cm_col * cms_per_row + cm_row]); + } + } + } +#else + [[unroll]] for (uint i = 0; i < BK; i++) { // Load from shared into cache [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { [[unroll]] for (uint j = 0; j < TM; j++) { - cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * (BK+1) + i]; + cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + i]; } } [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) { [[unroll]] for (uint j = 0; j < TN; j++) { - cache_b[wsic * TN + j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * (BK+1) + i]; + cache_b[wsic * TN + j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * SHMEM_STRIDE + i]; } } @@ -468,12 +537,13 @@ void main() { [[unroll]] for (uint cc = 0; cc < TN; cc++) { [[unroll]] for (uint cr = 0; cr < TM; cr++) { const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr; - sums[sums_idx] = fma(float(cache_a[wsir * TM + cr]), float(cache_b[wsic * TN + cc]), sums[sums_idx]); + sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr]), ACC_TYPE(cache_b[wsic * TN + cc]), sums[sums_idx]); } } } } } +#endif barrier(); } @@ -485,6 +555,54 @@ void main() { const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z; #endif +#ifdef COOPMAT +#ifdef MUL_MAT_ID + [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { + [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { + coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + + [[unroll]] for (uint col = 0; col < BN; col += storestride) { + const uint row_i = dc + cm_col * TN + col + store_c; + if (row_i >= _ne1) break; + + const u16vec2 row_idx = row_ids[row_i]; + + data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]); + } + } + } +#else + const bool is_aligned = p.stride_d % 4 == 0; // Assumption: D_TYPE == float + + [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { + [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { + const bool is_in_bounds = dr + (cm_row + 1) * TM <= p.M && dc + (cm_col + 1) * TN <= p.N; + + if (is_aligned && is_in_bounds) { + // Full coopMat is within bounds and stride_d is aligned with 16B + coopmat cm_dtype = coopmat(sums[cm_col * cms_per_row + cm_row]); + coopMatStore(cm_dtype, data_d, offsets + (dc + cm_col * TN) * p.stride_d + dr + cm_row * TM, p.stride_d, gl_CooperativeMatrixLayoutColumnMajor); + } else if (is_in_bounds) { + // Full coopMat is within bounds, but stride_d is not aligned + coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + + [[unroll]] for (uint col = 0; col < TN; col += storestride) { + data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]); + } + } else if (dr + cm_row * TM < p.M && dc + cm_col * TN < p.N) { + // Partial coopMat is within bounds + coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + + [[unroll]] for (uint col = 0; col < TN; col += storestride) { + if (dr + cm_row * TM + store_r < p.M && dc + cm_col * TN + col + store_c < p.N) { + data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]); + } + } + } + } + } +#endif // MUL_MAT_ID +#else [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) { [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { @@ -496,7 +614,7 @@ void main() { if (row_i >= _ne1) break; const u16vec2 row_idx = row_ids[row_i]; -#endif +#endif // MUL_MAT_ID [[unroll]] for (uint cr = 0; cr < TM; cr++) { #ifdef MUL_MAT_ID data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]); @@ -504,9 +622,10 @@ void main() { if (dr_warp + cr < p.M && dc_warp + cc < p.N) { data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]); } -#endif +#endif // MUL_MAT_ID } } } } +#endif // COOPMAT } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp b/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp index 74630dc7f..495f966bd 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp @@ -16,6 +16,5 @@ void main() { if (i >= p.KX) { return; } - - data_d[i] = D_TYPE(tanh(data_a[i])); + data_d[i] = D_TYPE(1. - 2. / (exp(2.*data_a[i]) + 1.)); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat2_support.comp b/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat2_support.comp new file mode 100644 index 000000000..28eb24e11 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat2_support.comp @@ -0,0 +1,7 @@ +#version 460 + +#extension GL_NV_cooperative_matrix2 : require + +void main() +{ +} 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 4716e2c80..bc6fca506 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -60,6 +60,7 @@ const std::vector type_names = { "iq4_nl" }; +namespace { void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str) { #ifdef _WIN32 HANDLE stdout_read, stdout_write; @@ -198,8 +199,8 @@ static uint32_t compile_count = 0; static std::mutex compile_count_mutex; static std::condition_variable compile_count_cond; -void string_to_spv_func(const std::string& _name, const std::string& in_fname, const std::map& defines, bool fp16 = true, bool coopmat2 = false, bool f16acc = false) { - std::string name = _name + (f16acc ? "_f16acc" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32")); +void string_to_spv_func(const std::string& _name, const std::string& in_fname, const std::map& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false) { + std::string name = _name + (f16acc ? "_f16acc" : "") + (coopmat ? "_coopmat" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32")); std::string out_fname = join_paths(output_dir, name + ".spv"); std::string in_path = join_paths(input_dir, in_fname); @@ -258,7 +259,7 @@ std::map merge_maps(const std::map> compiles; -void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map& defines, bool fp16 = true, bool coopmat2 = false, bool f16acc = false) { +void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false) { { // wait until fewer than N compiles are in progress. // 16 is an arbitrary limit, the goal is to avoid "failed to create pipe" errors. @@ -269,10 +270,10 @@ void string_to_spv(const std::string& _name, const std::string& in_fname, const } compile_count++; } - compiles.push_back(std::async(string_to_spv_func, _name, in_fname, defines, fp16, coopmat2, f16acc)); + compiles.push_back(std::async(string_to_spv_func, _name, in_fname, defines, fp16, coopmat, coopmat2, f16acc)); } -void matmul_shaders(bool fp16, bool matmul_id, bool coopmat2, bool f16acc) { +void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool f16acc) { std::string load_vec = coopmat2 ? "1" : fp16 ? "8" : "4"; std::string aligned_b_type_f32 = coopmat2 ? "float" : fp16 ? "mat2x4" : "vec4"; std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4"; @@ -291,14 +292,20 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat2, bool f16acc) { base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float"; + if (coopmat) { + base_dict["COOPMAT"] = "1"; + } + + base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float"; + std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp"; // Shaders with f16 B_TYPE - string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat2, f16acc); - string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat2, f16acc); + string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); - string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat2, f16acc); - string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat2, f16acc); + string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); for (const auto& tname : type_names) { std::string data_a_key = "DATA_A_" + to_uppercase(tname); @@ -307,12 +314,12 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat2, bool f16acc) { // For aligned matmul loads std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16") ? load_vec : "2"; - string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat2, f16acc); - string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); if (tname != "f16" && tname != "f32") { - string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat2, f16acc); - string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); } } } @@ -322,28 +329,27 @@ void process_shaders() { std::map base_dict = {{"FLOAT_TYPE", "float"}}; // matmul - for (const auto& fp16 : {false, true}) { - for (const auto& matmul_id : {false, true}) { - for (const auto& coopmat2 : {false, true}) { - for (const auto& f16acc : {false, true}) { -#if !defined(VK_NV_cooperative_matrix2) - if (coopmat2) { - continue; - } + for (const auto& matmul_id : {false, true}) { + // No coopmats + // fp32 + matmul_shaders(false, matmul_id, false, false, false); + + // fp16, fp32acc and fp16acc + matmul_shaders(true, matmul_id, false, false, false); + matmul_shaders(true, matmul_id, false, false, true); + + // Coopmat, fp32acc and fp16acc + matmul_shaders(true, matmul_id, true, false, false); + matmul_shaders(true, matmul_id, true, false, true); + +#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + // Coopmat2, fp32acc and fp16acc + matmul_shaders(true, matmul_id, false, true, false); + matmul_shaders(true, matmul_id, false, true, true); #endif - if (coopmat2 && !fp16) { - continue; - } - if (!coopmat2 && f16acc) { - continue; - } - matmul_shaders(fp16, matmul_id, coopmat2, f16acc); - } - } - } } -#if defined(VK_NV_cooperative_matrix2) +#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) // flash attention for (const auto& f16acc : {false, true}) { std::string acctype = f16acc ? "float16_t" : "float"; @@ -355,11 +361,11 @@ void process_shaders() { if (tname == "f16") { string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp", - merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, true, f16acc); + merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, false, true, f16acc); } else { std::string data_a_key = "DATA_A_" + to_uppercase(tname); string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp", - merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, true, f16acc); + merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, true, f16acc); } } } @@ -524,6 +530,7 @@ void write_output_files() { fclose(hdr); fclose(src); } +} int main(int argc, char** argv) { std::map args; diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 2c338dee5..058941c7a 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -8,7 +8,10 @@ // FIXME: required here for quantization functions #include "ggml-quants.h" -#include "ggml-aarch64.h" + +#ifdef GGML_USE_CPU_HBM +#include +#endif #if defined(_MSC_VER) || defined(__MINGW32__) #include // using malloc.h with MSC/MINGW @@ -788,32 +791,23 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = { .to_float = (ggml_to_float_t) ggml_bf16_to_fp32_row, .from_float_ref = (ggml_from_float_t) ggml_fp32_to_bf16_row_ref, }, - [GGML_TYPE_Q4_0_4_4] = { - .type_name = "q4_0_4x4", - .blck_size = QK4_0, - .blck_size_interleave = 4, - .type_size = sizeof(block_q4_0), - .is_quantized = true, - .to_float = NULL, - .from_float_ref = NULL, + [31] = { // GGML_TYPE_Q4_0_4_4 + .type_name = "TYPE_Q4_0_4_4 REMOVED, use Q4_0 with runtime repacking", + .blck_size = 0, + .type_size = 0, + .is_quantized = false, }, - [GGML_TYPE_Q4_0_4_8] = { - .type_name = "q4_0_4x8", - .blck_size = QK4_0, - .blck_size_interleave = 8, - .type_size = sizeof(block_q4_0), - .is_quantized = true, - .to_float = NULL, - .from_float_ref = NULL, + [32] = { // GGML_TYPE_Q4_0_4_8 + .type_name = "TYPE_Q4_0_4_8 REMOVED, use Q4_0 with runtime repacking", + .blck_size = 0, + .type_size = 0, + .is_quantized = false, }, - [GGML_TYPE_Q4_0_8_8] = { - .type_name = "q4_0_8x8", - .blck_size = QK4_0, - .blck_size_interleave = 8, - .type_size = sizeof(block_q4_0), - .is_quantized = true, - .to_float = NULL, - .from_float_ref = NULL, + [33] = { // GGML_TYPE_Q4_0_8_8 + .type_name = "TYPE_Q4_0_8_8 REMOVED, use Q4_0 with runtime repacking", + .blck_size = 0, + .type_size = 0, + .is_quantized = false, }, [GGML_TYPE_TQ1_0] = { .type_name = "tq1_0", @@ -831,14 +825,23 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = { .to_float = (ggml_to_float_t) dequantize_row_tq2_0, .from_float_ref = (ggml_from_float_t) quantize_row_tq2_0_ref, }, - [GGML_TYPE_IQ4_NL_4_4] = { - .type_name = "iq4_nl_4x4", - .blck_size = QK4_NL, - .blck_size_interleave = 4, - .type_size = sizeof(block_iq4_nl), - .is_quantized = true, - .to_float = NULL, - .from_float_ref = NULL, + [36] = { // GGML_TYPE_IQ4_NL_4_4 + .type_name = "TYPE_IQ4_NL_4_4 REMOVED, use IQ4_NL with runtime repacking", + .blck_size = 0, + .type_size = 0, + .is_quantized = false, + }, + [37] = { // GGML_TYPE_IQ4_NL_4_8 + .type_name = "TYPE_IQ4_NL_4_8 REMOVED, use IQ4_NL with runtime repacking", + .blck_size = 0, + .type_size = 0, + .is_quantized = false, + }, + [38] = { // GGML_TYPE_IQ4_NL_8_8 + .type_name = "TYPE_IQ4_NL_8_8 REMOVED, use IQ4_NL with runtime repacking", + .blck_size = 0, + .type_size = 0, + .is_quantized = false, }, }; @@ -1270,9 +1273,6 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_IQ4_XS: wtype = GGML_TYPE_IQ4_XS; break; case GGML_FTYPE_MOSTLY_IQ3_S: wtype = GGML_TYPE_IQ3_S; break; case GGML_FTYPE_MOSTLY_IQ2_S: wtype = GGML_TYPE_IQ2_S; break; - case GGML_FTYPE_MOSTLY_Q4_0_4_4: wtype = GGML_TYPE_Q4_0_4_4; break; - case GGML_FTYPE_MOSTLY_Q4_0_4_8: wtype = GGML_TYPE_Q4_0_4_8; break; - case GGML_FTYPE_MOSTLY_Q4_0_8_8: wtype = GGML_TYPE_Q4_0_8_8; break; case GGML_FTYPE_UNKNOWN: wtype = GGML_TYPE_COUNT; break; case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break; } @@ -6304,9 +6304,6 @@ size_t ggml_quantize_chunk( case GGML_TYPE_IQ1_M: result = quantize_iq1_m (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_NL: result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_IQ4_XS: result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; - case GGML_TYPE_Q4_0_4_4: result = quantize_q4_0_4x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; - case GGML_TYPE_Q4_0_4_8: result = quantize_q4_0_4x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; - case GGML_TYPE_Q4_0_8_8: result = quantize_q4_0_8x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; case GGML_TYPE_F16: { size_t elemsize = sizeof(ggml_fp16_t); @@ -6838,7 +6835,16 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p (int64_t) info->ne[2] * (int64_t) info->ne[3]; - if (ggml_blck_size(info->type) == 0 || ne % ggml_blck_size(info->type) != 0) { + if (ggml_blck_size(info->type) == 0 ) { + // this tensor type support have been removed: + fprintf(stderr, "%s: tensor '%s' of type %d: %s\n", + __func__, info->name.data, (int) info->type, ggml_type_name(info->type)); + fclose(file); + gguf_free(ctx); + return NULL; + } + + if (ne % ggml_blck_size(info->type) != 0) { fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n", __func__, info->name.data, (int) info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type)); fclose(file); diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 703199fcb..4c8710b39 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -761,6 +761,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_NORM, MODEL_TENSOR.ATTN_Q, MODEL_TENSOR.ATTN_K, @@ -1432,9 +1433,6 @@ class GGMLQuantizationType(IntEnum): F64 = 28 IQ1_M = 29 BF16 = 30 - Q4_0_4_4 = 31 - Q4_0_4_8 = 32 - Q4_0_8_8 = 33 TQ1_0 = 34 TQ2_0 = 35 @@ -1478,9 +1476,9 @@ class LlamaFileType(IntEnum): MOSTLY_IQ4_XS = 30 # except 1d tensors MOSTLY_IQ1_M = 31 # except 1d tensors MOSTLY_BF16 = 32 # except 1d tensors - MOSTLY_Q4_0_4_4 = 33 # except 1d tensors - MOSTLY_Q4_0_4_8 = 34 # except 1d tensors - MOSTLY_Q4_0_8_8 = 35 # except 1d tensors + # MOSTLY_Q4_0_4_4 = 33 # removed from gguf files, use Q4_0 and runtime repack + # MOSTLY_Q4_0_4_8 = 34 # removed from gguf files, use Q4_0 and runtime repack + # MOSTLY_Q4_0_8_8 = 35 # removed from gguf files, use Q4_0 and runtime repack MOSTLY_TQ1_0 = 36 # except 1d tensors MOSTLY_TQ2_0 = 37 # except 1d tensors @@ -1556,9 +1554,6 @@ GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = { GGMLQuantizationType.F64: (1, 8), GGMLQuantizationType.IQ1_M: (256, QK_K // 8 + QK_K // 16 + QK_K // 32), GGMLQuantizationType.BF16: (1, 2), - GGMLQuantizationType.Q4_0_4_4:(32, 2 + 16), - GGMLQuantizationType.Q4_0_4_8:(32, 2 + 16), - GGMLQuantizationType.Q4_0_8_8:(32, 2 + 16), GGMLQuantizationType.TQ1_0: (256, 2 + 4 * 13), GGMLQuantizationType.TQ2_0: (256, 2 + 64), } diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 1b6a3f4ad..f0a7b6478 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -146,6 +146,7 @@ class TensorNameMap: # Attention query MODEL_TENSOR.ATTN_Q: ( "model.layers.{bid}.self_attn.q_proj", # llama-hf nemotron olmoe olmo2 + "model.layers.{bid}.self_attn.q_proj_no_perm", # llama-custom "layers.{bid}.attention.wq", # llama-pth "encoder.layer.{bid}.attention.self.query", # bert "transformer.h.{bid}.attn.q_proj", # gpt-j @@ -158,6 +159,7 @@ class TensorNameMap: # Attention key MODEL_TENSOR.ATTN_K: ( "model.layers.{bid}.self_attn.k_proj", # llama-hf nemotron olmoe olmo2 + "model.layers.{bid}.self_attn.k_proj_no_perm", # llama-custom "layers.{bid}.attention.wk", # llama-pth "encoder.layer.{bid}.attention.self.key", # bert "transformer.h.{bid}.attn.k_proj", # gpt-j diff --git a/include/llama.h b/include/llama.h index aa5880096..35db40886 100644 --- a/include/llama.h +++ b/include/llama.h @@ -172,9 +172,9 @@ extern "C" { LLAMA_FTYPE_MOSTLY_IQ4_XS = 30, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ1_M = 31, // except 1d tensors LLAMA_FTYPE_MOSTLY_BF16 = 32, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34, // except 1d tensors - LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35, // except 1d tensors + //LLAMA_FTYPE_MOSTLY_Q4_0_4_4 = 33, // removed from gguf files, use Q4_0 and runtime repack + //LLAMA_FTYPE_MOSTLY_Q4_0_4_8 = 34, // removed from gguf files, use Q4_0 and runtime repack + //LLAMA_FTYPE_MOSTLY_Q4_0_8_8 = 35, // removed from gguf files, use Q4_0 and runtime repack LLAMA_FTYPE_MOSTLY_TQ1_0 = 36, // except 1d tensors LLAMA_FTYPE_MOSTLY_TQ2_0 = 37, // except 1d tensors diff --git a/models/ggml-vocab-roberta-bpe.gguf.inp b/models/ggml-vocab-roberta-bpe.gguf.inp new file mode 100644 index 000000000..9baf7d77a --- /dev/null +++ b/models/ggml-vocab-roberta-bpe.gguf.inp @@ -0,0 +1,112 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ + +__ggml_vocab_test__ + +__ggml_vocab_test__ + +__ggml_vocab_test__ + +__ggml_vocab_test__ + +__ggml_vocab_test__ + + +__ggml_vocab_test__ + + + +__ggml_vocab_test__ + + + + +__ggml_vocab_test__ + + +__ggml_vocab_test__ +Hello world +__ggml_vocab_test__ + Hello world +__ggml_vocab_test__ +Hello World +__ggml_vocab_test__ + Hello World +__ggml_vocab_test__ + Hello World! +__ggml_vocab_test__ +Hello, world! +__ggml_vocab_test__ + Hello, world! +__ggml_vocab_test__ + this is 🦙.cpp +__ggml_vocab_test__ +w048 7tuijk dsdfhu +__ggml_vocab_test__ +нещо на Български +__ggml_vocab_test__ +កាន់តែពិសេសអាចខលចេញ +__ggml_vocab_test__ +🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token) +__ggml_vocab_test__ +Hello +__ggml_vocab_test__ + Hello +__ggml_vocab_test__ + Hello +__ggml_vocab_test__ + Hello +__ggml_vocab_test__ + Hello +__ggml_vocab_test__ + Hello + Hello +__ggml_vocab_test__ + ( +__ggml_vocab_test__ + + = +__ggml_vocab_test__ +' era +__ggml_vocab_test__ +Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ +__ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ +3 +__ggml_vocab_test__ +33 +__ggml_vocab_test__ +333 +__ggml_vocab_test__ +3333 +__ggml_vocab_test__ +33333 +__ggml_vocab_test__ +333333 +__ggml_vocab_test__ +3333333 +__ggml_vocab_test__ +33333333 +__ggml_vocab_test__ +333333333 +__ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ + + + + + + + + + + + +🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````""""......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL +__ggml_vocab_test__ diff --git a/models/ggml-vocab-roberta-bpe.gguf.out b/models/ggml-vocab-roberta-bpe.gguf.out new file mode 100644 index 000000000..f181ac3dc --- /dev/null +++ b/models/ggml-vocab-roberta-bpe.gguf.out @@ -0,0 +1,46 @@ + 2550 204 18430 377 + 597 2768 298 8564 + + 1437 + 1437 1437 + 1437 1437 1437 + 50117 + 50118 + 50140 + 50140 50118 + 50117 50118 + 31414 232 + 20920 232 + 31414 623 + 20920 623 + 20920 623 328 + 31414 6 232 328 + 20920 6 232 328 + 42 16 8103 18164 27 4 49317 + 605 40976 262 10109 18474 385 29 36807 6455 + 36765 25482 22063 23171 34251 18697 10809 26161 18697 3602 22063 27969 40966 25417 15264 26161 24269 36709 41171 35328 + 1376 17772 7471 1376 17772 19002 1376 17772 9085 1376 4333 13859 1376 17772 9357 1376 4333 9264 1376 17772 25448 1376 17772 18400 1376 17772 4333 1376 4333 10172 1376 17772 4333 1376 17772 7258 1376 17772 19002 1376 17772 5782 1376 17772 10172 1376 17772 3726 1376 17772 5782 1376 4333 10172 1376 17772 23171 + 6569 15113 7471 36 21113 43 17841 19002 17 8384 6569 14285 4958 12605 36 34654 2841 4203 354 10146 26511 1070 43 36174 5782 36 8338 21554 14 34 63 308 19233 43 + 31414 + 20920 + 1437 20920 + 1437 1437 20920 + 1437 1437 1437 20920 + 1437 1437 1437 20920 50118 1437 1437 1437 20920 + 36 + 50118 5457 + 108 3567 + 31414 6 1423 108 1250 328 1336 32 47 17841 10172 17487 47876 3602 48617 15264 46537 11423 27326 48494 8210 49233 1558 1570 27761 49429 43251 10809 17772 + 32376 12846 + 246 + 3103 + 25631 + 46152 + 3103 25631 + 46152 3103 + 46152 25631 + 46152 46152 + 46152 3103 25631 + 347 1376 2023 12410 102 16376 1376 2023 6382 90 + 9553 5954 + 50118 1437 50140 1437 50140 50118 1437 50117 1437 50117 50117 1437 50117 50118 1437 1437 50118 1437 1437 1437 50118 1437 1437 1437 1437 50118 1437 1437 1437 1437 1437 50118 6569 15113 7471 36 21113 43 17841 19002 17 8384 6569 14285 4958 12605 36 34654 2841 4203 354 10146 26511 1070 43 36174 5782 8103 18164 27 6569 18164 27 155 2357 30242 155 25631 30242 3103 30242 25631 30242 46152 30242 3103 25631 155 4 246 155 7586 246 155 734 246 25974 17772 7471 1376 17772 19002 1376 17772 9085 1376 4333 13859 1376 17772 9357 1376 4333 9264 1376 17772 25448 1376 17772 18400 1376 17772 4333 1376 4333 10172 1376 17772 4333 1376 17772 7258 1376 17772 19002 1376 17772 5782 18636 10172 17487 47876 3602 48617 15264 46537 11423 27326 48494 8210 49233 1558 1570 27761 49429 43251 10809 17772 36738 48332 47463 18697 10809 25482 22063 23171 34251 18697 10809 26161 18697 3602 22063 27969 40966 25417 15264 26161 24269 36709 41171 35328 128 49690 108 49972 49519 12905 48149 48149 43796 32376 12846 27282 28749 38 348 57 128 41042 37 18 89 6 128 4629 47 686 116 128 448 45 686 38 581 146 24 6 128 495 47 101 103 6845 116 166 108 30660 10 108 462 574 diff --git a/src/llama.cpp b/src/llama.cpp index 7b43359de..14c8d9679 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -4578,9 +4578,6 @@ struct llama_model_loader { case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break; case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break; case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break; - case GGML_TYPE_Q4_0_4_4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_4; break; - case GGML_TYPE_Q4_0_4_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_8; break; - case GGML_TYPE_Q4_0_8_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_8_8; break; default: { LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max)); @@ -5344,9 +5341,6 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 bpw"; case LLAMA_FTYPE_MOSTLY_IQ3_M: return "IQ3_S mix - 3.66 bpw"; - case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4"; - case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: return "Q4_0_4_8"; - case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: return "Q4_0_8_8"; default: return "unknown, may not work"; } @@ -18367,10 +18361,6 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) { new_type = GGML_TYPE_IQ3_S; } - else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || - new_type == GGML_TYPE_Q4_0_8_8) { - new_type = GGML_TYPE_Q4_0; - } else if (ftype == LLAMA_FTYPE_MOSTLY_TQ1_0 || ftype == LLAMA_FTYPE_MOSTLY_TQ2_0) { new_type = GGML_TYPE_Q4_K; } @@ -18693,9 +18683,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_MOSTLY_IQ4_XS: default_type = GGML_TYPE_IQ4_XS; break; case LLAMA_FTYPE_MOSTLY_IQ3_S: default_type = GGML_TYPE_IQ3_S; break; case LLAMA_FTYPE_MOSTLY_IQ3_M: default_type = GGML_TYPE_IQ3_S; break; - case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: default_type = GGML_TYPE_Q4_0_4_4; break; - case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: default_type = GGML_TYPE_Q4_0_4_8; break; - case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: default_type = GGML_TYPE_Q4_0_8_8; break; default: throw std::runtime_error(format("invalid output file type %d\n", ftype)); } @@ -19034,14 +19021,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s f32_data = (float *) f32_conv_buf.data(); } - int chunk_size_multiplier = 1; - if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || new_type == GGML_TYPE_Q4_0_8_8) { - if ((new_type == GGML_TYPE_Q4_0_8_8) && (tensor->ne[1] % 8 != 0)) new_type = GGML_TYPE_Q4_0; - else if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0; - if (new_type == GGML_TYPE_Q4_0_8_8) chunk_size_multiplier = 8; - else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8) chunk_size_multiplier = 4; - } - LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type)); fflush(stdout); @@ -19054,8 +19033,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s const int64_t nrows = tensor->ne[1]; static const int64_t min_chunk_size = 32 * 512; - const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row)) * - chunk_size_multiplier; + const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row)); const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1]; const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;