diff --git a/.devops/main-intel.Dockerfile b/.devops/main-intel.Dockerfile index 572e5d8ea..274b91b71 100644 --- a/.devops/main-intel.Dockerfile +++ b/.devops/main-intel.Dockerfile @@ -10,14 +10,12 @@ WORKDIR /app COPY . . -RUN mkdir build && \ - cd build && \ - if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \ +RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \ echo "LLAMA_SYCL_F16 is set" && \ export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \ fi && \ - cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \ - cmake --build . --config Release --target main + cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \ + cmake --build build --config Release --target main FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime diff --git a/.devops/main-vulkan.Dockerfile b/.devops/main-vulkan.Dockerfile index bca460365..6c2b2ed5b 100644 --- a/.devops/main-vulkan.Dockerfile +++ b/.devops/main-vulkan.Dockerfile @@ -14,10 +14,8 @@ RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key # Build it WORKDIR /app COPY . . -RUN mkdir build && \ - cd build && \ - cmake .. -DLLAMA_VULKAN=1 && \ - cmake --build . --config Release --target main +RUN cmake -B build -DLLAMA_VULKAN=1 && \ + cmake --build build --config Release --target main # Clean up WORKDIR / diff --git a/.devops/server-intel.Dockerfile b/.devops/server-intel.Dockerfile index 304487335..a8e451fa9 100644 --- a/.devops/server-intel.Dockerfile +++ b/.devops/server-intel.Dockerfile @@ -10,14 +10,12 @@ WORKDIR /app COPY . . -RUN mkdir build && \ - cd build && \ - if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \ +RUN if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \ echo "LLAMA_SYCL_F16 is set" && \ export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \ fi && \ - cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \ - cmake --build . --config Release --target server + cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \ + cmake --build build --config Release --target server FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime diff --git a/.devops/server-vulkan.Dockerfile b/.devops/server-vulkan.Dockerfile index 7e5a5283b..6e757e171 100644 --- a/.devops/server-vulkan.Dockerfile +++ b/.devops/server-vulkan.Dockerfile @@ -18,10 +18,8 @@ RUN apt-get update && \ # Build it WORKDIR /app COPY . . -RUN mkdir build && \ - cd build && \ - cmake .. -DLLAMA_VULKAN=1 -DLLAMA_CURL=1 && \ - cmake --build . --config Release --target server +RUN cmake -B build -DLLAMA_VULKAN=1 -DLLAMA_CURL=1 && \ + cmake --build build --config Release --target server # Clean up WORKDIR / diff --git a/.flake8 b/.flake8 index 18fba2c15..bc41c2290 100644 --- a/.flake8 +++ b/.flake8 @@ -1,3 +1,4 @@ [flake8] max-line-length = 125 -ignore = W503 +ignore = E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704,W503 +exclude = examples/*,examples/*/**,*/**/__init__.py,scripts/gen-unicode-data.py,tests/test-tokenizer-0.py diff --git a/.github/workflows/bench.yml b/.github/workflows/bench.yml index 6dc91d27b..3e968d179 100644 --- a/.github/workflows/bench.yml +++ b/.github/workflows/bench.yml @@ -96,9 +96,7 @@ jobs: id: cmake_build run: | set -eux - mkdir build - cd build - cmake .. \ + cmake -B build \ -DLLAMA_NATIVE=OFF \ -DLLAMA_BUILD_SERVER=ON \ -DLLAMA_CURL=ON \ @@ -109,7 +107,7 @@ jobs: -DLLAMA_FATAL_WARNINGS=OFF \ -DLLAMA_ALL_WARNINGS=OFF \ -DCMAKE_BUILD_TYPE=Release; - cmake --build . --config Release -j $(nproc) --target server + cmake --build build --config Release -j $(nproc) --target server - name: Download the dataset id: download_dataset diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 50f76db3c..2d747e688 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -593,6 +593,63 @@ jobs: run: | make swift + windows-msys2: + runs-on: windows-latest + + strategy: + fail-fast: false + matrix: + include: + - { sys: UCRT64, env: ucrt-x86_64, build: Release } + - { sys: CLANG64, env: clang-x86_64, build: Release } + + steps: + - name: Clone + uses: actions/checkout@v4 + + - name: Setup ${{ matrix.sys }} + uses: msys2/setup-msys2@v2 + with: + update: true + msystem: ${{matrix.sys}} + install: >- + base-devel + mingw-w64-${{matrix.env}}-toolchain + mingw-w64-${{matrix.env}}-cmake + mingw-w64-${{matrix.env}}-openblas + + - name: Build using make + shell: msys2 {0} + run: | + make -j $(nproc) + + - name: Clean after building using make + shell: msys2 {0} + run: | + make clean + + - name: Build using make w/ OpenBLAS + shell: msys2 {0} + run: | + make LLAMA_OPENBLAS=1 -j $(nproc) + + - name: Build using CMake + shell: msys2 {0} + run: | + cmake -B build + cmake --build build --config ${{ matrix.build }} -j $(nproc) + + - name: Clean after building using CMake + shell: msys2 {0} + run: | + rm -rf build + + - name: Build using CMake w/ OpenBLAS + shell: msys2 {0} + run: | + cmake -B build -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS + cmake --build build --config ${{ matrix.build }} -j $(nproc) + windows-latest-cmake: runs-on: windows-latest diff --git a/.github/workflows/close-issue.yml b/.github/workflows/close-issue.yml index 7f21daec0..69c9f4f69 100644 --- a/.github/workflows/close-issue.yml +++ b/.github/workflows/close-issue.yml @@ -12,7 +12,7 @@ jobs: steps: - uses: actions/stale@v5 with: - exempt-issue-labels: "refactor,help wanted,good first issue,research" + exempt-issue-labels: "refactor,help wanted,good first issue,research,bug" days-before-issue-stale: 30 days-before-issue-close: 14 stale-issue-label: "stale" diff --git a/.github/workflows/python-lint.yml b/.github/workflows/python-lint.yml index 5be17f157..a8d46f31d 100644 --- a/.github/workflows/python-lint.yml +++ b/.github/workflows/python-lint.yml @@ -20,5 +20,4 @@ jobs: - name: flake8 Lint uses: py-actions/flake8@v2 with: - ignore: "E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704,W503" - exclude: "examples/*,examples/*/**,*/**/__init__.py,convert-hf-to-gguf-update.py" + plugins: "flake8-no-print" diff --git a/.github/workflows/server.yml b/.github/workflows/server.yml index 79cd7d643..afac89c5b 100644 --- a/.github/workflows/server.yml +++ b/.github/workflows/server.yml @@ -94,15 +94,13 @@ jobs: - name: Build id: cmake_build run: | - mkdir build - cd build - cmake .. \ + cmake -B build \ -DLLAMA_NATIVE=OFF \ -DLLAMA_BUILD_SERVER=ON \ -DLLAMA_CURL=ON \ -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \ -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ; - cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server + cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target server - name: Tests @@ -143,10 +141,8 @@ jobs: - name: Build id: cmake_build run: | - mkdir build - cd build - cmake .. -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include" - cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS} --target server + cmake -B build -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include" + cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target server - name: Python setup id: setup_python diff --git a/.gitignore b/.gitignore index 60f9d1f8d..50ae0973a 100644 --- a/.gitignore +++ b/.gitignore @@ -2,6 +2,7 @@ *.a *.so *.gguf +*.gguf.json *.bin *.exe *.dll diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index 65796fe2e..91d791628 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -3,13 +3,14 @@ exclude: prompts/.*.txt repos: - repo: https://github.com/pre-commit/pre-commit-hooks - rev: v3.2.0 + rev: v4.6.0 hooks: - id: trailing-whitespace - id: end-of-file-fixer - id: check-yaml - id: check-added-large-files - repo: https://github.com/PyCQA/flake8 - rev: 6.0.0 + rev: 7.0.0 hooks: - id: flake8 + additional_dependencies: [flake8-no-print] diff --git a/Makefile b/Makefile index 16341b089..c568dd008 100644 --- a/Makefile +++ b/Makefile @@ -39,6 +39,17 @@ ifndef UNAME_M UNAME_M := $(shell uname -m) endif +# In GNU make default CXX is g++ instead of c++. Let's fix that so that users +# of non-gcc compilers don't have to provide g++ alias or wrapper. +DEFCC := cc +DEFCXX := c++ +ifeq ($(origin CC),default) +CC := $(DEFCC) +endif +ifeq ($(origin CXX),default) +CXX := $(DEFCXX) +endif + # Mac OS + Arm can report x86_64 # ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789 ifeq ($(UNAME_S),Darwin) @@ -66,11 +77,10 @@ test: $(TEST_TARGETS) ./$$test_target $(CURDIR)/models/ggml-vocab-llama-bpe.gguf; \ ./$$test_target $(CURDIR)/models/ggml-vocab-phi-3.gguf; \ ./$$test_target $(CURDIR)/models/ggml-vocab-falcon.gguf; \ - ./$$test_target $(CURDIR)/models/ggml-vocab-deepseek-coder.gguf; \ - ./$$test_target $(CURDIR)/models/ggml-vocab-deepseek-llm.gguf; \ ./$$test_target $(CURDIR)/models/ggml-vocab-bert-bge.gguf; \ ./$$test_target $(CURDIR)/models/ggml-vocab-starcoder.gguf; \ ./$$test_target $(CURDIR)/models/ggml-vocab-gpt-2.gguf; \ + ./$$test_target $(CURDIR)/models/ggml-vocab-refact.gguf; \ elif [ "$$test_target" = "tests/test-tokenizer-1-spm" ]; then \ continue; \ elif [ "$$test_target" = "tests/test-tokenizer-1-bpe" ]; then \ diff --git a/README-sycl.md b/README-sycl.md index dc98c7b3e..cfa248a95 100644 --- a/README-sycl.md +++ b/README-sycl.md @@ -185,9 +185,8 @@ Upon a successful installation, SYCL is enabled for the available intel devices, ```sh git clone https://github.com/oneapi-src/oneMKL cd oneMKL -mkdir -p buildWithCublas && cd buildWithCublas -cmake ../ -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON -DTARGET_DOMAINS=blas -make +cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON -DTARGET_DOMAINS=blas +cmake --build buildWithCublas --config Release ``` @@ -227,16 +226,15 @@ Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA devic source /opt/intel/oneapi/setvars.sh # Build LLAMA with MKL BLAS acceleration for intel GPU -mkdir -p build && cd build # Option 1: Use FP32 (recommended for better performance in most cases) -cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx +cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx # Option 2: Use FP16 -cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON +cmake -B build -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON -#build all binary -cmake --build . --config Release -j -v +# build all binary +cmake --build build --config Release -j -v ``` #### Nvidia GPU @@ -248,16 +246,15 @@ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_ export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR # Build LLAMA with Nvidia BLAS acceleration through SYCL -mkdir -p build && cd build # Option 1: Use FP32 (recommended for better performance in most cases) -cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx +cmake -B build -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx # Option 2: Use FP16 -cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON +cmake -B build -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON -#build all binary -cmake --build . --config Release -j -v +# build all binary +cmake --build build --config Release -j -v ``` @@ -412,17 +409,15 @@ b. Download & install mingw-w64 make for Windows provided by w64devkit On the oneAPI command line window, step into the llama.cpp main directory and run the following: ``` -mkdir -p build -cd build @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force # Option 1: Use FP32 (recommended for better performance in most cases) -cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release +cmake -B build -G "MinGW Makefiles" -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release # Option 2: Or FP16 -cmake -G "MinGW Makefiles" .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON +cmake -B build -G "MinGW Makefiles" -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON -make -j +cmake --build build --config Release -j ``` Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions: diff --git a/README.md b/README.md index cc667f592..514ef3af1 100644 --- a/README.md +++ b/README.md @@ -20,7 +20,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) ### Hot topics -- **MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387** +- **BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920** +- MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387 - Model sharding instructions using `gguf-split` https://github.com/ggerganov/llama.cpp/discussions/6404 - Fix major bug in Metal batched inference https://github.com/ggerganov/llama.cpp/pull/6225 - Multi-GPU pipeline parallelism support https://github.com/ggerganov/llama.cpp/pull/6017 @@ -307,6 +308,8 @@ In order to build llama.cpp you have three different options. make ``` + **Note**: for `Debug` builds, run `make LLAMA_DEBUG=1` + - On Windows: 1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases). @@ -321,12 +324,26 @@ In order to build llama.cpp you have three different options. - Using `CMake`: ```bash - mkdir build - cd build - cmake .. - cmake --build . --config Release + cmake -B build + cmake --build build --config Release ``` + **Note**: for `Debug` builds, there are two cases: + + - Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag): + + ```bash + cmake -B build -DCMAKE_BUILD_TYPE=Debug + cmake --build build + ``` + + - Multi-config generators (`-G` param set to Visual Studio, XCode...): + + ```bash + cmake -B build -G "Xcode" + cmake --build build --config Debug + ``` + - Using `Zig` (version 0.11 or later): Building for optimization levels and CPU features can be accomplished using standard build arguments, for example AVX2, FMA, F16C, @@ -438,10 +455,8 @@ Building the program with BLAS support may lead to some performance improvements - Using `CMake` on Linux: ```bash - mkdir build - cd build - cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS - cmake --build . --config Release + cmake -B build -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS + cmake --build build --config Release ``` - #### BLIS @@ -461,11 +476,9 @@ Building the program with BLAS support may lead to some performance improvements - Using manual oneAPI installation: By default, `LLAMA_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DLLAMA_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps: ```bash - mkdir build - cd build source /opt/intel/oneapi/setvars.sh # You can skip this step if in oneapi-basekit docker image, only required for manual installation - cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_NATIVE=ON - cmake --build . --config Release + cmake -B build -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_NATIVE=ON + cmake --build build --config Release ``` - Using oneAPI docker image: @@ -486,10 +499,8 @@ Building the program with BLAS support may lead to some performance improvements - Using `CMake`: ```bash - mkdir build - cd build - cmake .. -DLLAMA_CUDA=ON - cmake --build . --config Release + cmake -B build -DLLAMA_CUDA=ON + cmake --build build --config Release ``` The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance: @@ -516,8 +527,8 @@ Building the program with BLAS support may lead to some performance improvements - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU): ```bash CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ \ - cmake -H. -Bbuild -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \ - && cmake --build build -- -j 16 + cmake -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \ + && cmake --build build --config Release -- -j 16 ``` On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON"`. However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs). @@ -563,15 +574,14 @@ Building the program with BLAS support may lead to some performance improvements ```sh git clone --recurse-submodules https://github.com/KhronosGroup/OpenCL-SDK.git - mkdir OpenCL-SDK/build - cd OpenCL-SDK/build - cmake .. -DBUILD_DOCS=OFF \ + cd OpenCL-SDK + cmake -B build -DBUILD_DOCS=OFF \ -DBUILD_EXAMPLES=OFF \ -DBUILD_TESTING=OFF \ -DOPENCL_SDK_BUILD_SAMPLES=OFF \ -DOPENCL_SDK_TEST_SAMPLES=OFF - cmake --build . --config Release - cmake --install . --prefix /some/path + cmake --build build + cmake --install build --prefix /some/path ``` @@ -593,23 +603,23 @@ Building the program with BLAS support may lead to some performance improvements ```cmd set OPENCL_SDK_ROOT="C:/OpenCL-SDK-v2023.04.17-Win-x64" git clone https://github.com/CNugteren/CLBlast.git - mkdir CLBlast\build - cd CLBlast\build - cmake .. -DBUILD_SHARED_LIBS=OFF -DOVERRIDE_MSVC_FLAGS_TO_MT=OFF -DTUNERS=OFF -DOPENCL_ROOT=%OPENCL_SDK_ROOT% -G "Visual Studio 17 2022" -A x64 - cmake --build . --config Release - cmake --install . --prefix C:/CLBlast + cd CLBlast + cmake -B build -DBUILD_SHARED_LIBS=OFF -DOVERRIDE_MSVC_FLAGS_TO_MT=OFF -DTUNERS=OFF -DOPENCL_ROOT=%OPENCL_SDK_ROOT% -G "Visual Studio 17 2022" -A x64 + cmake --build build --config Release + cmake --install build --prefix C:/CLBlast ``` + (note: `--config Release` at build time is the default and only relevant for Visual Studio builds - or multi-config Ninja builds) + -
Unix: ```sh git clone https://github.com/CNugteren/CLBlast.git - mkdir CLBlast/build - cd CLBlast/build - cmake .. -DBUILD_SHARED_LIBS=OFF -DTUNERS=OFF - cmake --build . --config Release - cmake --install . --prefix /some/path + cd CLBlast + cmake -B build -DBUILD_SHARED_LIBS=OFF -DTUNERS=OFF + cmake --build build --config Release + cmake --install build --prefix /some/path ``` Where `/some/path` is where the built library will be installed (default is `/usr/local`). @@ -623,21 +633,17 @@ Building the program with BLAS support may lead to some performance improvements ``` - CMake (Unix): ```sh - mkdir build - cd build - cmake .. -DLLAMA_CLBLAST=ON -DCLBlast_DIR=/some/path - cmake --build . --config Release + cmake -B build -DLLAMA_CLBLAST=ON -DCLBlast_DIR=/some/path + cmake --build build --config Release ``` - CMake (Windows): ```cmd set CL_BLAST_CMAKE_PKG="C:/CLBlast/lib/cmake/CLBlast" git clone https://github.com/ggerganov/llama.cpp cd llama.cpp - mkdir build - cd build - cmake .. -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=%CL_BLAST_CMAKE_PKG% -G "Visual Studio 17 2022" -A x64 - cmake --build . --config Release - cmake --install . --prefix C:/LlamaCPP + cmake -B build -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=%CL_BLAST_CMAKE_PKG% -G "Visual Studio 17 2022" -A x64 + cmake --build build --config Release + cmake --install build --prefix C:/LlamaCPP ``` ##### Running Llama with CLBlast @@ -693,10 +699,8 @@ Building the program with BLAS support may lead to some performance improvements Then, build llama.cpp using the cmake command below: ```bash - mkdir -p build - cd build - cmake .. -DLLAMA_VULKAN=1 - cmake --build . --config Release + cmake -B build -DLLAMA_VULKAN=1 + cmake --build build --config Release # Test the output binary (with "-ngl 33" to offload all layers to GPU) ./bin/main -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4 diff --git a/ci/run.sh b/ci/run.sh index da05f0d48..bf21b6b31 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -161,7 +161,7 @@ function gg_run_test_scripts_debug { set -e # TODO: too slow, run on dedicated node - (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log + #(cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log #(cd ./examples/quantize && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log set +e @@ -336,7 +336,8 @@ function gg_run_open_llama_3b_v2 { (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log - (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log function check_ppl { qnt="$1" @@ -517,7 +518,10 @@ function gg_run_open_llama_7b_v2 { (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log - (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/save-load-state -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/save-load-state -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log function check_ppl { qnt="$1" diff --git a/common/common.cpp b/common/common.cpp index aa494291d..243b88abf 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -67,7 +67,6 @@ #include #endif #define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083 -#define LLAMA_CURL_MAX_HEADER_LENGTH 256 #endif // LLAMA_USE_CURL using json = nlohmann::ordered_json; @@ -893,7 +892,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa invalid_param = true; return true; } - params.image = argv[i]; + params.image.emplace_back(argv[i]); return true; } if (arg == "-i" || arg == "--interactive") { @@ -948,6 +947,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa params.cont_batching = true; return true; } + if (arg == "-fa" || arg == "--flash-attn") { + params.flash_attn = true; + return true; + } if (arg == "--color") { params.use_color = true; return true; @@ -1324,6 +1327,29 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa return false; } +void gpt_params_handle_model_default(gpt_params & params) { + if (!params.hf_repo.empty()) { + // short-hand to avoid specifying --hf-file -> default it to --model + if (params.hf_file.empty()) { + if (params.model.empty()) { + throw std::invalid_argument("error: --hf-repo requires either --hf-file or --model\n"); + } + params.hf_file = params.model; + } else if (params.model.empty()) { + params.model = "models/" + string_split(params.hf_file, '/').back(); + } + } else if (!params.model_url.empty()) { + if (params.model.empty()) { + auto f = string_split(params.model_url, '#').front(); + f = string_split(f, '?').front(); + f = string_split(f, '/').back(); + params.model = "models/" + f; + } + } else if (params.model.empty()) { + params.model = DEFAULT_MODEL_PATH; + } +} + bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { bool invalid_param = false; std::string arg; @@ -1352,10 +1378,7 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n"); } - // short-hand to avoid specifying --hf-file -> default it to --model - if (!params.hf_repo.empty() && params.hf_file.empty()) { - params.hf_file = params.model; - } + gpt_params_handle_model_default(params); if (params.escape) { process_escapes(params.prompt); @@ -1494,8 +1517,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" -ns N, --sequences N number of sequences to decode (default: %d)\n", params.n_sequences); printf(" -ps N, --p-split N speculative decoding split probability (default: %.1f)\n", (double)params.p_split); printf(" -cb, --cont-batching enable continuous batching (a.k.a dynamic batching) (default: disabled)\n"); + printf(" -fa, --flash-attn enable Flash Attention (default: %s)\n", params.flash_attn ? "enabled" : "disabled"); printf(" --mmproj MMPROJ_FILE path to a multimodal projector file for LLaVA. see examples/llava/README.md\n"); - printf(" --image IMAGE_FILE path to an image file. use with multimodal models\n"); + printf(" --image IMAGE_FILE path to an image file. use with multimodal models. Specify multiple times for batching\n"); if (llama_supports_mlock()) { printf(" --mlock force system to keep model in RAM rather than swapping or compressing\n"); } @@ -1548,7 +1572,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --control-vector-layer-range START END\n"); printf(" layer range to apply the control vector(s) to, start and end inclusive\n"); printf(" -m FNAME, --model FNAME\n"); - printf(" model path (default: %s)\n", params.model.c_str()); + printf(" model path (default: models/$filename with filename from --hf-file or --model-url if set, otherwise %s)\n", DEFAULT_MODEL_PATH); printf(" -md FNAME, --model-draft FNAME\n"); printf(" draft model for speculative decoding (default: unused)\n"); printf(" -mu MODEL_URL, --model-url MODEL_URL\n"); @@ -1866,6 +1890,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param cparams.cb_eval = params.cb_eval; cparams.cb_eval_user_data = params.cb_eval_user_data; cparams.offload_kqv = !params.no_kv_offload; + cparams.flash_attn = params.flash_attn; cparams.type_k = kv_cache_type_from_str(params.cache_type_k); cparams.type_v = kv_cache_type_from_str(params.cache_type_v); @@ -1896,59 +1921,75 @@ void llama_batch_add( #ifdef LLAMA_USE_CURL -static bool llama_download_file(CURL * curl, const char * url, const char * path) { +static bool starts_with(const std::string & str, const std::string & prefix) { + // While we wait for C++20's std::string::starts_with... + return str.rfind(prefix, 0) == 0; +} + +static bool llama_download_file(const std::string & url, const std::string & path) { + + // Initialize libcurl + std::unique_ptr curl(curl_easy_init(), &curl_easy_cleanup); + if (!curl) { + fprintf(stderr, "%s: error initializing libcurl\n", __func__); + return false; + } + bool force_download = false; // Set the URL, allow to follow http redirection - curl_easy_setopt(curl, CURLOPT_URL, url); - curl_easy_setopt(curl, CURLOPT_FOLLOWLOCATION, 1L); + curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str()); + curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L); #if defined(_WIN32) // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of // operating system. Currently implemented under MS-Windows. - curl_easy_setopt(curl, CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA); + curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA); #endif // Check if the file already exists locally struct stat model_file_info; - auto file_exists = (stat(path, &model_file_info) == 0); + auto file_exists = (stat(path.c_str(), &model_file_info) == 0); - // If the file exists, check for ${path_model}.etag or ${path_model}.lastModified files - char etag[LLAMA_CURL_MAX_HEADER_LENGTH] = {0}; - char etag_path[PATH_MAX] = {0}; - snprintf(etag_path, sizeof(etag_path), "%s.etag", path); - - char last_modified[LLAMA_CURL_MAX_HEADER_LENGTH] = {0}; - char last_modified_path[PATH_MAX] = {0}; - snprintf(last_modified_path, sizeof(last_modified_path), "%s.lastModified", path); + // If the file exists, check its JSON metadata companion file. + std::string metadata_path = path + ".json"; + nlohmann::json metadata; + std::string etag; + std::string last_modified; if (file_exists) { - auto * f_etag = fopen(etag_path, "r"); - if (f_etag) { - if (!fgets(etag, sizeof(etag), f_etag)) { - fprintf(stderr, "%s: unable to read file %s\n", __func__, etag_path); - } else { - fprintf(stderr, "%s: previous file found %s: %s\n", __func__, etag_path, etag); + // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block). + std::ifstream metadata_in(metadata_path); + if (metadata_in.good()) { + try { + metadata_in >> metadata; + fprintf(stderr, "%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str()); + if (metadata.contains("url") && metadata["url"].is_string()) { + auto previous_url = metadata["url"].get(); + if (previous_url != url) { + fprintf(stderr, "%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str()); + return false; + } + } + if (metadata.contains("etag") && metadata["etag"].is_string()) { + etag = metadata["etag"]; + } + if (metadata.contains("lastModified") && metadata["lastModified"].is_string()) { + last_modified = metadata["lastModified"]; + } + } catch (const nlohmann::json::exception & e) { + fprintf(stderr, "%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what()); + return false; } - fclose(f_etag); - } - - auto * f_last_modified = fopen(last_modified_path, "r"); - if (f_last_modified) { - if (!fgets(last_modified, sizeof(last_modified), f_last_modified)) { - fprintf(stderr, "%s: unable to read file %s\n", __func__, last_modified_path); - } else { - fprintf(stderr, "%s: previous file found %s: %s\n", __func__, last_modified_path, - last_modified); - } - fclose(f_last_modified); } + } else { + fprintf(stderr, "%s: no previous model file found %s\n", __func__, path.c_str()); } // Send a HEAD request to retrieve the etag and last-modified headers struct llama_load_model_from_url_headers { - char etag[LLAMA_CURL_MAX_HEADER_LENGTH] = {0}; - char last_modified[LLAMA_CURL_MAX_HEADER_LENGTH] = {0}; + std::string etag; + std::string last_modified; }; llama_load_model_from_url_headers headers; { @@ -1956,38 +1997,37 @@ static bool llama_download_file(CURL * curl, const char * url, const char * path auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t { llama_load_model_from_url_headers *headers = (llama_load_model_from_url_headers *) userdata; - // Convert header field name to lowercase - for (size_t i = 0; i < n_items && buffer[i] != ':'; ++i) { - buffer[i] = tolower(buffer[i]); - } + static std::regex header_regex("([^:]+): (.*)\r\n"); + static std::regex etag_regex("ETag", std::regex_constants::icase); + static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase); - const char * etag_prefix = "etag: "; - if (strncmp(buffer, etag_prefix, strlen(etag_prefix)) == 0) { - strncpy(headers->etag, buffer + strlen(etag_prefix), n_items - strlen(etag_prefix) - 2); // Remove CRLF - } - - const char * last_modified_prefix = "last-modified: "; - if (strncmp(buffer, last_modified_prefix, strlen(last_modified_prefix)) == 0) { - strncpy(headers->last_modified, buffer + strlen(last_modified_prefix), - n_items - strlen(last_modified_prefix) - 2); // Remove CRLF + std::string header(buffer, n_items); + std::smatch match; + if (std::regex_match(header, match, header_regex)) { + const std::string & key = match[1]; + const std::string & value = match[2]; + if (std::regex_match(key, match, etag_regex)) { + headers->etag = value; + } else if (std::regex_match(key, match, last_modified_regex)) { + headers->last_modified = value; + } } return n_items; }; - curl_easy_setopt(curl, CURLOPT_NOBODY, 1L); // will trigger the HEAD verb - curl_easy_setopt(curl, CURLOPT_NOPROGRESS, 1L); // hide head request progress - curl_easy_setopt(curl, CURLOPT_HEADERFUNCTION, static_cast(header_callback)); - curl_easy_setopt(curl, CURLOPT_HEADERDATA, &headers); + curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb + curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress + curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast(header_callback)); + curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers); - CURLcode res = curl_easy_perform(curl); + CURLcode res = curl_easy_perform(curl.get()); if (res != CURLE_OK) { - curl_easy_cleanup(curl); fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res)); return false; } long http_code = 0; - curl_easy_getinfo(curl, CURLINFO_RESPONSE_CODE, &http_code); + curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code); if (http_code != 200) { // HEAD not supported, we don't know if the file has changed // force trigger downloading @@ -1996,28 +2036,30 @@ static bool llama_download_file(CURL * curl, const char * url, const char * path } } - // If the ETag or the Last-Modified headers are different: trigger a new download - bool should_download = !file_exists - || force_download - || (strlen(headers.etag) > 0 && strcmp(etag, headers.etag) != 0) - || (strlen(headers.last_modified) > 0 && strcmp(last_modified, headers.last_modified) != 0); + bool should_download = !file_exists || force_download; + if (!should_download) { + if (!etag.empty() && etag != headers.etag) { + fprintf(stderr, "%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str()); + should_download = true; + } else if (!last_modified.empty() && last_modified != headers.last_modified) { + fprintf(stderr, "%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str()); + should_download = true; + } + } if (should_download) { - char path_temporary[PATH_MAX] = {0}; - snprintf(path_temporary, sizeof(path_temporary), "%s.downloadInProgress", path); + std::string path_temporary = path + ".downloadInProgress"; if (file_exists) { - fprintf(stderr, "%s: deleting previous downloaded file: %s\n", __func__, path); - if (remove(path) != 0) { - curl_easy_cleanup(curl); - fprintf(stderr, "%s: unable to delete file: %s\n", __func__, path); + fprintf(stderr, "%s: deleting previous downloaded file: %s\n", __func__, path.c_str()); + if (remove(path.c_str()) != 0) { + fprintf(stderr, "%s: unable to delete file: %s\n", __func__, path.c_str()); return false; } } // Set the output file - auto * outfile = fopen(path_temporary, "wb"); + std::unique_ptr outfile(fopen(path_temporary.c_str(), "wb"), fclose); if (!outfile) { - curl_easy_cleanup(curl); - fprintf(stderr, "%s: error opening local file for writing: %s\n", __func__, path); + fprintf(stderr, "%s: error opening local file for writing: %s\n", __func__, path.c_str()); return false; } @@ -2025,12 +2067,12 @@ static bool llama_download_file(CURL * curl, const char * url, const char * path auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t { return fwrite(data, size, nmemb, (FILE *)fd); }; - curl_easy_setopt(curl, CURLOPT_NOBODY, 0L); - curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, static_cast(write_callback)); - curl_easy_setopt(curl, CURLOPT_WRITEDATA, outfile); + curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L); + curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast(write_callback)); + curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get()); // display download progress - curl_easy_setopt(curl, CURLOPT_NOPROGRESS, 0L); + curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L); // helper function to hide password in URL auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string { @@ -2049,51 +2091,34 @@ static bool llama_download_file(CURL * curl, const char * url, const char * path // start the download fprintf(stderr, "%s: downloading from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__, - llama_download_hide_password_in_url(url).c_str(), path, headers.etag, headers.last_modified); - auto res = curl_easy_perform(curl); + llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str()); + auto res = curl_easy_perform(curl.get()); if (res != CURLE_OK) { - fclose(outfile); - curl_easy_cleanup(curl); fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res)); return false; } long http_code = 0; - curl_easy_getinfo (curl, CURLINFO_RESPONSE_CODE, &http_code); + curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code); if (http_code < 200 || http_code >= 400) { - fclose(outfile); - curl_easy_cleanup(curl); fprintf(stderr, "%s: invalid http status code received: %ld\n", __func__, http_code); return false; } - // Clean up - fclose(outfile); + // Causes file to be closed explicitly here before we rename it. + outfile.reset(); - // Write the new ETag to the .etag file - if (strlen(headers.etag) > 0) { - auto * etag_file = fopen(etag_path, "w"); - if (etag_file) { - fputs(headers.etag, etag_file); - fclose(etag_file); - fprintf(stderr, "%s: file etag saved %s: %s\n", __func__, etag_path, headers.etag); - } - } + // Write the updated JSON metadata file. + metadata.update({ + {"url", url}, + {"etag", headers.etag}, + {"lastModified", headers.last_modified} + }); + std::ofstream(metadata_path) << metadata.dump(4); + fprintf(stderr, "%s: file metadata saved: %s\n", __func__, metadata_path.c_str()); - // Write the new lastModified to the .etag file - if (strlen(headers.last_modified) > 0) { - auto * last_modified_file = fopen(last_modified_path, "w"); - if (last_modified_file) { - fputs(headers.last_modified, last_modified_file); - fclose(last_modified_file); - fprintf(stderr, "%s: file last modified saved %s: %s\n", __func__, last_modified_path, - headers.last_modified); - } - } - - if (rename(path_temporary, path) != 0) { - curl_easy_cleanup(curl); - fprintf(stderr, "%s: unable to rename file: %s to %s\n", __func__, path_temporary, path); + if (rename(path_temporary.c_str(), path.c_str()) != 0) { + fprintf(stderr, "%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str()); return false; } } @@ -2111,15 +2136,7 @@ struct llama_model * llama_load_model_from_url( return NULL; } - // Initialize libcurl - auto * curl = curl_easy_init(); - - if (!curl) { - fprintf(stderr, "%s: error initializing libcurl\n", __func__); - return NULL; - } - - if (!llama_download_file(curl, model_url, path_model)) { + if (!llama_download_file(model_url, path_model)) { return NULL; } @@ -2133,7 +2150,6 @@ struct llama_model * llama_load_model_from_url( auto * ctx_gguf = gguf_init_from_file(path_model, gguf_params); if (!ctx_gguf) { fprintf(stderr, "\n%s: failed to load input GGUF from %s\n", __func__, path_model); - curl_easy_cleanup(curl); return NULL; } @@ -2145,8 +2161,6 @@ struct llama_model * llama_load_model_from_url( gguf_free(ctx_gguf); } - curl_easy_cleanup(curl); - if (n_split > 1) { char split_prefix[PATH_MAX] = {0}; char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0}; @@ -2177,11 +2191,7 @@ struct llama_model * llama_load_model_from_url( char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0}; llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split); - auto * curl = curl_easy_init(); - bool res = llama_download_file(curl, split_url, split_path); - curl_easy_cleanup(curl); - - return res; + return llama_download_file(split_url, split_path); }, idx)); } @@ -2668,7 +2678,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau); fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta); fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false"); - fprintf(stream, "model: %s # default: models/7B/ggml-model.bin\n", params.model.c_str()); + fprintf(stream, "model: %s # default: %s\n", params.model.c_str(), DEFAULT_MODEL_PATH); fprintf(stream, "model_draft: %s # default:\n", params.model_draft.c_str()); fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false"); fprintf(stream, "n_gpu_layers: %d # default: -1\n", params.n_gpu_layers); @@ -2703,6 +2713,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l fprintf(stream, "seed: %u # default: -1 (random seed)\n", params.seed); fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false"); fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false"); + fprintf(stream, "flash_attn: %s # default: false\n", params.flash_attn ? "true" : "false"); fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp); const std::vector tensor_split_vector(params.tensor_split, params.tensor_split + llama_max_devices()); diff --git a/common/common.h b/common/common.h index eea63a114..9252a4b63 100644 --- a/common/common.h +++ b/common/common.h @@ -31,6 +31,8 @@ fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET); \ } while(0) +#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf" + // build info extern int LLAMA_BUILD_NUMBER; extern char const *LLAMA_COMMIT; @@ -92,7 +94,7 @@ struct gpt_params { // // sampling parameters struct llama_sampling_params sparams; - std::string model = "models/7B/ggml-model-f16.gguf"; // model path + std::string model = ""; // model path std::string model_draft = ""; // draft model for speculative decoding std::string model_alias = "unknown"; // model alias std::string model_url = ""; // model url to download @@ -133,7 +135,7 @@ struct gpt_params { bool multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt size_t multiple_choice_tasks = 0; // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed - bool kl_divergence = false; // compute KL-divergence + bool kl_divergence = false; // compute KL divergence bool random_prompt = false; // do not randomize prompt if none provided bool use_color = false; // use color to distinguish generations and inputs @@ -148,6 +150,7 @@ struct gpt_params { bool multiline_input = false; // reverse the usage of `\` bool simple_io = false; // improves compatibility with subprocesses and limited consoles bool cont_batching = true; // insert new sequences for decoding on-the-fly + bool flash_attn = false; // flash attention bool input_prefix_bos = false; // prefix BOS to user inputs, preceding input_prefix bool ignore_eos = false; // ignore generated EOS tokens @@ -167,10 +170,12 @@ struct gpt_params { std::string cache_type_v = "f16"; // KV cache data type for the V // multimodal models (see examples/llava) - std::string mmproj = ""; // path to multimodal projector - std::string image = ""; // path to an image file + std::string mmproj = ""; // path to multimodal projector + std::vector image; // path to image file(s) }; +void gpt_params_handle_model_default(gpt_params & params); + bool parse_kv_override(const char * data, std::vector & overrides); bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params); diff --git a/common/log.h b/common/log.h index 2b2f0e455..6934c57b2 100644 --- a/common/log.h +++ b/common/log.h @@ -234,7 +234,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std:: // INTERNAL, DO NOT USE // USE LOG() INSTEAD // -#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) +#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__) #define LOG_IMPL(str, ...) \ do { \ if (LOG_TARGET != nullptr) \ @@ -257,7 +257,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std:: // INTERNAL, DO NOT USE // USE LOG_TEE() INSTEAD // -#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) +#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__) #define LOG_TEE_IMPL(str, ...) \ do { \ if (LOG_TARGET != nullptr) \ diff --git a/common/sampling.cpp b/common/sampling.cpp index f24665501..cc83600d9 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -68,7 +68,7 @@ void llama_sampling_reset(llama_sampling_context * ctx) { void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) { if (seed == LLAMA_DEFAULT_SEED) { - seed = time(NULL); + seed = std::random_device{}(); } ctx->rng.seed(seed); } diff --git a/convert-hf-to-gguf-update.py b/convert-hf-to-gguf-update.py index 1c559c3f6..917a4469d 100644 --- a/convert-hf-to-gguf-update.py +++ b/convert-hf-to-gguf-update.py @@ -21,6 +21,7 @@ # TODO: automate the update of convert-hf-to-gguf.py # +import logging import os import requests import sys @@ -28,12 +29,18 @@ import json from hashlib import sha256 from enum import IntEnum, auto +from transformers import AutoTokenizer + +logging.basicConfig(level=logging.DEBUG) +logger = logging.getLogger("convert-hf-to-gguf-update") + class TOKENIZER_TYPE(IntEnum): SPM = auto() BPE = auto() WPM = auto() + # TODO: this string has to exercise as much pre-tokenizer functionality as possible # will be updated with time - contributions welcome chktxt = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (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' @@ -41,36 +48,39 @@ chktxt = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍ if len(sys.argv) == 2: token = sys.argv[1] else: - print("Usage: python convert-hf-to-gguf-update.py ") + logger.info("Usage: python convert-hf-to-gguf-update.py ") sys.exit(1) # TODO: add models here, base models preferred models = [ - { "name": "llama-spm", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/meta-llama/Llama-2-7b-hf", }, - { "name": "llama-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Meta-Llama-3-8B", }, - { "name": "phi-3", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct", }, - { "name": "deepseek-llm", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-llm-7b-base", }, - { "name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", }, - { "name": "falcon", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", }, - { "name": "bert-bge", "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", }, - { "name": "mpt", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", }, - { "name": "starcoder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", }, - { "name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", }, - ] + {"name": "llama-spm", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/meta-llama/Llama-2-7b-hf", }, + {"name": "llama-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Meta-Llama-3-8B", }, + {"name": "phi-3", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct", }, + {"name": "deepseek-llm", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-llm-7b-base", }, + {"name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", }, + {"name": "falcon", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", }, + {"name": "bert-bge", "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", }, + {"name": "mpt", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", }, + {"name": "starcoder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", }, + {"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", }, + {"name": "refact", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", }, +] # make directory "models/tokenizers" if it doesn't exist if not os.path.exists("models/tokenizers"): os.makedirs("models/tokenizers") + def download_file_with_auth(url, token, save_path): headers = {"Authorization": f"Bearer {token}"} response = requests.get(url, headers=headers) if response.status_code == 200: with open(save_path, 'wb') as f: f.write(response.content) - print(f"File {save_path} downloaded successfully") + logger.info(f"File {save_path} downloaded successfully") else: - print(f"Failed to download file. Status code: {response.status_code}") + logger.info(f"Failed to download file. Status code: {response.status_code}") + # download the tokenizer models for model in models: @@ -81,10 +91,10 @@ for model in models: if not os.path.exists(f"models/tokenizers/{name}"): os.makedirs(f"models/tokenizers/{name}") else: - print(f"Directory models/tokenizers/{name} already exists - skipping") + logger.info(f"Directory models/tokenizers/{name} already exists - skipping") continue - print(f"Downloading {name} to models/tokenizers/{name}") + logger.info(f"Downloading {name} to models/tokenizers/{name}") url = f"{repo}/raw/main/config.json" save_path = f"models/tokenizers/{name}/config.json" @@ -115,76 +125,82 @@ for model in models: continue # create the tokenizer - from transformers import AutoTokenizer tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}") chktok = tokenizer.encode(chktxt) chkhsh = sha256(str(chktok).encode()).hexdigest() - print(f"model: {name}") - print(f"tokt: {tokt}") - print(f"repo: {model['repo']}") - print(f"chktok: {chktok}") - print(f"chkhsh: {chkhsh}") + logger.info(f"model: {name}") + logger.info(f"tokt: {tokt}") + logger.info(f"repo: {model['repo']}") + logger.info(f"chktok: {chktok}") + logger.info(f"chkhsh: {chkhsh}") # print the "pre_tokenizer" content from the tokenizer.json - with open(f"models/tokenizers/{name}/tokenizer.json", "r") as f: + with open(f"models/tokenizers/{name}/tokenizer.json", "r", encoding="utf-8") as f: cfg = json.load(f) pre_tokenizer = cfg["pre_tokenizer"] - print("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4)) + logger.info("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4)) - print(f"\n") + logger.info("") src_ifs += f" if chkhsh == \"{chkhsh}\":\n" src_ifs += f" # ref: {model['repo']}\n" src_ifs += f" res = \"{name}\"\n" -src_func = "" -src_func += " def get_vocab_base_pre(self, tokenizer) -> str:\n" -src_func += " # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that\n" -src_func += " # is specific for the BPE pre-tokenizer used by the model\n" -src_func += " # we will use this unique identifier to write a \"tokenizer.ggml.pre\" entry in the GGUF file which we can\n" -src_func += " # use in llama.cpp to implement the same pre-tokenizer\n" -src_func += "\n" -src_func += f" chktxt = {repr(chktxt)}\n" -src_func += "\n" -src_func += " chktok = tokenizer.encode(chktxt)\n" -src_func += " chkhsh = sha256(str(chktok).encode()).hexdigest()\n" -src_func += "\n" -src_func += " print(f\"chktok: {chktok}\")\n" -src_func += " print(f\"chkhsh: {chkhsh}\")\n" -src_func += "\n" -src_func += " res = None\n" -src_func += "\n" -src_func += " # NOTE: if you get an error here, you need to add the model to the if-elif chain below\n" -src_func += " # don't do this manually - use the convert-hf-to-gguf-update.py script!\n" -src_func += f"{src_ifs}\n" -src_func += " if res is None:\n" -src_func += " print(\"\\n\")\n" -src_func += " print(\"**************************************************************************************\")\n" -src_func += " print(\"** WARNING: The BPE pre-tokenizer was not recognized!\")\n" -src_func += " print(\"** This means that it was not added yet or you are using an older version.\")\n" -src_func += " print(\"** Check convert-hf-to-gguf-update.py and update it accordingly.\")\n" -src_func += " print(\"**\")\n" -src_func += " print(f\"** chkhsh: {chkhsh}\")\n" -src_func += " print(\"**************************************************************************************\")\n" -src_func += " print(\"\\n\")\n" -src_func += " raise NotImplementedError(\"BPE pre-tokenizer was not recognized - update get_vocab_base_pre()\")\n" -src_func += "\n" -src_func += " print(f\"tokenizer.ggml.pre: {res}\")\n" -src_func += " print(f\"chkhsh: {chkhsh}\")\n" -src_func += "\n" -src_func += " return res\n" +src_func = f""" + def get_vocab_base_pre(self, tokenizer) -> str: + # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that + # is specific for the BPE pre-tokenizer used by the model + # we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can + # use in llama.cpp to implement the same pre-tokenizer -print(src_func) + chktxt = {repr(chktxt)} -print("\n") -print("!!! Copy-paste the function above into convert-hf-to-gguf.py !!!") -print("\n") + chktok = tokenizer.encode(chktxt) + chkhsh = sha256(str(chktok).encode()).hexdigest() + + logger.debug(f"chktok: {{chktok}}") + logger.debug(f"chkhsh: {{chkhsh}}") + + res = None + + # NOTE: if you get an error here, you need to update the convert-hf-to-gguf-update.py script + # or pull the latest version of the model from Huggingface + # don't edit the hashes manually! +{src_ifs} + if res is None: + logger.warning("\\n") + logger.warning("**************************************************************************************") + logger.warning("** WARNING: The BPE pre-tokenizer was not recognized!") + logger.warning("** There are 2 possible reasons for this:") + logger.warning("** - the model has not been added to convert-hf-to-gguf-update.py yet") + logger.warning("** - the pre-tokenization config has changed upstream") + logger.warning("** Check your model files and convert-hf-to-gguf-update.py and update them accordingly.") + logger.warning("** ref: https://github.com/ggerganov/llama.cpp/pull/6920") + logger.warning("**") + logger.warning(f"** chkhsh: {{chkhsh}}") + logger.warning("**************************************************************************************") + logger.warning("\\n") + raise NotImplementedError("BPE pre-tokenizer was not recognized - update get_vocab_base_pre()") + + logger.debug(f"tokenizer.ggml.pre: {{repr(res)}}") + logger.debug(f"chkhsh: {{chkhsh}}") + + return res +""" + +print(src_func) # noqa: NP100 + +logger.info("\n") +logger.info("!!! Copy-paste the function above into convert-hf-to-gguf.py !!!") +logger.info("\n") # generate tests for each tokenizer model tests = [ + "ied 4 ½ months", + "Führer", "", " ", " ", @@ -246,10 +262,9 @@ for model in models: tokt = model["tokt"] # create the tokenizer - from transformers import AutoTokenizer tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}") - with open(f"models/ggml-vocab-{name}.gguf.inp", "w") as f: + with open(f"models/ggml-vocab-{name}.gguf.inp", "w", encoding="utf-8") as f: for text in tests: f.write(f"{text}") f.write("\n__ggml_vocab_test__\n") @@ -261,15 +276,15 @@ for model in models: f.write(f" {r}") f.write("\n") - print(f"Tests for {name} written in ./models/ggml-vocab-{name}.gguf.*") + logger.info(f"Tests for {name} written in ./models/ggml-vocab-{name}.gguf.*") # generate commands for creating vocab files -print("\nRun the following commands to generate the vocab files for testing:\n") +logger.info("\nRun the following commands to generate the vocab files for testing:\n") for model in models: name = model["name"] - print(f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") + print(f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only") # noqa: NP100 -print("\n") +logger.info("\n") diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 0c49fb3c0..88c16676b 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -2,6 +2,7 @@ from __future__ import annotations +import logging import argparse import contextlib import json @@ -26,6 +27,8 @@ import gguf from convert import LlamaHfVocab, permute +logger = logging.getLogger("hf-to-gguf") + ###### MODEL DEFINITIONS ###### @@ -76,7 +79,7 @@ class Model(ABC): def get_tensors(self) -> Iterator[tuple[str, Tensor]]: for part_name in self.part_names: - print(f"gguf: loading model part '{part_name}'") + logger.info(f"gguf: loading model part '{part_name}'") ctx: ContextManager[Any] if self.is_safetensors: from safetensors import safe_open @@ -95,42 +98,42 @@ class Model(ABC): if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None: self.gguf_writer.add_context_length(n_ctx) - print(f"gguf: context length = {n_ctx}") + logger.info(f"gguf: context length = {n_ctx}") n_embd = self.find_hparam(["hidden_size", "n_embd"]) self.gguf_writer.add_embedding_length(n_embd) - print(f"gguf: embedding length = {n_embd}") + logger.info(f"gguf: embedding length = {n_embd}") if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None: self.gguf_writer.add_feed_forward_length(n_ff) - print(f"gguf: feed forward length = {n_ff}") + logger.info(f"gguf: feed forward length = {n_ff}") n_head = self.find_hparam(["num_attention_heads", "n_head"]) self.gguf_writer.add_head_count(n_head) - print(f"gguf: head count = {n_head}") + logger.info(f"gguf: head count = {n_head}") if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None: self.gguf_writer.add_head_count_kv(n_head_kv) - print(f"gguf: key-value head count = {n_head_kv}") + logger.info(f"gguf: key-value head count = {n_head_kv}") if (rope_theta := self.hparams.get("rope_theta")) is not None: self.gguf_writer.add_rope_freq_base(rope_theta) - print(f"gguf: rope theta = {rope_theta}") + logger.info(f"gguf: rope theta = {rope_theta}") if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None: self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps) - print(f"gguf: rms norm epsilon = {f_rms_eps}") + logger.info(f"gguf: rms norm epsilon = {f_rms_eps}") if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None: self.gguf_writer.add_layer_norm_eps(f_norm_eps) - print(f"gguf: layer norm epsilon = {f_norm_eps}") + logger.info(f"gguf: layer norm epsilon = {f_norm_eps}") if (n_experts := self.hparams.get("num_local_experts")) is not None: self.gguf_writer.add_expert_count(n_experts) - print(f"gguf: expert count = {n_experts}") + logger.info(f"gguf: expert count = {n_experts}") if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None: self.gguf_writer.add_expert_used_count(n_experts_used) - print(f"gguf: experts used count = {n_experts_used}") + logger.info(f"gguf: experts used count = {n_experts_used}") self.gguf_writer.add_file_type(self.ftype) - print(f"gguf: file type = {self.ftype}") + logger.info(f"gguf: file type = {self.ftype}") def write_tensors(self): block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer"))) @@ -151,8 +154,7 @@ class Model(ABC): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -169,7 +171,7 @@ class Model(ABC): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -274,13 +276,14 @@ class Model(ABC): chktok = tokenizer.encode(chktxt) chkhsh = sha256(str(chktok).encode()).hexdigest() - print(f"chktok: {chktok}") - print(f"chkhsh: {chkhsh}") + logger.debug(f"chktok: {chktok}") + logger.debug(f"chkhsh: {chkhsh}") res = None - # NOTE: if you get an error here, you need to add the model to the if-elif chain below - # don't do this manually - use the convert-hf-to-gguf-update.py script! + # NOTE: if you get an error here, you need to update the convert-hf-to-gguf-update.py script + # or pull the latest version of the model from Huggingface + # don't edit the hashes manually! if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5": # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B res = "llama-bpe" @@ -305,21 +308,27 @@ class Model(ABC): if chkhsh == "3ce83efda5659b07b1ad37ca97ca5797ea4285d9b9ab0dc679e4a720c9da7454": # ref: https://huggingface.co/openai-community/gpt2 res = "gpt-2" + if chkhsh == "6221ad2852e85ce96f791f476e0b390cf9b474c9e3d1362f53a24a06dc8220ff": + # ref: https://huggingface.co/smallcloudai/Refact-1_6-base + res = "refact" if res is None: - print("\n") - print("**************************************************************************************") - print("** WARNING: The BPE pre-tokenizer was not recognized!") - print("** This means that it was not added yet or you are using an older version.") - print("** Check convert-hf-to-gguf-update.py and update it accordingly.") - print("**") - print(f"** chkhsh: {chkhsh}") - print("**************************************************************************************") - print("\n") + logger.warning("\n") + logger.warning("**************************************************************************************") + logger.warning("** WARNING: The BPE pre-tokenizer was not recognized!") + logger.warning("** There are 2 possible reasons for this:") + logger.warning("** - the model has not been added to convert-hf-to-gguf-update.py yet") + logger.warning("** - the pre-tokenization config has changed upstream") + logger.warning("** Check your model files and convert-hf-to-gguf-update.py and update them accordingly.") + logger.warning("** ref: https://github.com/ggerganov/llama.cpp/pull/6920") + logger.warning("**") + logger.warning(f"** chkhsh: {chkhsh}") + logger.warning("**************************************************************************************") + logger.warning("\n") raise NotImplementedError("BPE pre-tokenizer was not recognized - update get_vocab_base_pre()") - print(f"tokenizer.ggml.pre: {res}") - print(f"chkhsh: {chkhsh}") + logger.debug(f"tokenizer.ggml.pre: {repr(res)}") + logger.debug(f"chkhsh: {chkhsh}") return res @@ -435,9 +444,7 @@ class Model(ABC): if vocab_size > len(tokens): pad_count = vocab_size - len(tokens) - print( - f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]" - ) + logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]") for i in range(1, pad_count + 1): tokens.append(f"[PAD{i}]") scores.append(-1000.0) @@ -549,7 +556,7 @@ class BloomModel(Model): ), axis=0, ) - print("re-format attention.linear_qkv.weight") + logger.info("re-format attention.linear_qkv.weight") elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name): qkv_bias = data.reshape((n_head, 3, n_embed // n_head)) data = np.concatenate( @@ -560,13 +567,12 @@ class BloomModel(Model): ), axis=0, ) - print("re-format attention.linear_qkv.bias") + logger.info("re-format attention.linear_qkv.bias") # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -583,13 +589,13 @@ class BloomModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"=> {new_name}, shape = {data.shape}, {old_dtype} --> {data.dtype}") + logger.info(f"=> {new_name}, shape = {data.shape}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) if not has_lm_head and name == "word_embeddings.weight": self.gguf_writer.add_tensor("output.weight", data) - print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}") + logger.info(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}") @Model.register("MPTForCausalLM") @@ -649,8 +655,7 @@ class MPTModel(Model): else: new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -667,7 +672,7 @@ class MPTModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -693,8 +698,7 @@ class OrionModel(Model): elif "model_max_length" in self.hparams: ctx_length = self.hparams["model_max_length"] else: - print("gguf: can not find ctx length parameter.") - sys.exit() + raise ValueError("gguf: can not find ctx length parameter.") self.gguf_writer.add_file_type(self.ftype) self.gguf_writer.add_name(self.dir_model.name) @@ -732,8 +736,7 @@ class OrionModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -750,7 +753,7 @@ class OrionModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -775,8 +778,7 @@ class BaichuanModel(Model): elif "model_max_length" in self.hparams: ctx_length = self.hparams["model_max_length"] else: - print("gguf: can not find ctx length parameter.") - sys.exit() + raise ValueError("gguf: can not find ctx length parameter.") self.gguf_writer.add_name(self.dir_model.name) self.gguf_writer.add_source_hf_repo(hf_repo) @@ -805,7 +807,7 @@ class BaichuanModel(Model): for i in range(block_count): if (w := model_kv.get(f"model.layers.{i}.self_attn.W_pack.weight")) is not None: - print(f"Unpacking and permuting layer {i}") + logger.info(f"Unpacking and permuting layer {i}") model_kv[f"model.layers.{i}.self_attn.q_proj.weight"] = \ self._reverse_hf_permute_part(w, 0, head_count, head_count) model_kv[f"model.layers.{i}.self_attn.k_proj.weight"] = \ @@ -830,8 +832,7 @@ class BaichuanModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -848,7 +849,7 @@ class BaichuanModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor: @@ -933,8 +934,7 @@ class XverseModel(Model): elif "model_max_length" in self.hparams: ctx_length = self.hparams["model_max_length"] else: - print("gguf: can not find ctx length parameter.") - sys.exit() + raise ValueError("gguf: can not find ctx length parameter.") self.gguf_writer.add_name(self.dir_model.name) self.gguf_writer.add_source_hf_repo(hf_repo) @@ -983,8 +983,7 @@ class XverseModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -1001,7 +1000,7 @@ class XverseModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor: @@ -1088,8 +1087,7 @@ class FalconModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -1106,7 +1104,7 @@ class FalconModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -1193,8 +1191,7 @@ class RefactModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight",)) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -1211,7 +1208,7 @@ class RefactModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -1260,10 +1257,9 @@ class PersimmonModel(Model): data = data_torch.to(torch.float32).squeeze().numpy() new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -1328,8 +1324,7 @@ class StableLMModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -1346,7 +1341,7 @@ class StableLMModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and not new_name.endswith("_norm.weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.debug(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -1362,8 +1357,7 @@ class StableLMModel(Model): merged_name = f"model.layers.{bid}.self_attn.{layer_name}.weight" new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") if self.ftype == 1 and data_dtype == np.float16 and (n_dims == 1 or new_name.endswith("_norm.weight")): data = data.astype(np.float32) @@ -1371,7 +1365,7 @@ class StableLMModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and not new_name.endswith("_norm.weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}") + logger.debug(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -1423,7 +1417,7 @@ class LlamaModel(Model): experts = dict() for name, data_torch in self.get_tensors(): # we don't need these - if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")): + if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")): continue old_dtype = data_torch.dtype @@ -1476,10 +1470,9 @@ class LlamaModel(Model): new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") - print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) continue @@ -1487,8 +1480,7 @@ class LlamaModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -1505,7 +1497,7 @@ class LlamaModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -1580,10 +1572,9 @@ class GrokModel(Model): new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") - print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) continue @@ -1591,8 +1582,7 @@ class GrokModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -1609,7 +1599,7 @@ class GrokModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -1642,7 +1632,7 @@ class DbrxModel(Model): self.gguf_writer.add_layer_norm_eps(1e-5) self.gguf_writer.add_file_type(self.ftype) - print(f"gguf: file type = {self.ftype}") + logger.info(f"gguf: file type = {self.ftype}") def write_tensors(self): block_count = self.hparams.get("n_layers") @@ -1685,8 +1675,7 @@ class DbrxModel(Model): # https://huggingface.co/databricks/dbrx-instruct/blob/main/model.safetensors.index.json#L15 new_name = tensor_map.get_name(name if not experts else name + ".weight", try_suffixes=(".weight",)) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -1694,8 +1683,7 @@ class DbrxModel(Model): # Most of the codebase that takes in 1D tensors only handles F32 tensors # and most of the outputs tensors are F32. if data_dtype != np.float32 and n_dims == 1: - print(f"Can not map tensor {name!r}: all 1D tensors must be F32") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}: all 1D tensors must be F32") # if f32 desired, convert any float16 to float32 if self.ftype == 0 and data_dtype == np.float16: @@ -1705,7 +1693,7 @@ class DbrxModel(Model): if self.ftype == 1 and data_dtype == np.float32 and n_dims > 1: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}") + logger.debug(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -1767,8 +1755,7 @@ class MiniCPMModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -1785,7 +1772,7 @@ class MiniCPMModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -1851,8 +1838,7 @@ class QwenModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -1869,7 +1855,7 @@ class QwenModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -1946,10 +1932,9 @@ class Qwen2MoeModel(Model): new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") - print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}") + logger.debug(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) continue @@ -1957,8 +1942,7 @@ class Qwen2MoeModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -1975,7 +1959,7 @@ class Qwen2MoeModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}") + logger.debug(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -2020,8 +2004,7 @@ class GPT2Model(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -2038,13 +2021,13 @@ class GPT2Model(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) # note: GPT2 output is tied to (same as) wte in original model if new_name == "token_embd.weight": - print(f"output.weight, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"output.weight, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor("output.weight", data) @@ -2083,8 +2066,7 @@ class Phi3MiniModel(Model): tokenizer_path = self.dir_model / 'tokenizer.model' if not tokenizer_path.is_file(): - print(f'Error: Missing {tokenizer_path}', file=sys.stderr) - sys.exit(1) + raise ValueError(f'Error: Missing {tokenizer_path}') tokenizer = SentencePieceProcessor(str(tokenizer_path)) @@ -2122,7 +2104,7 @@ class Phi3MiniModel(Model): for key in added_tokens_json: token_id = added_tokens_json[key] if (token_id >= vocab_size): - print(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') + logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') continue tokens[token_id] = key.encode("utf-8") @@ -2204,8 +2186,7 @@ class PlamoModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") # shuffle for broadcasting of gqa in ggml_mul_mat if new_name.endswith("attn_q.weight"): @@ -2236,7 +2217,7 @@ class PlamoModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -2282,8 +2263,7 @@ class CodeShellModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -2300,13 +2280,13 @@ class CodeShellModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) if not has_lm_head and name == "transformer.wte.weight": self.gguf_writer.add_tensor("output.weight", data) - print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}") + logger.info(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}") @Model.register("InternLM2ForCausalLM") @@ -2328,7 +2308,7 @@ class InternLM2Model(Model): toktypes: list[int] = [] if not tokenizer_path.is_file(): - print(f'Error: Missing {tokenizer_path}', file=sys.stderr) + logger.error(f'Error: Missing {tokenizer_path}') sys.exit(1) sentencepiece_model = model.ModelProto() @@ -2345,7 +2325,7 @@ class InternLM2Model(Model): if text == b"\x00": # (TODO): fixme # Hack here and replace the \x00 characters. - print(f"InternLM2 convert token '{text}' to '🐉'!") + logger.debug(f"InternLM2 convert token '{text}' to '🐉'!") text = "🐉" toktype = SentencePieceTokenTypes.NORMAL @@ -2386,7 +2366,7 @@ class InternLM2Model(Model): # TODO: this is a hack, should be fixed # https://github.com/ggerganov/llama.cpp/pull/6745#issuecomment-2067687048 special_vocab.special_token_ids["eos"] = self._try_get_sft_eos(tokenizer) - print(f"Replace eos:{old_eos} with a special token:{special_vocab.special_token_ids['eos']} \ + logger.warning(f"Replace eos:{old_eos} with a special token:{special_vocab.special_token_ids['eos']} \ in chat mode so that the conversation can end normally.") special_vocab.add_to_gguf(self.gguf_writer) @@ -2431,8 +2411,7 @@ in chat mode so that the conversation can end normally.") # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -2449,7 +2428,7 @@ in chat mode so that the conversation can end normally.") if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) def write_tensors(self): @@ -2560,8 +2539,11 @@ class BertModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) data = data_torch.squeeze().numpy() n_dims = len(data.shape) @@ -2577,7 +2559,7 @@ class BertModel(Model): # if f32 desired, convert any float16 to float32 new_dtype = np.float32 - print(f"{new_name}, n_dims = {n_dims}, {data_torch.dtype} --> {new_dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {data_torch.dtype} --> {new_dtype}") if data.dtype != new_dtype: data = data.astype(new_dtype) @@ -2656,7 +2638,7 @@ class GemmaModel(Model): # lm_head is not used in llama.cpp, while autoawq will include this tensor in model # To prevent errors, skip loading lm_head.weight. if name == "lm_head.weight": - print(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.") + logger.debug(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.") continue old_dtype = data_torch.dtype @@ -2673,8 +2655,7 @@ class GemmaModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -2685,7 +2666,7 @@ class GemmaModel(Model): if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -2713,7 +2694,7 @@ class MambaModel(Model): else: # Use the GPT-NeoX tokenizer when no tokenizer files are present tokenizer_path = Path(sys.path[0]) / "models" / "ggml-vocab-gpt-neox.gguf" - print(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'") + logger.warning(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'") neox_reader = gguf.GGUFReader(tokenizer_path, "r") field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL) @@ -2785,17 +2766,16 @@ class MambaModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") if name.endswith(".A_log"): - print("A_log --> A ==> " + new_name) + logger.debug("A_log --> A ==> " + new_name) data_torch = -torch.exp(data_torch) # assuming token_embd.weight is seen before output.weight if tok_embd is not None and new_name == output_name: if torch.equal(tok_embd, data_torch): - print(f"{output_name} is equivalent to {tok_embd_name}, omitting") + logger.debug(f"{output_name} is equivalent to {tok_embd_name}, omitting") continue if new_name == tok_embd_name: tok_embd = data_torch @@ -2818,7 +2798,7 @@ class MambaModel(Model): if self.ftype == 1 and data_dtype == np.float32 and new_weight_name.endswith((".ssm_in", ".ssm_out", "token_embd", "output")) and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -2877,8 +2857,7 @@ class OlmoModel(Model): # map tensor names new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: - print(f"Can not map tensor {name!r}") - sys.exit() + raise ValueError(f"Can not map tensor {name!r}") n_dims = len(data.shape) data_dtype = data.dtype @@ -2895,7 +2874,7 @@ class OlmoModel(Model): if self.ftype == 1 and data_dtype == np.float32 and n_dims == 2: data = data.astype(np.float16) - print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + logger.info(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") self.gguf_writer.add_tensor(new_name, data) @@ -2928,6 +2907,7 @@ def parse_args() -> argparse.Namespace: ) parser.add_argument("--use-temp-file", action="store_true", help="use the tempfile library while processing (helpful when running out of memory, process killed)") parser.add_argument("--model-name", type=str, default=None, help="name of the model") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") return parser.parse_args() @@ -2935,6 +2915,8 @@ def parse_args() -> argparse.Namespace: def main() -> None: args = parse_args() + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + dir_model = args.model if args.awq_path: @@ -2943,15 +2925,15 @@ def main() -> None: tmp_model_path = args.model / "weighted_model" dir_model = tmp_model_path if tmp_model_path.is_dir(): - print(f"{tmp_model_path} exists as a weighted model.") + logger.info(f"{tmp_model_path} exists as a weighted model.") else: tmp_model_path.mkdir(parents=True, exist_ok=True) - print("Saving new weighted model ...") + logger.info("Saving new weighted model ...") add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path)) - print(f"Saved weighted model at {tmp_model_path}.") + logger.info(f"Saved weighted model at {tmp_model_path}.") if not dir_model.is_dir(): - print(f'Error: {args.model} is not a directory', file=sys.stderr) + logger.error(f'Error: {args.model} is not a directory') sys.exit(1) ftype_map = { @@ -2965,7 +2947,7 @@ def main() -> None: # output in the same directory as the model by default fname_out = dir_model / f'ggml-model-{args.outtype}.gguf' - print(f"Loading model: {dir_model.name}") + logger.info(f"Loading model: {dir_model.name}") hparams = Model.load_hparams(dir_model) @@ -2973,20 +2955,20 @@ def main() -> None: model_class = Model.from_model_architecture(hparams["architectures"][0]) model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file) - print("Set model parameters") + logger.info("Set model parameters") model_instance.set_gguf_parameters() - print("Set model tokenizer") + logger.info("Set model tokenizer") model_instance.set_vocab() if args.vocab_only: - print(f"Exporting model vocab to '{fname_out}'") + logger.info(f"Exporting model vocab to '{fname_out}'") model_instance.write_vocab() else: - print(f"Exporting model to '{fname_out}'") + logger.info(f"Exporting model to '{fname_out}'") model_instance.write() - print(f"Model successfully exported to '{fname_out}'") + logger.info(f"Model successfully exported to '{fname_out}'") if __name__ == '__main__': diff --git a/convert-llama-ggml-to-gguf.py b/convert-llama-ggml-to-gguf.py index 5354b748b..9349de3b3 100755 --- a/convert-llama-ggml-to-gguf.py +++ b/convert-llama-ggml-to-gguf.py @@ -1,6 +1,7 @@ #!/usr/bin/env python3 from __future__ import annotations +import logging import argparse import os import struct @@ -14,6 +15,8 @@ if 'NO_LOCAL_GGUF' not in os.environ: sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) import gguf +logger = logging.getLogger("ggml-to-gguf") + class GGMLFormat(IntEnum): GGML = 0 @@ -125,7 +128,6 @@ class Tensor: self.start_offset = offset self.len_bytes = n_bytes offset += n_bytes - # print(n_dims, name_len, dtype, self.dims, self.name, pad) return offset - orig_offset @@ -175,7 +177,7 @@ class GGMLModel: offset += self.validate_header(data, offset) hp = Hyperparameters() offset += hp.load(data, offset) - print(f'* File format: {self.file_format.name}v{self.format_version} with ftype {hp.ftype.name}') + logger.info(f'* File format: {self.file_format.name}v{self.format_version} with ftype {hp.ftype.name}') self.validate_conversion(hp.ftype) vocab = Vocab(load_scores = self.file_format > GGMLFormat.GGML) offset += vocab.load(data, offset, hp.n_vocab) @@ -215,12 +217,12 @@ class GGMLToGGUF: if float(hp.n_head) / float(x) == gqa: n_kv_head = x assert n_kv_head is not None, "Couldn't determine n_kv_head from GQA param" - print(f'- Guessed n_kv_head = {n_kv_head} based on GQA {cfg.gqa}') + logger.info(f'- Guessed n_kv_head = {n_kv_head} based on GQA {cfg.gqa}') self.n_kv_head = n_kv_head self.name_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.LLAMA, ggml_model.hyperparameters.n_layer) def save(self): - print('* Preparing to save GGUF file') + logger.info('* Preparing to save GGUF file') gguf_writer = gguf.GGUFWriter( self.cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], @@ -230,11 +232,11 @@ class GGMLToGGUF: if self.special_vocab is not None: self.special_vocab.add_to_gguf(gguf_writer) self.add_tensors(gguf_writer) - print(" gguf: write header") + logger.info(" gguf: write header") gguf_writer.write_header_to_file() - print(" gguf: write metadata") + logger.info(" gguf: write metadata") gguf_writer.write_kv_data_to_file() - print(" gguf: write tensors") + logger.info(" gguf: write tensors") gguf_writer.write_tensors_to_file() gguf_writer.close() @@ -250,7 +252,7 @@ class GGMLToGGUF: name = cfg.name if cfg.name is not None else cfg.input.name except UnicodeDecodeError: name = None - print('* Adding model parameters and KV items') + logger.info('* Adding model parameters and KV items') if name is not None: gguf_writer.add_name(name) gguf_writer.add_description(desc) @@ -287,7 +289,7 @@ class GGMLToGGUF: toktypes = [] if self.vocab_override is not None: vo = self.vocab_override - print('* Adding vocab item(s)') + logger.info('* Adding vocab item(s)') for (idx, (vbytes, score, ttype)) in enumerate(vo.all_tokens()): tokens.append(vbytes) scores.append(score) @@ -299,7 +301,7 @@ class GGMLToGGUF: if len(toktypes) > 0: gguf_writer.add_token_types(toktypes) return - print(f'* Adding {hp.n_vocab} vocab item(s)') + logger.info(f'* Adding {hp.n_vocab} vocab item(s)') assert len(self.model.vocab.items) >= 3, 'Cannot handle unexpectedly short model vocab' for (tokid, (vbytes, vscore)) in enumerate(self.model.vocab.items): tt = 1 # Normal @@ -334,7 +336,7 @@ class GGMLToGGUF: def add_tensors(self, gguf_writer): tensor_map = self.name_map data = self.data - print(f'* Adding {len(self.model.tensors)} tensor(s)') + logger.info(f'* Adding {len(self.model.tensors)} tensor(s)') for tensor in self.model.tensors: name = str(tensor.name, 'UTF-8') mapped_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) @@ -344,7 +346,6 @@ class GGMLToGGUF: temp = tempdims[1] tempdims[1] = tempdims[0] tempdims[0] = temp - # print(f'+ {tensor.name} | {mapped_name} {tensor.dims} :: {tempdims}') gguf_writer.add_tensor( mapped_name, data[tensor.start_offset:tensor.start_offset + tensor.len_bytes], @@ -401,33 +402,35 @@ def handle_args(): help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir") parser.add_argument("--vocabtype", default="spm,hfft", help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm,hfft)") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") return parser.parse_args() def main(): cfg = handle_args() - print(f'* Using config: {cfg}') - print('\n=== WARNING === Be aware that this conversion script is best-effort. Use a native GGUF model if possible. === WARNING ===\n') + logging.basicConfig(level=logging.DEBUG if cfg.verbose else logging.INFO) + logger.info(f'* Using config: {cfg}') + logger.warning('=== WARNING === Be aware that this conversion script is best-effort. Use a native GGUF model if possible. === WARNING ===') if cfg.model_metadata_dir is None and (cfg.gqa == 1 or cfg.eps == '5.0e-06'): - print('- Note: If converting LLaMA2, specifying "--eps 1e-5" is required. 70B models also need "--gqa 8".') + logger.info('- Note: If converting LLaMA2, specifying "--eps 1e-5" is required. 70B models also need "--gqa 8".') data = np.memmap(cfg.input, mode = 'r') model = GGMLModel() - print('* Scanning GGML input file') + logger.info('* Scanning GGML input file') offset = model.load(data, 0) # noqa - print(f'* GGML model hyperparameters: {model.hyperparameters}') + logger.info(f'* GGML model hyperparameters: {model.hyperparameters}') vocab_override = None params_override = None special_vocab = None if cfg.model_metadata_dir is not None: (params_override, vocab_override, special_vocab) = handle_metadata(cfg, model.hyperparameters) - print('!! Note: When overriding params the --gqa, --eps and --context-length options are ignored.') - print(f'* Overriding params: {params_override}') - print(f'* Overriding vocab: {vocab_override}') - print(f'* Special vocab: {special_vocab}') + logger.info('!! Note: When overriding params the --gqa, --eps and --context-length options are ignored.') + logger.info(f'* Overriding params: {params_override}') + logger.info(f'* Overriding vocab: {vocab_override}') + logger.info(f'* Special vocab: {special_vocab}') else: - print('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n') + logger.warning('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n') if model.file_format == GGMLFormat.GGML: - print('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!') + logger.info('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!') converter = GGMLToGGUF( model, data, cfg, params_override = params_override, @@ -435,7 +438,7 @@ def main(): special_vocab = special_vocab ) converter.save() - print(f'* Successful completion. Output saved to: {cfg.output}') + logger.info(f'* Successful completion. Output saved to: {cfg.output}') if __name__ == '__main__': diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py index 9a9936dec..39536feb9 100755 --- a/convert-lora-to-ggml.py +++ b/convert-lora-to-ggml.py @@ -1,6 +1,7 @@ #!/usr/bin/env python3 from __future__ import annotations +import logging import json import os import struct @@ -15,6 +16,8 @@ if 'NO_LOCAL_GGUF' not in os.environ: sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) import gguf +logger = logging.getLogger("lora-to-gguf") + NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1} @@ -48,11 +51,9 @@ def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_ty if __name__ == '__main__': if len(sys.argv) < 2: - print(f"Usage: python {sys.argv[0]} [arch]") - print( - "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'" - ) - print(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)") + logger.info(f"Usage: python {sys.argv[0]} [arch]") + logger.info("Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'") + logger.info(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)") sys.exit(1) input_json = os.path.join(sys.argv[1], "adapter_config.json") @@ -70,7 +71,7 @@ if __name__ == '__main__': arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama" if arch_name not in gguf.MODEL_ARCH_NAMES.values(): - print(f"Error: unsupported architecture {arch_name}") + logger.error(f"Error: unsupported architecture {arch_name}") sys.exit(1) arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)] @@ -80,21 +81,21 @@ if __name__ == '__main__': params = json.load(f) if params["peft_type"] != "LORA": - print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA") + logger.error(f"Error: unsupported adapter type {params['peft_type']}, expected LORA") sys.exit(1) if params["fan_in_fan_out"] is True: - print("Error: param fan_in_fan_out is not supported") + logger.error("Error: param fan_in_fan_out is not supported") sys.exit(1) if params["bias"] is not None and params["bias"] != "none": - print("Error: param bias is not supported") + logger.error("Error: param bias is not supported") sys.exit(1) # TODO: these seem to be layers that have been trained but without lora. # doesn't seem widely used but eventually should be supported if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0: - print("Error: param modules_to_save is not supported") + logger.error("Error: param modules_to_save is not supported") sys.exit(1) with open(output_path, "wb") as fout: @@ -125,13 +126,13 @@ if __name__ == '__main__': suffix = k[-len(lora_suffixes[0]):] k = k[: -len(lora_suffixes[0])] else: - print(f"Error: unrecognized tensor name {orig_k}") + logger.error(f"Error: unrecognized tensor name {orig_k}") sys.exit(1) tname = name_map.get_name(k) if tname is None: - print(f"Error: could not map tensor name {orig_k}") - print(" Note: the arch parameter must be specified if the model is not llama") + logger.error(f"Error: could not map tensor name {orig_k}") + logger.error(" Note: the arch parameter must be specified if the model is not llama") sys.exit(1) if suffix == ".lora_A.weight": @@ -141,8 +142,8 @@ if __name__ == '__main__': else: assert False - print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB") + logger.info(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB") write_tensor_header(fout, tname, t.shape, t.dtype) t.tofile(fout) - print(f"Converted {input_json} and {input_model} to {output_path}") + logger.info(f"Converted {input_json} and {input_model} to {output_path}") diff --git a/convert-persimmon-to-gguf.py b/convert-persimmon-to-gguf.py index aba575426..07dcade74 100755 --- a/convert-persimmon-to-gguf.py +++ b/convert-persimmon-to-gguf.py @@ -1,6 +1,7 @@ #!/usr/bin/env python3 from __future__ import annotations +import logging import argparse import os import sys @@ -14,6 +15,8 @@ if 'NO_LOCAL_GGUF' not in os.environ: sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) import gguf +logger = logging.getLogger("persimmon-to-gguf") + def _flatten_dict(dct, tensors, prefix=None): assert isinstance(dct, dict) @@ -30,9 +33,9 @@ def _flatten_dict(dct, tensors, prefix=None): def _get_sentencepiece_tokenizer_info(dir_model: Path): tokenizer_path = dir_model / 'adept_vocab.model' - print('gguf: getting sentencepiece tokenizer from', tokenizer_path) + logger.info('getting sentencepiece tokenizer from', tokenizer_path) tokenizer = SentencePieceProcessor(str(tokenizer_path)) - print('gguf: adding tokens') + logger.info('adding tokens') tokens: list[bytes] = [] scores: list[float] = [] toktypes: list[int] = [] @@ -67,8 +70,10 @@ def main(): parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") parser.add_argument("--ckpt-path", type=Path, help="path to persimmon checkpoint .pt file") parser.add_argument("--model-dir", type=Path, help="directory containing model e.g. 8b_chat_model_release") - parser.add_argument("--adept-inference-dir", type=str, help="path to adept-inference code directory") + parser.add_argument("--adept-inference-dir", type=str, help="path to adept-inference code directory") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") args = parser.parse_args() + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) sys.path.append(str(args.adept_inference_dir)) persimmon_model = torch.load(args.ckpt_path) hparams = persimmon_model['args'] @@ -107,7 +112,7 @@ def main(): gguf_writer.add_eos_token_id(71013) tensor_map = gguf.get_tensor_name_map(arch, block_count) - print(tensor_map) + logger.info(tensor_map) for name in tensors.keys(): data_torch = tensors[name] if name.endswith(".self_attention.rotary_emb.inv_freq"): @@ -117,22 +122,21 @@ def main(): data = data_torch.to(torch.float32).squeeze().numpy() new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) if new_name is None: - print("Can not map tensor '" + name + "'") - sys.exit() + raise ValueError(f"Can not map tensor '{name}'") + n_dims = len(data.shape) - print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) + logger.debug(f"{new_name}, n_dims = {str(n_dims)}, {str(old_dtype)} --> {str(data.dtype)}") gguf_writer.add_tensor(new_name, data) - print("gguf: write header") + logger.info("gguf: write header") gguf_writer.write_header_to_file() - print("gguf: write metadata") + logger.info("gguf: write metadata") gguf_writer.write_kv_data_to_file() - print("gguf: write tensors") + logger.info("gguf: write tensors") gguf_writer.write_tensors_to_file() gguf_writer.close() - print(f"gguf: model successfully exported to '{args.outfile}'") - print("") + logger.info(f"gguf: model successfully exported to '{args.outfile}'") if __name__ == '__main__': diff --git a/convert.py b/convert.py index 1c700cf6a..7f0b6b749 100755 --- a/convert.py +++ b/convert.py @@ -1,6 +1,7 @@ #!/usr/bin/env python3 from __future__ import annotations +import logging import argparse import concurrent.futures import enum @@ -35,6 +36,8 @@ import gguf if TYPE_CHECKING: from typing_extensions import Self, TypeAlias +logger = logging.getLogger("convert") + if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'): faulthandler.register(signal.SIGUSR1) @@ -643,7 +646,6 @@ class LlamaHfVocab(Vocab): def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray: - # print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) ) if n_head_kv is not None and n_head != n_head_kv: n_head = n_head_kv return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) @@ -1033,12 +1035,12 @@ def check_vocab_size(params: Params, vocab: BaseVocab, pad_vocab: bool = False) # Check for a vocab size mismatch if params.n_vocab == vocab.vocab_size: - print("Ignoring added_tokens.json since model matches vocab size without it.") + logger.warning("Ignoring added_tokens.json since model matches vocab size without it.") return if pad_vocab and params.n_vocab > vocab.vocab_size: pad_count = params.n_vocab - vocab.vocab_size - print( + logger.debug( f"Padding vocab with {pad_count} token(s) - through " ) for i in range(1, pad_count + 1): @@ -1166,7 +1168,7 @@ class OutputFile: elapsed = time.time() - start size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape) padi = len(str(len(model))) - print( + logger.info( f"[{i + 1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}" ) self.gguf.write_tensor_data(ndarray) @@ -1281,12 +1283,12 @@ def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> # HF models permut or pack some of the tensors, so we need to undo that for i in itertools.count(): if f"model.layers.{i}.self_attn.q_proj.weight" in model: - print(f"Permuting layer {i}") + logger.debug(f"Permuting layer {i}") tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head, params.n_head) tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head, params.n_head_kv) # tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"] elif f"model.layers.{i}.self_attn.W_pack.weight" in model: - print(f"Unpacking and permuting layer {i}") + logger.debug(f"Unpacking and permuting layer {i}") tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head) tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 1, params.n_head, params.n_head_kv) tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = part_lazy (model[f"model.layers.{i}.self_attn.W_pack.weight"], 2) @@ -1299,15 +1301,15 @@ def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None) if name_new is None: if skip_unknown: - print(f"Unexpected tensor name: {name} - skipping") + logger.warning(f"Unexpected tensor name: {name} - skipping") continue raise ValueError(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)") if tensor_type in should_skip: - print(f"skipping tensor {name_new}") + logger.debug(f"skipping tensor {name_new}") continue - print(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}") + logger.debug(f"{name:48s} -> {name_new:40s} | {lazy_tensor.data_type.name:6s} | {lazy_tensor.shape}") out[name_new] = lazy_tensor return out @@ -1372,7 +1374,7 @@ def load_some_model(path: Path) -> ModelPlus: paths = find_multifile_paths(path) models_plus: list[ModelPlus] = [] for path in paths: - print(f"Loading model file {path}") + logger.info(f"Loading model file {path}") models_plus.append(lazy_load_file(path)) model_plus = merge_multifile_models(models_plus) @@ -1413,7 +1415,7 @@ class VocabFactory: else: raise FileNotFoundError(f"Could not find a tokenizer matching any of {vocab_types}") - print(f"Loaded vocab file {vocab.fname_tokenizer!r}, type {vocab.name!r}") + logger.info(f"Loaded vocab file {vocab.fname_tokenizer!r}, type {vocab.name!r}") return vocab def load_vocab(self, vocab_types: list[str] | None, model_parent_path: Path) -> tuple[BaseVocab, gguf.SpecialVocab]: @@ -1438,19 +1440,19 @@ def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path: }[file_type] ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf" if ret in model_paths: - sys.stderr.write( + logger.error( f"Error: Default output path ({ret}) would overwrite the input. " - "Please explicitly specify a path using --outfile.\n") + "Please explicitly specify a path using --outfile.") sys.exit(1) return ret def do_dump_model(model_plus: ModelPlus) -> None: - print(f"model_plus.paths = {model_plus.paths!r}") - print(f"model_plus.format = {model_plus.format!r}") - print(f"model_plus.vocab = {model_plus.vocab!r}") + print(f"model_plus.paths = {model_plus.paths!r}") # noqa: NP100 + print(f"model_plus.format = {model_plus.format!r}") # noqa: NP100 + print(f"model_plus.vocab = {model_plus.vocab!r}") # noqa: NP100 for name, lazy_tensor in model_plus.model.items(): - print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}") + print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}") # noqa: NP100 def main(args_in: list[str] | None = None) -> None: @@ -1473,8 +1475,18 @@ def main(args_in: list[str] | None = None) -> None: parser.add_argument("--big-endian", action="store_true", help="model is executed on big endian machine") parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides") parser.add_argument("--skip-unknown", action="store_true", help="skip unknown tensor names instead of failing") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") args = parser.parse_args(args_in) + + if args.verbose: + logging.basicConfig(level=logging.DEBUG) + elif args.dump_single or args.dump: + # Avoid printing anything besides the dump output + logging.basicConfig(level=logging.WARNING) + else: + logging.basicConfig(level=logging.INFO) + if args.no_vocab and args.vocab_only: raise ValueError("--vocab-only does not make sense with --no-vocab") @@ -1491,6 +1503,7 @@ def main(args_in: list[str] | None = None) -> None: if args.dump: do_dump_model(model_plus) return + endianess = gguf.GGUFEndian.LITTLE if args.big_endian: endianess = gguf.GGUFEndian.BIG @@ -1513,7 +1526,7 @@ def main(args_in: list[str] | None = None) -> None: "q8_0": GGMLFileType.MostlyQ8_0, }[args.outtype] - print(f"params = {params}") + logger.info(f"params = {params}") model_parent_path = model_plus.paths[0].parent vocab_path = Path(args.vocab_dir or args.model or model_parent_path) @@ -1528,15 +1541,14 @@ def main(args_in: list[str] | None = None) -> None: outfile = args.outfile OutputFile.write_vocab_only(outfile, params, vocab, special_vocab, endianess=endianess, pad_vocab=args.pad_vocab) - print(f"Wrote {outfile}") + logger.info(f"Wrote {outfile}") return if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab: vocab = model_plus.vocab - print(f"Vocab info: {vocab}") - print(f"Special vocab info: {special_vocab}") - + logger.info(f"Vocab info: {vocab}") + logger.info(f"Special vocab info: {special_vocab}") model = model_plus.model model = convert_model_names(model, params, args.skip_unknown) ftype = pick_output_type(model, args.outtype) @@ -1544,11 +1556,11 @@ def main(args_in: list[str] | None = None) -> None: outfile = args.outfile or default_outfile(model_plus.paths, ftype) params.ftype = ftype - print(f"Writing {outfile}, format {ftype}") + logger.info(f"Writing {outfile}, format {ftype}") OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab) - print(f"Wrote {outfile}") + logger.info(f"Wrote {outfile}") if __name__ == '__main__': diff --git a/examples/batched-bench/batched-bench.cpp b/examples/batched-bench/batched-bench.cpp index 1e34de620..2924d8116 100644 --- a/examples/batched-bench/batched-bench.cpp +++ b/examples/batched-bench/batched-bench.cpp @@ -32,7 +32,7 @@ int main(int argc, char ** argv) { gpt_params params; if (argc == 1 || argv[1][0] == '-') { - printf("usage: %s MODEL_PATH [N_KV_MAX] [N_BATCH] [N_UBATCH] [IS_PP_SHARED] [NGL] \n" , argv[0]); + printf("usage: %s MODEL_PATH [N_KV_MAX] [N_BATCH] [N_UBATCH] [FATTN] [IS_PP_SHARED] [NGL] \n" , argv[0]); printf(" , and PL are comma-separated lists of numbers without spaces\n\n"); printf(" example: %s ggml-model-f16.gguf 2048 2048 512 0 999 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]); return 1 ; @@ -41,6 +41,7 @@ int main(int argc, char ** argv) { int n_kv_max = 2048; int n_batch = 2048; int n_ubatch = 512; + bool flash_attn = false; int is_pp_shared = 0; int n_gpu_layers = 0; @@ -66,23 +67,27 @@ int main(int argc, char ** argv) { } if (argc >= 6) { - is_pp_shared = std::atoi(argv[5]); + flash_attn = std::atoi(argv[5]); } if (argc >= 7) { - n_gpu_layers = std::atoi(argv[6]); + is_pp_shared = std::atoi(argv[6]); } if (argc >= 8) { - n_pp = parse_list(argv[7]); + n_gpu_layers = std::atoi(argv[7]); } if (argc >= 9) { - n_tg = parse_list(argv[8]); + n_pp = parse_list(argv[8]); } if (argc >= 10) { - n_pl = parse_list(argv[9]); + n_tg = parse_list(argv[9]); + } + + if (argc >= 11) { + n_pl = parse_list(argv[10]); } // init LLM @@ -108,10 +113,11 @@ int main(int argc, char ** argv) { llama_context_params ctx_params = llama_context_default_params(); - ctx_params.seed = 1234; - ctx_params.n_ctx = n_kv_max; - ctx_params.n_batch = n_batch; - ctx_params.n_ubatch = n_ubatch; + ctx_params.seed = 1234; + ctx_params.n_ctx = n_kv_max; + ctx_params.n_batch = n_batch; + ctx_params.n_ubatch = n_ubatch; + ctx_params.flash_attn = flash_attn; ctx_params.n_threads = params.n_threads; ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch; @@ -169,7 +175,7 @@ int main(int argc, char ** argv) { } LOG_TEE("\n"); - LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, n_batch, n_ubatch, is_pp_shared, n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch); + LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, n_batch, n_ubatch, flash_attn, is_pp_shared, n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch); LOG_TEE("\n"); LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s"); diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index 8b532c8b6..95c3095dd 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -174,6 +174,7 @@ struct cmd_params { std::vector split_mode; std::vector main_gpu; std::vector no_kv_offload; + std::vector flash_attn; std::vector> tensor_split; std::vector use_mmap; std::vector embeddings; @@ -195,6 +196,7 @@ static const cmd_params cmd_params_defaults = { /* split_mode */ {LLAMA_SPLIT_MODE_LAYER}, /* main_gpu */ {0}, /* no_kv_offload */ {false}, + /* flash_attn */ {false}, /* tensor_split */ {std::vector(llama_max_devices(), 0.0f)}, /* use_mmap */ {true}, /* embeddings */ {false}, @@ -220,6 +222,7 @@ static void print_usage(int /* argc */, char ** argv) { printf(" -sm, --split-mode (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str()); printf(" -mg, --main-gpu (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str()); printf(" -nkvo, --no-kv-offload <0|1> (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str()); + printf(" -fa, --flash-attn <0|1> (default: %s)\n", join(cmd_params_defaults.flash_attn, ",").c_str()); printf(" -mmp, --mmap <0|1> (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str()); printf(" -embd, --embeddings <0|1> (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str()); printf(" -ts, --tensor-split (default: 0)\n"); @@ -393,6 +396,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { } auto p = split(argv[i], split_delim); params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end()); + } else if (arg == "-fa" || arg == "--flash-attn") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = split(argv[i], split_delim); + params.flash_attn.insert(params.flash_attn.end(), p.begin(), p.end()); } else if (arg == "-mmp" || arg == "--mmap") { if (++i >= argc) { invalid_param = true; @@ -477,6 +487,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { if (params.split_mode.empty()) { params.split_mode = cmd_params_defaults.split_mode; } if (params.main_gpu.empty()) { params.main_gpu = cmd_params_defaults.main_gpu; } if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; } + if (params.flash_attn.empty()) { params.flash_attn = cmd_params_defaults.flash_attn; } if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; } if (params.use_mmap.empty()) { params.use_mmap = cmd_params_defaults.use_mmap; } if (params.embeddings.empty()) { params.embeddings = cmd_params_defaults.embeddings; } @@ -498,6 +509,7 @@ struct cmd_params_instance { llama_split_mode split_mode; int main_gpu; bool no_kv_offload; + bool flash_attn; std::vector tensor_split; bool use_mmap; bool embeddings; @@ -532,6 +544,7 @@ struct cmd_params_instance { cparams.type_k = type_k; cparams.type_v = type_v; cparams.offload_kqv = !no_kv_offload; + cparams.flash_attn = flash_attn; cparams.embeddings = embeddings; return cparams; @@ -554,6 +567,7 @@ static std::vector get_cmd_params_instances(const cmd_param for (const auto & tk : params.type_k) for (const auto & tv : params.type_v) for (const auto & nkvo : params.no_kv_offload) + for (const auto & fa : params.flash_attn) for (const auto & nt : params.n_threads) { for (const auto & n_prompt : params.n_prompt) { if (n_prompt == 0) { @@ -572,6 +586,7 @@ static std::vector get_cmd_params_instances(const cmd_param /* .split_mode = */ sm, /* .main_gpu = */ mg, /* .no_kv_offload= */ nkvo, + /* .flash_attn = */ fa, /* .tensor_split = */ ts, /* .use_mmap = */ mmp, /* .embeddings = */ embd, @@ -596,6 +611,7 @@ static std::vector get_cmd_params_instances(const cmd_param /* .split_mode = */ sm, /* .main_gpu = */ mg, /* .no_kv_offload= */ nkvo, + /* .flash_attn = */ fa, /* .tensor_split = */ ts, /* .use_mmap = */ mmp, /* .embeddings = */ embd, @@ -633,6 +649,7 @@ struct test { llama_split_mode split_mode; int main_gpu; bool no_kv_offload; + bool flash_attn; std::vector tensor_split; bool use_mmap; bool embeddings; @@ -657,6 +674,7 @@ struct test { split_mode = inst.split_mode; main_gpu = inst.main_gpu; no_kv_offload = inst.no_kv_offload; + flash_attn = inst.flash_attn; tensor_split = inst.tensor_split; use_mmap = inst.use_mmap; embeddings = inst.embeddings; @@ -731,7 +749,7 @@ struct test { "n_batch", "n_ubatch", "n_threads", "type_k", "type_v", "n_gpu_layers", "split_mode", - "main_gpu", "no_kv_offload", + "main_gpu", "no_kv_offload", "flash_attn", "tensor_split", "use_mmap", "embeddings", "n_prompt", "n_gen", "test_time", "avg_ns", "stddev_ns", @@ -753,7 +771,7 @@ struct test { } if (field == "cuda" || field == "opencl" || field == "vulkan" || field == "kompute" || field == "metal" || field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" || - field == "use_mmap" || field == "embeddings") { + field == "flash_attn" || field == "use_mmap" || field == "embeddings") { return BOOL; } if (field == "avg_ts" || field == "stddev_ts") { @@ -787,7 +805,7 @@ struct test { std::to_string(n_batch), std::to_string(n_ubatch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v), std::to_string(n_gpu_layers), split_mode_str(split_mode), - std::to_string(main_gpu), std::to_string(no_kv_offload), + std::to_string(main_gpu), std::to_string(no_kv_offload), std::to_string(flash_attn), tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings), std::to_string(n_prompt), std::to_string(n_gen), test_time, std::to_string(avg_ns()), std::to_string(stdev_ns()), @@ -955,6 +973,9 @@ struct markdown_printer : public printer { if (field == "no_kv_offload") { return "nkvo"; } + if (field == "flash_attn") { + return "fa"; + } if (field == "use_mmap") { return "mmap"; } @@ -1001,6 +1022,9 @@ struct markdown_printer : public printer { if (params.no_kv_offload.size() > 1 || params.no_kv_offload != cmd_params_defaults.no_kv_offload) { fields.emplace_back("no_kv_offload"); } + if (params.flash_attn.size() > 1 || params.flash_attn != cmd_params_defaults.flash_attn) { + fields.emplace_back("flash_attn"); + } if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) { fields.emplace_back("tensor_split"); } diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp index a44c6cd76..157a680b5 100644 --- a/examples/llava/llava-cli.cpp +++ b/examples/llava/llava-cli.cpp @@ -113,11 +113,11 @@ struct llava_context { }; static void show_additional_info(int /*argc*/, char ** argv) { - LOG_TEE("\n example usage: %s -m --mmproj --image [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]); + LOG_TEE("\n example usage: %s -m --mmproj --image --image [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]); LOG_TEE(" note: a lower temperature value like 0.1 is recommended for better quality.\n"); } -static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params) { +static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params, const std::string & fname) { // load and preprocess the image llava_image_embed * embed = NULL; @@ -133,9 +133,9 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para } params->prompt = remove_image_from_prompt(prompt); } else { - embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, params->image.c_str()); + embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, fname.c_str()); if (!embed) { - LOG_TEE("%s: is %s really an image file?\n", __func__, params->image.c_str()); + fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str()); return NULL; } } @@ -207,17 +207,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_ printf("\n"); } - -static struct llava_context * llava_init(gpt_params * params) { - const char * clip_path = params->mmproj.c_str(); - - auto prompt = params->prompt; - if (prompt.empty()) { - prompt = "describe the image in detail."; - } - - auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1); - +static struct llama_model * llava_init(gpt_params * params) { llama_backend_init(); llama_numa_init(params->numa); @@ -228,6 +218,19 @@ static struct llava_context * llava_init(gpt_params * params) { LOG_TEE("%s: error: unable to load model\n" , __func__); return NULL; } + return model; +} + +static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) { + const char * clip_path = params->mmproj.c_str(); + + auto prompt = params->prompt; + if (prompt.empty()) { + prompt = "describe the image in detail."; + } + + auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1); + llama_context_params ctx_params = llama_context_params_from_gpt_params(*params); ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings @@ -286,24 +289,30 @@ int main(int argc, char ** argv) { show_additional_info(argc, argv); return 1; } - - auto ctx_llava = llava_init(¶ms); - if (ctx_llava == NULL) { - LOG_TEE("%s: error: failed to init llava\n", __func__); + auto model = llava_init(¶ms); + if (model == NULL) { + fprintf(stderr, "%s: error: failed to init llava model\n", __func__); return 1; } - auto image_embed = load_image(ctx_llava, ¶ms); - if (!image_embed) { - return 1; + for (auto & image : params.image) { + auto ctx_llava = llava_init_context(¶ms, model); + + auto image_embed = load_image(ctx_llava, ¶ms, image); + if (!image_embed) { + std::cerr << "error: failed to load image " << image << ". Terminating\n\n"; + return 1; + } + + // process the prompt + process_prompt(ctx_llava, image_embed, ¶ms, params.prompt); + + llama_print_timings(ctx_llava->ctx_llama); + llava_image_embed_free(image_embed); + ctx_llava->model = NULL; + llava_free(ctx_llava); } + llama_free_model(model); - // process the prompt - process_prompt(ctx_llava, image_embed, ¶ms, params.prompt); - - llama_print_timings(ctx_llava->ctx_llama); - - llava_image_embed_free(image_embed); - llava_free(ctx_llava); return 0; } diff --git a/examples/main-cmake-pkg/README.md b/examples/main-cmake-pkg/README.md index f599fbaec..edf20d8db 100644 --- a/examples/main-cmake-pkg/README.md +++ b/examples/main-cmake-pkg/README.md @@ -17,11 +17,9 @@ In this case, CLBlast was already installed so the CMake package is referenced i ```cmd git clone https://github.com/ggerganov/llama.cpp cd llama.cpp -mkdir build -cd build -cmake .. -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=C:/CLBlast/lib/cmake/CLBlast -G "Visual Studio 17 2022" -A x64 -cmake --build . --config Release -cmake --install . --prefix C:/LlamaCPP +cmake -B build -DBUILD_SHARED_LIBS=OFF -DLLAMA_CLBLAST=ON -DCMAKE_PREFIX_PATH=C:/CLBlast/lib/cmake/CLBlast -G "Visual Studio 17 2022" -A x64 +cmake --build build --config Release +cmake --install build --prefix C:/LlamaCPP ``` ### Build main-cmake-pkg @@ -29,9 +27,7 @@ cmake --install . --prefix C:/LlamaCPP ```cmd cd ..\examples\main-cmake-pkg -mkdir build -cd build -cmake .. -DBUILD_SHARED_LIBS=OFF -DCMAKE_PREFIX_PATH="C:/CLBlast/lib/cmake/CLBlast;C:/LlamaCPP/lib/cmake/Llama" -G "Visual Studio 17 2022" -A x64 -cmake --build . --config Release -cmake --install . --prefix C:/MyLlamaApp +cmake -B build -DBUILD_SHARED_LIBS=OFF -DCMAKE_PREFIX_PATH="C:/CLBlast/lib/cmake/CLBlast;C:/LlamaCPP/lib/cmake/Llama" -G "Visual Studio 17 2022" -A x64 +cmake --build build --config Release +cmake --install build --prefix C:/MyLlamaApp ``` diff --git a/examples/main/README.md b/examples/main/README.md index 649f4e0f3..e7a38743c 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -66,7 +66,7 @@ main.exe -m models\7B\ggml-model.bin --ignore-eos -n -1 --random-prompt In this section, we cover the most commonly used options for running the `main` program with the LLaMA models: -- `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`). +- `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`; inferred from `--model-url` if set). - `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g https://huggingface.co/ggml-org/models/resolve/main/phi-2/ggml-model-q4_0.gguf). - `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses. - `-ins, --instruct`: Run the program in instruction mode, which is particularly useful when working with Alpaca models. diff --git a/examples/main/main.cpp b/examples/main/main.cpp index a74d4d9c7..eabbc2db3 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -324,7 +324,7 @@ int main(int argc, char ** argv) { log_tostr(embd_inp.empty()), n_matching_session_tokens, embd_inp.size(), session_tokens.size(), embd_inp.size()); // if we will use the cache for the full prompt without reaching the end of the cache, force - // reevaluation of the last token token to recalculate the cached logits + // reevaluation of the last token to recalculate the cached logits if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) { LOGLN("recalculate the cached logits (do): session_tokens.resize( %zu )", embd_inp.size() - 1); @@ -544,7 +544,7 @@ int main(int argc, char ** argv) { // if we run out of context: // - take the n_keep first tokens from the original prompt (via n_past) // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches - if (n_past + (int) embd.size() + std::max(0, guidance_offset) > n_ctx) { + if (n_past + (int) embd.size() + std::max(0, guidance_offset) >= n_ctx) { if (params.n_predict == -2) { LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict); break; diff --git a/examples/perplexity/README.md b/examples/perplexity/README.md index 1a8c0dd64..c5e2bc5de 100644 --- a/examples/perplexity/README.md +++ b/examples/perplexity/README.md @@ -1,8 +1,118 @@ -# perplexity +# Perplexity -TODO +The `perplexity` example can be used to calculate the so-called perplexity value of a language model over a given text corpus. +Perplexity measures how well the model can predict the next token with lower values being better. +Note that perplexity is **not** directly comparable between models, especially if they use different tokenizers. +Also note that finetunes typically result in a higher perplexity value even though the human-rated quality of outputs increases. + +Within llama.cpp the perplexity of base models is used primarily to judge the quality loss from e.g. quantized models vs. FP16. +The convention among contributors is to use the Wikitext-2 test set for testing unless noted otherwise (can be obtained with `scripts/get-wikitext-2.sh`). + +By default only the mean perplexity value and the corresponding uncertainty is calculated. +The uncertainty is determined empirically by assuming a Gaussian distribution of the "correct" logits per and then applying error propagation. + +More statistics can be obtained by recording the logits from the FP16 version of a model. +To do this, supply `perplexity` with `--kl-divergence-base path/to/logit/binary/file.kld`. +The program will then record all logits and save them to the provided path in binary format. +**The logit file will be very large, 11 GiB for LLaMA 2 or 37 GiB for LLaMA 3 when using the Wikitext-2 test set.** +Once you have the file, supply `perplexity` with the quantized model, the logits file via `--kl-divergence-base`, +and finally the `--kl-divergence` argument to indicate that the program should calculate the so-called Kullback-Leibler divergence. +This is a measure of how similar the FP16 and the quantized logit distributions are with a value of 0 indicating that the distribution are the same. +The uncertainty on the mean KL divergence is calculated by assuming the KL divergence per token follows a Gaussian distribution. + +In addition to the KL divergence the following statistics are calculated with `--kl-divergence`: + +* Ratio of mean FP16 PPL and quantized PPL. Uncertainty is estimated on logits, then propagated. The logarithm of this metric is also calculated and printed, it is 0 if the logit distributions are the same. +* Difference of mean FP16 PPL and quantized PPL. Uncertainty is estimated on logits, then propagated. +* Mean change in "correct" token probability. Positive values mean the model gets better at prediction, negative values mean it gets worse. +* Pearson correlation coefficient of the "correct" token probabilites between models. +* Percentiles of change in "correct" token probability. Positive values mean the model gets better at prediction, negative values mean it gets worse. Can be used to judge noise vs. quality loss from quantization. If the percentiles are symmetric then the quantization is essentially just adding noise. If the negative values are significantly larger than the positive values then this indicates that the model is actually becoming worse from the quantization. +* The root mean square of the change in token probabilities. If you were to assume that the quantization simply causes Gaussian noise on the token probabilities then this would be the standard deviation of said noise. The uncertainty on the value is calculated that the change in token probabilities follows a Gaussian distribution. Related discussion: https://github.com/ggerganov/llama.cpp/discussions/2875 . +* Same top p: Percentage of how often the token was assigned the highest probabilites by both models. The uncertainty is calculated from the Gaussian approximation of the binomial distribution. + +## LLaMA 3 8b Scoreboard + +Results are sorted by Kullback-Leibler divergence relative to FP16. +The "WT" importance matrices were created using varying numbers of Wikitext tokens and can be found [here](https://huggingface.co/JohannesGaessler/llama.cpp_importance_matrices/blob/main/imatrix-llama_3-8b-f16-2.7m_tokens.dat). + +| Quantization | imatrix | Model size [GiB] | PPL | ΔPPL | KLD | Mean Δp | RMS Δp | +|--------------|---------|------------------|------------------------|------------------------|-----------------------|-------------------|------------------| +| f16 | None | 14.97 | 6.233160 ± 0.037828 | - | - | - | - | +| q8_0 | None | 7.96 | 6.234284 ± 0.037878 | 0.002650 ± 0.001006 | 0.001355 ± 0.000006 | -0.019 ± 0.003 % | 1.198 ± 0.007 % | +| q6_K | None | 6.14 | 6.253382 ± 0.038078 | 0.021748 ± 0.001852 | 0.005452 ± 0.000035 | -0.007 ± 0.006 % | 2.295 ± 0.019 % | +| q5_K_M | None | 5.33 | 6.288607 ± 0.038338 | 0.056974 ± 0.002598 | 0.010762 ± 0.000079 | -0.114 ± 0.008 % | 3.160 ± 0.031 % | +| q5_K_S | None | 5.21 | 6.336598 ± 0.038755 | 0.104964 ± 0.003331 | 0.016595 ± 0.000122 | -0.223 ± 0.010 % | 3.918 ± 0.036 % | +| q5_1 | None | 5.65 | 6.337857 ± 0.038677 | 0.106223 ± 0.003476 | 0.018045 ± 0.000139 | -0.287 ± 0.011 % | 4.123 ± 0.039 % | +| q5_0 | None | 5.21 | 6.363224 ± 0.038861 | 0.131591 ± 0.003894 | 0.022239 ± 0.000166 | -0.416 ± 0.012 % | 4.634 ± 0.043 % | +| q4_K_M | WT 10m | 4.58 | 6.382937 ± 0.039055 | 0.151303 ± 0.004429 | 0.028152 ± 0.000240 | -0.389 ± 0.014 % | 5.251 ± 0.049 % | +| q4_K_M | None | 4.58 | 6.407115 ± 0.039119 | 0.175482 ± 0.004620 | 0.031273 ± 0.000238 | -0.596 ± 0.014 % | 5.519 ± 0.050 % | +| q4_K_S | WT 10m | 4.37 | 6.409697 ± 0.039189 | 0.178064 ± 0.004744 | 0.031951 ± 0.000259 | -0.531 ± 0.015 % | 5.645 ± 0.051 % | +| iq4_NL | WT 10m | 4.35 | 6.455593 ± 0.039630 | 0.223959 ± 0.005201 | 0.035742 ± 0.000288 | -0.590 ± 0.016 % | 5.998 ± 0.054 % | +| iq4_XS | WT 10m | 4.14 | 6.459705 ± 0.039595 | 0.228071 ± 0.005207 | 0.036334 ± 0.000284 | -0.668 ± 0.016 % | 6.044 ± 0.054 % | +| q4_K_S | None | 4.37 | 6.500529 ± 0.039778 | 0.268895 ± 0.005638 | 0.043136 ± 0.000314 | -0.927 ± 0.017 % | 6.562 ± 0.055 % | +| q4_1 | None | 4.78 | 6.682737 ± 0.041285 | 0.451103 ± 0.008030 | 0.071683 ± 0.000505 | -0.927 ± 0.017 % | 8.512 ± 0.063 % | +| q4_0 | None | 4.34 | 6.700147 ± 0.041226 | 0.468514 ± 0.007951 | 0.071940 ± 0.000491 | -1.588 ± 0.022 % | 8.434 ± 0.061 % | +| q3_K_L | WT 10m | 4.03 | 6.671223 ± 0.041427 | 0.439590 ± 0.008154 | 0.073077 ± 0.000529 | -0.940 ± 0.023 % | 8.662 ± 0.064 % | +| q3_K_M | WT 10m | 3.74 | 6.734255 ± 0.041838 | 0.502622 ± 0.008901 | 0.084358 ± 0.000588 | -1.198 ± 0.024 % | 9.292 ± 0.065 % | +| q3_K_L | None | 4.03 | 6.787876 ± 0.042104 | 0.556242 ± 0.009171 | 0.087176 ± 0.000614 | -1.532 ± 0.025 % | 9.432 ± 0.067 % | +| q3_K_M | None | 3.74 | 6.888498 ± 0.042669 | 0.656864 ± 0.010071 | 0.101913 ± 0.000677 | -1.990 ± 0.026 % | 10.203 ± 0.068 % | +| iq3_M | WT 10m | 3.53 | 6.898327 ± 0.041643 | 0.666694 ± 0.009449 | 0.102534 ± 0.000663 | -3.178 ± 0.026 % | 10.513 ± 0.066 % | +| iq3_S | WT 10m | 3.42 | 6.965501 ± 0.042406 | 0.733867 ± 0.010245 | 0.111278 ± 0.000710 | -3.066 ± 0.027 % | 10.845 ± 0.068 % | +| iq3_XS | WT 10m | 3.28 | 7.163043 ± 0.043772 | 0.931409 ± 0.012084 | 0.138693 ± 0.000857 | -3.667 ± 0.031 % | 12.148 ± 0.070 % | +| iq3_XXS | WT 10m | 3.05 | 7.458436 ± 0.046404 | 1.226803 ± 0.015234 | 0.183625 ± 0.001042 | -3.918 ± 0.035 % | 13.836 ± 0.074 % | +| q3_K_S | WT 10m | 3.41 | 7.602878 ± 0.046848 | 1.371244 ± 0.015688 | 0.199821 ± 0.001008 | -5.046 ± 0.037 % | 14.980 ± 0.070 % | +| q3_K_S | None | 3.41 | 7.863786 ± 0.048885 | 1.632152 ± 0.017733 | 0.228217 ± 0.001079 | -5.604 ± 0.038 % | 15.541 ± 0.070 % | +| iq2_M | WT 10m | 2.74 | 8.600799 ± 0.055124 | 2.369166 ± 0.025244 | 0.325989 ± 0.00160 | -6.463 ± 0.046 % | 18.519 ± 0.080 % | +| q2_K | WT 10k | 2.96 | 8.652290 ± 0.055572 | 2.420657 ± 0.025587 | 0.331393 ± 0.001562 | -6.606 ± 0.046 % | 18.790 ± 0.078 % | +| q2_K | WT 100k | 2.96 | 8.641993 ± 0.055406 | 2.410359 ± 0.025495 | 0.331672 ± 0.001569 | -6.628 ± 0.047 % | 18.856 ± 0.078 % | +| q2_K | WT 10m | 2.96 | 8.647825 ± 0.055610 | 2.416191 ± 0.025683 | 0.332223 ± 0.001572 | -6.500 ± 0.047 % | 18.881 ± 0.078 % | +| q2_K | WT 1m | 2.96 | 8.674365 ± 0.055743 | 2.442732 ± 0.025843 | 0.335308 ± 0.001576 | -6.634 ± 0.047 % | 19.009 ± 0.079 % | +| q2_K | WT 1k | 2.96 | 8.682605 ± 0.055916 | 2.450972 ± 0.026069 | 0.337093 ± 0.001596 | -6.596 ± 0.047 % | 18.977 ± 0.079 % | +| q2_K_S | WT 10m | 2.96 | 9.323778 ± 0.061551 | 3.092145 ± 0.031914 | 0.403360 ± 0.001787 | -7.131 ± 0.049 % | 20.050 ± 0.081 % | +| q2_K_S | WT 1m | 2.96 | 9.329321 ± 0.061378 | 3.097688 ± 0.031816 | 0.403590 ± 0.001797 | -7.289 ± 0.049 % | 20.123 ± 0.081 % | +| q2_K_S | WT 100k | 2.96 | 9.362973 ± 0.061740 | 3.131339 ± 0.032169 | 0.408367 ± 0.001802 | -7.198 ± 0.050 % | 20.132 ± 0.081 % | +| q2_K_S | WT 10k | 2.96 | 9.376479 ± 0.062045 | 3.144846 ± 0.032464 | 0.408662 ± 0.001819 | -7.141 ± 0.050 % | 20.120 ± 0.081 % | +| q2_K_S | WT 1k | 2.96 | 9.415200 ± 0.062475 | 3.183567 ± 0.032993 | 0.415865 ± 0.001846 | -7.153 ± 0.050 % | 20.311 ± 0.082 % | +| iq2_S | WT 10m | 2.56 | 9.650781 ± 0.063209 | 3.419148 ± 0.034017 | 0.439197 ± 0.001976 | -8.319 ± 0.052 % | 21.491 ± 0.083 % | +| q2_K | None | 2.96 | 9.751568 ± 0.063312 | 3.519934 ± 0.033863 | 0.445132 ± 0.001835 | -9.123 ± 0.051 % | 21.421 ± 0.079 % | +| iq2_XS | WT 10m | 2.43 | 10.761424 ± 0.071056 | 4.529791 ± 0.042229 | 0.546290 ± 0.002133 | -10.576 ± 0.056 % | 23.872 ± 0.082 % | +| iq2_XXS | WT 10m | 2.24 | 14.091782 ± 0.098396 | 7.860148 ± 0.070752 | 0.812022 ± 0.002741 | -14.363 ± 0.065 % | 28.576 ± 0.084 % | +| iq1_M | WT 10m | 2.01 | 25.493722 ± 0.177903 | 19.262089 ± 0.152396 | 1.393084 ± 0.003529 | -24.672 ± 0.077 % | 38.287 ± 0.084 % | +| iq1_S | WT 1m | 1.88 | 58.097760 ± 0.438604 | 51.866126 ± 0.416604 | 2.211278 ± 0.004688 | -32.471 ± 0.087 % | 46.418 ± 0.085 % | +| iq1_S | WT 1k | 1.88 | 58.267851 ± 0.446208 | 52.036218 ± 0.424373 | 2.214858 ± 0.004778 | -31.880 ± 0.089 % | 46.330 ± 0.086 % | +| iq1_S | WT 100k | 1.88 | 58.581498 ± 0.453145 | 52.349864 ± 0.431360 | 2.220834 ± 0.004818 | -32.261 ± 0.089 % | 46.002 ± 0.086 % | +| iq1_S | WT 10m | 1.88 | 60.694593 ± 0.471290 | 54.462959 ± 0.449644 | 2.254554 ± 0.004868 | -31.973 ± 0.088 % | 46.271 ± 0.086 % | +| iq1_S | WT 10k | 1.88 | 63.221324 ± 0.493077 | 56.989691 ± 0.471423 | 2.293527 ± 0.004885 | -32.261 ± 0.089 % | 46.562 ± 0.086 % | + +There seems to be no consistent improvement from using more Wikitext tokens for the importance matrix. +K-quants score better on mean Δp than the legacy quants than e.g. KL divergence would suggest. + +## LLaMA 2 vs. LLaMA 3 Quantization comparison + +| Metric | L2 7b q2_K | L3 8b q2_K | L2 7b q4_K_M | L3 8b q4_K_M | L2 7b q6_K | L3 8b q6_K | L2 7b q8_0 | L3 8b q8_0 | +|-----------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------| +| Mean PPL | 5.794552 ± 0.032298 | 9.751568 ± 0.063312 | 5.877078 ± 0.032781 | 6.407115 ± 0.039119 | 5.808494 ± 0.032425 | 6.253382 ± 0.038078 | 5.798542 ± 0.032366 | 6.234284 ± 0.037878 | +| Mean PPL ratio | 1.107955 ± 0.001427 | 1.564849 ± 0.004525 | 1.014242 ± 0.000432 | 1.028160 ± 0.000723 | 1.002406 ± 0.000191 | 1.003490 ± 0.000296 | 1.000689 ± 0.000107 | 1.000425 ± 0.000161 | +| Mean ΔPPL | 0.625552 ± 0.008725 | 3.519934 ± 0.033863 | 0.082526 ± 0.002530 | 0.175482 ± 0.004620 | 0.013941 ± 0.001110 | 0.021748 ± 0.001852 | 0.003990 ± 0.000624 | 0.002650 ± 0.001006 | +| PPL correlation | 97.36% | 89.62% | 99.71% | 99.34% | 99.94% | 99.88% | 99.98% | 99.96% | +| Mean KLD | 0.108903 ± 0.000645 | 0.445132 ± 0.001835 | 0.012686 ± 0.000079 | 0.031273 ± 0.000238 | 0.002098 ± 0.000014 | 0.005452 ± 0.000035 | 0.000369 ± 0.000007 | 0.001355 ± 0.000006 | +| Mean Δp | -2.710 ± 0.023 % | -9.123 ± 0.051 % | -0.416 ± 0.008 % | -0.596 ± 0.014 % | -0.035 ± 0.003 % | -0.007 ± 0.006 % | -0.005 ± 0.002 % | -0.019 ± 0.003 % | +| Maximum Δp | 85.136% | 94.268% | 45.209% | 95.054% | 23.593% | 53.601% | 43.925% | 28.734% | +| 99.9% Δp | 37.184% | 50.003% | 17.461% | 27.084% | 7.798% | 13.613% | 3.387% | 6.402% | +| 99.0% Δp | 18.131% | 25.875% | 7.798% | 12.084% | 3.838% | 6.407% | 1.867% | 3.544% | +| Median Δp | -0.391% | -2.476% | -0.026% | -0.024% | -0.001% | 0.000% | -0.000% | -0.000% | +| 1.0% Δp | -39.762% | -87.173% | -11.433% | -19.567% | -4.222% | -6.767% | -1.862% | -3.698% | +| 0.1% Δp | -79.002% | -98.897% | -26.433% | -56.054% | -9.091% | -16.584% | -3.252% | -6.579% | +| Minimum Δp | -99.915% | -99.965% | -83.383% | -98.699% | -43.142% | -68.487% | -9.343% | -24.301% | +| RMS Δp | 9.762 ± 0.053 % | 21.421 ± 0.079 % | 3.252 ± 0.024 % | 5.519 ± 0.050 % | 1.339 ± 0.010 % | 2.295 ± 0.019 % | 0.618 ± 0.011 % | 1.198 ± 0.007 % | +| Same top p | 85.584 ± 0.086 % | 71.138 ± 0.119 % | 94.665 ± 0.055 % | 91.901 ± 0.072 % | 97.520 ± 0.038 % | 96.031 ± 0.051 % | 98.846 ± 0.026 % | 97.674 ± 0.040 % | + + +## Old Numbers + +
+Llama 2 70B Scoreboard -## Llama 2 70B Scorechart | Quantization | Model size (GiB) | Perplexity | Delta to fp16 | |--------------|------------------|------------|---------------| | Q4_0 | 36.20 | 3.5550 | 3.61% | @@ -18,3 +128,5 @@ TODO | Q5_K_M | 45.41 | 3.4451 | 0.40% | | Q6_K | 52.70 | 3.4367 | 0.16% | | fp16 | 128.5 | 3.4313 | - | + +
diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 9a3374fdc..db6e0949d 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -216,17 +216,22 @@ static void process_logits(std::ostream& out, int n_vocab, const float * logits, } struct kl_divergence_result { - double sum_nll = 0; - double sum_nll2 = 0; - double sum_kld = 0; - double sum_kld2 = 0; - double sum_nll_diff = 0; - double sum_nll_diff2 = 0; - size_t n_same_top = 0; - size_t count = 0; + double sum_nll = 0.0; + double sum_nll2 = 0.0; + double sum_nll_base = 0.0; + double sum_nll_base2 = 0.0; + double sum_nll_nll_base = 0.0; + double sum_kld = 0.0; + double sum_kld2 = 0.0; + double sum_p_diff = 0.0; + double sum_p_diff2 = 0.0; + double sum_p_diff4 = 0.0; + float max_p_diff = 0.0f; + size_t n_same_top = 0.0; + size_t count = 0.0; }; -static double log_softmax(int n_vocab, const float * logits, const uint16_t * base_log_prob, int tok, kl_divergence_result & kld) { +static std::pair log_softmax(int n_vocab, const float * logits, const uint16_t * base_log_prob, int tok, kl_divergence_result & kld) { float max_logit = logits[0]; int imax = 0; for (int i = 1; i < n_vocab; ++i) { @@ -244,12 +249,17 @@ static double log_softmax(int n_vocab, const float * logits, const uint16_t * ba const float scale = d[0]; const float min_log_prob = d[1]; base_log_prob += 4; - float nll = max_logit + log_sum_exp - logits[tok]; + + const float nll = max_logit + log_sum_exp - logits[tok]; kld.sum_nll += nll; kld.sum_nll2 += nll*nll; - nll += (scale*base_log_prob[tok] + min_log_prob); - kld.sum_nll_diff += nll; - kld.sum_nll_diff2 += nll*nll; + + const float nll_base = -(scale*base_log_prob[tok] + min_log_prob); + kld.sum_nll_base += nll_base; + kld.sum_nll_base2 += nll_base*nll_base; + + kld.sum_nll_nll_base += nll*nll_base; + max_logit += log_sum_exp; double sum = 0; int imax_base = -1; @@ -269,34 +279,50 @@ static double log_softmax(int n_vocab, const float * logits, const uint16_t * ba kld.sum_kld2 += sum*sum; ++kld.count; if (imax == imax_base) ++kld.n_same_top; - return sum; + + const float p_base = expf(-nll_base); + const float p = expf(-nll); + const float p_diff = p - p_base; + kld.sum_p_diff += p_diff; + const double p_diff2 = p_diff*p_diff; + kld.sum_p_diff2 += p_diff2; + kld.sum_p_diff4 += p_diff2*p_diff2; + kld.max_p_diff = std::max(kld.max_p_diff, std::fabs(p_diff)); + + return std::make_pair(sum, p_diff); } static void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, std::vector & workers, const std::vector & base_log_probs, kl_divergence_result & kld, - float * kld_values) { + float * kld_values, float * p_diff_values) { std::mutex mutex; const int nv = 2*((n_vocab + 1)/2) + 4; int counter = 0; - auto compute = [&mutex, &counter, &base_log_probs, &kld, n_vocab, logits, tokens, n_token, nv, kld_values] () { + auto compute = [&mutex, &counter, &base_log_probs, &kld, n_vocab, logits, tokens, n_token, nv, kld_values, p_diff_values] () { kl_divergence_result local_kld; while (true) { std::unique_lock lock(mutex); int i = counter++; if (i >= n_token) { - kld.sum_nll += local_kld.sum_nll; - kld.sum_nll2 += local_kld.sum_nll2; - kld.sum_kld += local_kld.sum_kld; - kld.sum_kld2 += local_kld.sum_kld2; - kld.sum_nll_diff += local_kld.sum_nll_diff; - kld.sum_nll_diff2 += local_kld.sum_nll_diff2; - kld.n_same_top += local_kld.n_same_top; - kld.count += local_kld.count; + kld.sum_nll += local_kld.sum_nll; + kld.sum_nll2 += local_kld.sum_nll2; + kld.sum_nll_base += local_kld.sum_nll_base; + kld.sum_nll_base2 += local_kld.sum_nll_base2; + kld.sum_nll_nll_base += local_kld.sum_nll_nll_base; + kld.sum_kld += local_kld.sum_kld; + kld.sum_kld2 += local_kld.sum_kld2; + kld.sum_p_diff += local_kld.sum_p_diff; + kld.sum_p_diff2 += local_kld.sum_p_diff2; + kld.sum_p_diff4 += local_kld.sum_p_diff4; + kld.n_same_top += local_kld.n_same_top; + kld.max_p_diff = std::max(kld.max_p_diff, local_kld.max_p_diff); + kld.count += local_kld.count; break; } lock.unlock(); - double v = log_softmax(n_vocab, logits + i*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld); - kld_values[i] = (float)v; + std::pair v = log_softmax(n_vocab, logits + i*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld); + kld_values[i] = (float)v.first; + p_diff_values[i] = v.second; } }; for (auto & w : workers) { @@ -1711,7 +1737,8 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1); std::vector log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv); - std::vector kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk); + std::vector kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk); + std::vector p_diff_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk); std::vector logits; if (num_batches > 1) { logits.reserve(n_ctx * n_vocab); @@ -1728,9 +1755,18 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { df = df > 0 && count > 10 ? sqrt(df/(count-1)) : 0.; return std::make_pair(f, df); }; + auto covariance = [] (double suma, double sumb, double sumab, size_t count) { + if (count < 10) { + return 0.0; + } + double var = sumab/count - (suma/count)*(sumb/count); + var /= count - 1; + return var; + }; kl_divergence_result kld; - auto kld_ptr = kld_values.data(); + auto kld_ptr = kld_values.data(); + auto p_diff_ptr = p_diff_values.data(); for (int i = 0; i < n_chunk; ++i) { const int start = i * n_ctx; @@ -1785,24 +1821,42 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { } fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0); - printf("\nchunk PPL ln(PPL(Q)/PPL(base)) KL-Divergence Same top\n"); + printf("\nchunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p\n"); } const int first = n_ctx/2; const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx); process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, - workers, log_probs_uint16, kld, kld_ptr); - kld_ptr += n_ctx - 1 - first; + workers, log_probs_uint16, kld, kld_ptr, p_diff_ptr); + p_diff_ptr += n_ctx - 1 - first; + kld_ptr += n_ctx - 1 - first; - auto ppl = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count); - auto log_ppl_ratio = mean_and_uncertainty(kld.sum_nll_diff, kld.sum_nll_diff2, kld.count); - auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count); - auto p_top = 1.*kld.n_same_top/kld.count; - auto d_p_top = sqrt(p_top*(1 - p_top)/(kld.count - 1)); + printf("%4d", i+1); - printf("%4d %10.4lf %10.5lf ± %10.5f %10.5f ± %10.5lf %.5f ± %.5f\n", i+1, exp(ppl.first), - log_ppl_ratio.first, log_ppl_ratio.second, kl_div.first, kl_div.second, - p_top, d_p_top); + auto log_ppl = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count); + const double ppl_val = exp(log_ppl.first); + const double ppl_unc = ppl_val * log_ppl.second; // ppl_unc = sqrt( (dexp(x) / dx) ** 2 * log_ppl.second ** 2 ) + printf(" %9.4lf ± %9.4lf", ppl_val, ppl_unc); + + auto log_ppl_base = mean_and_uncertainty(kld.sum_nll_base, kld.sum_nll_base2, kld.count); + const double log_ppl_cov = covariance(kld.sum_nll, kld.sum_nll_base, kld.sum_nll_nll_base, kld.count); + const double log_ppl_ratio_val = log_ppl.first - log_ppl_base.first; + const double log_ppl_ratio_unc = sqrt(log_ppl.second*log_ppl.second + log_ppl_base.second*log_ppl_base.second - 2.0*log_ppl_cov); + printf(" %10.5lf ± %10.5lf", log_ppl_ratio_val, log_ppl_ratio_unc); + + auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count); + printf(" %10.5lf ± %10.5lf", kl_div.first, kl_div.second); + + auto p_diff_mse = mean_and_uncertainty(kld.sum_p_diff2, kld.sum_p_diff4, kld.count); + const double p_diff_rms_val = sqrt(p_diff_mse.first); + const double p_diff_rms_unc = 0.5/p_diff_rms_val * p_diff_mse.second; + printf(" %6.3lf ± %6.3lf %%", 100.0*p_diff_rms_val, 100.0*p_diff_rms_unc); + + double p_top_val = 1.*kld.n_same_top/kld.count; + double p_top_unc = sqrt(p_top_val*(1 - p_top_val)/(kld.count - 1)); + printf(" %6.3lf ± %6.3lf %%", 100.0*p_top_val, 100.0*p_top_unc); + + printf("\n"); fflush(stdout); @@ -1813,31 +1867,97 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { if (kld.count < 100) return; // we do not wish to do statistics on so few values std::sort(kld_values.begin(), kld_values.end()); + std::sort(p_diff_values.begin(), p_diff_values.end()); - printf("===== KL-divergence statistics\n"); + printf("====== Perplexity statistics ======\n"); + + auto log_ppl = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count); + const double ppl_val = exp(log_ppl.first); + const double ppl_unc = ppl_val * log_ppl.second; // ppl_unc = sqrt( (dexp(x) / dx) ** 2 * log_ppl.second ** 2 ) + printf("Mean PPL(Q) : %10.6lf ± %10.6lf\n", ppl_val, ppl_unc); + + auto log_ppl_base = mean_and_uncertainty(kld.sum_nll_base, kld.sum_nll_base2, kld.count); + const double ppl_base_val = exp(log_ppl_base.first); + const double ppl_base_unc = ppl_base_val * log_ppl_base.second; // ppl_base_unc = sqrt( (dexp(x) / dx) ** 2 * log_ppl_base.second ** 2 ) + printf("Mean PPL(base) : %10.6lf ± %10.6lf\n", ppl_base_val, ppl_base_unc); + + const double log_ppl_cov = covariance(kld.sum_nll, kld.sum_nll_base, kld.sum_nll_nll_base, kld.count); + // printf("Cov(ln(PPL(Q)), ln(PPL(base))): %10.6lf\n", log_ppl_cov); + const double log_ppl_cor = log_ppl_cov / (log_ppl.second*log_ppl_base.second); + printf("Cor(ln(PPL(Q)), ln(PPL(base))): %6.2lf%%\n", 100.0*log_ppl_cor); + + const double log_ppl_ratio_val = log_ppl.first - log_ppl_base.first; + const double log_ppl_ratio_unc = sqrt(log_ppl.second*log_ppl.second + log_ppl_base.second*log_ppl_base.second - 2.0*log_ppl_cov); + printf("Mean ln(PPL(Q)/PPL(base)) : %10.6lf ± %10.6lf\n", log_ppl_ratio_val, log_ppl_ratio_unc); + + const double ppl_ratio_val = exp(log_ppl_ratio_val); + const double ppl_ratio_unc = ppl_ratio_val * log_ppl_ratio_unc; // ppl_ratio_unc = sqrt( (dexp(x) / dx) ** 2 * log_ppl_ratio.second ** 2 ) + printf("Mean PPL(Q)/PPL(base) : %10.6lf ± %10.6lf\n", ppl_ratio_val, ppl_ratio_unc); + + const double ppl_cov = ppl_val * ppl_base_val * log_ppl_cov; + const double ppl_diff_val = ppl_val - ppl_base_val; + const double ppl_diff_unc = sqrt(ppl_unc*ppl_unc + ppl_base_unc*ppl_base_unc - 2.0*ppl_cov); + printf("Mean PPL(Q)-PPL(base) : %10.6lf ± %10.6lf\n", ppl_diff_val, ppl_diff_unc); + + printf("\n"); + + printf("====== KL divergence statistics ======\n"); auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count); - printf("Average: %10.6f ±%10.6lf\n", kl_div.first, kl_div.second); + printf("Mean KLD: %10.6lf ± %10.6lf\n", kl_div.first, kl_div.second); auto kld_median = kld_values.size()%2 == 0 ? 0.5f*(kld_values[kld_values.size()/2] + kld_values[kld_values.size()/2-1]) : kld_values[kld_values.size()/2]; - printf("Median : %10.6f\n", kld_median); - auto percentile = [&kld_values] (float fraction) { - if (fraction <= 0) return kld_values.front(); - if (fraction >= 1) return kld_values.back(); - float p = fraction*(kld_values.size() - 1); + auto percentile = [] (std::vector values, float fraction) { + if (fraction <= 0) return values.front(); + if (fraction >= 1) return values.back(); + float p = fraction*(values.size() - 1); size_t ip = size_t(p); p -= ip; - return (1 - p)*kld_values[ip] + p*kld_values[std::min(ip+1, kld_values.size()-1)]; + return (1 - p)*values[ip] + p*values[std::min(ip+1, values.size()-1)]; }; - printf("Maximum: %10.6f\n", kld_values.back()); - printf("KLD_99 : %10.6f\n", percentile(0.99f)); - printf("KLD_95 : %10.6f\n", percentile(0.95f)); - printf("KLD_90 : %10.6f\n", percentile(0.90f)); + printf("Maximum KLD: %10.6f\n", kld_values.back()); + printf("99.9%% KLD: %10.6f\n", percentile(kld_values, 0.999f)); + printf("99.0%% KLD: %10.6f\n", percentile(kld_values, 0.990f)); + printf("99.0%% KLD: %10.6f\n", percentile(kld_values, 0.990f)); + printf("Median KLD: %10.6f\n", kld_median); + printf("10.0%% KLD: %10.6f\n", percentile(kld_values, 0.100f)); + printf(" 5.0%% KLD: %10.6f\n", percentile(kld_values, 0.050f)); + printf(" 1.0%% KLD: %10.6f\n", percentile(kld_values, 0.010f)); + printf("Minimum KLD: %10.6f\n", kld_values.front()); - printf("Minimum: %10.6f\n", kld_values.front()); - printf("KLD_01 : %10.6f\n", percentile(0.01f)); - printf("KLD_05 : %10.6f\n", percentile(0.05f)); - printf("KLD_10 : %10.6f\n", percentile(0.10f)); + printf("\n"); + + printf("====== Token probability statistics ======\n"); + + auto p_diff = mean_and_uncertainty(kld.sum_p_diff, kld.sum_p_diff2, kld.count); + printf("Mean Δp: %6.3lf ± %5.3lf %%\n", 100.0*p_diff.first, 100.0*p_diff.second); + + auto p_diff_median = p_diff_values.size()%2 == 0 ? 0.5f*(p_diff_values[p_diff_values.size()/2] + p_diff_values[p_diff_values.size()/2-1]) + : p_diff_values[p_diff_values.size()/2]; + + printf("Maximum Δp: %6.3lf%%\n", 100.0*p_diff_values.back()); + printf("99.9%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.999f)); + printf("99.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.990f)); + printf("95.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.950f)); + printf("90.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.900f)); + printf("75.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.750f)); + printf("Median Δp: %6.3lf%%\n", 100.0*p_diff_median); + printf("25.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.250f)); + printf("10.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.100f)); + printf(" 5.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.050f)); + printf(" 1.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.010f)); + printf(" 0.1%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.001f)); + printf("Minimum Δp: %6.3lf%%\n", 100.0*p_diff_values.front()); + + auto p_diff_mse = mean_and_uncertainty(kld.sum_p_diff2, kld.sum_p_diff4, kld.count); + // printf("MSE Δp : %10.6lf ± %10.6lf\n", p_diff_mse.first, p_diff_mse.second); + + const double p_diff_rms_val = sqrt(p_diff_mse.first); + const double p_diff_rms_unc = 0.5/p_diff_rms_val * p_diff_mse.second; + printf("RMS Δp : %6.3lf ± %5.3lf %%\n", 100.0*p_diff_rms_val, 100.0*p_diff_rms_unc); + + const double same_top_p = 1.0*kld.n_same_top/kld.count; + printf("Same top p: %6.3lf ± %5.3lf %%\n", 100.0*same_top_p, 100.0*sqrt(same_top_p*(1.0 - same_top_p)/(kld.count - 1))); } diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp index 1d05f1391..746df8446 100644 --- a/examples/quantize-stats/quantize-stats.cpp +++ b/examples/quantize-stats/quantize-stats.cpp @@ -23,7 +23,7 @@ #endif struct quantize_stats_params { - std::string model = "models/7B/ggml-model-f16.gguf"; + std::string model = DEFAULT_MODEL_PATH; bool verbose = false; bool per_layer_stats = false; bool print_histogram = false; diff --git a/examples/server/README.md b/examples/server/README.md index 918ac1295..b96a4444a 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -74,15 +74,18 @@ page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/ - Using `make`: ```bash - make + make server ``` - Using `CMake`: ```bash - cmake --build . --config Release + cmake -B build + cmake --build build --config Release -t server ``` + Binary is at `./build/bin/server` + ## Build with SSL `server` can also be built with SSL support using OpenSSL 3 @@ -99,10 +102,8 @@ page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/ - Using `CMake`: ```bash - mkdir build - cd build - cmake .. -DLLAMA_SERVER_SSL=ON - make server + cmake -B build -DLLAMA_SERVER_SSL=ON + cmake --build build --config Release -t server ``` ## Quick Start diff --git a/examples/server/bench/bench.py b/examples/server/bench/bench.py index 6ca637bdd..86c5de101 100644 --- a/examples/server/bench/bench.py +++ b/examples/server/bench/bench.py @@ -268,6 +268,7 @@ def start_server_background(args): server_args.extend(['--defrag-thold', "0.1"]) server_args.append('--cont-batching') server_args.append('--metrics') + server_args.append('--flash-attn') server_args.extend(['--log-format', "text"]) args = [str(arg) for arg in [server_path, *server_args]] print(f"bench: starting server with: {' '.join(args)}") diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 2760aea8f..ff0814b2f 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1383,9 +1383,10 @@ struct server_context { if (!slot.params.stream && slot.stopped_word) { const std::vector stop_word_toks = llama_tokenize(ctx, slot.stopping_word, false); + size_t safe_offset = std::min(slot.generated_token_probs.size(), stop_word_toks.size()); probs = std::vector( slot.generated_token_probs.begin(), - slot.generated_token_probs.end() - stop_word_toks.size()); + slot.generated_token_probs.end() - safe_offset); } else { probs = std::vector( slot.generated_token_probs.begin(), @@ -2353,7 +2354,7 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co printf(" disable KV offload\n"); } printf(" -m FNAME, --model FNAME\n"); - printf(" model path (default: %s)\n", params.model.c_str()); + printf(" model path (default: models/$filename with filename from --hf-file or --model-url if set, otherwise %s)\n", DEFAULT_MODEL_PATH); printf(" -mu MODEL_URL, --model-url MODEL_URL\n"); printf(" model download url (default: unused)\n"); printf(" -hfr REPO, --hf-repo REPO\n"); @@ -2377,6 +2378,7 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co printf(" --embeddings enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled"); printf(" -np N, --parallel N number of slots for process requests (default: %d)\n", params.n_parallel); printf(" -cb, --cont-batching enable continuous batching (a.k.a dynamic batching) (default: enabled)\n"); + printf(" -fa, --flash-attn enable Flash Attention (default: %s)\n", params.flash_attn ? "enabled" : "disabled"); printf(" -spf FNAME, --system-prompt-file FNAME\n"); printf(" set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n"); printf(" -ctk TYPE, --cache-type-k TYPE\n"); @@ -2742,6 +2744,8 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams, params.embedding = true; } else if (arg == "-cb" || arg == "--cont-batching") { params.cont_batching = true; + } else if (arg == "-fa" || arg == "--flash-attn") { + params.flash_attn = true; } else if (arg == "-np" || arg == "--parallel") { if (++i >= argc) { invalid_param = true; @@ -2835,6 +2839,8 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams, } } + gpt_params_handle_model_default(params); + if (!params.kv_overrides.empty()) { params.kv_overrides.emplace_back(); params.kv_overrides.back().key[0] = 0; diff --git a/examples/server/tests/features/embeddings.feature b/examples/server/tests/features/embeddings.feature index dcf1434f9..6f163ce04 100644 --- a/examples/server/tests/features/embeddings.feature +++ b/examples/server/tests/features/embeddings.feature @@ -5,7 +5,7 @@ Feature: llama.cpp server Background: Server startup Given a server listening on localhost:8080 And a model url https://huggingface.co/ggml-org/models/resolve/main/bert-bge-small/ggml-model-f16.gguf - And a model file ggml-model-f16.gguf + And a model file bert-bge-small.gguf And a model alias bert-bge-small And 42 as server seed And 2 slots diff --git a/examples/server/tests/features/results.feature b/examples/server/tests/features/results.feature index f17120f7b..aa0b8d0c6 100644 --- a/examples/server/tests/features/results.feature +++ b/examples/server/tests/features/results.feature @@ -7,44 +7,16 @@ Feature: Results And a model file tinyllamas/split/stories15M-00001-of-00003.gguf from HF repo ggml-org/models And a model file test-model-00001-of-00003.gguf And 128 as batch size - And 256 KV cache size + And 1024 KV cache size And 128 max tokens to predict - - Scenario Outline: Multi users completion - Given slots And continuous batching + + Scenario Outline: consistent results with same seed + Given slots Then the server is starting Then the server is healthy - Given 42 as seed - And a prompt: - """ - Write a very long story about AI. - """ - - Given 42 as seed - And a prompt: - """ - Write a very long story about AI. - """ - - Given 42 as seed - And a prompt: - """ - Write a very long story about AI. - """ - - Given 42 as seed - And a prompt: - """ - Write a very long story about AI. - """ - - Given 42 as seed - And a prompt: - """ - Write a very long story about AI. - """ + Given 4 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 42 Given concurrent completion requests Then the server is busy @@ -55,3 +27,55 @@ Feature: Results | n_slots | | 1 | | 2 | + + Scenario Outline: different results with different seed + Given slots + Then the server is starting + Then the server is healthy + + Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 42 + Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 43 + Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 44 + Given 1 prompts "Title: Little Red Riding Hood But In Space\n\nSummary:" with seed 45 + + Given concurrent completion requests + Then the server is busy + Then the server is idle + And all slots are idle + Then all predictions are different + Examples: + | n_slots | + | 1 | + | 2 | + + Scenario Outline: consistent results with same seed and varying batch size + Given 4 slots + And temperature + # And 0 as draft + Then the server is starting + Then the server is healthy + + Given 1 prompts "Write a very long story about AI." with seed 42 + And concurrent completion requests + # Then the server is busy # Not all slots will be utilized. + Then the server is idle + And all slots are idle + + Given prompts "Write a very long story about AI." with seed 42 + And concurrent completion requests + # Then the server is busy # Not all slots will be utilized. + Then the server is idle + And all slots are idle + + Then all predictions are equal + Examples: + | n_parallel | temp | + | 1 | 0.0 | + | 2 | 0.0 | + | 4 | 0.0 | + | 1 | 1.0 | + # FIXME: These tests fail on master. The problem seems to be the unified KV cache. + # See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227 + # and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574 . + # | 2 | 1.0 | + # | 4 | 1.0 | diff --git a/examples/server/tests/features/steps/steps.py b/examples/server/tests/features/steps/steps.py index f71e0d706..b8dbef21d 100644 --- a/examples/server/tests/features/steps/steps.py +++ b/examples/server/tests/features/steps/steps.py @@ -65,6 +65,7 @@ def step_server_config(context, server_fqdn, server_port): context.server_seed = None context.user_api_key = None context.response_format = None + context.temperature = None context.tasks_result = [] context.concurrent_tasks = [] @@ -232,15 +233,17 @@ async def step_all_slots_status(context, expected_slot_status_string): @async_run_until_complete async def step_request_completion(context, api_error): expect_api_error = api_error == 'raised' + seeds = await completions_seed(context, num_seeds=1) completion = await request_completion(context.prompts.pop(), + seeds[0] if seeds is not None else seeds, context.base_url, debug=context.debug, n_predict=context.n_predict, cache_prompt=context.cache_prompt, id_slot=context.id_slot, - seed=await completions_seed(context), expect_api_error=expect_api_error, - user_api_key=context.user_api_key) + user_api_key=context.user_api_key, + temperature=context.temperature) context.tasks_result.append(completion) if context.debug: print(f"Completion response: {completion}") @@ -269,6 +272,15 @@ async def step_predictions_equal(context): context.tasks_result = [] +@step('all predictions are different') +@async_run_until_complete +async def step_predictions_equal(context): + n_completions = await gather_tasks_results(context) + assert n_completions >= 2, "need at least 2 completions" + assert_all_predictions_different(context.tasks_result) + context.tasks_result = [] + + @step('the completion is truncated') def step_assert_completion_truncated(context): step_assert_completion_truncated(context, '') @@ -311,6 +323,11 @@ def step_response_format(context, response_format): context.response_format = json.loads(response_format) +@step('{temperature:f} temperature') +def step_temperature(context, temperature): + context.temperature = temperature + + @step('streaming is {enable_streaming}') def step_streaming(context, enable_streaming): context.enable_streaming = enable_streaming == 'enabled' @@ -353,7 +370,10 @@ def step_n_ubatch(context, n_ubatch): @step('{seed:d} as seed') def step_seed(context, seed): - context.seed = seed + if context.seed is None: + context.seed = [seed] + else: + context.seed.append(seed) @step('a prefix prompt') @@ -413,7 +433,9 @@ async def step_oai_chat_completions(context, api_error): if context.debug: print(f"Submitting OAI compatible completions request...") expect_api_error = api_error == 'raised' + seeds = await completions_seed(context, num_seeds=1), completion = await oai_chat_completions(context.prompts.pop(), + seeds[0] if seeds is not None else seeds, context.system_prompt, context.base_url, '/v1/chat', @@ -429,8 +451,6 @@ async def step_oai_chat_completions(context, api_error): response_format=context.response_format if hasattr(context, 'response_format') else None, - seed=await completions_seed(context), - user_api_key=context.user_api_key if hasattr(context, 'user_api_key') else None, @@ -457,20 +477,31 @@ def step_a_prompt_prompt(context, prompt): context.n_prompts = len(context.prompts) +@step('{num_prompts:d} prompts {prompt} with seed {seed:d}') +def step_many_prompts(context, num_prompts, prompt, seed): + if context.seed is None: + context.seed = [] + for _ in range(num_prompts): + context.seed.append(seed) + context.prompts.append(prompt) + context.n_prompts = len(context.prompts) + + @step('concurrent completion requests') @async_run_until_complete() async def step_concurrent_completion_requests(context): - await concurrent_requests(context, - request_completion, - # prompt is inserted automatically - context.base_url, - debug=context.debug, - prompt_prefix=context.prompt_prefix, - prompt_suffix=context.prompt_suffix, - n_predict=context.n_predict if hasattr(context, 'n_predict') else None, - seed=await completions_seed(context), - user_api_key=context.user_api_key if hasattr(context, - 'user_api_key') else None) + await concurrent_requests( + context, + request_completion, + # prompt is inserted automatically + context.base_url, + debug=context.debug, + prompt_prefix=context.prompt_prefix, + prompt_suffix=context.prompt_suffix, + n_predict=context.n_predict if hasattr(context, 'n_predict') else None, + user_api_key=context.user_api_key if hasattr(context, 'user_api_key') else None, + temperature=context.temperature, + ) @step('concurrent OAI completions requests') @@ -490,7 +521,6 @@ async def step_oai_chat_completions(context): if hasattr(context, 'enable_streaming') else None, response_format=context.response_format if hasattr(context, 'response_format') else None, - seed=await completions_seed(context), user_api_key=context.user_api_key if hasattr(context, 'user_api_key') else None) @@ -512,10 +542,6 @@ async def step_oai_chat_completions(context): if hasattr(context, 'enable_streaming') else None, response_format=context.response_format if hasattr(context, 'response_format') else None, - seed=context.seed - if hasattr(context, 'seed') else - context.server_seed - if hasattr(context, 'server_seed') else None, user_api_key=context.user_api_key if hasattr(context, 'user_api_key') else None) @@ -544,7 +570,7 @@ async def all_prompts_are_predicted(context, expected_predicted_n=None): @async_run_until_complete async def step_compute_embedding(context): context.n_prompts = 1 - context.embeddings = await request_embedding(context_text(context), base_url=context.base_url) + context.embeddings = await request_embedding(context_text(context), None, base_url=context.base_url) @step('all embeddings are the same') @@ -585,7 +611,7 @@ def step_assert_embeddings(context): @async_run_until_complete async def step_oai_compute_embeddings(context): context.n_prompts = 1 - context.embeddings = await request_oai_embeddings(context_text(context), + context.embeddings = await request_oai_embeddings(context_text(context), None, base_url=context.base_url, user_api_key=context.user_api_key, model=context.model) @@ -594,7 +620,7 @@ async def step_oai_compute_embeddings(context): @step('an OAI compatible embeddings computation request for multiple inputs') @async_run_until_complete async def step_oai_compute_embeddings_multiple_inputs(context): - context.embeddings = await request_oai_embeddings(context.prompts, + context.embeddings = await request_oai_embeddings(context.prompts, None, base_url=context.base_url, user_api_key=context.user_api_key, model=context.model) @@ -740,8 +766,9 @@ async def concurrent_requests(context, f_completion, *args, **kwargs): if context.debug: print(f"starting {context.n_prompts} concurrent completion requests...") assert context.n_prompts > 0 + seeds = await completions_seed(context) for prompt_no in range(context.n_prompts): - shifted_args = [context.prompts.pop(), *args] + shifted_args = [context.prompts.pop(), seeds[prompt_no], *args] context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs))) await asyncio.sleep(0.1) @@ -781,6 +808,7 @@ def step_server_responds_with_status_code(context, status_code): async def request_completion(prompt, + seed, base_url, debug=False, prompt_prefix=None, @@ -788,9 +816,9 @@ async def request_completion(prompt, n_predict=None, cache_prompt=False, id_slot=None, - seed=None, expect_api_error=None, - user_api_key=None): + user_api_key=None, + temperature=None): if debug: print(f"Sending completion request: {prompt}") origin = "my.super.domain" @@ -811,7 +839,8 @@ async def request_completion(prompt, "n_predict": n_predict if n_predict is not None else -1, "cache_prompt": cache_prompt, "id_slot": id_slot, - "seed": seed if seed is not None else 42 + "seed": seed if seed is not None else 42, + "temperature": temperature if temperature is not None else "0.8f", }, headers=headers, timeout=3600) as response: @@ -824,6 +853,7 @@ async def request_completion(prompt, async def oai_chat_completions(user_prompt, + seed, system_prompt, base_url, base_path, @@ -833,7 +863,6 @@ async def oai_chat_completions(user_prompt, n_predict=None, enable_streaming=None, response_format=None, - seed=None, user_api_key=None, expect_api_error=None): if debug: @@ -952,7 +981,7 @@ async def oai_chat_completions(user_prompt, return completion_response -async def request_embedding(content, base_url=None): +async def request_embedding(content, seed, base_url=None): async with aiohttp.ClientSession() as session: async with session.post(f'{base_url}/embedding', json={ @@ -963,7 +992,7 @@ async def request_embedding(content, base_url=None): return [response_json['embedding']] -async def request_oai_embeddings(input, +async def request_oai_embeddings(input, seed, base_url=None, user_api_key=None, model=None, async_client=False): # openai client always expects an api_key @@ -1036,21 +1065,31 @@ def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re f' {n_predicted} <> {expected_predicted_n}') def assert_all_predictions_equal(completion_responses): - content_0 = completion_responses[0]['content'] - if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON': - print(f"content 0: {content_0}") + for i, response_i in enumerate(completion_responses): + content_i = response_i['content'] + print(f"content {i}: {content_i}") + for i, response_i in enumerate(completion_responses): + content_i = response_i['content'] + for j, response_j in enumerate(completion_responses): + if i == j: + continue + content_j = response_j['content'] + assert content_i == content_j, "contents not equal" - i = 1 - for response in completion_responses[1:]: - content = response['content'] - if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON': - print(f"content {i}: {content}") - - assert content == content_0, "contents not equal" - - i += 1 +def assert_all_predictions_different(completion_responses): + if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON': + for i, response_i in enumerate(completion_responses): + content_i = response_i['content'] + print(f"content {i}: {content_i}") + for i, response_i in enumerate(completion_responses): + content_i = response_i['content'] + for j, response_j in enumerate(completion_responses): + if i == j: + continue + content_j = response_j['content'] + assert content_i != content_j, "contents not different" async def gather_tasks_results(context): @@ -1145,9 +1184,22 @@ def assert_slots_status(slots, expected_slots): f" = {expected[key]} != {slot[key]}") -async def completions_seed(context): - return context.seed if hasattr(context, 'seed') and context.seed is not None \ - else context.server_seed if hasattr(context, 'server_seed') else None +async def completions_seed(context, num_seeds=None): + if hasattr(context, "seed") and context.seed is not None: + assert len(context.seed) == context.n_prompts + if num_seeds is None: + num_seeds = context.n_prompts + assert num_seeds <= context.n_prompts + seeds = context.seed[:num_seeds] + context.seed = context.seed[num_seeds:] if num_seeds < context.n_prompts else None + return seeds + + if hasattr(context, "server_seed") and context.server_seed is not None: + if num_seeds is None: + return [context.server_seed] * context.n_prompts + else: + return [context.server_seed] * num_seeds + return None def context_text(context): diff --git a/ggml-cuda.cu b/ggml-cuda.cu index d277104d1..c30554f0c 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -14,6 +14,7 @@ #include "ggml-cuda/cpy.cuh" #include "ggml-cuda/diagmask.cuh" #include "ggml-cuda/dmmv.cuh" +#include "ggml-cuda/fattn.cuh" #include "ggml-cuda/getrows.cuh" #include "ggml-cuda/im2col.cuh" #include "ggml-cuda/mmq.cuh" @@ -140,6 +141,7 @@ static ggml_cuda_device_info ggml_cuda_init() { info.devices[id].cc = 100*prop.major + 10*prop.minor; #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) info.devices[id].smpb = prop.sharedMemPerBlock; + info.devices[id].nsm = prop.multiProcessorCount; } for (int id = 0; id < info.device_count; ++id) { @@ -2290,6 +2292,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_ARGSORT: ggml_cuda_op_argsort(ctx, dst); break; + case GGML_OP_FLASH_ATTN_EXT: + ggml_cuda_flash_attn_ext(ctx, dst); + break; default: return false; } @@ -2564,6 +2569,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons case GGML_OP_ARANGE: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_LEAKY_RELU: + case GGML_OP_FLASH_ATTN_EXT: return true; default: return false; diff --git a/ggml-cuda/common.cuh b/ggml-cuda/common.cuh index 481065b2a..b2627b7b4 100644 --- a/ggml-cuda/common.cuh +++ b/ggml-cuda/common.cuh @@ -137,11 +137,13 @@ #define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__) #define WARP_SIZE 32 -#define CUDART_HMAX 11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed) +#define CUDART_HMAX 11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed) +#define CUDART_HMASK 12000 // CUDA 12.0, min. ver. for half2 -> uint mask comparisons #define CC_PASCAL 600 #define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products #define CC_VOLTA 700 +#define CC_AMPERE 800 #define CC_OFFSET_AMD 1000000 #define CC_RDNA1 (CC_OFFSET_AMD + 1010) #define CC_RDNA2 (CC_OFFSET_AMD + 1030) @@ -271,7 +273,6 @@ static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) { return a; } -#ifdef GGML_CUDA_F16 static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) { #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL #pragma unroll @@ -284,7 +285,6 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) { NO_DEVICE_CODE; #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL } -#endif // GGML_CUDA_F16 static __device__ __forceinline__ float warp_reduce_max(float x) { #pragma unroll @@ -294,19 +294,60 @@ static __device__ __forceinline__ float warp_reduce_max(float x) { return x; } -//static __device__ __forceinline__ half2 warp_reduce_max(half2 x) { -//#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX -//#pragma unroll -// for (int mask = 16; mask > 0; mask >>= 1) { -// x = __hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32)); -// } -// return x; -//#else -// GGML_UNUSED(x); -// NO_DEVICE_CODE; -//#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX -//} +static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) { +#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) +#if CUDART_VERSION >= CUDART_HMAX + return __hmax(a, b); +#else + return __half2float(a) > __half2float(b) ? a : b; +#endif // CUDART_VERSION >= CUDART_HMAX + +#else + GGML_UNUSED(a); + GGML_UNUSED(b); + NO_DEVICE_CODE; +#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX +} +static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) { +#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) + +#if CUDART_VERSION >= CUDART_HMAX + return __hmax2(a, b); +#else + half2 ret; + reinterpret_cast(ret.x) = __low2float(a) > __low2float(b) ? __low2half(a) : __low2half(b); + reinterpret_cast(ret.y) = __high2float(a) > __high2float(b) ? __high2half(a) : __high2half(b); + return ret; +#endif // CUDART_VERSION >= CUDART_HMAX + +#else + GGML_UNUSED(a); + GGML_UNUSED(b); + NO_DEVICE_CODE; +#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX +} + +static __device__ __forceinline__ half2 warp_reduce_max(half2 x) { +#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32)); + } + return x; +#else + GGML_UNUSED(x); + NO_DEVICE_CODE; +#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL +} + +#if CUDART_VERSION < CUDART_HMASK +static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) { + const uint32_t mask_low = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b))); + const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b))); + return mask_low | mask_high; +} +#endif // CUDART_VERSION < 12000 #if defined(GGML_USE_HIPBLAS) #define __CUDA_ARCH__ 1300 @@ -391,6 +432,11 @@ static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) { } #endif // defined(GGML_USE_HIPBLAS) +#define FP16_AVAILABLE defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) ? \ + defined(RDNA1) || defined(RDNA2) || defined(RDNA3) : __CUDA_ARCH__ >= CC_PASCAL + +#define FP16_MMA_AVAILABLE !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA + // TODO: move to ggml-common.h static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; @@ -404,6 +450,7 @@ struct ggml_cuda_device_info { struct cuda_device_info { int cc; // compute capability + int nsm; // number of streaming multiprocessors size_t smpb; // max. shared memory per block bool vmm; // virtual memory support size_t vmm_granularity; // granularity of virtual memory diff --git a/ggml-cuda/convert.cu b/ggml-cuda/convert.cu index b15e35782..75e50c985 100644 --- a/ggml-cuda/convert.cu +++ b/ggml-cuda/convert.cu @@ -5,16 +5,16 @@ template static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) { - const int64_t i = 2*(blockDim.x*blockIdx.x + threadIdx.x); + const int64_t i = (int64_t)2*(blockDim.x*blockIdx.x + threadIdx.x); if (i >= k) { return; } const int64_t ib = i/qk; // block index - const int iqs = (i%qk)/qr; // quant index - const int iybs = i - i%qk; // y block start index - const int y_offset = qr == 1 ? 1 : qk/2; + const int64_t iqs = (i%qk)/qr; // quant index + const int64_t iybs = i - i%qk; // y block start index + const int64_t y_offset = qr == 1 ? 1 : qk/2; // dequantize dfloat2 v; @@ -29,7 +29,7 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h #if __CUDA_ARCH__ >= CC_PASCAL constexpr int nint = CUDA_Q8_0_NE_ALIGN/sizeof(int) + WARP_SIZE; - const int i0 = CUDA_Q8_0_NE_ALIGN*blockIdx.x; + const int64_t i0 = CUDA_Q8_0_NE_ALIGN*blockIdx.x; const int * x0 = ((int *) vx) + blockIdx.x * nint; half2 * y2 = (half2 *) (y + i0); @@ -73,9 +73,9 @@ static __global__ void dequantize_block_q4_0(const void * __restrict__ vx, dst_t const int64_t i = blockIdx.x; // assume 32 threads - const int tid = threadIdx.x; - const int il = tid/8; - const int ir = tid%8; + const int64_t tid = threadIdx.x; + const int64_t il = tid/8; + const int64_t ir = tid%8; const int64_t ib = 8*i + ir; if (ib >= nb32) { return; @@ -101,9 +101,9 @@ static __global__ void dequantize_block_q4_1(const void * __restrict__ vx, dst_t const int64_t i = blockIdx.x; // assume 32 threads - const int tid = threadIdx.x; - const int il = tid/8; - const int ir = tid%8; + const int64_t tid = threadIdx.x; + const int64_t il = tid/8; + const int64_t ir = tid%8; const int64_t ib = 8*i + ir; if (ib >= nb32) { return; @@ -127,14 +127,14 @@ static __global__ void dequantize_block_q4_1(const void * __restrict__ vx, dst_t template static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; const block_q2_K * x = (const block_q2_K *) vx; - const int tid = threadIdx.x; + const int64_t tid = threadIdx.x; #if QK_K == 256 - const int n = tid/32; - const int l = tid - 32*n; - const int is = 8*n + l/16; + const int64_t n = tid/32; + const int64_t l = tid - 32*n; + const int64_t is = 8*n + l/16; const uint8_t q = x[i].qs[32*n + l]; dst_t * y = yy + i*QK_K + 128*n; @@ -146,8 +146,8 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4); y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4); #else - const int is = tid/16; // 0 or 1 - const int il = tid%16; // 0...15 + const int64_t is = tid/16; // 0 or 1 + const int64_t il = tid%16; // 0...15 const uint8_t q = x[i].qs[il] >> (2*is); dst_t * y = yy + i*QK_K + 16*is + il; float dall = __low2half(x[i].dm); @@ -161,19 +161,19 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t template static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; const block_q3_K * x = (const block_q3_K *) vx; #if QK_K == 256 - const int r = threadIdx.x/4; - const int tid = r/2; - const int is0 = r%2; - const int l0 = 16*is0 + 4*(threadIdx.x%4); - const int n = tid / 4; - const int j = tid - 4*n; + const int64_t r = threadIdx.x/4; + const int64_t tid = r/2; + const int64_t is0 = r%2; + const int64_t l0 = 16*is0 + 4*(threadIdx.x%4); + const int64_t n = tid / 4; + const int64_t j = tid - 4*n; uint8_t m = 1 << (4*n + j); - int is = 8*n + 2*j + is0; + int64_t is = 8*n + 2*j + is0; int shift = 2*j; int8_t us = is < 4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) : @@ -189,11 +189,11 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)); #else - const int tid = threadIdx.x; - const int is = tid/16; // 0 or 1 - const int il = tid%16; // 0...15 - const int im = il/8; // 0...1 - const int in = il%8; // 0...7 + const int64_t tid = threadIdx.x; + const int64_t is = tid/16; // 0 or 1 + const int64_t il = tid%16; // 0...15 + const int64_t im = il/8; // 0...1 + const int64_t in = il%8; // 0...7 dst_t * y = yy + i*QK_K + 16*is + il; @@ -227,15 +227,15 @@ template static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { const block_q4_K * x = (const block_q4_K *) vx; - const int i = blockIdx.x; + const int64_t i = blockIdx.x; #if QK_K == 256 // assume 32 threads - const int tid = threadIdx.x; - const int il = tid/8; - const int ir = tid%8; - const int is = 2*il; - const int n = 4; + const int64_t tid = threadIdx.x; + const int64_t il = tid/8; + const int64_t ir = tid%8; + const int64_t is = 2*il; + const int64_t n = 4; dst_t * y = yy + i*QK_K + 64*il + n*ir; @@ -254,7 +254,7 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t y[l +32] = d2 * (q[l] >> 4) - m2; } #else - const int tid = threadIdx.x; + const int64_t tid = threadIdx.x; const uint8_t * q = x[i].qs; dst_t * y = yy + i*QK_K; const float d = (float)x[i].dm[0]; @@ -268,14 +268,14 @@ template static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) { const block_q5_K * x = (const block_q5_K *) vx; - const int i = blockIdx.x; + const int64_t i = blockIdx.x; #if QK_K == 256 // assume 64 threads - this is very slightly better than the one below - const int tid = threadIdx.x; - const int il = tid/16; // il is in 0...3 - const int ir = tid%16; // ir is in 0...15 - const int is = 2*il; // is is in 0...6 + const int64_t tid = threadIdx.x; + const int64_t il = tid/16; // il is in 0...3 + const int64_t ir = tid%16; // ir is in 0...15 + const int64_t is = 2*il; // is is in 0...6 dst_t * y = yy + i*QK_K + 64*il + 2*ir; @@ -298,11 +298,11 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2; y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2; #else - const int tid = threadIdx.x; + const int64_t tid = threadIdx.x; const uint8_t q = x[i].qs[tid]; - const int im = tid/8; // 0...3 - const int in = tid%8; // 0...7 - const int is = tid/16; // 0 or 1 + const int64_t im = tid/8; // 0...3 + const int64_t in = tid%8; // 0...7 + const int64_t is = tid/16; // 0 or 1 const uint8_t h = x[i].qh[in] >> im; const float d = x[i].d; dst_t * y = yy + i*QK_K + tid; @@ -359,13 +359,13 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t template static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; const block_iq2_xxs * x = (const block_iq2_xxs *) vx; - const int tid = threadIdx.x; + const int64_t tid = threadIdx.x; #if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 + const int64_t il = tid/8; // 0...3 + const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; const uint16_t * q2 = x[i].qs + 4*ib; const uint8_t * aux8 = (const uint8_t *)q2; @@ -383,13 +383,13 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds template static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; const block_iq2_xs * x = (const block_iq2_xs *) vx; - const int tid = threadIdx.x; + const int64_t tid = threadIdx.x; #if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 + const int64_t il = tid/8; // 0...3 + const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; const uint16_t * q2 = x[i].qs + 4*ib; const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511)); @@ -405,13 +405,13 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst template static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; const block_iq2_s * x = (const block_iq2_s *) vx; - const int tid = threadIdx.x; + const int64_t tid = threadIdx.x; #if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 + const int64_t il = tid/8; // 0...3 + const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300))); const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; @@ -426,13 +426,13 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_ template static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; const block_iq3_xxs * x = (const block_iq3_xxs *) vx; - const int tid = threadIdx.x; + const int64_t tid = threadIdx.x; #if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 + const int64_t il = tid/8; // 0...3 + const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; const uint8_t * q3 = x[i].qs + 8*ib; const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib; @@ -454,13 +454,13 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds template static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; const block_iq3_s * x = (const block_iq3_s *) vx; - const int tid = threadIdx.x; + const int64_t tid = threadIdx.x; #if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 + const int64_t il = tid/8; // 0...3 + const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; const uint8_t * qs = x[i].qs + 8*ib; const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256))); @@ -480,13 +480,13 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_ template static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; const block_iq1_s * x = (const block_iq1_s *) vx; - const int tid = threadIdx.x; + const int64_t tid = threadIdx.x; #if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 + const int64_t il = tid/8; // 0...3 + const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA; const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1); @@ -506,18 +506,18 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_ template static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; const block_iq1_m * x = (const block_iq1_m *) vx; - const int tid = threadIdx.x; + const int64_t tid = threadIdx.x; #if QK_K == 256 - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 + const int64_t il = tid/8; // 0...3 + const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 8*il; const uint16_t * sc = (const uint16_t *)x[i].scales; iq1m_scale_t scale; scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); - const int ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4); + const int64_t ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4); const float d = (float)scale.f16 * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1); const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA; uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32; @@ -537,12 +537,12 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_ template static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL); - const int tid = threadIdx.x; - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 + const int64_t tid = threadIdx.x; + const int64_t il = tid/8; // 0...3 + const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 4*il; const uint8_t * q4 = x[ib].qs + 4*il; const float d = (float)x[ib].d; @@ -556,12 +556,12 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst #if QK_K != 64 template static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) { - const int i = blockIdx.x; + const int64_t i = blockIdx.x; const block_iq4_xs * x = (const block_iq4_xs *)vx; - const int tid = threadIdx.x; - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 + const int64_t tid = threadIdx.x; + const int64_t il = tid/8; // 0...3 + const int64_t ib = tid%8; // 0...7 dst_t * y = yy + i*QK_K + 32*ib + 4*il; const uint8_t * q4 = x[i].qs + 16*ib + 4*il; const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32); diff --git a/ggml-cuda/fattn.cu b/ggml-cuda/fattn.cu new file mode 100644 index 000000000..c8a11d173 --- /dev/null +++ b/ggml-cuda/fattn.cu @@ -0,0 +1,944 @@ +#include "common.cuh" +#include "fattn.cuh" + +#include + +#if FP16_MMA_AVAILABLE +#include +#endif + +#define FATTN_KQ_STRIDE 256 +#define HALF_MAX_HALF __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction. +#define SOFTMAX_FTZ_THRESHOLD -20.0f // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs. + +template // D == head size +__launch_bounds__(((D + WARP_SIZE - 1) / WARP_SIZE)*WARP_SIZE, 1) +static __global__ void flash_attn_vec_ext_f16( + const char * __restrict__ Q, + const char * __restrict__ K, + const char * __restrict__ V, + const char * __restrict__ mask, + float * __restrict__ dst, + float2 * __restrict__ dst_meta, + const float scale, + const int ne00, + const int ne01, + const int ne02, + const int ne03, + const int ne10, + const int ne11, + const int ne12, + const int ne13, + const int ne31, + const int nb31, + const int nb01, + const int nb02, + const int nb03, + const int nb11, + const int nb12, + const int nb13, + const int ne0, + const int ne1, + const int ne2, + const int ne3) { +#if FP16_AVAILABLE + //In this kernel Q, K, V are matrices while i, j, k are matrix indices. + + const int ic = blockIdx.x / parallel_blocks; // Index of the Q/QKV column to work on. + const int ip = blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel. + + const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix. + const float2 * Q_f2 = (const float2 *) (Q + nb02* blockIdx.y + nb01*ic); + const half2 * K_h2 = (const half2 *) (K + nb12*(blockIdx.y / gqa_ratio)); + const half * V_h = (const half *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape + const half * maskh = (const half *) mask + ne11*ic; + + const int stride_KV = nb11 / sizeof(half); + const int stride_KV2 = nb11 / sizeof(half2); + + constexpr int nwarps = (D + WARP_SIZE - 1) / WARP_SIZE; + const int tid = WARP_SIZE*threadIdx.y + threadIdx.x; + __builtin_assume(tid < nwarps*WARP_SIZE); + + __shared__ half KQ[nwarps*WARP_SIZE]; + KQ[tid] = -INFINITY; + half2 * KQ2 = (half2 *) KQ; + + half kqmax = -HALF_MAX_HALF; + half kqsum = 0.0f; + + __shared__ half kqmax_shared[WARP_SIZE]; + __shared__ half kqsum_shared[WARP_SIZE]; + if (threadIdx.y == 0) { + kqmax_shared[threadIdx.x] = -HALF_MAX_HALF; + kqsum_shared[threadIdx.x] = 0.0f; + } + __syncthreads(); + + // Convert Q to half2 and store in registers: + half2 Q_h2[(D/2 + WARP_SIZE - 1) / WARP_SIZE]; +#pragma unroll + for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) { + const int i = i0 + threadIdx.x; + if (i0 + WARP_SIZE > D/2 && i >= D/2) { + break; + } + + Q_h2[i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(Q_f2[i].x, Q_f2[i].y); + } + + half2 VKQ = make_half2(0.0f, 0.0f); // Each thread calculates a single VKQ value. + + const int k_start = parallel_blocks == 1 ? 0 : ip*D; + for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) { + // Calculate KQ tile and keep track of new maximum KQ values: + half kqmax_new = kqmax; +#pragma unroll + for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) { + const int i_KQ = i_KQ_0 + threadIdx.y; + + if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) { + break; + } + + half2 sum2 = make_half2(0.0f, 0.0f); +#pragma unroll + for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) { + const int k_KQ = k_KQ_0 + threadIdx.x; + if (k_KQ_0 + WARP_SIZE > D/2 && k_KQ >= D/2) { + break; + } + + const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ]; + sum2 += K_ik * Q_h2[k_KQ_0/WARP_SIZE]; + } + + sum2 = warp_reduce_sum(sum2); + half sum = __low2half(sum2) + __high2half(sum2); + sum += mask ? maskh[k_VKQ_0 + i_KQ] : __float2half(0.0f); + kqmax_new = ggml_cuda_hmax(kqmax_new, sum); + if (threadIdx.x == 0) { + KQ[i_KQ] = sum; + } + } + + kqmax_new = warp_reduce_max(kqmax_new); + if (threadIdx.x == 0) { + kqmax_shared[threadIdx.y] = kqmax_new; + } + __syncthreads(); + kqmax_new = kqmax_shared[threadIdx.x]; + kqmax_new = warp_reduce_max(kqmax_new); + + const half KQ_max_scale = hexp(kqmax - kqmax_new); + kqmax = kqmax_new; + + const half val = hexp(KQ[tid] - kqmax); + kqsum = kqsum*KQ_max_scale + val; + KQ[tid] = val; + + VKQ *= __half2half2(KQ_max_scale); + + __syncthreads(); + + if (tid < D) { +#pragma unroll + for (int k0 = 0; k0 < D; k0 += 2) { + if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) { + break; + } + + half2 V_k; + reinterpret_cast(V_k.x) = V_h[(k_VKQ_0 + k0 + 0)*stride_KV + tid]; + reinterpret_cast(V_k.y) = V_h[(k_VKQ_0 + k0 + 1)*stride_KV + tid]; + VKQ += V_k*KQ2[k0/2]; + } + } + + __syncthreads(); + } + + if (tid >= D) { + kqsum = 0.0f; + } + + kqsum = warp_reduce_sum(kqsum); + if (threadIdx.x == 0) { + kqsum_shared[threadIdx.y] = kqsum; + } + __syncthreads(); + kqsum = kqsum_shared[threadIdx.x]; + kqsum = warp_reduce_sum(kqsum); + + if (tid >= D) { + return; + } + + half dst_val = (__low2half(VKQ) + __high2half(VKQ)); + if (parallel_blocks == 1) { + dst_val /= kqsum; + } + dst[D*gridDim.y*blockIdx.x + D*blockIdx.y + tid] = dst_val; + + if (parallel_blocks == 1 || tid != 0) { + return; + } + dst_meta[ic*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax, kqsum); +#else + NO_DEVICE_CODE; +#endif // FP16_AVAILABLE +} + +// D == head size, VKQ_stride == num VKQ rows calculated in parallel: +template +__launch_bounds__(nwarps*WARP_SIZE, 1) +static __global__ void flash_attn_ext_f16( + const char * __restrict__ Q, + const char * __restrict__ K, + const char * __restrict__ V, + const char * __restrict__ mask, + float * __restrict__ dst, + float2 * __restrict__ dst_meta, + const float scale, + const int ne00, + const int ne01, + const int ne02, + const int ne03, + const int ne10, + const int ne11, + const int ne12, + const int ne13, + const int ne31, + const int nb31, + const int nb01, + const int nb02, + const int nb03, + const int nb11, + const int nb12, + const int nb13, + const int ne0, + const int ne1, + const int ne2, + const int ne3) { +#if FP16_MMA_AVAILABLE + //In this kernel Q, K, V are matrices while i, j, k are matrix indices. + + const int ic0 = ncols*(blockIdx.x / parallel_blocks); // Index of the first Q/QKV column to work on. + const int ip = blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel. + + static_assert(D <= FATTN_KQ_STRIDE, "D must be <= FATTN_KQ_STRIDE."); + static_assert(ncols == 8 || ncols % 16 == 0, "ncols must be 8 or a multiple of 16."); + constexpr int frag_m = ncols == 8 ? 32 : 16; + constexpr int frag_n = ncols == 8 ? 8 : 16; + static_assert(D % frag_m == 0, "If ncols == 8 then D % frag_m must be 0."); + typedef nvcuda::wmma::fragment frag_a_K; + typedef nvcuda::wmma::fragment frag_a_V; + typedef nvcuda::wmma::fragment frag_b; + typedef nvcuda::wmma::fragment frag_c_KQ; + typedef nvcuda::wmma::fragment frag_c_VKQ; + + constexpr int KQ_stride_tc = nwarps*frag_m; // Number of KQ rows calculated in parallel. + constexpr int VKQ_ratio = KQ_stride_tc/VKQ_stride; // Number of parallel VKQ accumulators needed to keep all warps busy. + static_assert(VKQ_ratio <= nwarps, "VKQ_ratio must be <= nwarps."); + + // Pad internal representation of KQ, KQV to reduce shared memory bank conflicts: + constexpr int D_padded = D + 8; + constexpr int kqs_padded = FATTN_KQ_STRIDE + 8; + constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half); + + const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix. + const float * Q_f = (const float *) (Q + nb02* blockIdx.y + nb01*ic0); + const half * K_h = (const half *) (K + nb12*(blockIdx.y / gqa_ratio)); + const half * V_h = (const half *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape + const half * maskh = (const half *) mask + (nb31/sizeof(half))* ic0; + const half2 * mask2 = (const half2 *) mask + (nb31/sizeof(half))*(ic0/2); + + const int stride_Q = nb01 / sizeof(float); + const int stride_KV = nb11 / sizeof(half); + + frag_b Q_b[D/16][ncols/frag_n]; + + // A single buffer for temporarily holding tiles of KQ and VKQ parts: + constexpr int mem_KQ = ncols*kqs_padded*kqar; + constexpr int mem_VKQ_parts = VKQ_ratio*ncols*D_padded; + __shared__ half KQ[mem_KQ >= mem_VKQ_parts ? mem_KQ : mem_VKQ_parts]; + float * KQ_f = (float *) KQ; + half2 * KQ2 = (half2 *) KQ; + + float KQ_rowsum_f[ncols/nwarps] = {0.0f}; + float KQ_max_f[ncols/nwarps]; + float KQ_max_scale_f[ncols/nwarps] = {0.0f}; + +#pragma unroll + for (int j = 0; j < ncols/nwarps; ++j) { + KQ_max_f[j] = -FLT_MAX/2.0f; + } + + half2 KQ_rowsum_h2[ncols/nwarps] = {{0.0f, 0.0f}}; + half2 KQ_max_h2[ncols/nwarps]; + half2 KQ_max_scale_h2[ncols/nwarps] = {{0.0f, 0.0f}}; + +#pragma unroll + for (int j = 0; j < ncols/nwarps; ++j) { + KQ_max_h2[j] = make_half2(-HALF_MAX_HALF, -HALF_MAX_HALF); + } + + __shared__ half VKQ[ncols*D_padded]; // Accumulator for final VKQ slice. + half2 * VKQ2 = (half2 *) VKQ; +#pragma unroll + for (int j0 = 0; j0 < ncols; j0 += nwarps) { + const int j = j0 + threadIdx.y; +#pragma unroll + for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) { + const int i = i0 + threadIdx.x; + if (i0 + WARP_SIZE > D/2 && i >= D/2) { + break; + } + VKQ2[j*(D_padded/2) + i] = make_half2(0.0f, 0.0f); + } + } + + // Convert Q to half and apply scale, temporarily store in KQ: +#pragma unroll + for (int j0 = 0; j0 < ncols; j0 += nwarps) { + const int j = j0 + threadIdx.y; +#pragma unroll + for (int i0 = 0; i0 < D; i0 += WARP_SIZE) { + const int i = i0 + threadIdx.x; + if (i0 + WARP_SIZE > D && i >= D) { + break; + } + KQ[j*D_padded + i] = ic0 + j < ne01 ? Q_f[j*stride_Q + i] * scale : 0.0f; + } + } + + __syncthreads(); + + // Load Q into tensor core fragments/registers since it will be used frequently: +#pragma unroll + for (int i0 = 0; i0 < D; i0 += 16) { +#pragma unroll + for (int j0 = 0; j0 < ncols; j0 += frag_n) { + nvcuda::wmma::load_matrix_sync(Q_b[i0/16][j0/frag_n], KQ + j0*D_padded + i0, D_padded); + } + } + + __syncthreads(); + + // Iterate over ne11 == previous tokens: + for (int k_VKQ_0 = ip*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE) { + // Calculate tile of KQ: +#pragma unroll + for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE; i_KQ_0 += KQ_stride_tc) { + frag_c_KQ KQ_c[ncols/frag_n]; +#pragma unroll + for (int j = 0; j < ncols/frag_n; ++j) { + nvcuda::wmma::fill_fragment(KQ_c[j], 0.0f); + } +#pragma unroll + for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 16) { + frag_a_K K_a; + nvcuda::wmma::load_matrix_sync(K_a, K_h + (k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV); +#pragma unroll + for (int j = 0; j < ncols/frag_n; ++j) { + nvcuda::wmma::mma_sync(KQ_c[j], K_a, Q_b[k_KQ_0/16][j], KQ_c[j]); + } + } +#pragma unroll + for (int j0 = 0; j0 < ncols; j0 += frag_n) { + nvcuda::wmma::store_matrix_sync((KQ_acc_t *) KQ + j0*kqs_padded + i_KQ_0 + frag_m*threadIdx.y, KQ_c[j0/frag_n], kqs_padded, nvcuda::wmma::mem_col_major); + } + } + + __syncthreads(); + + // Calculate softmax for each KQ column using the current max. value. + // The divisor is stored in KQ_rowsum and will be applied at the end. +#pragma unroll + for (int j0 = 0; j0 < ncols; j0 += nwarps) { + const int j = j0 + threadIdx.y; + + if (std::is_same::value) { + float KQ_f_tmp[FATTN_KQ_STRIDE / WARP_SIZE]; +#pragma unroll + for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) { + const int k = k0 + threadIdx.x; + + KQ_f_tmp[k0/WARP_SIZE] = KQ_f[j*kqs_padded + k]; + } + + float KQ_max_new = KQ_max_f[j0/nwarps]; +#pragma unroll + for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) { + const int k = k0 + threadIdx.x; + + KQ_f_tmp[k0/WARP_SIZE] += mask ? __half2float(maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f; + KQ_max_new = max(KQ_max_new, KQ_f_tmp[k0/WARP_SIZE]); + } + KQ_max_new = warp_reduce_max(KQ_max_new); + + const float diff = KQ_max_f[j0/nwarps] - KQ_max_new; + KQ_max_scale_f[j0/nwarps] = expf(diff); + if (diff <= SOFTMAX_FTZ_THRESHOLD) { + KQ_max_scale_f[j0/nwarps] = 0.0f; + } + KQ_max_f[j0/nwarps] = KQ_max_new; + + float KQ_rowsum_add = 0.0f; +#pragma unroll + for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) { + const int k = k0 + threadIdx.x; + + const float diff = KQ_f_tmp[k0/WARP_SIZE] - KQ_max_f[j0/nwarps]; + KQ_f_tmp[k0/WARP_SIZE] = expf(diff); + if (diff <= SOFTMAX_FTZ_THRESHOLD) { + KQ_f_tmp[k0/WARP_SIZE] = 0.0f; + } + KQ_rowsum_add += KQ_f_tmp[k0/WARP_SIZE]; + KQ[j*(kqar*kqs_padded) + k] = KQ_f_tmp[k0/WARP_SIZE]; + } + KQ_rowsum_add = warp_reduce_sum(KQ_rowsum_add); + + // Scale previous KQ_rowsum to account for a potential increase in KQ_max: + KQ_rowsum_f[j0/nwarps] = KQ_max_scale_f[j0/nwarps]*KQ_rowsum_f[j0/nwarps] + KQ_rowsum_add; + } else { + half2 KQ2_tmp[FATTN_KQ_STRIDE/(2*WARP_SIZE)]; +#pragma unroll + for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) { + const int k = k0 + threadIdx.x; + + KQ2_tmp[k0/WARP_SIZE] = KQ2[j*(kqs_padded/2) + k]; + } + + half2 KQ_max_new = KQ_max_h2[j0/nwarps]; +#pragma unroll + for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) { + const int k = k0 + threadIdx.x; + + KQ2_tmp[k0/WARP_SIZE] += mask ? mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f); + KQ_max_new = ggml_cuda_hmax2(KQ_max_new, KQ2_tmp[k0/WARP_SIZE]); + } + KQ_max_new = __half2half2(warp_reduce_max(ggml_cuda_hmax(__low2half(KQ_max_new), __high2half(KQ_max_new)))); + const half2 diff = KQ_max_h2[j0/nwarps] - KQ_max_new; + KQ_max_scale_h2[j0/nwarps] = h2exp(diff); + const uint32_t ftz_mask = __hgt2_mask(diff, make_half2(SOFTMAX_FTZ_THRESHOLD, SOFTMAX_FTZ_THRESHOLD)); + *((uint32_t *) &KQ_max_scale_h2[j0/nwarps]) &= ftz_mask; + KQ_max_h2[j0/nwarps] = KQ_max_new; + + half2 KQ_rowsum_add = make_half2(0.0f, 0.0f); +#pragma unroll + for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) { + const int k = k0 + threadIdx.x; + + const half2 diff = KQ2_tmp[k0/WARP_SIZE] - KQ_max_h2[j0/nwarps]; + KQ2_tmp[k0/WARP_SIZE] = h2exp(diff); + const uint32_t ftz_mask = __hgt2_mask(diff, make_half2(SOFTMAX_FTZ_THRESHOLD, SOFTMAX_FTZ_THRESHOLD)); + *((uint32_t *) &KQ2_tmp[k0/WARP_SIZE]) &= ftz_mask; + KQ_rowsum_add += KQ2_tmp[k0/WARP_SIZE]; + KQ2[j*(kqs_padded/2) + k] = KQ2_tmp[k0/WARP_SIZE]; + } + KQ_rowsum_add = warp_reduce_sum(KQ_rowsum_add); + + // Scale previous KQ_rowsum to account for a potential increase in KQ_max: + KQ_rowsum_h2[j0/nwarps] = KQ_max_scale_h2[j0/nwarps]*KQ_rowsum_h2[j0/nwarps] + KQ_rowsum_add; + } + } + + __syncthreads(); + + frag_b KQ_b[FATTN_KQ_STRIDE/(VKQ_ratio*16)][ncols/frag_n]; +#pragma unroll + for (int j0 = 0; j0 < ncols; j0 += frag_n) { +#pragma unroll + for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += VKQ_ratio*16) { + const int k = k0 + (threadIdx.y % VKQ_ratio)*16; + nvcuda::wmma::load_matrix_sync( + KQ_b[k0/(VKQ_ratio*16)][j0/frag_n], + KQ + j0*(kqar*kqs_padded) + k, + kqar*kqs_padded); + } + } + + frag_c_VKQ VKQ_c[D/VKQ_stride][ncols/frag_n]; +#pragma unroll + for (int i_VKQ_0 = 0; i_VKQ_0 < D; i_VKQ_0 += VKQ_stride) { +#pragma unroll + for (int j = 0; j < ncols/frag_n; ++j) { + nvcuda::wmma::fill_fragment(VKQ_c[i_VKQ_0/VKQ_stride][j], 0.0f); + } + +#pragma unroll + for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += VKQ_ratio*16) { + const int k = k0 + (threadIdx.y % VKQ_ratio)*16; + + frag_a_V v_a; + nvcuda::wmma::load_matrix_sync(v_a, V_h + (k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV); +#pragma unroll + for (int j = 0; j < ncols/frag_n; ++j) { + nvcuda::wmma::mma_sync(VKQ_c[i_VKQ_0/VKQ_stride][j], v_a, KQ_b[k0/(VKQ_ratio*16)][j], VKQ_c[i_VKQ_0/VKQ_stride][j]); + } + } + } + + __syncthreads(); + + const int offset_k = (threadIdx.y % VKQ_ratio) * (ncols*D_padded); +#pragma unroll + for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += VKQ_stride) { +#pragma unroll + for (int j0 = 0; j0 < ncols; j0 += frag_n) { + nvcuda::wmma::store_matrix_sync( + KQ + offset_k + j0*D_padded + i_KQ_0 + frag_m*(threadIdx.y/VKQ_ratio), + VKQ_c[i_KQ_0/VKQ_stride][j0/frag_n], + D_padded, nvcuda::wmma::mem_col_major); + } + } + + __syncthreads(); + +#pragma unroll + for (int j0 = 0; j0 < ncols; j0 += nwarps) { + const int j = j0 + threadIdx.y; + + half2 VKQ_scale; + if (std::is_same::value) { + VKQ_scale = make_half2(KQ_max_scale_f[j0/nwarps], KQ_max_scale_f[j0/nwarps]); + } else { + VKQ_scale = KQ_max_scale_h2[j0/nwarps]; + } + +#pragma unroll + for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) { + const int i = i0 + threadIdx.x; + if (i0 + WARP_SIZE > D/2 && i >= D/2) { + break; + } + + half2 VKQ_add = make_half2(0.0f, 0.0f); +#pragma unroll + for (int l = 0; l < VKQ_ratio; ++l) { + VKQ_add += KQ2[l*(ncols*D_padded/2) + j*(D_padded/2) + i]; + } + VKQ2[j*(D_padded/2) + i] = VKQ_scale*VKQ2[j*(D_padded/2) + i] + VKQ_add; + } + } + + __syncthreads(); + } + +#pragma unroll + for (int j0 = 0; j0 < ncols; j0 += nwarps) { + const int j_VKQ = j0 + threadIdx.y; + if (ic0 + j_VKQ >= ne01) { + return; + } + const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip; + + float KQ_rowsum_j; + if (std::is_same::value) { + KQ_rowsum_j = KQ_rowsum_f[j0/nwarps]; + } else { + KQ_rowsum_j = __low2float(KQ_rowsum_h2[j0/nwarps]) + __high2float(KQ_rowsum_h2[j0/nwarps]); + } + +#pragma unroll + for (int i0 = 0; i0 < D; i0 += WARP_SIZE) { + const int i = i0 + threadIdx.x; + if (i0 + WARP_SIZE > D && i >= D) { + break; + } + float dst_val = VKQ[j_VKQ*D_padded + i]; + if (parallel_blocks == 1) { + dst_val /= KQ_rowsum_j; + } + dst[j_dst*gridDim.y*D + blockIdx.y*D + i] = dst_val; + } + + if (parallel_blocks == 1 || threadIdx.x != 0) { + continue; + } + + float2 dst_meta_val; + if (std::is_same::value) { + dst_meta_val.x = KQ_max_f[j0/nwarps]; + } else { + dst_meta_val.x = __low2float(KQ_max_h2[j0/nwarps]); + } + dst_meta_val.y = KQ_rowsum_j; + dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = dst_meta_val; + } +#else + NO_DEVICE_CODE; +#endif // FP16_MMA_AVAILABLE +} + +template // D == head size +__launch_bounds__(D, 1) +static __global__ void flash_attn_combine_results( + const float * __restrict__ VKQ_parts, + const float2 * __restrict__ VKQ_meta, + float * __restrict__ dst) { +#if FP16_AVAILABLE + VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x; + VKQ_meta += parallel_blocks * gridDim.y*blockIdx.x; + dst += D * gridDim.y*blockIdx.x; + + const int tid = threadIdx.x; + __builtin_assume(tid < D); + + __shared__ float2 meta[parallel_blocks]; + if (tid < 2*parallel_blocks) { + ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid]; + } + + __syncthreads(); + + float kqmax = meta[0].x; +#pragma unroll + for (int l = 1; l < parallel_blocks; ++l) { + kqmax = max(kqmax, meta[l].x); + } + + float VKQ_numerator = 0.0f; + float VKQ_denominator = 0.0f; +#pragma unroll + for (int l = 0; l < parallel_blocks; ++l) { + const float diff = meta[l].x - kqmax; + const float KQ_max_scale = expf(diff); + const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD); + *((uint32_t *) &KQ_max_scale) &= ftz_mask; + + VKQ_numerator += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid]; + VKQ_denominator += KQ_max_scale * meta[l].y; + } + + dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator; +#else + NO_DEVICE_CODE; +#endif // FP16_AVAILABLE +} + +constexpr int get_max_power_of_2(int x) { + return x % 2 == 0 ? 2*get_max_power_of_2(x/2) : 1; +} + +static_assert(get_max_power_of_2(1) == 1, "Test failed."); +static_assert(get_max_power_of_2(2) == 2, "Test failed."); +static_assert(get_max_power_of_2(4) == 4, "Test failed."); +static_assert(get_max_power_of_2(6) == 2, "Test failed."); + +// Number of VKQ rows calculated in parallel: +constexpr int get_VKQ_stride(int D, int nwarps, int frag_m) { + return (get_max_power_of_2(D/frag_m) < nwarps ? get_max_power_of_2(D/frag_m) : nwarps)*frag_m; +} + +static_assert(get_VKQ_stride(128, 1, 32) == 32, "Test failed."); +static_assert(get_VKQ_stride(128, 2, 32) == 64, "Test failed."); +static_assert(get_VKQ_stride(128, 4, 32) == 128, "Test failed."); +static_assert(get_VKQ_stride( 64, 1, 32) == 32, "Test failed."); +static_assert(get_VKQ_stride( 64, 2, 32) == 64, "Test failed."); +static_assert(get_VKQ_stride( 64, 4, 32) == 64, "Test failed."); +static_assert(get_VKQ_stride( 80, 1, 16) == 16, "Test failed."); +static_assert(get_VKQ_stride( 80, 2, 16) == 16, "Test failed."); +static_assert(get_VKQ_stride( 80, 4, 16) == 16, "Test failed."); + +template void launch_fattn_vec_f16( + const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask, + ggml_cuda_pool & pool, cudaStream_t main_stream +) { + ggml_cuda_pool_alloc dst_tmp(pool); + ggml_cuda_pool_alloc dst_tmp_meta(pool); + + if (parallel_blocks > 1) { + dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV)); + dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV)); + } + + constexpr int nwarps = (D + WARP_SIZE - 1) / WARP_SIZE; + const dim3 block_dim(WARP_SIZE, nwarps, 1); + const dim3 blocks_num(parallel_blocks*Q->ne[1], Q->ne[2], Q->ne[3]); + const int shmem = 0; + + float scale; + memcpy(&scale, KQV->op_params, sizeof(float)); + + flash_attn_vec_ext_f16 + <<>> ( + (const char *) Q->data, + (const char *) K->data, + (const char *) V->data, + mask ? ((const char *) mask->data) : nullptr, + parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr, + scale, + Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3], + K->ne[0], K->ne[1], K->ne[2], K->ne[3], + mask ? mask->ne[1] : 0, mask ? mask->nb[1] : 0, + Q->nb[1], Q->nb[2], Q->nb[3], + K->nb[1], K->nb[2], K->nb[3], + KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3] + ); + CUDA_CHECK(cudaGetLastError()); + + if (parallel_blocks == 1) { + return; + } + + const dim3 block_dim_combine(D, 1, 1); + const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z); + const int shmem_combine = 0; + + flash_attn_combine_results + <<>> + (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data); + CUDA_CHECK(cudaGetLastError()); +} + +template void launch_fattn_f16_impl( + const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask, + ggml_cuda_pool & pool, cudaStream_t main_stream +) { + ggml_cuda_pool_alloc dst_tmp(pool); + ggml_cuda_pool_alloc dst_tmp_meta(pool); + + if (parallel_blocks > 1) { + dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV)); + dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV)); + } + + constexpr int frag_m = (cols_per_block) == 8 && (D) % 32 == 0 ? 32 : 16; + const dim3 block_dim(WARP_SIZE, nwarps, 1); + const dim3 blocks_num(parallel_blocks*(Q->ne[1] + cols_per_block - 1) / cols_per_block, Q->ne[2], Q->ne[3]); + const int shmem = 0; + + float scale; + memcpy(&scale, KQV->op_params, sizeof(float)); + + flash_attn_ext_f16 + <<>> ( + (const char *) Q->data, + (const char *) K->data, + (const char *) V->data, + mask ? ((const char *) mask->data) : nullptr, + (parallel_blocks) == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr, + scale, + Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3], + K->ne[0], K->ne[1], K->ne[2], K->ne[3], + mask ? mask->ne[1] : 0, mask ? mask->nb[1] : 0, + Q->nb[1], Q->nb[2], Q->nb[3], + K->nb[1], K->nb[2], K->nb[3], + KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3] + ); + CUDA_CHECK(cudaGetLastError()); + + if ((parallel_blocks) == 1) { + return; + } + + const dim3 block_dim_combine(D, 1, 1); + const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z); + const int shmem_combine = 0; + + flash_attn_combine_results + <<>> + (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data); + CUDA_CHECK(cudaGetLastError()); +} + +template void launch_fattn_f16( + const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask, + const int nsm, ggml_cuda_pool & pool, cudaStream_t main_stream +) { + const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3]; + + if (4*blocks_num_pb1 < 2*nsm) { + launch_fattn_f16_impl(Q, K, V, KQV, mask, pool, main_stream); + return; + } + if (2*blocks_num_pb1 < 2*nsm) { + launch_fattn_f16_impl(Q, K, V, KQV, mask, pool, main_stream); + return; + } + launch_fattn_f16_impl(Q, K, V, KQV, mask, pool, main_stream); +} + +void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * Q = dst->src[0]; + const ggml_tensor * K = dst->src[1]; + const ggml_tensor * V = dst->src[2]; + + const ggml_tensor * mask = dst->src[3]; + + ggml_tensor * KQV = dst; + + GGML_ASSERT(Q->type == GGML_TYPE_F32); + GGML_ASSERT(K->type == GGML_TYPE_F16); + GGML_ASSERT(V->type == GGML_TYPE_F16); + GGML_ASSERT(KQV->type == GGML_TYPE_F32); + + GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16); + GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) && + "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big"); + + GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding."); + + ggml_cuda_set_device(ctx.device); + + const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm; + + const int32_t precision = KQV->op_params[1]; + + if (precision != GGML_PREC_DEFAULT) { + if (Q->ne[1] <= 32 || Q->ne[0] > 128) { + constexpr int cols_per_block = 16; + constexpr int nwarps = 4; + switch (Q->ne[0]) { + case 64: + launch_fattn_f16< 64, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 80: + launch_fattn_f16< 80, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 96: + launch_fattn_f16< 96, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 112: + launch_fattn_f16<112, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 128: + launch_fattn_f16<128, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 256: + launch_fattn_f16<256, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + default: + GGML_ASSERT(false); + break; + } + } else { + constexpr int cols_per_block = 32; + constexpr int nwarps = 4; + switch (Q->ne[0]) { + case 64: + launch_fattn_f16< 64, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 80: + launch_fattn_f16< 80, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 96: + launch_fattn_f16< 96, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 112: + launch_fattn_f16<112, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 128: + launch_fattn_f16<128, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + // case 256: + // launch_fattn_f16<256, cols_per_block, nwarps, float>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + // break; + default: + GGML_ASSERT(false); + break; + } + } + return; + } + + if (Q->ne[1] == 1 && Q->ne[0] % (2*WARP_SIZE) == 0) { + constexpr int parallel_blocks = 4; + switch (Q->ne[0]) { + case 64: + launch_fattn_vec_f16< 64, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream()); + break; + case 128: + launch_fattn_vec_f16<128, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream()); + break; + case 256: + launch_fattn_vec_f16<256, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream()); + break; + default: + GGML_ASSERT(false); + break; + } + return; + } + + if (Q->ne[1] <= 8 && Q->ne[0] % WARP_SIZE == 0) { + constexpr int cols_per_block = 8; + constexpr int nwarps = 4; + switch (Q->ne[0]) { + case 64: + launch_fattn_f16< 64, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 96: + launch_fattn_f16< 96, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 128: + launch_fattn_f16<128, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 256: + launch_fattn_f16<256, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + default: + GGML_ASSERT(false); + break; + } + return; + } + + if (Q->ne[1] <= 32) { + constexpr int cols_per_block = 16; + constexpr int nwarps = 4; + switch (Q->ne[0]) { + case 64: + launch_fattn_f16< 64, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 80: + launch_fattn_f16< 80, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 96: + launch_fattn_f16< 96, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 112: + launch_fattn_f16<112, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 128: + launch_fattn_f16<128, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 256: + launch_fattn_f16<256, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + default: + GGML_ASSERT(false); + break; + } + return; + } + + constexpr int cols_per_block = 32; + constexpr int nwarps = 4; + switch (Q->ne[0]) { + case 64: + launch_fattn_f16< 64, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 80: + launch_fattn_f16< 80, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 96: + launch_fattn_f16< 96, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 112: + launch_fattn_f16<112, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 128: + launch_fattn_f16<128, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + case 256: + launch_fattn_f16<256, cols_per_block, nwarps, half>(Q, K, V, KQV, mask, nsm, ctx.pool(), ctx.stream()); + break; + default: + GGML_ASSERT(false); + break; + } + return; +} diff --git a/ggml-cuda/fattn.cuh b/ggml-cuda/fattn.cuh new file mode 100644 index 000000000..ad3ca7a8d --- /dev/null +++ b/ggml-cuda/fattn.cuh @@ -0,0 +1,3 @@ +#include "common.cuh" + +void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml-cuda/softmax.cu b/ggml-cuda/softmax.cu index 9bda18e58..6ed225999 100644 --- a/ggml-cuda/softmax.cu +++ b/ggml-cuda/softmax.cu @@ -1,7 +1,17 @@ #include "softmax.cuh" -template -static __global__ void soft_max_f32(const float * x, const float * mask, const float * pos, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) { +template +static __device__ __forceinline__ float t2f32(T val) { + return (float) val; +} + +template <> +__device__ float __forceinline__ t2f32(half val) { + return __half2float(val); +} + +template +static __global__ void soft_max_f32(const float * x, const T * mask, const T * pos, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) { const int ncols = ncols_template == 0 ? ncols_par : ncols_template; const int tid = threadIdx.x; @@ -28,7 +38,7 @@ static __global__ void soft_max_f32(const float * x, const float * mask, const f extern __shared__ float data_soft_max_f32[]; float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication // shared memory buffer to cache values between iterations: - float * vals = vals_smem ? buf_iw + WARP_SIZE : dst + rowx*ncols; + float * vals = vals_smem ? buf_iw + WARP_SIZE : dst + (int64_t)rowx*ncols; float max_val = -INFINITY; @@ -40,10 +50,10 @@ static __global__ void soft_max_f32(const float * x, const float * mask, const f break; } - const int ix = rowx*ncols + col; - const int iy = rowy*ncols + col; + const int64_t ix = (int64_t)rowx*ncols + col; + const int64_t iy = (int64_t)rowy*ncols + col; - const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f) + (pos ? slope*pos[col] : 0.0f); + const float val = x[ix]*scale + (mask ? t2f32(mask[iy]) : 0.0f) + (pos ? slope*t2f32(pos[col]) : 0.0f); vals[col] = val; max_val = max(max_val, val); @@ -109,12 +119,13 @@ static __global__ void soft_max_f32(const float * x, const float * mask, const f return; } - const int idst = rowx*ncols + col; + const int64_t idst = (int64_t)rowx*ncols + col; dst[idst] = vals[col] * inv_sum; } } -static void soft_max_f32_cuda(const float * x, const float * mask, const float * pos, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) { +template +static void soft_max_f32_cuda(const float * x, const T * mask, const T * pos, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) { int nth = WARP_SIZE; while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2; const dim3 block_dims(nth, 1, 1); @@ -167,15 +178,19 @@ static void soft_max_f32_cuda(const float * x, const float * mask, const float * void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * src0 = dst->src[0]; const ggml_tensor * src1 = dst->src[1]; + const ggml_tensor * src2 = dst->src[2]; + const float * src0_d = (const float *)src0->data; - const float * src1_d = src1 ? (const float *)src1->data : nullptr; + const void * src1_d = src1 ? (const void *)src1->data : nullptr; + float * dst_d = (float *)dst->data; cudaStream_t stream = ctx.stream(); GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); - GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional + GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional + GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F16 || src2->type == GGML_TYPE_F32); // src2 contains positions and it is optional const int64_t ne00 = src0->ne[0]; const int64_t nrows_x = ggml_nrows(src0); @@ -188,14 +203,25 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float)); // positions tensor - float * src2_dd = nullptr; + void * src2_d = nullptr; - ggml_tensor * src2 = dst->src[2]; const bool use_src2 = src2 != nullptr; if (use_src2) { - src2_dd = (float *)src2->data; + src2_d = (void *)src2->data; } - soft_max_f32_cuda(src0_d, src1_d, src2_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream); + const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16) || (src2 && src2->type == GGML_TYPE_F16); + + if (use_f16) { + const half * src1_dd = (const half *)src1_d; + const half * src2_dd = (const half *)src2_d; + + soft_max_f32_cuda(src0_d, src1_dd, src2_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream); + } else { + const float * src1_dd = (const float *)src1_d; + const float * src2_dd = (const float *)src2_d; + + soft_max_f32_cuda(src0_d, src1_dd, src2_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream); + } } diff --git a/ggml-impl.h b/ggml-impl.h index 2087f7ded..94a1cc668 100644 --- a/ggml-impl.h +++ b/ggml-impl.h @@ -313,7 +313,7 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) #endif // defined(__ARM_NEON) -#if defined(__ARM_NEON) && !defined(__MSC_VER) +#if defined(__ARM_NEON) && !defined(_MSC_VER) #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) diff --git a/ggml-kompute.cpp b/ggml-kompute.cpp index 407062e6f..9a469821d 100644 --- a/ggml-kompute.cpp +++ b/ggml-kompute.cpp @@ -1427,6 +1427,7 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml for (int i = node_start; i < node_end; ++i) { struct ggml_tensor * src0 = gf->nodes[i]->src[0]; struct ggml_tensor * src1 = gf->nodes[i]->src[1]; + struct ggml_tensor * src2 = gf->nodes[i]->src[2]; GGML_UNUSED(src2); struct ggml_tensor * dst = gf->nodes[i]; GGML_ASSERT(dst->data != nullptr); @@ -1559,6 +1560,12 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml { float scale; memcpy(&scale, dst->op_params, sizeof(float)); + +#pragma message("TODO: add ggml_vk_soft_max() F16/F32 src1 and src2 support") +#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021") + GGML_ASSERT(!src1 || src1t == GGML_TYPE_F32); + GGML_ASSERT(src2 == nullptr); + ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale); } break; case GGML_OP_DIAG_MASK_INF: diff --git a/ggml-metal.m b/ggml-metal.m index 9cb421988..017b72ce9 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -46,8 +46,10 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, GGML_METAL_KERNEL_TYPE_SILU, GGML_METAL_KERNEL_TYPE_SILU_4, - GGML_METAL_KERNEL_TYPE_SOFT_MAX, - GGML_METAL_KERNEL_TYPE_SOFT_MAX_4, + GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, + GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4, + GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32, + GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4, GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, @@ -177,6 +179,14 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, GGML_METAL_KERNEL_TYPE_CPY_F32_F16, GGML_METAL_KERNEL_TYPE_CPY_F32_F32, GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, @@ -443,7 +453,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { } /* - GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \ + GGML_METAL_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \ (int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \ (int) kernel->pipeline.threadExecutionWidth); \ */ @@ -459,172 +469,182 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { return NULL; \ } \ } else { \ - GGML_METAL_LOG_WARN("%s: skipping %-32s (not supported)\n", __func__, "kernel_"#name); \ + GGML_METAL_LOG_WARN("%s: skipping %-40s (not supported)\n", __func__, "kernel_"#name); \ } // simd_sum and simd_max requires MTLGPUFamilyApple7 - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD, add, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW, add_row, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL, mul, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW, mul_row, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW, div_row, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE, scale, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4, scale_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP, clamp, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH, tanh, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU, relu, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU, gelu, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4, gelu_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK, gelu_quick, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4, silu_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX, soft_max, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_4, soft_max_4, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, diag_mask_inf, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, diag_mask_inf_8, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16, get_rows_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0, get_rows_q4_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1, get_rows_q4_1, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0, get_rows_q5_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1, get_rows_q5_1, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0, get_rows_q8_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K, get_rows_q2_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K, get_rows_q3_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K, get_rows_q4_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K, get_rows_q5_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K, get_rows_q6_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS, get_rows_iq2_xxs, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS, get_rows_iq2_xs, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS, get_rows_iq3_xxs, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S, get_rows_iq3_s, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S, get_rows_iq2_s, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S, get_rows_iq1_s, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M, get_rows_iq1_m, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL, get_rows_iq4_nl, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, get_rows_iq4_xs, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32, mul_mv_iq1_m_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, ctx->support_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, ctx->support_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, ctx->support_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32, mul_mv_id_iq1_m_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, ctx->support_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32, mul_mm_q5_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32, mul_mm_q6_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, mul_mm_iq1_m_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, ctx->support_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32, rope_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16, rope_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32, alibi_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16, im2col_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32, im2col_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, timestep_embedding_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32, arange_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, argsort_f32_i32_desc, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, leaky_relu_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, cpy_f32_q4_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, cpy_f32_q4_1, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, cpy_f32_q5_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, cpy_f32_q5_1, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, cpy_f32_iq4_nl, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32, cpy_f16_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT, concat, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR, sqr, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD, add, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW, add_row, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL, mul, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW, mul_row, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW, div_row, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE, scale, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4, scale_4, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP, clamp, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH, tanh, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU, relu, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU, gelu, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4, gelu_4, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK, gelu_quick, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4, silu_4, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, soft_max_f16, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4, soft_max_f16_4, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32, soft_max_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4, soft_max_f32_4, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, diag_mask_inf, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, diag_mask_inf_8, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16, get_rows_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0, get_rows_q4_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1, get_rows_q4_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0, get_rows_q5_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1, get_rows_q5_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0, get_rows_q8_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K, get_rows_q2_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K, get_rows_q3_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K, get_rows_q4_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K, get_rows_q5_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K, get_rows_q6_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS, get_rows_iq2_xxs, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS, get_rows_iq2_xs, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS, get_rows_iq3_xxs, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S, get_rows_iq3_s, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S, get_rows_iq2_s, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S, get_rows_iq1_s, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M, get_rows_iq1_m, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL, get_rows_iq4_nl, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, get_rows_iq4_xs, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32, mul_mv_iq1_m_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, ctx->support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, ctx->support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, ctx->support_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32, mul_mv_id_iq1_m_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, ctx->support_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32, mul_mm_q5_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32, mul_mm_q6_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, mul_mm_iq1_m_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, ctx->support_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32, rope_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16, rope_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32, alibi_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16, im2col_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32, im2col_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, timestep_embedding_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32, arange_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, argsort_f32_i32_desc, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, leaky_relu_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, flash_attn_ext_f16_h64, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, flash_attn_ext_f16_h80, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, cpy_f32_q4_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, cpy_f32_q4_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, cpy_f32_q5_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, cpy_f32_q5_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, cpy_f32_iq4_nl, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32, cpy_f16_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT, concat, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR, sqr, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true); } [metal_library release]; @@ -743,6 +763,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_ARGSORT: case GGML_OP_LEAKY_RELU: + case GGML_OP_FLASH_ATTN_EXT: return true; case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: @@ -1326,20 +1347,33 @@ static enum ggml_status ggml_metal_graph_compute( } break; case GGML_OP_SOFT_MAX: { + GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); + GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F16 || src2->type == GGML_TYPE_F32); + int nth = 32; // SIMD width id pipeline = nil; + const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16) || (src2 && src2->type == GGML_TYPE_F16); + if (ne00%4 == 0) { while (nth < ne00/4 && nth < 256) { nth *= 2; } - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_4].pipeline; + if (use_f16) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4].pipeline; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4].pipeline; + } } else { while (nth < ne00 && nth < 1024) { nth *= 2; } - pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX].pipeline; + if (use_f16) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16].pipeline; + } else { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32].pipeline; + } } float scale; @@ -2503,6 +2537,161 @@ static enum ggml_status ggml_metal_graph_compute( [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; + case GGML_OP_FLASH_ATTN_EXT: + { + GGML_ASSERT(ne00 % 4 == 0); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + + struct ggml_tensor * src3 = gf->nodes[i]->src[3]; + + GGML_ASSERT(ggml_are_same_shape(src1, src2)); + GGML_ASSERT(src3); + + size_t offs_src3 = 0; + + id id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil; + + GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16); + GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) && + "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big"); + + const int64_t ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30); + const int64_t ne31 = src3 ? src3->ne[1] : 0; + const int64_t ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32); + const int64_t ne33 = src3 ? src3->ne[3] : 0; GGML_UNUSED(ne33); + + const uint64_t nb30 = src3 ? src3->nb[0] : 0; GGML_UNUSED(nb30); + const uint64_t nb31 = src3 ? src3->nb[1] : 0; + const uint64_t nb32 = src3 ? src3->nb[2] : 0; GGML_UNUSED(nb32); + const uint64_t nb33 = src3 ? src3->nb[3] : 0; GGML_UNUSED(nb33); + + const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t); + + float scale; + memcpy(&scale, dst->op_params, sizeof(float)); + + id pipeline = nil; + + bool use_vec_kernel = false; + + if (ne01 >= 4 || (ne00%128 != 0)) { + switch (ne00) { + case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break; + case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break; + case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break; + case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break; + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break; + case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break; + default: + { + GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_METAL_LOG_ERROR("add template specialization for this size\n"); + GGML_ASSERT(false && "add template specialization for this size"); + } + } + } else { + use_vec_kernel = true; + + switch (ne00) { + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break; + case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break; + default: + { + GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_METAL_LOG_ERROR("add template specialization for this size\n"); + GGML_ASSERT(false && "add template specialization for this size"); + } + } + } + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2]; + [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:4]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:5]; + [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:6]; + [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:7]; + [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:8]; + [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:9]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:10]; + [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:11]; + [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:12]; + [encoder setBytes:&ne10 length:sizeof( int64_t) atIndex:13]; + [encoder setBytes:&ne11 length:sizeof( int64_t) atIndex:14]; + [encoder setBytes:&ne12 length:sizeof( int64_t) atIndex:15]; + [encoder setBytes:&ne13 length:sizeof( int64_t) atIndex:16]; + [encoder setBytes:&nb10 length:sizeof(uint64_t) atIndex:17]; + [encoder setBytes:&nb11 length:sizeof(uint64_t) atIndex:18]; + [encoder setBytes:&nb12 length:sizeof(uint64_t) atIndex:19]; + [encoder setBytes:&nb13 length:sizeof(uint64_t) atIndex:20]; + [encoder setBytes:&ne31 length:sizeof( int64_t) atIndex:21]; + [encoder setBytes:&nb31 length:sizeof(uint64_t) atIndex:22]; + [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:23]; + [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:24]; + [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:25]; + [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:26]; + [encoder setBytes:&scale length:sizeof( float) atIndex:27]; + + if (!use_vec_kernel) { + // half8x8 kernel + const int64_t nqptg = 8; // queries per threadgroup !! sync with kernel template arguments !! + const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !! + + GGML_ASSERT(nqptg <= 32); + GGML_ASSERT(nqptg % 8 == 0); + GGML_ASSERT(ncpsg % 32 == 0); + + int64_t nsgmax = 2; + + while (true) { + const size_t smem = nqptg*(ne00 + 2*nsgmax*(ncpsg + nqptg))*(sizeof(float)/2); + if (smem > ctx->device.maxThreadgroupMemoryLength) { + break; + } + nsgmax *= 2; + } + nsgmax /= 2; + + // simdgroups per threadgroup (a.k.a. warps) + const int64_t nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4; + + const size_t smem = nqptg*(ne00 + 2*nsg*(ncpsg + nqptg))*(sizeof(float)/2); + + //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength); + GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength); + + [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0]; + + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)]; + } else { + // half1x4 kernel + const int64_t nqptg = 1; // queries per threadgroup !! sync with kernel template arguments !! + const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !! + + GGML_ASSERT(nqptg <= 32); + GGML_ASSERT(nqptg % 1 == 0); + GGML_ASSERT(ncpsg % 32 == 0); + + // simdgroups per threadgroup (a.k.a. warps) + const int64_t nsgt = MAX(2, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)); + + int64_t nsg = 1; + while (nsg <= nsgt) { + nsg *= 2; + } + nsg /= 2; + + const size_t smem = (nqptg*(ne00 + 2*nsg*(ncpsg + nqptg)) + nsg*ne00)*(sizeof(float)/2); + + //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength); + GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength); + [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0]; + + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)]; + } + } break; case GGML_OP_DUP: case GGML_OP_CPY: case GGML_OP_CONT: @@ -2590,6 +2779,11 @@ static enum ggml_status ggml_metal_graph_compute( MTLCommandBufferStatus status = [command_buffer status]; if (status != MTLCommandBufferStatusCompleted) { GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status); + if (status == MTLCommandBufferStatusError) { + NSString * error_code = [command_buffer error].localizedDescription; + GGML_METAL_LOG_INFO("error: %s\n", [error_code UTF8String]); + } + return GGML_STATUS_FAILED; } } @@ -2706,10 +2900,13 @@ GGML_CALL static const char * ggml_backend_metal_buffer_type_get_name(ggml_backe UNUSED(buft); } -static void ggml_backend_metal_log_allocated_size(id device) { +static void ggml_backend_metal_log_allocated_size(id device, size_t size_aligned) { +#ifndef GGML_METAL_NDEBUG #if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15) if (@available(macOS 10.12, iOS 16.0, *)) { - GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)", + GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)", + __func__, + size_aligned / 1024.0 / 1024.0, device.currentAllocatedSize / 1024.0 / 1024.0, device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); @@ -2719,10 +2916,15 @@ static void ggml_backend_metal_log_allocated_size(id device) { GGML_METAL_LOG_INFO("\n"); } } else { - GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0); + GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f)\n", + __func__, + size_aligned / 1024.0 / 1024.0, + device.currentAllocatedSize / 1024.0 / 1024.0); } +#endif #endif UNUSED(device); + UNUSED(size_aligned); } GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { @@ -2756,8 +2958,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buff return NULL; } - GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0); - ggml_backend_metal_log_allocated_size(device); + //ggml_backend_metal_log_allocated_size(device, size_aligned); return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size); } @@ -2844,7 +3045,7 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, return false; } - GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0); + ggml_backend_metal_log_allocated_size(device, size_aligned); ++ctx->n_buffers; } else { @@ -2867,7 +3068,8 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, return false; } - GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, offs = %12ld", __func__, size_step_aligned / 1024.0 / 1024.0, i); + ggml_backend_metal_log_allocated_size(device, size_step_aligned); + if (i + size_step < size) { GGML_METAL_LOG_INFO("\n"); } @@ -2876,8 +3078,6 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, } } - ggml_backend_metal_log_allocated_size(device); - return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size); } diff --git a/ggml-metal.metal b/ggml-metal.metal index 191880af1..3d4276ae0 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -352,11 +352,12 @@ kernel void kernel_sum_rows( dst_row[0] = row_sum; } +template kernel void kernel_soft_max( - device const float * src0, - device const float * src1, - device const float * src2, - device float * dst, + device const char * src0, + device const char * src1, + device const char * src2, + device char * dst, constant int64_t & ne00, constant int64_t & ne01, constant int64_t & ne02, @@ -375,10 +376,10 @@ kernel void kernel_soft_max( const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01; const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01); - device const float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; - device const float * pmask = src1 != src0 ? src1 + i01*ne00 : nullptr; - device const float * ppos = src2 != src0 ? src2 : nullptr; - device float * pdst = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; + device const float * psrc0 = (device const float *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); + device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*ne00 : nullptr; + device const T * ppos = src2 != src0 ? (device const T *) src2 : nullptr; + device float * pdst = (device float *) dst + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); float slope = 0.0f; @@ -456,11 +457,12 @@ kernel void kernel_soft_max( } } +template kernel void kernel_soft_max_4( - device const float * src0, - device const float * src1, - device const float * src2, - device float * dst, + device const char * src0, + device const char * src1, + device const char * src2, + device char * dst, constant int64_t & ne00, constant int64_t & ne01, constant int64_t & ne02, @@ -479,10 +481,10 @@ kernel void kernel_soft_max_4( const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01; const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01); - device const float4 * psrc4 = (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); - device const float4 * pmask = src1 != src0 ? (device const float4 *)(src1 + i01*ne00) : nullptr; - device const float4 * ppos = src2 != src0 ? (device const float4 *)(src2) : nullptr; - device float4 * pdst4 = (device float4 *)(dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); + device const float4 * psrc4 = (device const float4 *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4; + device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*ne00/4 : nullptr; + device const T * ppos = src2 != src0 ? (device const T *) src2 : nullptr; + device float4 * pdst4 = (device float4 *) dst + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4; float slope = 0.0f; @@ -499,7 +501,7 @@ kernel void kernel_soft_max_4( float4 lmax4 = -INFINITY; for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) { - lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)); + lmax4 = fmax(lmax4, psrc4[i00]*scale + (float4)((pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f))); } const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3])); @@ -525,7 +527,7 @@ kernel void kernel_soft_max_4( // parallel sum float4 lsum4 = 0.0f; for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) { - const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val); + const float4 exp_psrc4 = exp((psrc4[i00]*scale + (float4)((pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f))) - max_val); lsum4 += exp_psrc4; pdst4[i00] = exp_psrc4; } @@ -562,6 +564,14 @@ kernel void kernel_soft_max_4( } } +typedef decltype(kernel_soft_max) kernel_soft_max_t; +typedef decltype(kernel_soft_max_4) kernel_soft_max_4_t; + +template [[host_name("kernel_soft_max_f16")]] kernel kernel_soft_max_t kernel_soft_max; +template [[host_name("kernel_soft_max_f32")]] kernel kernel_soft_max_t kernel_soft_max; +template [[host_name("kernel_soft_max_f16_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4; +template [[host_name("kernel_soft_max_f32_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4; + kernel void kernel_diag_mask_inf( device const float * src0, device float * dst, @@ -2084,6 +2094,632 @@ kernel void kernel_leaky_relu_f32( dst[tpig] = src0[tpig] > 0.0f ? src0[tpig] : src0[tpig] * slope; } +typedef void (flash_attn_ext_f16_t)( + device const char * q, + device const char * k, + device const char * v, + device const char * mask, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant int64_t & ne03, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant uint64_t & nb03, + constant int64_t & ne10, + constant int64_t & ne11, + constant int64_t & ne12, + constant int64_t & ne13, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant uint64_t & nb12, + constant uint64_t & nb13, + constant int64_t & ne31, + constant uint64_t & nb31, + constant int64_t & ne0, + constant int64_t & ne1, + constant int64_t & ne2, + constant int64_t & ne3, + constant float & scale, + threadgroup half * shared, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]], + ushort tiisg[[thread_index_in_simdgroup]], + ushort sgitg[[simdgroup_index_in_threadgroup]]); + +// ref: https://arxiv.org/pdf/2307.08691.pdf +template // head size, queries per threadgroup, cache items per threadgroup +kernel void kernel_flash_attn_ext_f16( + device const char * q, + device const char * k, + device const char * v, + device const char * mask, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant int64_t & ne03, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant uint64_t & nb03, + constant int64_t & ne10, + constant int64_t & ne11, + constant int64_t & ne12, + constant int64_t & ne13, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant uint64_t & nb12, + constant uint64_t & nb13, + constant int64_t & ne31, + constant uint64_t & nb31, + constant int64_t & ne0, + constant int64_t & ne1, + constant int64_t & ne2, + constant int64_t & ne3, + constant float & scale, + threadgroup half * shared [[threadgroup(0)]], + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]], + ushort tiisg[[thread_index_in_simdgroup]], + ushort sgitg[[simdgroup_index_in_threadgroup]]) { + const short nsg = ntg.y; // number of simdgroups + + const short iq3 = tgpig[2]; + const short iq2 = tgpig[1]; + const short iq1 = tgpig[0]*Q; + + const short D4 = D/4; + const short D8 = D/8; + const short Q8 = Q/8; + const short NW = N_SIMDWIDTH; + const short SH = (C + Q); // shared memory per simdgroup in (half) + + const short T = D + 2*nsg*SH; // shared memory size per query in (half) + const short TF = T/2; // shared memory size per query in (float) + const short T4 = T/4; // shared memory size per query in (half4) + + threadgroup half * sq = (threadgroup half *) (shared + 0*D); // holds the query data + threadgroup half4 * sq4 = (threadgroup half4 *) (shared + 0*D); // same as above but in half4 + threadgroup float * ss = (threadgroup float *) (shared + 2*sgitg*SH + 1*D); // scratch buffer for attention and diagonal matrix + + // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper) + simdgroup_half8x8 lo[D8]; + + // load heads from Q to shared memory + for (short j = sgitg; j < Q; j += nsg) { + device const float4 * q4 = (device const float4 *) ((device const char *) q + ((iq1 + j)*nb01 + iq2*nb02 + iq3*nb03)); + + for (short i = tiisg; i < D4; i += NW) { + if (iq1 + j < ne01) { + sq4[j*T4 + i] = (half4) q4[i]; + } else { + sq4[j*T4 + i] = 0.0h; + } + } + } + + // zero out lo + for (short i = 0; i < D8; ++i) { + lo[i] = make_filled_simdgroup_matrix(0.0h); + } + + // zero out shared memory SH + for (short j = 0; j < Q; ++j) { + for (short i = tiisg; i < SH; i += NW) { + ss[j*TF + i] = 0.0f; + } + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + { + float S[Q] = { [0 ... Q-1] = 0.0h }; + float M[Q] = { [0 ... Q-1] = -FLT_MAX/2 }; + + // assume K and V are same shape + const short ne22 = ne12; + const short ne23 = ne13; + + const uint nb21 = nb11; + const uint nb22 = nb12; + const uint nb23 = nb13; + + // broadcast + const short rk2 = ne02/ne12; + const short rk3 = ne03/ne13; + + const short rv2 = ne02/ne22; + const short rv3 = ne03/ne23; + + // k indices + const short ik2 = iq2/rk2; + const short ik3 = iq3/rk3; + + // v indices + const short iv2 = iq2/rv2; + const short iv3 = iq3/rv3; + + // load the queries from shared memory into local memory + simdgroup_half8x8 mq[D8]; + + for (short i = 0; i < D8; ++i) { + simdgroup_load(mq[i], sq + i*8, T); + } + + // pointer to the mask + device const half * mp = (device const half *) (mask + iq1*nb31); + + // prepare diagonal scale matrix + simdgroup_float8x8 mscale(scale); + + // loop over the KV cache + // each simdgroup handles blocks of Q rows and C columns + for (int ic0 = 0; ic0 < ne11; ic0 += C*nsg) { + const int ic = ic0 + C*sgitg; + if (ic >= ne11) { + break; + } + + // Q*K^T + { + for (short cc = 0; cc < C/8; ++cc) { + simdgroup_float8x8 mqk = make_filled_simdgroup_matrix(0.h); + + device const half * pk = (device const half *) ((device const char *) k + ((ic + 8*cc)*nb11 + ik2*nb12 + ik3*nb13)); + + for (short i = 0; i < D8; ++i) { + simdgroup_half8x8 mk; + simdgroup_load(mk, pk + i*8, nb11/sizeof(half), 0, true); // transpose + + simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk); + } + + // mqk = mqk*scale + mask + simdgroup_half8x8 mm; + simdgroup_load(mm, mp + ic + 8*cc, nb31/sizeof(half), 0, false); + simdgroup_multiply_accumulate(mqk, mqk, mscale, mm); + + simdgroup_store(mqk, ss + 8*cc, TF, 0, false); + } + } + + // used to detect blocks full of -INF + float smax = -INFINITY; + + // online softmax + { + float ms[Q]; + + for (short j = 0; j < Q; ++j) { + const short p = tiisg; + + const float m = M[j]; + const float s = ss[j*TF + p]; + + smax = simd_max(max(smax, s)); + M[j] = simd_max(max(M[j], s)); + + ms[j] = exp(m - M[j]); + const float vs = exp(s - M[j]); + + S[j] = S[j]*ms[j] + simd_sum(vs); + + // the P matrix from the paper (Q rows, C columns) + ss[j*TF + p] = vs; + } + + // create a QxQ diagonal matrix for rescaling the output + if (tiisg < Q) { + ss[tiisg*TF + C + tiisg] = ms[tiisg]; + } + } + + // skip -INF blocks + if (smax == -INFINITY) { + continue; + } + + // O = diag(ms)*O + { + simdgroup_float8x8 mm; + simdgroup_load(mm, ss + C, TF, 0, false); + + for (short i = 0; i < D8; ++i) { + simdgroup_multiply(lo[i], mm, lo[i]); + } + } + + // O = O + (Q*K^T)*V + { + for (short cc = 0; cc < C/8; ++cc) { + device const half * pv = (device const half *) ((device const char *) v + ((ic + 8*cc)*nb21 + iv2*nb22 + iv3*nb23)); + + for (short i = 0; i < D8; ++i) { + simdgroup_half8x8 mk; + simdgroup_load(mk, pv + i*8, nb21/sizeof(half), 0, false); + + simdgroup_float8x8 mv; + simdgroup_load(mv, ss + 8*cc, TF, 0, false); + + simdgroup_multiply_accumulate(lo[i], mv, mk, lo[i]); + } + } + } + } + + // these are needed for reducing the results from the simdgroups (reuse the ss buffer) + for (short j = 0; j < Q; ++j) { + if (tiisg == 0) { + ss[j*TF + 0] = S[j]; + ss[j*TF + 1] = M[j]; + } + } + } + + // reduce the warps sequentially + for (short sg = 1; sg < nsg; ++sg) { + float S = { 0.0h }; + float M = { -FLT_MAX/2 }; + + threadgroup_barrier(mem_flags::mem_threadgroup); + + // each simdgroup stores its output to shared memory, reusing sq + if (sgitg == sg) { + for (short i = 0; i < D8; ++i) { + simdgroup_store(lo[i], sq + i*8, T, 0, false); + } + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + // the first simdgroup accumulates the results from the other simdgroups + if (sgitg == 0) { + for (short j = 0; j < Q; ++j) { + const float S0 = ss[j*TF + 0]; + const float S1 = ss[j*TF + sg*SH + 0]; + + const float M0 = ss[j*TF + 1]; + const float M1 = ss[j*TF + sg*SH + 1]; + + M = max(M0, M1); + + const float ms0 = exp(M0 - M); + const float ms1 = exp(M1 - M); + + S = S0*ms0 + S1*ms1; + + if (tiisg == 0) { + ss[j*TF + 0] = S; + ss[j*TF + 1] = M; + + ss[j*TF + C + j ] = ms0; + ss[j*TF + C + j + sg*SH] = ms1; + } + } + + // O_0 = diag(ms0)*O_0 + diag(ms1)*O_1 + { + simdgroup_half8x8 t; + simdgroup_float8x8 ms0; + simdgroup_float8x8 ms1; + + simdgroup_load(ms0, ss + C, TF, 0, false); + simdgroup_load(ms1, ss + C + sg*SH, TF, 0, false); + + for (short i = 0; i < D8; ++i) { + simdgroup_load (t, sq + i*8, T, 0, false); + simdgroup_multiply(t, ms1, t); + + simdgroup_multiply_accumulate(lo[i], ms0, lo[i], t); + } + } + } + } + + // store result to shared memory (reuse sq) + if (sgitg == 0) { + for (short i = 0; i < D8; ++i) { + simdgroup_store(lo[i], sq + i*8, T, 0, false); + } + } + + device float4 * dst4 = (device float4 *) dst; + + // final rescale with 1/S and store to global memory + if (sgitg == 0) { + for (short j = 0; j < Q && iq1 + j < ne01; ++j) { + const float S = ss[j*TF + 0]; + + for (short i = tiisg; i < D4; i += NW) { + dst4[(iq3*ne2*ne1 + iq2 + (iq1 + j)*ne1)*D4 + i] = (float4) sq4[j*T4 + i]/S; + } + } + } +} + +template [[host_name("kernel_flash_attn_ext_f16_h64" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<64>; +template [[host_name("kernel_flash_attn_ext_f16_h80" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<80>; +template [[host_name("kernel_flash_attn_ext_f16_h96" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<96>; +template [[host_name("kernel_flash_attn_ext_f16_h112")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<112>; +template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<128>; +template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<256>; + +template // head size, queries per threadgroup, cache items per threadgroup +kernel void kernel_flash_attn_ext_vec_f16( + device const char * q, + device const char * k, + device const char * v, + device const char * mask, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant int64_t & ne03, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant uint64_t & nb03, + constant int64_t & ne10, + constant int64_t & ne11, + constant int64_t & ne12, + constant int64_t & ne13, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant uint64_t & nb12, + constant uint64_t & nb13, + constant int64_t & ne31, + constant uint64_t & nb31, + constant int64_t & ne0, + constant int64_t & ne1, + constant int64_t & ne2, + constant int64_t & ne3, + constant float & scale, + threadgroup half * shared [[threadgroup(0)]], + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]], + ushort tiisg[[thread_index_in_simdgroup]], + ushort sgitg[[simdgroup_index_in_threadgroup]]) { + const short nsg = ntg.y; // number of simdgroups + + const short iq3 = tgpig[2]; + const short iq2 = tgpig[1]; + const short iq1 = tgpig[0]; + + const short D4 = D/4; + const short NW = N_SIMDWIDTH; + const short SH = (C + Q); // shared memory per simdgroup in (half) + + const short T = D + 2*nsg*SH; // shared memory size per query in (half) + + //threadgroup half * sq = (threadgroup half *) (shared + 0*D); // holds the query data + threadgroup half4 * sq4 = (threadgroup half4 *) (shared + 0*D); // same as above but in half4 + threadgroup float * ss = (threadgroup float *) (shared + 2*sgitg*SH + 1*D); // scratch buffer for attention and diagonal matrix + threadgroup float4 * ss4 = (threadgroup float4 *) (shared + 2*sgitg*SH + 1*D); // same as above but in half4 + threadgroup half4 * sr4 = (threadgroup half4 *) (shared + sgitg*D + 1*T); // scratch buffer for the results + + // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper) + half4 lo[D4/NW]; + + // load heads from Q to shared memory + device const float4 * q4 = (device const float4 *) ((device const char *) q + (iq1*nb01 + iq2*nb02 + iq3*nb03)); + + for (short i = tiisg; i < D4; i += NW) { + if (iq1 < ne01) { + sq4[i] = (half4) q4[i]; + } else { + sq4[i] = 0.0h; + } + } + + // zero out lo + for (short i = tiisg; i < D4; i += NW) { + lo[i/NW] = 0.0h; + } + + // zero out shared memory SH + for (short i = tiisg; i < SH/4; i += NW) { + ss4[i] = 0.0h; + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + { + float S = { 0.0h }; + float M = { -FLT_MAX/2 }; + + // assume K and V are same shape + const short ne22 = ne12; + const short ne23 = ne13; + + const uint nb21 = nb11; + const uint nb22 = nb12; + const uint nb23 = nb13; + + // broadcast + const short rk2 = ne02/ne12; + const short rk3 = ne03/ne13; + + const short rv2 = ne02/ne22; + const short rv3 = ne03/ne23; + + // k indices + const short ik2 = iq2 / rk2; + const short ik3 = iq3 / rk3; + + // v indices + const short iv2 = iq2 / rv2; + const short iv3 = iq3 / rv3; + + // load the queries from shared memory into local memory + half4 mq[D4]; + + for (short ii = 0; ii < D4; ii += NW) { + short i = ii + tiisg; + mq[i] = sq4[i]; + } + + // pointer to the mask + device const half4 * mp4 = (device const half4 *) (mask + iq1*nb31); + + // loop over the KV cache + // each simdgroup handles blocks of Q rows and C columns + for (int ic0 = 0; ic0 < ne11; ic0 += C*nsg) { + const int ic = ic0 + C*sgitg; + if (ic >= ne11) { + break; + } + + // Q*K^T + { +#pragma unroll + for (short cc = 0; cc < C/4; ++cc) { + float4 mqk = { 0.0h }; + + device const half4 * pk4 = (device const half4 *) ((device const char *) k + ((ic + 4*cc)*nb11 + ik2*nb12 + ik3*nb13)); + +#pragma unroll + for (short ii = 0; ii < D4; ii += NW) { + const short i = ii + tiisg; + + half4x4 mk; + mk[0] = pk4[i + 0*(nb11/8)]; + mk[1] = pk4[i + 1*(nb11/8)]; + mk[2] = pk4[i + 2*(nb11/8)]; + mk[3] = pk4[i + 3*(nb11/8)]; + + mqk += (float4) (mq[i] * mk); + } + + // reduce the results from the threads in the simdgroup + mqk += simd_shuffle_down(mqk, 16); + mqk += simd_shuffle_down(mqk, 8); + mqk += simd_shuffle_down(mqk, 4); + mqk += simd_shuffle_down(mqk, 2); + mqk += simd_shuffle_down(mqk, 1); + + // mqk = mqk*scale + mask + if (tiisg == 0) { + float4 mm = (float4) mp4[ic/4 + cc]; + mqk = mqk*scale + mm; + + ss4[cc] = mqk; + } + } + } + + // online softmax + { + const short p = tiisg; + + const float m = M; + const float s = ss[p]; + + M = simd_max(max(M, s)); + + const float ms = exp(m - M); + const float vs = exp(s - M); + + S = S*ms + simd_sum(vs); + + // the P matrix from the paper (Q rows, C columns) + ss[p] = vs; + + // O = diag(ms)*O +#pragma unroll + for (short ii = 0; ii < D4; ii += NW) { + const short i = ii + tiisg; + lo[i/NW] *= ms; + } + } + + // O = O + (Q*K^T)*V + { +#pragma unroll + for (short cc = 0; cc < C/4; ++cc) { + device const half4 * pv4 = (device const half4 *) ((device const char *) v + ((ic + 4*cc)*nb21 + iv2*nb22 + iv3*nb23)); + +#pragma unroll + for (short ii = 0; ii < D4; ii += NW) { + const short i = ii + tiisg; + + lo[i/NW] += pv4[i + 0*(nb21/8)] * ss[4*cc + 0]; + lo[i/NW] += pv4[i + 1*(nb21/8)] * ss[4*cc + 1]; + lo[i/NW] += pv4[i + 2*(nb21/8)] * ss[4*cc + 2]; + lo[i/NW] += pv4[i + 3*(nb21/8)] * ss[4*cc + 3]; + } + } + } + + } + + // these are needed for reducing the results from the simdgroups (reuse the ss buffer) + if (tiisg == 0) { + ss[0] = S; + ss[1] = M; + } + } + + // store results to shared memory + for (short ii = 0; ii < D4; ii += NW) { + short i = ii + tiisg; + sr4[i] = lo[ii/NW]; + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + // parallel reduce + for (short r = nsg/2; r > 0; r >>= 1) { + if (sgitg < r) { + const float S0 = ss[ 0]; + const float S1 = ss[r*SH + 0]; + + const float M0 = ss[ 1]; + const float M1 = ss[r*SH + 1]; + + const float M = max(M0, M1); + + const float ms0 = exp(M0 - M); + const float ms1 = exp(M1 - M); + + const float S = S0*ms0 + S1*ms1; + + if (tiisg == 0) { + ss[0] = S; + ss[1] = M; + } + + // O_0 = diag(ms0)*O_0 + diag(ms1)*O_1 + for (short ii = 0; ii < D4; ii += NW) { + short i = ii + tiisg; + sr4[i] = sr4[i]*ms0 + sr4[i + r*D4]*ms1; + } + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + } + + device float4 * dst4 = (device float4 *) dst; + + // final rescale with 1/S and store to global memory + if (sgitg == 0) { + const float S = ss[0]; + + for (short ii = 0; ii < D4; ii += NW) { + short i = ii + tiisg; + dst4[(iq3*ne2*ne1 + iq2 + (iq1)*ne1)*D4 + i] = (float4) sr4[i]/S; + } + } +} + +template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<128>; +template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<256>; + kernel void kernel_cpy_f16_f16( device const half * src0, device half * dst, diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp index a9b310243..57fe4ea3d 100644 --- a/ggml-sycl.cpp +++ b/ggml-sycl.cpp @@ -13416,11 +13416,16 @@ void print_device_detail(int id, sycl::device &device, std::string device_type) version += std::to_string(prop.get_minor_version()); device_type = std::regex_replace(device_type, std::regex("ext_oneapi_"), ""); + std::string name = std::string(prop.get_name()); + name = std::regex_replace(name, std::regex("\\(R\\)"), ""); + name = std::regex_replace(name, std::regex("\\(TM\\)"), ""); - fprintf(stderr, "|%2d|%18s|%45s|%10s|%11d|%8d|%7d|%15lu|\n", id, device_type.c_str(), - prop.get_name(), version.c_str(), prop.get_max_compute_units(), + auto global_mem_size = prop.get_global_mem_size()/1000000; + + fprintf(stderr, "|%2d|%19s|%39s|%7s|%7d|%8d|%5d|%6luM|%21s|\n", id, device_type.c_str(), + name.c_str(), version.c_str(), prop.get_max_compute_units(), prop.get_max_work_group_size(), prop.get_max_sub_group_size(), - prop.get_global_mem_size()); + global_mem_size, device.get_info().c_str()); } void ggml_backend_sycl_print_sycl_devices() { @@ -13428,9 +13433,10 @@ void ggml_backend_sycl_print_sycl_devices() { int device_count = dpct::dev_mgr::instance().device_count(); std::map DeviceNums; fprintf(stderr, "found %d SYCL devices:\n", device_count); - fprintf(stderr, "| | | |Compute |Max compute|Max work|Max sub| |\n"); - fprintf(stderr, "|ID| Device Type| Name|capability|units |group |group |Global mem size|\n"); - fprintf(stderr, "|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|\n"); + fprintf(stderr, "| | | | |Max | |Max |Global | |\n"); + fprintf(stderr, "| | | | |compute|Max work|sub |mem | |\n"); + fprintf(stderr, "|ID| Device Type| Name|Version|units |group |group|size | Driver version|\n"); + fprintf(stderr, "|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|\n"); for (int id = 0; id < device_count; ++id) { sycl::device device = dpct::dev_mgr::instance().get_device(id); sycl::backend backend = device.get_backend(); @@ -14738,7 +14744,12 @@ inline void ggml_sycl_op_soft_max(const ggml_tensor *src0, GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); + const ggml_tensor * src2 = dst->src[2]; + +#pragma message("TODO: add ggml_sycl_op_soft_max() F16 src1 and src2 support") +#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021") GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional + GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F32); // src2 contains positions and it is optional const int64_t ne00 = src0->ne[0]; const int64_t nrows_x = ggml_nrows(src0); @@ -14754,7 +14765,6 @@ inline void ggml_sycl_op_soft_max(const ggml_tensor *src0, float * src2_dd = nullptr; sycl_pool_alloc src2_f; - ggml_tensor * src2 = dst->src[2]; const bool use_src2 = src2 != nullptr; if (use_src2) { diff --git a/ggml-vulkan.cpp b/ggml-vulkan.cpp index 1736ab736..f712cdd5a 100644 --- a/ggml-vulkan.cpp +++ b/ggml-vulkan.cpp @@ -3178,6 +3178,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const } return nullptr; case GGML_OP_SOFT_MAX: +#pragma message("TODO: add ggml_vk_soft_max() F16 src1 and src2 support") +#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021") + GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); + GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F32); + if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && (src2 == nullptr || src2->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_soft_max_f32; } diff --git a/ggml.c b/ggml.c index 34eef23fc..74ecd5927 100644 --- a/ggml.c +++ b/ggml.c @@ -951,7 +951,7 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) { #define GGML_F16_VEC_ZERO GGML_F16x8_ZERO #define GGML_F16_VEC_SET1 GGML_F16x8_SET1 #define GGML_F16_VEC_LOAD(p, i) GGML_F16x8_LOAD(p) - #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE(p, r[i]) + #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((ggml_fp16_internal_t *)(p), r[i]) #define GGML_F16_VEC_FMA GGML_F16x8_FMA #define GGML_F16_VEC_ADD GGML_F16x8_ADD #define GGML_F16_VEC_MUL GGML_F16x8_MUL @@ -977,7 +977,7 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) { #define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO #define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1 #define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p) - #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE(p, r[i]) + #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE((ggml_fp16_internal_t *)(p), r[i]) #define GGML_F16_VEC_FMA GGML_F32Cx4_FMA #define GGML_F16_VEC_ADD GGML_F32Cx4_ADD #define GGML_F16_VEC_MUL GGML_F32Cx4_MUL @@ -1046,7 +1046,7 @@ do { \ // unlike _mm256_cvt intrinsics that require F16C, _mm512_cvt is defined in AVX512F // so F16C guard isn't required -#define GGML_F32Cx16_LOAD(x) _mm512_cvtph_ps(_mm256_loadu_si256((__m256i *)(x))) +#define GGML_F32Cx16_LOAD(x) _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(x))) #define GGML_F32Cx16_STORE(x, y) _mm256_storeu_si256((__m256i *)(x), _mm512_cvtps_ph(y, 0)) #define GGML_F32Cx16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a) @@ -1144,7 +1144,7 @@ do { \ #if defined(__F16C__) // the _mm256_cvt intrinsics require F16C -#define GGML_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((__m128i *)(x))) +#define GGML_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x))) #define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0)) #else static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) { @@ -1662,6 +1662,37 @@ inline static void ggml_vec_mad_f32(const int n, float * restrict y, const float #endif } +inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * restrict y, const ggml_fp16_t * restrict x, const float v) { +#if defined(GGML_SIMD) + const int np = (n & ~(GGML_F16_STEP - 1)); + + GGML_F16_VEC vx = GGML_F16_VEC_SET1(v); + + GGML_F16_VEC ax[GGML_F16_ARR]; + GGML_F16_VEC ay[GGML_F16_ARR]; + + for (int i = 0; i < np; i += GGML_F16_STEP) { + for (int j = 0; j < GGML_F16_ARR; j++) { + ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j); + ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j); + ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx); + + GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j); + } + } + + // leftovers + for (int i = np; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v); + } +#else + // scalar + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v); + } +#endif +} + // xs and vs are byte strides of x and v inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int vs, float * restrict y, const float * restrict xv, const float * restrict vv) { @@ -1746,6 +1777,35 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float v) { #endif } +inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) { +#if defined(GGML_SIMD) + const int np = (n & ~(GGML_F16_STEP - 1)); + + GGML_F16_VEC vx = GGML_F16_VEC_SET1(v); + + GGML_F16_VEC ay[GGML_F16_ARR]; + + for (int i = 0; i < np; i += GGML_F16_STEP) { + for (int j = 0; j < GGML_F16_ARR; j++) { + ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j); + ay[j] = GGML_F16_VEC_MUL(ay[j], vx); + + GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j); + } + } + + // leftovers + for (int i = np; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v); + } +#else + // scalar + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v); + } +#endif +} + inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, 0, x, 0, x, 0, 1); *s = sqrtf(*s); } inline static void ggml_vec_sqr_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i]; } inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); } @@ -2000,6 +2060,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "LEAKY_RELU", "FLASH_ATTN", + "FLASH_ATTN_EXT", "FLASH_FF", "FLASH_ATTN_BACK", "SSM_CONV", @@ -2026,7 +2087,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "CROSS_ENTROPY_LOSS_BACK", }; -static_assert(GGML_OP_COUNT == 76, "GGML_OP_COUNT != 76"); +static_assert(GGML_OP_COUNT == 77, "GGML_OP_COUNT != 77"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -2090,6 +2151,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "leaky_relu(x)", "flash_attn(x)", + "flash_attn_ext(x)", "flash_ff(x)", "flash_attn_back(x)", "ssm_conv(x)", @@ -2116,7 +2178,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "cross_entropy_loss_back(x,y)", }; -static_assert(GGML_OP_COUNT == 76, "GGML_OP_COUNT != 76"); +static_assert(GGML_OP_COUNT == 77, "GGML_OP_COUNT != 77"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -4559,6 +4621,8 @@ struct ggml_tensor * ggml_mul_mat( void ggml_mul_mat_set_prec( struct ggml_tensor * a, enum ggml_prec prec) { + GGML_ASSERT(a->op == GGML_OP_MUL_MAT); + const int32_t prec_i32 = (int32_t) prec; ggml_set_op_params_i32(a, 0, prec_i32); @@ -5397,17 +5461,23 @@ static struct ggml_tensor * ggml_soft_max_impl( GGML_ASSERT(ggml_is_contiguous(a)); if (mask) { + GGML_ASSERT(mask->type == GGML_TYPE_F16 || mask->type == GGML_TYPE_F32); GGML_ASSERT(ggml_is_contiguous(mask)); GGML_ASSERT(ggml_is_matrix(mask)); - GGML_ASSERT(ggml_can_repeat_rows(mask, a)); + GGML_ASSERT(mask->ne[0] == a->ne[0]); + GGML_ASSERT(mask->ne[1] >= a->ne[1]); } if (pos) { GGML_ASSERT(ggml_is_vector(pos)); - GGML_ASSERT(pos->type == GGML_TYPE_F32); + GGML_ASSERT(pos->type == GGML_TYPE_F16 || pos->type == GGML_TYPE_F32); GGML_ASSERT(pos->ne[0] == a->ne[0]); } + if (pos && mask) { + GGML_ASSERT(pos->type == mask->type); + } + if (max_bias > 0.0f) { GGML_ASSERT(pos); } @@ -6216,6 +6286,59 @@ struct ggml_tensor * ggml_flash_attn( return result; } +// ggml_flash_attn_ext + +struct ggml_tensor * ggml_flash_attn_ext( + struct ggml_context * ctx, + struct ggml_tensor * q, + struct ggml_tensor * k, + struct ggml_tensor * v, + struct ggml_tensor * mask, + float scale) { + GGML_ASSERT(ggml_can_mul_mat(k, q)); + // TODO: check if vT can be multiplied by (k*qT) + if (mask) { + GGML_ASSERT(ggml_is_contiguous(mask)); + GGML_ASSERT(mask->ne[2] == 1); + GGML_ASSERT(mask->ne[3] == 1); + GGML_ASSERT(mask->ne[1] >= GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD) && + "the Flash-Attention kernel requires the mask to be padded to GGML_KQ_MASK_PAD and at least n_queries big"); + //GGML_ASSERT(ggml_can_repeat_rows(mask, qk)); + } + + bool is_node = false; + + if (q->grad || k->grad || v->grad) { + is_node = true; + } + + // permute(0, 2, 1, 3) + int64_t ne[4] = { q->ne[0], q->ne[2], q->ne[1], q->ne[3] }; + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); + + float params[] = { scale }; + ggml_set_op_params(result, params, sizeof(params)); + + result->op = GGML_OP_FLASH_ATTN_EXT; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = q; + result->src[1] = k; + result->src[2] = v; + result->src[3] = mask; + + return result; +} + +void ggml_flash_attn_ext_set_prec( + struct ggml_tensor * a, + enum ggml_prec prec) { + GGML_ASSERT(a->op == GGML_OP_FLASH_ATTN_EXT); + + const int32_t prec_i32 = (int32_t) prec; + + ggml_set_op_params_i32(a, 1, prec_i32); // scale is on first pos +} + // ggml_flash_ff struct ggml_tensor * ggml_flash_ff( @@ -12255,7 +12378,7 @@ static void ggml_compute_forward_soft_max_f32( GGML_TENSOR_UNARY_OP_LOCALS - const int64_t ne11 = src1 ? src1->ne[1] : 1; + //const int64_t ne11 = src1 ? src1->ne[1] : 1; // TODO: is this supposed to be ceil instead of floor? // https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370 @@ -12278,19 +12401,31 @@ static void ggml_compute_forward_soft_max_f32( float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith; // when max_bias <= 0.0f, src2 is not used and we default it to src0 to avoid branching - float * pos = src2 ? (float *) src2->data : src0->data; + ggml_fp16_t * pos_f16 = src2 ? (ggml_fp16_t *) src2->data : src0->data; + float * pos_f32 = src2 ? (float *) src2->data : src0->data; + + const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16) || (src2 && src2->type == GGML_TYPE_F16); for (int i1 = ir0; i1 < ir1; i1++) { float * sp = (float *)((char *) src0->data + i1*src0->nb[1]); float * dp = (float *)((char *) dst->data + i1*dst->nb[1]); // broadcast the mask across rows - float * mp = src1 ? (float *)((char *) src1->data + (i1%ne11)*src1->nb[1]) : NULL; + ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data) + (i1%ne01)*ne00 : NULL; + float * mp_f32 = src1 ? (float *)((char *) src1->data) + (i1%ne01)*ne00 : NULL; ggml_vec_cpy_f32 (nc, wp, sp); ggml_vec_scale_f32(nc, wp, scale); - if (mp) { - ggml_vec_acc_f32(nc, wp, mp); + if (mp_f32) { + if (use_f16) { + for (int i = 0; i < nc; ++i) { + wp[i] += GGML_FP16_TO_FP32(mp_f16[i]); + } + } else { + for (int i = 0; i < nc; ++i) { + wp[i] += mp_f32[i]; + } + } } // ALiBi bias @@ -12298,8 +12433,14 @@ static void ggml_compute_forward_soft_max_f32( const uint32_t h = (i1/ne01)%ne02; // head const float slope = h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1); - for (int i = 0; i < nc; i++) { - wp[i] = wp[i] + slope*pos[i]; + if (use_f16) { + for (int i = 0; i < nc; ++i) { + wp[i] += slope*GGML_FP16_TO_FP32(pos_f16[i]); + } + } else { + for (int i = 0; i < nc; ++i) { + wp[i] += slope*pos_f32[i]; + } } } @@ -14569,6 +14710,198 @@ static void ggml_compute_forward_flash_attn( } } +// ggml_compute_forward_flash_attn_ext + +static void ggml_compute_forward_flash_attn_ext_f16( + const struct ggml_compute_params * params, + const struct ggml_tensor * q, + const struct ggml_tensor * k, + const struct ggml_tensor * v, + const struct ggml_tensor * mask, + struct ggml_tensor * dst) { + int64_t t0 = ggml_perf_time_us(); + UNUSED(t0); + + GGML_TENSOR_LOCALS(int64_t, neq, q, ne) + GGML_TENSOR_LOCALS(size_t, nbq, q, nb) + GGML_TENSOR_LOCALS(int64_t, nek, k, ne) + GGML_TENSOR_LOCALS(size_t, nbk, k, nb) + GGML_TENSOR_LOCALS(int64_t, nev, v, ne) + GGML_TENSOR_LOCALS(size_t, nbv, v, nb) + GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) + GGML_TENSOR_LOCALS(size_t, nb, dst, nb) + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t D = neq0; + const int64_t N = neq1; + + GGML_ASSERT(ne0 == D); + GGML_ASSERT(ne2 == N); + + GGML_ASSERT(nbq0 == sizeof(float)); + GGML_ASSERT(nbk0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nbv0 == sizeof(ggml_fp16_t)); + + GGML_ASSERT(neq0 == D); + GGML_ASSERT(nek0 == D); + GGML_ASSERT(nev0 == D); + + GGML_ASSERT(neq1 == N); + GGML_ASSERT(nev0 == D); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + // broadcast factors + const int64_t rk2 = neq2/nek2; + const int64_t rk3 = neq3/nek3; + + const int64_t rv2 = neq2/nev2; + const int64_t rv3 = neq3/nev3; + + if (params->type == GGML_TASK_TYPE_INIT) { + return; + } + + if (params->type == GGML_TASK_TYPE_FINALIZE) { + return; + } + + // parallelize by q rows using ggml_vec_dot_f32 + + // total rows in q + const int nr = neq1*neq2*neq3; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float scale = 1.0f; + memcpy(&scale, (float *) dst->op_params + 0, sizeof(float)); + + // loop over n_batch and n_head + for (int ir = ir0; ir < ir1; ++ir) { + // q indices + const int iq3 = ir/(neq2*neq1); + const int iq2 = (ir - iq3*neq2*neq1)/neq1; + const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1); + + float S = 0.0f; + float M = -INFINITY; + + float * V32 = (float *) params->wdata + ith*(2*D + CACHE_LINE_SIZE_F32); + ggml_fp16_t * Q16 = (ggml_fp16_t *) (V32); // reuse memory + ggml_fp16_t * V16 = (ggml_fp16_t *) (V32 + D); + + memset(V16, 0, D*sizeof(ggml_fp16_t)); + + const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1]) : NULL; + + // k indices + const int ik3 = iq3 / rk3; + const int ik2 = iq2 / rk2; + + // v indices + const int iv3 = iq3 / rv3; + const int iv2 = iq2 / rv2; + + // online softmax / attention + // loop over n_kv and n_head_kv + // ref: https://arxiv.org/pdf/2112.05682.pdf + for (int64_t ic = 0; ic < nek1; ++ic) { + const float mv = mp ? GGML_FP16_TO_FP32(mp[ic]) : 0.0f; + if (mv == -INFINITY) { + continue; + } + + float s; + + // convert Q to F16 in V32 + { + const float * pq = (const float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)); + + for (int64_t d = 0; d < D; ++d) { + Q16[d] = GGML_FP32_TO_FP16(pq[d]); + } + } + + ggml_vec_dot_f16(D, + &s, 0, + (ggml_fp16_t *) ((char *) k->data + ( ic*nbk1 + ik2*nbk2 + ik3*nbk3)), 0, + Q16, 0, 1); + + s = s*scale + mv; + + const float Mold = M; + + float ms = 1.0f; + float vs = 1.0f; + + if (s > M) { + M = s; + ms = expf(Mold - M); + + // V = V*expf(Mold - M) + ggml_vec_scale_f16(D, V16, ms); + } else { + vs = expf(s - M); + } + + const ggml_fp16_t * v16 = (const ggml_fp16_t *) ((char *) v->data + (ic*nbv1 + iv2*nbv2 + iv3*nbv3)); + + // V += v*expf(s - M) + ggml_vec_mad_f16(D, V16, v16, vs); + + S = S*ms + vs; + } + + // V /= S + for (int64_t d = 0; d < D; ++d) { + V32[d] = GGML_FP16_TO_FP32(V16[d])/S; + } + + // dst indices + const int i1 = iq1; + const int i2 = iq2; + const int i3 = iq3; + + // original + //memcpy((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3), V, nev0*sizeof(float)); + + // permute(0, 2, 1, 3) + memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, V32, nb1); + } +} + +static void ggml_compute_forward_flash_attn_ext( + const struct ggml_compute_params * params, + const struct ggml_tensor * q, + const struct ggml_tensor * k, + const struct ggml_tensor * v, + const struct ggml_tensor * mask, + struct ggml_tensor * dst) { + switch (dst->op_params[1]) { + case GGML_PREC_DEFAULT: + case GGML_PREC_F32: + { + // uses F32 accumulators + ggml_compute_forward_flash_attn_ext_f16(params, q, k, v, mask, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_flash_ff static void ggml_compute_forward_flash_ff_f16( @@ -16376,6 +16709,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm const bool masked = t != 0; ggml_compute_forward_flash_attn(params, masked, tensor); } break; + case GGML_OP_FLASH_ATTN_EXT: + { + ggml_compute_forward_flash_attn_ext(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], tensor); + } break; case GGML_OP_FLASH_FF: { ggml_compute_forward_flash_ff(params, tensor); @@ -17388,6 +17725,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor GGML_ASSERT(false); // TODO: not implemented } break; case GGML_OP_FLASH_ATTN: + case GGML_OP_FLASH_ATTN_EXT: { struct ggml_tensor * flash_grad = NULL; if (src0->grad || src1->grad || tensor->src[2]->grad) { @@ -18160,6 +18498,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads, int n_cur_ n_tasks = n_threads; } break; case GGML_OP_FLASH_ATTN: + case GGML_OP_FLASH_ATTN_EXT: { n_tasks = n_threads; } break; @@ -18563,6 +18902,12 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa cur += sizeof(float)*ne11*n_tasks; // this is overestimated by x2 } } break; + case GGML_OP_FLASH_ATTN_EXT: + { + const int64_t ne00 = node->src[0]->ne[0]; // D + + cur = 2*sizeof(float)*ne00*n_tasks; // 2x head size + } break; case GGML_OP_FLASH_FF: { if (node->src[1]->type == GGML_TYPE_F32) { @@ -20819,6 +21164,14 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p // TODO: return an error instead of crashing with GGML_ASSERT gguf_tensor_info_sanitize(info); + // make sure there is no duplicated tensor names + for (uint64_t j = 0; j < i; ++j) { + if (strcmp(info->name.data, ctx->infos[j].name.data) == 0) { + fprintf(stderr, "%s: duplicated tensor name %s\n", __func__, info->name.data); + ok = false; + } + } + if (!ok) { fprintf(stderr, "%s: failed to read tensor info\n", __func__); fclose(file); @@ -21355,6 +21708,10 @@ void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src) { void gguf_add_tensor( struct gguf_context * ctx, const struct ggml_tensor * tensor) { + if (gguf_find_tensor(ctx, tensor->name) != -1) { + GGML_ASSERT(false && "duplicated tensor name"); + } + const int idx = ctx->header.n_tensors; ctx->infos = realloc(ctx->infos, (idx + 1)*sizeof(struct gguf_tensor_info)); diff --git a/ggml.h b/ggml.h index 86e5a8dc5..a11795973 100644 --- a/ggml.h +++ b/ggml.h @@ -475,6 +475,7 @@ extern "C" { GGML_OP_LEAKY_RELU, GGML_OP_FLASH_ATTN, + GGML_OP_FLASH_ATTN_EXT, GGML_OP_FLASH_FF, GGML_OP_FLASH_ATTN_BACK, GGML_OP_SSM_CONV, @@ -1722,6 +1723,25 @@ extern "C" { struct ggml_tensor * v, bool masked); +#define GGML_KQ_MASK_PAD 32 + + // q: [n_embd, n_batch, n_head, 1] + // k: [n_embd, n_kv, n_head_kv, 1] + // v: [n_embd, n_kv, n_head_kv, 1] !! not transposed !! + // mask: [n_kv, n_batch_pad, 1, 1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !! + // res: [n_embd, n_head, n_batch, 1] !! permuted !! + GGML_API struct ggml_tensor * ggml_flash_attn_ext( + struct ggml_context * ctx, + struct ggml_tensor * q, + struct ggml_tensor * k, + struct ggml_tensor * v, + struct ggml_tensor * mask, + float scale); + + GGML_API void ggml_flash_attn_ext_set_prec( + struct ggml_tensor * a, + enum ggml_prec prec); + GGML_API struct ggml_tensor * ggml_flash_attn_back( struct ggml_context * ctx, struct ggml_tensor * q, diff --git a/ggml_vk_generate_shaders.py b/ggml_vk_generate_shaders.py index 5dd700963..1d9a0cc82 100644 --- a/ggml_vk_generate_shaders.py +++ b/ggml_vk_generate_shaders.py @@ -1,11 +1,14 @@ #!/usr/bin/env python +import logging import argparse import asyncio import os import sys from tempfile import gettempdir, NamedTemporaryFile +logger = logging.getLogger("ggml-vk-generate-shaders") + shader_f32 = """ #define FLOAT_TYPE float """ @@ -2498,7 +2501,7 @@ async def string_to_spv(name, code, defines, fp16=True): stdout, stderr = await proc.communicate() - print(" ".join(cmd)) + logger.info(" ".join(cmd)) if proc.returncode: raise RuntimeError(f"{name=} {f.name=} {stdout=} {stderr=}") @@ -2507,7 +2510,7 @@ async def string_to_spv(name, code, defines, fp16=True): cmd.extend([f"-D{key}={value}" for key, value in defines.items()]) code_with_lines = "\n".join([f"{i + 1}: {line}" for i, line in enumerate(preprocessed_code.splitlines())]) - print(f"ERROR compiling {name}\n\n{code_with_lines}\n\n{error}") + logger.error(f"cannot compile {name}\n\n{code_with_lines}\n\n{error}") f.close() os.remove(f.name) sys.exit(proc.returncode) @@ -2520,7 +2523,7 @@ async def string_to_spv(name, code, defines, fp16=True): async def main(): - print("ggml_vulkan: Generating and compiling shaders to SPIR-V") + logger.info("ggml_vulkan: Generating and compiling shaders to SPIR-V") tasks = [] @@ -2768,9 +2771,12 @@ if __name__ == "__main__": parser = argparse.ArgumentParser(description="GGML Vulkan Shader Generator") parser.add_argument("--glslc", help="Path to glslc") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") args = parser.parse_args() + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + if args.glslc: GLSLC = args.glslc diff --git a/gguf-py/examples/reader.py b/gguf-py/examples/reader.py index 62e0769da..d841048c6 100644 --- a/gguf-py/examples/reader.py +++ b/gguf-py/examples/reader.py @@ -1,8 +1,10 @@ #!/usr/bin/env python3 +import logging import sys from pathlib import Path from gguf.gguf_reader import GGUFReader +logger = logging.getLogger("reader") sys.path.insert(0, str(Path(__file__).parent.parent)) @@ -18,28 +20,28 @@ def read_gguf_file(gguf_file_path): reader = GGUFReader(gguf_file_path) # List all key-value pairs in a columnized format - print("Key-Value Pairs:") + print("Key-Value Pairs:") # noqa: NP100 max_key_length = max(len(key) for key in reader.fields.keys()) for key, field in reader.fields.items(): value = field.parts[field.data[0]] - print(f"{key:{max_key_length}} : {value}") - print("----") + print(f"{key:{max_key_length}} : {value}") # noqa: NP100 + print("----") # noqa: NP100 # List all tensors - print("Tensors:") + print("Tensors:") # noqa: NP100 tensor_info_format = "{:<30} | Shape: {:<15} | Size: {:<12} | Quantization: {}" - print(tensor_info_format.format("Tensor Name", "Shape", "Size", "Quantization")) - print("-" * 80) + print(tensor_info_format.format("Tensor Name", "Shape", "Size", "Quantization")) # noqa: NP100 + print("-" * 80) # noqa: NP100 for tensor in reader.tensors: shape_str = "x".join(map(str, tensor.shape)) size_str = str(tensor.n_elements) quantization_str = tensor.tensor_type.name - print(tensor_info_format.format(tensor.name, shape_str, size_str, quantization_str)) + print(tensor_info_format.format(tensor.name, shape_str, size_str, quantization_str)) # noqa: NP100 if __name__ == '__main__': if len(sys.argv) < 2: - print("Usage: reader.py ") + logger.info("Usage: reader.py ") sys.exit(1) gguf_file_path = sys.argv[1] read_gguf_file(gguf_file_path) diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 6d597bfd9..4f232e18d 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -1,6 +1,5 @@ from __future__ import annotations -import sys from enum import Enum, IntEnum, auto from typing import Any @@ -854,8 +853,7 @@ class GGUFValueType(IntEnum): return GGUFValueType.INT32 # TODO: need help with 64-bit types in Python else: - print("Unknown type:", type(val)) - sys.exit() + raise ValueError(f"Unknown type: {type(val)}") # Note: Does not support GGML_QKK_64 diff --git a/gguf-py/gguf/gguf_reader.py b/gguf-py/gguf/gguf_reader.py index cc2abab34..db8525d85 100644 --- a/gguf-py/gguf/gguf_reader.py +++ b/gguf-py/gguf/gguf_reader.py @@ -4,6 +4,7 @@ # from __future__ import annotations +import logging import os from collections import OrderedDict from typing import Any, Literal, NamedTuple, TypeVar, Union @@ -27,6 +28,7 @@ from gguf.constants import ( GGUFValueType, ) +logger = logging.getLogger(__name__) READER_SUPPORTED_VERSIONS = [2, GGUF_VERSION] @@ -142,7 +144,7 @@ class GGUFReader: # TODO: add option to generate error on duplicate keys # raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}') - print(f'Warning: Duplicate key {field.name} at offset {field.offset}') + logger.warning(f'Duplicate key {field.name} at offset {field.offset}') self.fields[field.name + '_{}'.format(field.offset)] = field else: self.fields[field.name] = field @@ -239,8 +241,14 @@ class GGUFReader: def _build_tensors(self, start_offs: int, fields: list[ReaderField]) -> None: tensors = [] + tensor_names = set() # keep track of name to prevent duplicated tensors for field in fields: _name_len, name_data, _n_dims, dims, raw_dtype, offset_tensor = field.parts + # check if there's any tensor having same name already in the list + tensor_name = str(bytes(name_data), encoding = 'utf-8') + if tensor_name in tensor_names: + raise ValueError(f'Found duplicated tensor with name {tensor_name}') + tensor_names.add(tensor_name) ggml_type = GGMLQuantizationType(raw_dtype[0]) n_elems = np.prod(dims) block_size, type_size = GGML_QUANT_SIZES[ggml_type] @@ -272,7 +280,7 @@ class GGUFReader: item_count = n_bytes item_type = np.uint8 tensors.append(ReaderTensor( - name = str(bytes(name_data), encoding = 'utf-8'), + name = tensor_name, tensor_type = ggml_type, shape = dims, n_elements = n_elems, diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 30a55819f..d9cfbf711 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -1,5 +1,6 @@ from __future__ import annotations +import logging import os import shutil import struct @@ -24,6 +25,8 @@ from .constants import ( TokenType, ) +logger = logging.getLogger(__name__) + class WriterState(Enum): EMPTY = auto() @@ -63,10 +66,11 @@ class GGUFWriter: self.kv_data_count = 0 self.ti_data = bytearray() self.ti_data_count = 0 + self.ti_names = set() self.use_temp_file = use_temp_file self.temp_file = None self.tensors = [] - print("gguf: This GGUF file is for {0} Endian only".format( + logger.info("gguf: This GGUF file is for {0} Endian only".format( "Big" if self.endianess == GGUFEndian.BIG else "Little", )) self.state = WriterState.EMPTY @@ -197,6 +201,10 @@ class GGUFWriter: if self.state is not WriterState.EMPTY: raise ValueError(f'Expected output file to be empty, got {self.state}') + if name in self.ti_names: + raise ValueError(f'Duplicated tensor name {name}') + self.ti_names.add(name) + encoded_name = name.encode("utf8") self.ti_data += self._pack("Q", len(encoded_name)) self.ti_data += encoded_name diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index 378eaecad..c97a78f39 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -1,13 +1,15 @@ from __future__ import annotations +import logging import json import os -import sys from pathlib import Path from typing import Any, Callable from .gguf_writer import GGUFWriter +logger = logging.getLogger(__name__) + class SpecialVocab: merges: list[str] @@ -40,38 +42,29 @@ class SpecialVocab: def add_to_gguf(self, gw: GGUFWriter, quiet: bool = False) -> None: if self.merges: if not quiet: - print(f'gguf: Adding {len(self.merges)} merge(s).') + logger.info(f'Adding {len(self.merges)} merge(s).') gw.add_token_merges(self.merges) elif self.load_merges: - print( - 'gguf: WARNING: Adding merges requested but no merges found, output may be non-functional.', - file = sys.stderr, - ) + logger.warning('Adding merges requested but no merges found, output may be non-functional.') for typ, tokid in self.special_token_ids.items(): id_handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None) if id_handler is None: - print( - f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping', - file = sys.stderr, - ) + logger.warning(f'No handler for special token type {typ} with id {tokid} - skipping') continue if not quiet: - print(f'gguf: Setting special token type {typ} to {tokid}') + logger.info(f'Setting special token type {typ} to {tokid}') id_handler(tokid) for typ, value in self.add_special_token.items(): add_handler: Callable[[bool], None] | None = getattr(gw, f'add_add_{typ}_token', None) if add_handler is None: - print( - f'gguf: WARNING: No handler for add_{typ}_token with value {value} - skipping', - file = sys.stderr, - ) + logger.warning(f'No handler for add_{typ}_token with value {value} - skipping') continue if not quiet: - print(f'gguf: Setting add_{typ}_token to {value}') + logger.info(f'Setting add_{typ}_token to {value}') add_handler(value) if self.chat_template is not None: if not quiet: - print(f'gguf: Setting chat_template to {self.chat_template}') + logger.info(f'Setting chat_template to {self.chat_template}') gw.add_chat_template(self.chat_template) def _load(self, path: Path) -> None: @@ -99,10 +92,7 @@ class SpecialVocab: continue parts = line.split(None, 3) if len(parts) != 2: - print( - f'gguf: WARNING: {merges_file.name}: Line {line_num}: Entry malformed, ignoring', - file = sys.stderr, - ) + logger.warning(f'{merges_file.name}: Line {line_num}: Entry malformed, ignoring') continue merges.append(f'{parts[0]} {parts[1]}') self.merges = merges @@ -118,10 +108,7 @@ class SpecialVocab: return self.special_token_ids[typ] = tid return - print( - f'gguf: WARNING: Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping', - file = sys.stderr, - ) + logger.warning(f'Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping') def _try_load_from_tokenizer_json(self, path: Path) -> bool: tokenizer_file = path / 'tokenizer.json' @@ -144,10 +131,7 @@ class SpecialVocab: if chat_template is None or isinstance(chat_template, (str, list)): self.chat_template = chat_template else: - print( - f'gguf: WARNING: Bad type for chat_template field in {tokenizer_config_file!r} - ignoring', - file = sys.stderr - ) + logger.warning(f'Bad type for chat_template field in {tokenizer_config_file!r} - ignoring') for typ in self.special_token_types: add_entry = tokenizer_config.get(f'add_{typ}_token') if isinstance(add_entry, bool): diff --git a/gguf-py/scripts/gguf-convert-endian.py b/gguf-py/scripts/gguf-convert-endian.py index 10a16ad06..b698af0fe 100755 --- a/gguf-py/scripts/gguf-convert-endian.py +++ b/gguf-py/scripts/gguf-convert-endian.py @@ -1,9 +1,11 @@ #!/usr/bin/env python3 from __future__ import annotations +import logging import argparse import os import sys +from tqdm import tqdm from pathlib import Path import numpy as np @@ -14,6 +16,8 @@ if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / import gguf +logger = logging.getLogger("gguf-convert-endian") + def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None: if np.uint32(1) == np.uint32(1).newbyteorder("<"): @@ -29,11 +33,11 @@ def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None else: file_endian = host_endian order = host_endian if args.order == "native" else args.order - print(f"* Host is {host_endian.upper()} endian, GGUF file seems to be {file_endian.upper()} endian") + logger.info(f"* Host is {host_endian.upper()} endian, GGUF file seems to be {file_endian.upper()} endian") if file_endian == order: - print(f"* File is already {order.upper()} endian. Nothing to do.") + logger.info(f"* File is already {order.upper()} endian. Nothing to do.") sys.exit(0) - print("* Checking tensors for conversion compatibility") + logger.info("* Checking tensors for conversion compatibility") for tensor in reader.tensors: if tensor.tensor_type not in ( gguf.GGMLQuantizationType.F32, @@ -41,51 +45,64 @@ def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None gguf.GGMLQuantizationType.Q8_0, ): raise ValueError(f"Cannot handle type {tensor.tensor_type.name} for tensor {repr(tensor.name)}") - print(f"* Preparing to convert from {file_endian.upper()} to {order.upper()}") + logger.info(f"* Preparing to convert from {file_endian.upper()} to {order.upper()}") if args.dry_run: return - print("\n*** Warning *** Warning *** Warning **") - print("* This conversion process may damage the file. Ensure you have a backup.") + logger.warning("*** Warning *** Warning *** Warning **") + logger.warning("* This conversion process may damage the file. Ensure you have a backup.") if order != host_endian: - print("* Requested endian differs from host, you will not be able to load the model on this machine.") - print("* The file will be modified immediately, so if conversion fails or is interrupted") - print("* the file will be corrupted. Enter exactly YES if you are positive you want to proceed:") + logger.warning("* Requested endian differs from host, you will not be able to load the model on this machine.") + logger.warning("* The file will be modified immediately, so if conversion fails or is interrupted") + logger.warning("* the file will be corrupted. Enter exactly YES if you are positive you want to proceed:") response = input("YES, I am sure> ") if response != "YES": - print("You didn't enter YES. Okay then, see ya!") + logger.warning("You didn't enter YES. Okay then, see ya!") sys.exit(0) - print(f"\n* Converting fields ({len(reader.fields)})") + logger.info(f"* Converting fields ({len(reader.fields)})") for idx, field in enumerate(reader.fields.values()): - print(f"- {idx:4}: Converting field {repr(field.name)}, part count: {len(field.parts)}") + logger.info(f"- {idx:4}: Converting field {repr(field.name)}, part count: {len(field.parts)}") for part in field.parts: part.byteswap(inplace=True) - print(f"\n* Converting tensors ({len(reader.tensors)})") - for idx, tensor in enumerate(reader.tensors): - print( - f" - {idx:4}: Converting tensor {repr(tensor.name)}, type={tensor.tensor_type.name}, " - f"elements={tensor.n_elements}... ", - end="", + logger.info(f"* Converting tensors ({len(reader.tensors)})") + + for idx, tensor in enumerate(pbar := tqdm(reader.tensors, desc="Converting tensor")): + log_message = ( + f"Converting tensor {repr(tensor.name)}, " + f"type={tensor.tensor_type.name}, " + f"elements={tensor.n_elements} " ) - tensor_type = tensor.tensor_type + + # Byte-swap each part of the tensor's field for part in tensor.field.parts: part.byteswap(inplace=True) - if tensor_type != gguf.GGMLQuantizationType.Q8_0: + + # Byte-swap tensor data if necessary + if tensor.tensor_type == gguf.GGMLQuantizationType.Q8_0: + # Handle Q8_0 tensor blocks (block_q8_0) + # Specific handling of block_q8_0 is required. + # Each block_q8_0 consists of an f16 delta (scaling factor) followed by 32 int8 quantizations. + + block_size = 34 # 34 bytes = + 32 * + + n_blocks = len(tensor.data) // block_size + for block_num in (inner_pbar := tqdm(range(n_blocks), desc="Byte-swapping Blocks", leave=False)): + block_offs = block_num * block_size + + # Byte-Swap f16 sized delta field + delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + # Byte-Swap Q8 weights + if block_num % 100000 == 0: + inner_pbar.set_description(f"Byte-swapping Blocks [{(n_blocks - block_num) // n_blocks}]") + + else: + # Handle other tensor types tensor.data.byteswap(inplace=True) - print() - continue - # A Q8_0 block consists of a f16 delta followed by 32 int8 quants, so 34 bytes - block_size = 34 - n_blocks = len(tensor.data) // block_size - for block_num in range(n_blocks): - block_offs = block_num * block_size - # I know I said f16, but it doesn't matter here - any simple 16 bit type works. - delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16) - delta.byteswap(inplace=True) - if block_num % 100000 == 0: - print(f"[{(n_blocks - block_num) // 1000}K]", end="") - sys.stdout.flush() - print() - print("* Completion") + + pbar.set_description(log_message) + + logger.info("* Completion") def main() -> None: @@ -102,8 +119,13 @@ def main() -> None: "--dry-run", action="store_true", help="Don't actually change anything", ) + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) - print(f'* Loading: {args.model}') + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + logger.info(f'* Loading: {args.model}') reader = gguf.GGUFReader(args.model, 'r' if args.dry_run else 'r+') convert_byteorder(reader, args) diff --git a/gguf-py/scripts/gguf-dump.py b/gguf-py/scripts/gguf-dump.py index dbf891508..2d3c3943f 100755 --- a/gguf-py/scripts/gguf-dump.py +++ b/gguf-py/scripts/gguf-dump.py @@ -1,6 +1,7 @@ #!/usr/bin/env python3 from __future__ import annotations +import logging import argparse import os import sys @@ -15,6 +16,8 @@ if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / from gguf import GGUFReader, GGUFValueType # noqa: E402 +logger = logging.getLogger("gguf-dump") + def get_file_host_endian(reader: GGUFReader) -> tuple[str, str]: host_endian = 'LITTLE' if np.uint32(1) == np.uint32(1).newbyteorder("<") else 'BIG' @@ -29,8 +32,8 @@ def get_file_host_endian(reader: GGUFReader) -> tuple[str, str]: # please see the comments in the modify_gguf.py example. def dump_metadata(reader: GGUFReader, args: argparse.Namespace) -> None: host_endian, file_endian = get_file_host_endian(reader) - print(f'* File is {file_endian} endian, script is running on a {host_endian} endian host.') - print(f'\n* Dumping {len(reader.fields)} key/value pair(s)') + print(f'* File is {file_endian} endian, script is running on a {host_endian} endian host.') # noqa: NP100 + print(f'* Dumping {len(reader.fields)} key/value pair(s)') # noqa: NP100 for n, field in enumerate(reader.fields.values(), 1): if not field.types: pretty_type = 'N/A' @@ -39,20 +42,21 @@ def dump_metadata(reader: GGUFReader, args: argparse.Namespace) -> None: pretty_type = '[' * nest_count + str(field.types[-1].name) + ']' * nest_count else: pretty_type = str(field.types[-1].name) - print(f' {n:5}: {pretty_type:10} | {len(field.data):8} | {field.name}', end = '') + + log_message = f' {n:5}: {pretty_type:10} | {len(field.data):8} | {field.name}' if len(field.types) == 1: curr_type = field.types[0] if curr_type == GGUFValueType.STRING: - print(' = {0}'.format(repr(str(bytes(field.parts[-1]), encoding='utf8')[:60])), end = '') + log_message += ' = {0}'.format(repr(str(bytes(field.parts[-1]), encoding='utf8')[:60])) elif field.types[0] in reader.gguf_scalar_to_np: - print(' = {0}'.format(field.parts[-1][0]), end = '') - print() + log_message += ' = {0}'.format(field.parts[-1][0]) + print(log_message) # noqa: NP100 if args.no_tensors: return - print(f'\n* Dumping {len(reader.tensors)} tensor(s)') + print(f'* Dumping {len(reader.tensors)} tensor(s)') # noqa: NP100 for n, tensor in enumerate(reader.tensors, 1): prettydims = ', '.join('{0:5}'.format(d) for d in list(tensor.shape) + [1] * (4 - len(tensor.shape))) - print(f' {n:5}: {tensor.n_elements:10} | {prettydims} | {tensor.tensor_type.name:7} | {tensor.name}') + print(f' {n:5}: {tensor.n_elements:10} | {prettydims} | {tensor.tensor_type.name:7} | {tensor.name}') # noqa: NP100 def dump_metadata_json(reader: GGUFReader, args: argparse.Namespace) -> None: @@ -103,10 +107,17 @@ def main() -> None: parser.add_argument("--no-tensors", action="store_true", help="Don't dump tensor metadata") parser.add_argument("--json", action="store_true", help="Produce JSON output") parser.add_argument("--json-array", action="store_true", help="Include full array values in JSON output (long)") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + if not args.json: - print(f'* Loading: {args.model}') + logger.info(f'* Loading: {args.model}') + reader = GGUFReader(args.model, 'r') + if args.json: dump_metadata_json(reader, args) else: diff --git a/gguf-py/scripts/gguf-set-metadata.py b/gguf-py/scripts/gguf-set-metadata.py index 3ebdfa898..e35b651b8 100755 --- a/gguf-py/scripts/gguf-set-metadata.py +++ b/gguf-py/scripts/gguf-set-metadata.py @@ -1,4 +1,5 @@ #!/usr/bin/env python3 +import logging import argparse import os import sys @@ -10,6 +11,8 @@ if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / from gguf import GGUFReader # noqa: E402 +logger = logging.getLogger("gguf-set-metadata") + def minimal_example(filename: str) -> None: reader = GGUFReader(filename, 'r+') @@ -41,36 +44,33 @@ def minimal_example(filename: str) -> None: def set_metadata(reader: GGUFReader, args: argparse.Namespace) -> None: field = reader.get_field(args.key) if field is None: - print(f'! Field {repr(args.key)} not found', file = sys.stderr) + logger.error(f'! Field {repr(args.key)} not found') sys.exit(1) # Note that field.types is a list of types. This is because the GGUF # format supports arrays. For example, an array of UINT32 would # look like [GGUFValueType.ARRAY, GGUFValueType.UINT32] handler = reader.gguf_scalar_to_np.get(field.types[0]) if field.types else None if handler is None: - print( - f'! This tool only supports changing simple values, {repr(args.key)} has unsupported type {field.types}', - file = sys.stderr, - ) + logger.error(f'! This tool only supports changing simple values, {repr(args.key)} has unsupported type {field.types}') sys.exit(1) current_value = field.parts[field.data[0]][0] new_value = handler(args.value) - print(f'* Preparing to change field {repr(args.key)} from {current_value} to {new_value}') + logger.info(f'* Preparing to change field {repr(args.key)} from {current_value} to {new_value}') if current_value == new_value: - print(f'- Key {repr(args.key)} already set to requested value {current_value}') + logger.info(f'- Key {repr(args.key)} already set to requested value {current_value}') sys.exit(0) if args.dry_run: sys.exit(0) if not args.force: - print('*** Warning *** Warning *** Warning **') - print('* Changing fields in a GGUF file can make it unusable. Proceed at your own risk.') - print('* Enter exactly YES if you are positive you want to proceed:') + logger.warning('*** Warning *** Warning *** Warning **') + logger.warning('* Changing fields in a GGUF file can make it unusable. Proceed at your own risk.') + logger.warning('* Enter exactly YES if you are positive you want to proceed:') response = input('YES, I am sure> ') if response != 'YES': - print("You didn't enter YES. Okay then, see ya!") + logger.info("You didn't enter YES. Okay then, see ya!") sys.exit(0) field.parts[field.data[0]][0] = new_value - print('* Field changed. Successful completion.') + logger.info('* Field changed. Successful completion.') def main() -> None: @@ -80,8 +80,13 @@ def main() -> None: parser.add_argument("value", type=str, help="Metadata value to set") parser.add_argument("--dry-run", action="store_true", help="Don't actually change anything") parser.add_argument("--force", action="store_true", help="Change the field without confirmation") + parser.add_argument("--verbose", action="store_true", help="increase output verbosity") + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) - print(f'* Loading: {args.model}') + + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + logger.info(f'* Loading: {args.model}') reader = GGUFReader(args.model, 'r' if args.dry_run else 'r+') set_metadata(reader, args) diff --git a/llama.cpp b/llama.cpp index 820ce9d07..7c7a06df3 100644 --- a/llama.cpp +++ b/llama.cpp @@ -108,7 +108,6 @@ #define LLAMA_MAX_NODES 8192 #define LLAMA_MAX_EXPERTS 60 - // // logging // @@ -1846,7 +1845,7 @@ struct llama_hparams { float f_logit_scale = 0.0f; bool causal_attn = true; - bool need_kq_pos = false; + bool use_alibi = false; // currently, we need KQ_pos data for ALiBi-based models enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_NONE; enum llama_rope_type rope_type = LLAMA_ROPE_TYPE_NONE; @@ -1936,6 +1935,7 @@ struct llama_cparams { bool embeddings; bool causal_attn; bool offload_kqv; + bool flash_attn; enum llama_pooling_type pooling_type; @@ -2039,8 +2039,8 @@ struct llama_kv_cache { bool has_shift = false; bool do_defrag = false; bool do_copy = false; - // with recurrent state models, a cell can hold the state for more than one past token - bool recurrent = false; + bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token + bool v_trans = true; // the value tensor is transposed // Note: The value of head isn't only used to optimize searching // for a free KV slot. llama_decode_internal also uses it, so it @@ -2339,11 +2339,14 @@ struct llama_context { static bool llama_kv_cache_init( struct llama_kv_cache & cache, - const llama_model & model, + const llama_context * ctx, ggml_type type_k, ggml_type type_v, uint32_t kv_size, bool offload) { + const llama_model & model = ctx->model; + const llama_cparams & cparams = ctx->cparams; + const struct llama_hparams & hparams = model.hparams; const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); @@ -2354,8 +2357,9 @@ static bool llama_kv_cache_init( // TODO: find a nicer way to add other recurrent model architectures cache.recurrent = model.arch == LLM_ARCH_MAMBA; + cache.v_trans = !cparams.flash_attn; - // TODO: support mixed reccurent Transformer architectues + // TODO: support mixed recurrent Transformer architectures // NOTE: (!a || b) is a logical implication (a -> b) GGML_ASSERT(!cache.recurrent || n_embd_k_gqa == hparams.n_embd_k_s()); GGML_ASSERT(!cache.recurrent || n_embd_v_gqa == hparams.n_embd_v_s()); @@ -2566,6 +2570,10 @@ static void llama_kv_cache_clear(struct llama_kv_cache & cache) { } cache.head = 0; cache.used = 0; + + for (auto & buf : cache.bufs) { + ggml_backend_buffer_clear(buf, 0); + } } static bool llama_kv_cache_seq_rm( @@ -3123,9 +3131,17 @@ struct llama_model_loader { fver = (enum llama_fver) gguf_get_version(meta); + std::set tensor_names; for (auto & w : weights) { n_elements += ggml_nelements(w.tensor); n_bytes += ggml_nbytes(w.tensor); + // make sure there is no duplicated tensor names + const std::string name(w.tensor->name); + auto found = tensor_names.find(name); + if (found != tensor_names.end()) { + throw std::runtime_error(format("invalid model: tensor '%s' is duplicated", w.tensor->name)); + } + tensor_names.insert(name); } LLAMA_LOG_INFO("%s: loaded meta data with %d key-value pairs and %d tensors from %s (version %s)\n", @@ -4186,7 +4202,7 @@ static void llm_load_hparams( model.ftype = ml.ftype; if (hparams.f_max_alibi_bias > 0.0f) { - hparams.need_kq_pos = true; + hparams.use_alibi = true; } hparams.rope_type = llama_rope_type(&model); @@ -4367,6 +4383,9 @@ static void llm_load_vocab( } else if ( tokenizer_pre == "gpt-2") { vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT2; + } else if ( + tokenizer_pre == "refact") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_REFACT; } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); } @@ -6195,37 +6214,47 @@ static struct ggml_tensor * llm_build_inp_embd( static void llm_build_kv_store( struct ggml_context * ctx, const llama_hparams & hparams, + const llama_cparams & cparams, const llama_kv_cache & kv, struct ggml_cgraph * graph, struct ggml_tensor * k_cur, struct ggml_tensor * v_cur, - int64_t n_ctx, int32_t n_tokens, int32_t kv_head, const llm_build_cb & cb, int64_t il) { + const int64_t n_ctx = cparams.n_ctx; + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); GGML_ASSERT(kv.size == n_ctx); - // compute the transposed [n_tokens, n_embd] V matrix - assert(v_cur->ne[0] == n_embd_v_gqa && v_cur->ne[1] == n_tokens); - struct ggml_tensor * v_cur_t = ggml_transpose(ctx, v_cur); - cb(v_cur_t, "v_cur_t", il); - struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k_l[il], n_tokens*n_embd_k_gqa, (ggml_row_size(kv.k_l[il]->type, n_embd_k_gqa))*kv_head); cb(k_cache_view, "k_cache_view", il); - struct ggml_tensor * v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_v_gqa, - ( n_ctx)*ggml_element_size(kv.v_l[il]), - (kv_head)*ggml_element_size(kv.v_l[il])); + // note: storing RoPE-ed version of K in the KV cache + ggml_build_forward_expand(graph, ggml_cpy(ctx, k_cur, k_cache_view)); + + assert(v_cur->ne[0] == n_embd_v_gqa && v_cur->ne[1] == n_tokens); + + struct ggml_tensor * v_cache_view = nullptr; + + if (cparams.flash_attn) { + v_cache_view = ggml_view_1d(ctx, kv.v_l[il], n_tokens*n_embd_v_gqa, + (kv_head)*ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa)); + } else { + // note: the V cache is transposed when not using flash attention + v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_v_gqa, + ( n_ctx)*ggml_element_size(kv.v_l[il]), + (kv_head)*ggml_element_size(kv.v_l[il])); + + v_cur = ggml_transpose(ctx, v_cur); + } cb(v_cache_view, "v_cache_view", il); - // important: storing RoPE-ed version of K in the KV cache! - ggml_build_forward_expand(graph, ggml_cpy(ctx, k_cur, k_cache_view)); - ggml_build_forward_expand(graph, ggml_cpy(ctx, v_cur_t, v_cache_view)); + ggml_build_forward_expand(graph, ggml_cpy(ctx, v_cur, v_cache_view)); } static struct ggml_tensor * llm_build_norm( @@ -6445,11 +6474,11 @@ static struct ggml_tensor * llm_build_moe_ffn( return moe_out; } -// if max_alibi_bias > 0 then apply ALiBi static struct ggml_tensor * llm_build_kqv( struct ggml_context * ctx, const llama_model & model, const llama_hparams & hparams, + const llama_cparams & cparams, const llama_kv_cache & kv, struct ggml_cgraph * graph, struct ggml_tensor * wo, @@ -6457,12 +6486,12 @@ static struct ggml_tensor * llm_build_kqv( struct ggml_tensor * q_cur, struct ggml_tensor * kq_mask, struct ggml_tensor * kq_pos, - int64_t n_ctx, int32_t n_tokens, int32_t n_kv, float kq_scale, const llm_build_cb & cb, int il) { + const int64_t n_ctx = cparams.n_ctx; const int64_t n_head = hparams.n_head; const int64_t n_head_kv = hparams.n_head_kv; const int64_t n_embd_head_k = hparams.n_embd_head_k; @@ -6480,72 +6509,100 @@ static struct ggml_tensor * llm_build_kqv( 0); cb(k, "k", il); - struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q); - cb(kq, "kq", il); + struct ggml_tensor * cur; - if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) { - // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs - // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847 - ggml_mul_mat_set_prec(kq, GGML_PREC_F32); - } + if (cparams.flash_attn) { + GGML_UNUSED(model); + GGML_UNUSED(n_ctx); - if (model.arch == LLM_ARCH_GROK) { - // need to do the following: - // multiply by attn_output_multiplyer of 0.08838834764831845 - // and then : - // kq = 30 * tanh(kq / 30) - // before the softmax below + // note: if this assert triggers, then some check has failed earlier + // the idea is to detect during context creation that ALiBi would be used and disable Flash Attention + GGML_ASSERT(kq_pos == nullptr && "ALiBi is not yet supported with Flash Attention"); - //try from phi2 - //ggml_mul_mat_set_prec(kq, GGML_PREC_F32); + // split cached v into n_head heads (not transposed) + struct ggml_tensor * v = + ggml_view_3d(ctx, kv.v_l[il], + n_embd_head_v, n_kv, n_head_kv, + ggml_row_size(kv.v_l[il]->type, n_embd_k_gqa), + ggml_row_size(kv.v_l[il]->type, n_embd_head_k), + 0); + cb(v, "v", il); - kq = ggml_tanh(ctx, ggml_scale(ctx, kq, 0.08838834764831845f/30.0f)); - kq = ggml_scale(ctx, kq, 30); - } + cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale); + + if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) { + ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32); + } + + cur = ggml_reshape_2d(ctx, cur, n_embd_head_k*n_head, n_tokens); + } else { + struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q); + cb(kq, "kq", il); + + if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) { + // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs + // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847 + ggml_mul_mat_set_prec(kq, GGML_PREC_F32); + } + + if (model.arch == LLM_ARCH_GROK) { + // need to do the following: + // multiply by attn_output_multiplyer of 0.08838834764831845 + // and then : + // kq = 30 * tanh(kq / 30) + // before the softmax below + + //try from phi2 + //ggml_mul_mat_set_prec(kq, GGML_PREC_F32); + + kq = ggml_tanh(ctx, ggml_scale(ctx, kq, 0.08838834764831845f/30.0f)); + kq = ggml_scale(ctx, kq, 30); + } #if defined(GGML_USE_KOMPUTE) #pragma message("TODO: ALiBi support in ggml_soft_max_ext is not implemented for Kompute") #pragma message(" Falling back to ggml_alibi(). Will become an error in Mar 2024") #pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5488") - if (hparams.f_max_alibi_bias > 0.0f) { - kq = ggml_scale(ctx, kq, kq_scale); - cb(kq, "kq_scaled", il); + if (hparams.use_alibi) { + kq = ggml_scale(ctx, kq, kq_scale); + cb(kq, "kq_scaled", il); - kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, hparams.f_max_alibi_bias); - cb(kq, "kq_scaled_alibi", il); + kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, hparams.f_max_alibi_bias); + cb(kq, "kq_scaled_alibi", il); - kq = ggml_add(ctx, kq, kq_mask); - cb(kq, "kq_masked", il); + kq = ggml_add(ctx, kq, kq_mask); + cb(kq, "kq_masked", il); - kq = ggml_soft_max(ctx, kq); - cb(kq, "kq_soft_max", il); - } else + kq = ggml_soft_max(ctx, kq); + cb(kq, "kq_soft_max", il); + } else #endif - { - kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_pos, kq_scale, hparams.f_max_alibi_bias); - cb(kq, "kq_soft_max_ext", il); + { + kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_pos, kq_scale, hparams.f_max_alibi_bias); + cb(kq, "kq_soft_max_ext", il); + } + + GGML_ASSERT(kv.size == n_ctx); + + // split cached v into n_head heads + struct ggml_tensor * v = + ggml_view_3d(ctx, kv.v_l[il], + n_kv, n_embd_head_v, n_head_kv, + ggml_element_size(kv.v_l[il])*n_ctx, + ggml_element_size(kv.v_l[il])*n_ctx*n_embd_head_v, + 0); + cb(v, "v", il); + + struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq); + cb(kqv, "kqv", il); + + struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_k*n_head, n_tokens); + cb(cur, "kqv_merged_cont", il); } - GGML_ASSERT(kv.size == n_ctx); - - // split cached v into n_head heads - struct ggml_tensor * v = - ggml_view_3d(ctx, kv.v_l[il], - n_kv, n_embd_head_v, n_head_kv, - ggml_element_size(kv.v_l[il])*n_ctx, - ggml_element_size(kv.v_l[il])*n_ctx*n_embd_head_v, - 0); - cb(v, "v", il); - - struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq); - cb(kqv, "kqv", il); - - struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3); - cb(kqv_merged, "kqv_merged", il); - - struct ggml_tensor * cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_k*n_head, n_tokens); - cb(cur, "kqv_merged_cont", il); - ggml_build_forward_expand(graph, cur); cur = ggml_mul_mat(ctx, wo, cur); @@ -6564,6 +6621,7 @@ static struct ggml_tensor * llm_build_kv( struct ggml_context * ctx, const llama_model & model, const llama_hparams & hparams, + const llama_cparams & cparams, const llama_kv_cache & kv, struct ggml_cgraph * graph, struct ggml_tensor * wo, @@ -6573,7 +6631,6 @@ static struct ggml_tensor * llm_build_kv( struct ggml_tensor * q_cur, struct ggml_tensor * kq_mask, struct ggml_tensor * kq_pos, - int64_t n_ctx, int32_t n_tokens, int32_t kv_head, int32_t n_kv, @@ -6587,12 +6644,12 @@ static struct ggml_tensor * llm_build_kv( ggml_build_forward_expand(graph, k_cur); ggml_build_forward_expand(graph, v_cur); - llm_build_kv_store(ctx, hparams, kv, graph, k_cur, v_cur, n_ctx, n_tokens, kv_head, cb, il); + llm_build_kv_store(ctx, hparams, cparams, kv, graph, k_cur, v_cur, n_tokens, kv_head, cb, il); struct ggml_tensor * cur; - cur = llm_build_kqv(ctx, model, hparams, kv, graph, wo, wo_b, - q_cur, kq_mask, kq_pos, n_ctx, n_tokens, n_kv, kq_scale, cb, il); + cur = llm_build_kqv(ctx, model, hparams, cparams, kv, graph, wo, wo_b, + q_cur, kq_mask, kq_pos, n_tokens, n_kv, kq_scale, cb, il); cb(cur, "kqv_out", il); return cur; @@ -6634,6 +6691,8 @@ struct llm_build_context { const int32_t kv_head; // index of where we store new KV data in the cache const int32_t n_orig_ctx; + const bool flash_attn; + const enum llama_pooling_type pooling_type; const enum llama_rope_type rope_type; @@ -6680,6 +6739,7 @@ struct llm_build_context { n_outputs (worst_case ? n_tokens : lctx.n_outputs), kv_head (worst_case ? (kv_self.recurrent ? 0 : kv_self.size - n_tokens) : kv_self.head), n_orig_ctx (cparams.n_yarn_orig_ctx), + flash_attn (cparams.flash_attn), pooling_type (cparams.pooling_type), rope_type (hparams.rope_type), cb (cb), @@ -6794,15 +6854,31 @@ struct llm_build_context { ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*id)); - ggml_tensor * view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il], - nm, n_embd_v_gqa, - ggml_row_size(kv_self.v_l[il]->type, kv_self.size), - ggml_row_size(kv_self.v_l[il]->type, i)); + ggml_tensor * view_v_src; + ggml_tensor * view_v_dst; - ggml_tensor * view_v_dst = ggml_view_2d(ctx0, kv_self.v_l[il], - nm, n_embd_v_gqa, - ggml_row_size(kv_self.v_l[il]->type, kv_self.size), - ggml_row_size(kv_self.v_l[il]->type, id)); + if (flash_attn) { + // NOTE: the V cache is not transposed when using flash attention + view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il], + n_embd_v_gqa, nm, + ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa), + ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*i)); + + view_v_dst = ggml_view_2d(ctx0, kv_self.v_l[il], + n_embd_v_gqa, nm, + ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa), + ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*id)); + } else { + view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il], + nm, n_embd_v_gqa, + ggml_row_size(kv_self.v_l[il]->type, kv_self.size), + ggml_row_size(kv_self.v_l[il]->type, i)); + + view_v_dst = ggml_view_2d(ctx0, kv_self.v_l[il], + nm, n_embd_v_gqa, + ggml_row_size(kv_self.v_l[il]->type, kv_self.size), + ggml_row_size(kv_self.v_l[il]->type, id)); + } ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_k_src, view_k_dst)); ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_v_src, view_v_dst)); @@ -6832,20 +6908,26 @@ struct llm_build_context { struct ggml_tensor * build_inp_KQ_mask(bool causal = true) { if (causal) { - lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, n_tokens); + lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); } else { - lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens); + lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); } cb(lctx.inp_KQ_mask, "KQ_mask", -1); ggml_set_input(lctx.inp_KQ_mask); - return lctx.inp_KQ_mask; + return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask, GGML_TYPE_F16) : lctx.inp_KQ_mask; } - struct ggml_tensor * build_inp_KQ_pos() { - lctx.inp_KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_kv); + struct ggml_tensor * build_inp_KQ_pos(bool causal = true) { + if (causal) { + lctx.inp_KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_kv); + } else { + // TODO: this will be needed for ALiBi-based BERT models + // https://github.com/ggerganov/llama.cpp/pull/6826 + lctx.inp_KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_tokens); + } cb(lctx.inp_KQ_pos, "KQ_pos", -1); ggml_set_input(lctx.inp_KQ_pos); - return lctx.inp_KQ_pos; + return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_pos, GGML_TYPE_F16) : lctx.inp_KQ_pos; } struct ggml_tensor * build_inp_mean() { @@ -6951,9 +7033,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -7091,9 +7173,9 @@ struct llm_build_context { cb(Qcur, "Qcur", il); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -7198,9 +7280,9 @@ struct llm_build_context { ext_factor, attn_factor, beta_fast, beta_slow ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -7318,9 +7400,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -7443,9 +7525,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f, cb, il); } if (il == n_layer - 1) { @@ -7595,9 +7677,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, NULL, + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -7707,9 +7789,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -7911,9 +7993,9 @@ struct llm_build_context { ); cb(Vcur, "Vcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Q, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Q, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -8007,9 +8089,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); cb(Qcur, "Qcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -8300,9 +8382,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -8431,14 +8513,15 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, - model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } else { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } } @@ -8580,9 +8663,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -8698,9 +8781,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -8811,9 +8894,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -8925,9 +9008,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -9080,9 +9163,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f, cb, il); } if (il == n_layer - 1) { @@ -9197,9 +9280,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, - model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il); + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f, cb, il); } if (il == n_layer - 1) { @@ -9310,9 +9393,9 @@ struct llm_build_context { ext_factor, attn_factor, beta_fast, beta_slow); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } struct ggml_tensor * sa_out = cur; @@ -9413,9 +9496,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -9520,9 +9603,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -9636,9 +9719,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -9753,9 +9836,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -9883,9 +9966,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -10004,9 +10087,9 @@ struct llm_build_context { ext_factor, attn_factor, beta_fast, beta_slow); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, NULL, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f, cb, il); } if (il == n_layer - 1) { @@ -10123,9 +10206,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -10413,9 +10496,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -10544,9 +10627,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, model.layers[il].wo, nullptr, - Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); } if (il == n_layer - 1) { @@ -10973,7 +11056,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } - if (hparams.need_kq_pos) { + // ALiBi requires the KQ_pos tensor to provide the sequence position of each token in the batch + // this allows to process multiple sequences in parallel with ALiBi-based models + if (hparams.use_alibi) { const int64_t n_kv = kv_self.n; GGML_ASSERT(lctx.inp_KQ_pos); @@ -11355,7 +11440,7 @@ static int llama_decode_internal( // a heuristic, to avoid attending the full cache if it is not yet utilized // after enough generations, the benefit from this heuristic disappears // if we start defragmenting the cache, the benefit from this will be more important - kv_self.n = std::min(kv_self.size, std::max(32u, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32))); + kv_self.n = std::min(kv_self.size, std::max(256u, GGML_PAD(llama_kv_cache_cell_max(kv_self), 256))); //kv_self.n = llama_kv_cache_cell_max(kv_self); } } @@ -11552,7 +11637,9 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { // each move requires 6*n_layer tensors (see build_defrag) // - source view, destination view, copy operation // - x2 for keys and values - const uint32_t max_moves = LLAMA_MAX_NODES/(6*n_layer); + //const uint32_t max_moves = LLAMA_MAX_NODES/(6*n_layer); + // TODO: tmp fix https://github.com/ggerganov/llama.cpp/issues/6685#issuecomment-2057579516 + const uint32_t max_moves = (LLAMA_MAX_NODES - 2*n_layer)/(6*n_layer); // determine which KV cells to move where // @@ -11868,7 +11955,7 @@ static bool llama_is_user_defined_token(const llama_vocab& vocab, llama_token id static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) { GGML_ASSERT(llama_vocab_get_type(vocab) != LLAMA_VOCAB_TYPE_NONE); GGML_ASSERT(llama_is_byte_token(vocab, id)); - const auto& token_data = vocab.id_to_token.at(id); + const auto & token_data = vocab.id_to_token.at(id); switch (llama_vocab_get_type(vocab)) { case LLAMA_VOCAB_TYPE_SPM: { auto buf = token_data.text.substr(3, 2); @@ -12128,14 +12215,13 @@ struct llm_tokenizer_bpe { "\\s?\\p{L}+", "\\s?\\p{P}+", "[一-龥ࠀ-一가-퟿]+", - "\\p{N}+", + "\\p{N}", }); break; case LLAMA_VOCAB_PRE_TYPE_FALCON: word_collection = unicode_regex_split(text, { "[\\p{P}\\$\\+<=>\\^~\\|]+", "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", - "\\p{N}+", "[0-9][0-9][0-9]", }); break; @@ -12151,6 +12237,12 @@ struct llm_tokenizer_bpe { }); break; case LLAMA_VOCAB_PRE_TYPE_STARCODER: + case LLAMA_VOCAB_PRE_TYPE_REFACT: + word_collection = unicode_regex_split(text, { + "\\p{N}", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + }); + break; case LLAMA_VOCAB_PRE_TYPE_GPT2: word_collection = unicode_regex_split(text, { "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", @@ -15159,6 +15251,7 @@ struct llama_context_params llama_context_default_params() { /*.logits_all =*/ false, /*.embeddings =*/ false, /*.offload_kqv =*/ true, + /*.flash_attn =*/ false, /*.abort_callback =*/ nullptr, /*.abort_callback_data =*/ nullptr, }; @@ -15325,6 +15418,7 @@ struct llama_context * llama_new_context_with_model( cparams.defrag_thold = params.defrag_thold; cparams.embeddings = params.embeddings; cparams.offload_kqv = params.offload_kqv; + cparams.flash_attn = params.flash_attn; cparams.pooling_type = params.pooling_type; cparams.n_ctx = params.n_ctx == 0 ? hparams.n_ctx_train : params.n_ctx; @@ -15332,12 +15426,20 @@ struct llama_context * llama_new_context_with_model( cparams.rope_freq_scale = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale; // this is necessary due to kv_self.n being padded later during inference - cparams.n_ctx = GGML_PAD(cparams.n_ctx, 32); + cparams.n_ctx = GGML_PAD(cparams.n_ctx, 256); // with causal attention, the batch size is limited by the context size cparams.n_batch = hparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch; - cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch); + // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask + // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext) + // ref: https://github.com/ggerganov/llama.cpp/pull/5021 + if (cparams.n_batch < GGML_KQ_MASK_PAD) { + LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD); + cparams.n_batch = GGML_KQ_MASK_PAD; + } + + cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch); cparams.n_yarn_orig_ctx = params.yarn_orig_ctx != 0 ? params.yarn_orig_ctx : hparams.n_yarn_orig_ctx != 0 ? hparams.n_yarn_orig_ctx : @@ -15369,6 +15471,23 @@ struct llama_context * llama_new_context_with_model( } } + if (cparams.flash_attn && hparams.use_alibi) { + LLAMA_LOG_WARN("%s: flash_attn is not yet compatible with ALiBi - forcing off\n", __func__); + cparams.flash_attn = false; + } + + if (cparams.flash_attn && model->arch == LLM_ARCH_GROK) { + LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__); + cparams.flash_attn = false; + } + +#ifdef GGML_USE_HIPBLAS + if (cparams.flash_attn) { + LLAMA_LOG_WARN("%s: flash_attn is not yet compatible with HIPBLAS builds - forcing off\n", __func__); + cparams.flash_attn = false; + } +#endif + if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); } @@ -15376,6 +15495,7 @@ struct llama_context * llama_new_context_with_model( LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, cparams.n_ctx); LLAMA_LOG_INFO("%s: n_batch = %u\n", __func__, cparams.n_batch); LLAMA_LOG_INFO("%s: n_ubatch = %u\n", __func__, cparams.n_ubatch); + LLAMA_LOG_INFO("%s: flash_attn = %d\n", __func__, cparams.flash_attn); LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base); LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale); @@ -15504,7 +15624,7 @@ struct llama_context * llama_new_context_with_model( } ctx->backends.push_back(ctx->backend_cpu); - if (!llama_kv_cache_init(ctx->kv_self, ctx->model, type_k, type_v, kv_size, cparams.offload_kqv)) { + if (!llama_kv_cache_init(ctx->kv_self, ctx, type_k, type_v, kv_size, cparams.offload_kqv)) { LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__); llama_free(ctx); return nullptr; @@ -16103,6 +16223,7 @@ size_t llama_state_get_size(const struct llama_context * ctx) { const size_t s_kv_head = sizeof(uint32_t); const size_t s_kv_size = sizeof(uint32_t); const size_t s_kv_used = sizeof(uint32_t); + const size_t s_v_trans = sizeof(uint32_t); const size_t s_kv = ctx->kv_self.total_size(); const size_t s_kv_cell = sizeof(llama_pos) + sizeof(size_t) + cparams.n_seq_max*sizeof(llama_seq_id); const size_t s_kv_cells = ctx->kv_self.size * s_kv_cell; @@ -16120,10 +16241,14 @@ size_t llama_state_get_size(const struct llama_context * ctx) { + s_kv_head + s_kv_size + s_kv_used + + s_v_trans + s_kv + s_kv_cells ); + // on session change it is very likely that the state size has changed - so we need to update this function + static_assert(LLAMA_SESSION_VERSION == 6, "So you just bumped the session version - good. But did you remember to update llama_state_get_size?"); + return s_total; } @@ -16269,11 +16394,13 @@ static void llama_state_get_data_internal(struct llama_context * ctx, llama_data const uint32_t kv_size = kv_self.size; const size_t kv_buf_size = kv_self.total_size() / (kv_size ? kv_size : 1) * kv_head; const uint32_t kv_used = kv_self.used; + const uint32_t v_trans = kv_self.v_trans ? 1 : 0; data_ctx->write(&kv_buf_size, sizeof(kv_buf_size)); data_ctx->write(&kv_head, sizeof(kv_head)); data_ctx->write(&kv_size, sizeof(kv_size)); data_ctx->write(&kv_used, sizeof(kv_used)); + data_ctx->write(&v_trans, sizeof(v_trans)); if (kv_buf_size) { const size_t pre_kv_buf_size = data_ctx->get_size_written(); @@ -16286,7 +16413,7 @@ static void llama_state_get_data_internal(struct llama_context * ctx, llama_data ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), 0, tmp_buf.size()); data_ctx->write(tmp_buf.data(), tmp_buf.size()); - if (kv_self.recurrent) { + if (kv_self.recurrent || !kv_self.v_trans) { // v is contiguous for recurrent models // TODO: use other tensors for state models than k and v const size_t v_size = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*kv_head); @@ -16419,11 +16546,15 @@ size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) { uint32_t kv_head; uint32_t kv_size; uint32_t kv_used; + uint32_t v_trans; memcpy(&kv_buf_size, inp, sizeof(kv_buf_size)); inp += sizeof(kv_buf_size); memcpy(&kv_head, inp, sizeof(kv_head)); inp += sizeof(kv_head); memcpy(&kv_size, inp, sizeof(kv_size)); inp += sizeof(kv_size); memcpy(&kv_used, inp, sizeof(kv_used)); inp += sizeof(kv_used); + memcpy(&v_trans, inp, sizeof(v_trans)); inp += sizeof(v_trans); + + GGML_ASSERT(kv_self.v_trans == (bool) v_trans); // incompatible V transposition if (kv_self.size != kv_size) { // the KV cache needs to be big enough to load all the KV cells from the saved state @@ -16433,6 +16564,8 @@ size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) { __func__, kv_head, kv_size, kv_self.size); } + llama_kv_cache_clear(ctx); + if (kv_buf_size) { const size_t pre_kv_buf_size = inp - src; @@ -16444,7 +16577,7 @@ size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) { ggml_backend_tensor_set(kv_self.k_l[il], inp, 0, k_size); inp += k_size; - if (kv_self.recurrent) { + if (kv_self.recurrent || !kv_self.v_trans) { // v is contiguous for recurrent models // TODO: use other tensors for state models than k and v const size_t v_size = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*kv_head); @@ -16466,8 +16599,6 @@ size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src) { GGML_ASSERT(kv_buf_size == inp - src - pre_kv_buf_size); } - llama_kv_cache_clear(ctx); - ctx->kv_self.head = kv_head; ctx->kv_self.used = kv_used; @@ -16727,28 +16858,49 @@ static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llam } } - // For the values, they are transposed, so we also need the element size and get the element ranges from each row - const uint32_t kv_size = kv_self.size; - for (int il = 0; il < (int)n_layer; ++il) { - // Write value type - const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; - data_ctx.write(&v_type_i, sizeof(v_type_i)); + // TODO: simplify, reduce copy-paste + if (!kv_self.v_trans) { + for (int il = 0; il < (int)n_layer; ++il) { + // Write value type + const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; + data_ctx.write(&v_type_i, sizeof(v_type_i)); - // Write element size - const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type); - data_ctx.write(&v_size_el, sizeof(v_size_el)); + // Write row size of value + const size_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa); + data_ctx.write(&v_size_row, sizeof(v_size_row)); - // For each row, we get the element values of each cell - for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { - // Read each range of cells of v_size_el length each into tmp_buf and write out + // Read each range of cells of v_size length each into tmp_buf and write out for (const auto & range : cell_ranges) { const size_t range_size = range.second - range.first; - const size_t src_offset = (range.first + j * kv_size) * v_size_el; - tmp_buf.resize(range_size * v_size_el); - ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), src_offset, tmp_buf.size()); + tmp_buf.resize(range_size * v_size_row); + ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), range.first * v_size_row, range_size * v_size_row); data_ctx.write(tmp_buf.data(), tmp_buf.size()); } } + } else { + // For the values, they are transposed, so we also need the element size and get the element ranges from each row + const uint32_t kv_size = kv_self.size; + for (int il = 0; il < (int)n_layer; ++il) { + // Write value type + const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; + data_ctx.write(&v_type_i, sizeof(v_type_i)); + + // Write element size + const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type); + data_ctx.write(&v_size_el, sizeof(v_size_el)); + + // For each row, we get the element values of each cell + for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { + // Read each range of cells of v_size_el length each into tmp_buf and write out + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + const size_t src_offset = (range.first + j * kv_size) * v_size_el; + tmp_buf.resize(range_size * v_size_el); + ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), src_offset, tmp_buf.size()); + data_ctx.write(tmp_buf.data(), tmp_buf.size()); + } + } + } } return data_ctx.get_size_written(); @@ -16873,41 +17025,75 @@ size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, } } - // For each layer, read the values for each cell (transposed) - for (int il = 0; il < (int)n_layer; ++il) { - // Read type of value - int32_t v_type_i_ref; - memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref)); - inp += sizeof(v_type_i_ref); - const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; - if (v_type_i != v_type_i_ref) { - llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); - LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); - return 0; - } + // TODO: simplify, reduce copy-paste + if (!kv_self.v_trans) { + for (int il = 0; il < (int)n_layer; ++il) { + // Read type of value + int32_t v_type_i_ref; + memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref)); + inp += sizeof(v_type_i_ref); + const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; + if (v_type_i != v_type_i_ref) { + llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); + LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); + return 0; + } - // Read element size of value - size_t v_size_el_ref; - memcpy(&v_size_el_ref, inp, sizeof(v_size_el_ref)); - inp += sizeof(v_size_el_ref); - const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type); - if (v_size_el != v_size_el_ref) { - llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); - LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, v_size_el_ref, il); - return 0; - } + // Read row size of value + size_t v_size_row_ref; + memcpy(&v_size_row_ref, inp, sizeof(v_size_row_ref)); + inp += sizeof(v_size_row_ref); + const size_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa); + if (v_size_row != v_size_row_ref) { + llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); + LLAMA_LOG_ERROR("%s: mismatched value row size (%zu != %zu, layer %d)\n", __func__, v_size_row, v_size_row_ref, il); + return 0; + } - if (cell_count) { - // For each row in the transposed matrix, read the values for the whole cell range - for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { - const size_t dst_offset = (kv_head + j * kv_size) * v_size_el; - ggml_backend_tensor_set(kv_self.v_l[il], inp, dst_offset, cell_count * v_size_el); - inp += cell_count * v_size_el; + if (cell_count) { + // Read and set the values for the whole cell range + ggml_backend_tensor_set(kv_self.v_l[il], inp, kv_head * v_size_row, cell_count * v_size_row); + inp += cell_count * v_size_row; + } + } + } else { + // For each layer, read the values for each cell (transposed) + for (int il = 0; il < (int)n_layer; ++il) { + // Read type of value + int32_t v_type_i_ref; + memcpy(&v_type_i_ref, inp, sizeof(v_type_i_ref)); + inp += sizeof(v_type_i_ref); + const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; + if (v_type_i != v_type_i_ref) { + llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); + LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); + return 0; + } + + // Read element size of value + size_t v_size_el_ref; + memcpy(&v_size_el_ref, inp, sizeof(v_size_el_ref)); + inp += sizeof(v_size_el_ref); + const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type); + if (v_size_el != v_size_el_ref) { + llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); + LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, v_size_el_ref, il); + return 0; + } + + if (cell_count) { + // For each row in the transposed matrix, read the values for the whole cell range + for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { + const size_t dst_offset = (kv_head + j * kv_size) * v_size_el; + ggml_backend_tensor_set(kv_self.v_l[il], inp, dst_offset, cell_count * v_size_el); + inp += cell_count * v_size_el; + } } } } const size_t nread = inp - src; + return nread; } @@ -17288,9 +17474,10 @@ int32_t llama_tokenize( static std::string llama_decode_text(const std::string & text) { std::string decoded_text; - auto unicode_sequences = unicode_cpts_from_utf8(text); - for (auto & unicode_sequence : unicode_sequences) { - decoded_text += unicode_utf8_to_byte(unicode_cpt_to_utf8(unicode_sequence)); + + const auto cpts = unicode_cpts_from_utf8(text); + for (const auto cpt : cpts) { + decoded_text += unicode_utf8_to_byte(unicode_cpt_to_utf8(cpt)); } return decoded_text; @@ -17704,9 +17891,9 @@ const char * llama_print_system_info(void) { s += "VSX = " + std::to_string(ggml_cpu_has_vsx()) + " | "; s += "MATMUL_INT8 = " + std::to_string(ggml_cpu_has_matmul_int8()) + " | "; #ifdef GGML_USE_LLAMAFILE - s += "LAMMAFILE = 1 | "; + s += "LLAMAFILE = 1 | "; #else - s += "LAMMAFILE = 0 | "; + s += "LLAMAFILE = 0 | "; #endif return s.c_str(); diff --git a/llama.h b/llama.h index 30835de5f..5ff04c1d4 100644 --- a/llama.h +++ b/llama.h @@ -40,7 +40,7 @@ #define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq' #define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN -#define LLAMA_SESSION_VERSION 5 +#define LLAMA_SESSION_VERSION 6 #define LLAMA_STATE_SEQ_MAGIC LLAMA_FILE_MAGIC_GGSQ #define LLAMA_STATE_SEQ_VERSION 1 @@ -79,6 +79,7 @@ extern "C" { LLAMA_VOCAB_PRE_TYPE_MPT = 5, LLAMA_VOCAB_PRE_TYPE_STARCODER = 6, LLAMA_VOCAB_PRE_TYPE_GPT2 = 7, + LLAMA_VOCAB_PRE_TYPE_REFACT = 8, }; // note: these values should be synchronized with ggml_rope @@ -171,7 +172,7 @@ extern "C" { bool sorted; } llama_token_data_array; - typedef bool (*llama_progress_callback)(float progress, void *ctx); + typedef bool (*llama_progress_callback)(float progress, void * user_data); // Input data for llama_decode // A llama_batch object can contain input about one or many sequences @@ -287,6 +288,7 @@ extern "C" { bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) bool embeddings; // if true, extract embeddings (together with logits) bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU + bool flash_attn; // whether to use flash attention // Abort callback // if it returns true, execution of llama_decode() will be aborted @@ -542,7 +544,7 @@ extern "C" { // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx); - // Clear the KV cache + // Clear the KV cache - both cell info is erased and KV data is zeroed LLAMA_API void llama_kv_cache_clear( struct llama_context * ctx); diff --git a/models/ggml-vocab-bert-bge.gguf.inp b/models/ggml-vocab-bert-bge.gguf.inp index 0389f00c7..0a89107c6 100644 --- a/models/ggml-vocab-bert-bge.gguf.inp +++ b/models/ggml-vocab-bert-bge.gguf.inp @@ -1,3 +1,7 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ __ggml_vocab_test__ diff --git a/models/ggml-vocab-bert-bge.gguf.out b/models/ggml-vocab-bert-bge.gguf.out index 969552e1c..e4a76cdb0 100644 --- a/models/ggml-vocab-bert-bge.gguf.out +++ b/models/ggml-vocab-bert-bge.gguf.out @@ -1,3 +1,5 @@ + 29464 2094 1018 1092 2706 + 11865 17875 diff --git a/models/ggml-vocab-deepseek-coder.gguf.inp b/models/ggml-vocab-deepseek-coder.gguf.inp index 0389f00c7..0a89107c6 100644 --- a/models/ggml-vocab-deepseek-coder.gguf.inp +++ b/models/ggml-vocab-deepseek-coder.gguf.inp @@ -1,3 +1,7 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ __ggml_vocab_test__ diff --git a/models/ggml-vocab-deepseek-coder.gguf.out b/models/ggml-vocab-deepseek-coder.gguf.out index 8ef585c78..9ccc560d6 100644 --- a/models/ggml-vocab-deepseek-coder.gguf.out +++ b/models/ggml-vocab-deepseek-coder.gguf.out @@ -1,3 +1,5 @@ + 1050 207 19 207 19192 4217 + 37 32009 71 6247 207 243 diff --git a/models/ggml-vocab-deepseek-llm.gguf.inp b/models/ggml-vocab-deepseek-llm.gguf.inp index 0389f00c7..0a89107c6 100644 --- a/models/ggml-vocab-deepseek-llm.gguf.inp +++ b/models/ggml-vocab-deepseek-llm.gguf.inp @@ -1,3 +1,7 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ __ggml_vocab_test__ diff --git a/models/ggml-vocab-deepseek-llm.gguf.out b/models/ggml-vocab-deepseek-llm.gguf.out index 0ea9d66e3..fd94b896d 100644 --- a/models/ggml-vocab-deepseek-llm.gguf.out +++ b/models/ggml-vocab-deepseek-llm.gguf.out @@ -1,3 +1,5 @@ + 1052 207 19 207 19109 4223 + 37 100014 71 6245 207 243 diff --git a/models/ggml-vocab-falcon.gguf.inp b/models/ggml-vocab-falcon.gguf.inp index 0389f00c7..0a89107c6 100644 --- a/models/ggml-vocab-falcon.gguf.inp +++ b/models/ggml-vocab-falcon.gguf.inp @@ -1,3 +1,7 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ __ggml_vocab_test__ diff --git a/models/ggml-vocab-falcon.gguf.out b/models/ggml-vocab-falcon.gguf.out index cb8da7b1a..209b04cda 100644 --- a/models/ggml-vocab-falcon.gguf.out +++ b/models/ggml-vocab-falcon.gguf.out @@ -1,3 +1,5 @@ + 878 204 31 3068 133 2137 + 28611 132 30042 204 258 diff --git a/models/ggml-vocab-gpt-2.gguf.inp b/models/ggml-vocab-gpt-2.gguf.inp index 0389f00c7..0a89107c6 100644 --- a/models/ggml-vocab-gpt-2.gguf.inp +++ b/models/ggml-vocab-gpt-2.gguf.inp @@ -1,3 +1,7 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ __ggml_vocab_test__ diff --git a/models/ggml-vocab-gpt-2.gguf.out b/models/ggml-vocab-gpt-2.gguf.out index 9986f38e4..78430f0d3 100644 --- a/models/ggml-vocab-gpt-2.gguf.out +++ b/models/ggml-vocab-gpt-2.gguf.out @@ -1,3 +1,5 @@ + 798 604 25208 1933 + 37 9116 71 11751 220 220 220 diff --git a/models/ggml-vocab-llama-bpe.gguf.inp b/models/ggml-vocab-llama-bpe.gguf.inp index 0389f00c7..0a89107c6 100644 --- a/models/ggml-vocab-llama-bpe.gguf.inp +++ b/models/ggml-vocab-llama-bpe.gguf.inp @@ -1,3 +1,7 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ __ggml_vocab_test__ diff --git a/models/ggml-vocab-llama-bpe.gguf.out b/models/ggml-vocab-llama-bpe.gguf.out index 4d903e1cd..1f00e3812 100644 --- a/models/ggml-vocab-llama-bpe.gguf.out +++ b/models/ggml-vocab-llama-bpe.gguf.out @@ -1,3 +1,5 @@ + 1142 220 19 220 27154 4038 + 37 51853 261 220 256 diff --git a/models/ggml-vocab-llama-spm.gguf.inp b/models/ggml-vocab-llama-spm.gguf.inp index 0389f00c7..0a89107c6 100644 --- a/models/ggml-vocab-llama-spm.gguf.inp +++ b/models/ggml-vocab-llama-spm.gguf.inp @@ -1,3 +1,7 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ __ggml_vocab_test__ diff --git a/models/ggml-vocab-llama-spm.gguf.out b/models/ggml-vocab-llama-spm.gguf.out index 15d00b106..9c3327cb5 100644 --- a/models/ggml-vocab-llama-spm.gguf.out +++ b/models/ggml-vocab-llama-spm.gguf.out @@ -1,3 +1,5 @@ + 474 287 29871 29946 29871 30226 7378 + 383 4000 261 259 1678 diff --git a/models/ggml-vocab-mpt.gguf.inp b/models/ggml-vocab-mpt.gguf.inp index 0389f00c7..0a89107c6 100644 --- a/models/ggml-vocab-mpt.gguf.inp +++ b/models/ggml-vocab-mpt.gguf.inp @@ -1,3 +1,7 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ __ggml_vocab_test__ diff --git a/models/ggml-vocab-mpt.gguf.out b/models/ggml-vocab-mpt.gguf.out index 1f4b0eb3a..d8d0fe909 100644 --- a/models/ggml-vocab-mpt.gguf.out +++ b/models/ggml-vocab-mpt.gguf.out @@ -1,3 +1,5 @@ + 728 577 24142 2607 + 39 26288 6554 209 50276 diff --git a/models/ggml-vocab-phi-3.gguf.inp b/models/ggml-vocab-phi-3.gguf.inp index 0389f00c7..0a89107c6 100644 --- a/models/ggml-vocab-phi-3.gguf.inp +++ b/models/ggml-vocab-phi-3.gguf.inp @@ -1,3 +1,7 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ __ggml_vocab_test__ diff --git a/models/ggml-vocab-phi-3.gguf.out b/models/ggml-vocab-phi-3.gguf.out index 15d00b106..9c3327cb5 100644 --- a/models/ggml-vocab-phi-3.gguf.out +++ b/models/ggml-vocab-phi-3.gguf.out @@ -1,3 +1,5 @@ + 474 287 29871 29946 29871 30226 7378 + 383 4000 261 259 1678 diff --git a/models/ggml-vocab-refact.gguf b/models/ggml-vocab-refact.gguf index 8f26cfb76..52afcf01a 100644 Binary files a/models/ggml-vocab-refact.gguf and b/models/ggml-vocab-refact.gguf differ diff --git a/models/ggml-vocab-refact.gguf.inp b/models/ggml-vocab-refact.gguf.inp new file mode 100644 index 000000000..0a89107c6 --- /dev/null +++ b/models/ggml-vocab-refact.gguf.inp @@ -0,0 +1,106 @@ +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__ +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__ + + + + + + + + + + + +🚀 (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-refact.gguf.out b/models/ggml-vocab-refact.gguf.out new file mode 100644 index 000000000..06b15c090 --- /dev/null +++ b/models/ggml-vocab-refact.gguf.out @@ -0,0 +1,43 @@ + 4833 225 38 225 143 140 17723 + 56 2006 3935 265 + + 225 + 261 + 264 + 202 + 203 + 478 + 2831 + 15773 + 8279 5788 + 12000 5788 + 8279 10896 + 12000 10896 + 12000 10896 19 + 8279 30 5788 19 + 12000 30 5788 19 + 458 438 5945 118 252 32 3766 + 105 34 38 42 225 41 102 1707 12530 10180 1479 8278 + 39862 8372 1039 9446 40242 13852 2053 8949 12531 1520 10700 + 14574 227 14574 133 14574 246 30457 238 14574 242 30457 229 14574 249 14574 134 14574 258 30457 228 14574 258 14574 114 14574 133 14574 232 14574 228 14574 254 14574 232 30457 228 14574 236 + 3807 253 227 308 4382 27 18458 133 46113 44967 123 13868 308 12565 19775 33071 40824 733 27 41889 308 2585 22680 688 1401 2819 4369 2404 27 + 8279 + 12000 + 225 12000 + 261 12000 + 264 12000 + 264 12000 284 12000 + 308 + 203 280 + 25 34666 + 8279 30 533 25 464 19 4971 884 844 18458 228 1018 4982 13368 2909 9513 17827 35 37 35 38 35 39 35 11873 47838 + 37 + 37 37 + 37 37 37 + 37 37 37 37 + 37 37 37 37 37 + 37 37 37 37 37 37 + 37 37 37 37 37 37 37 + 37 37 37 37 37 37 37 37 + 37 37 37 37 37 37 37 37 37 + 334 719 8878 202 10885 4222 16104 28570 203 3807 253 227 308 4382 27 18458 133 46113 44967 123 13868 308 12565 19775 33071 40824 733 27 41889 5945 118 252 3807 118 252 225 37 225 37 37 225 37 37 37 225 37 37 37 37 225 37 37 37 37 37 225 37 37 37 37 37 37 225 37 37 37 37 37 37 37 225 37 37 37 37 37 37 37 37 225 37 32 37 225 37 497 37 225 37 1179 37 225 14574 227 14574 133 14574 246 30457 238 14574 242 30457 229 14574 249 14574 134 14574 258 30457 228 14574 258 14574 114 14574 133 14574 232 36628 228 1018 4982 13368 2909 9513 17827 35 37 35 38 35 39 35 11873 47838 20921 16623 13028 8372 1039 9446 40242 13852 2053 8949 12531 1520 10700 5881 9592 13299 914 31753 31359 9163 3202 35472 10397 439 4763 2583 330 102 1455 938 1182 2017 30 330 613 844 3654 49 330 63 646 3654 439 4621 1930 561 30 330 54 844 2124 1629 35993 49 2688 25 7709 312 25 94 62 diff --git a/models/ggml-vocab-starcoder.gguf.inp b/models/ggml-vocab-starcoder.gguf.inp index 0389f00c7..0a89107c6 100644 --- a/models/ggml-vocab-starcoder.gguf.inp +++ b/models/ggml-vocab-starcoder.gguf.inp @@ -1,3 +1,7 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ __ggml_vocab_test__ diff --git a/models/ggml-vocab-starcoder.gguf.out b/models/ggml-vocab-starcoder.gguf.out index cd04254af..ccb55c7fe 100644 --- a/models/ggml-vocab-starcoder.gguf.out +++ b/models/ggml-vocab-starcoder.gguf.out @@ -1,3 +1,5 @@ + 4850 244 57 244 162 159 17722 + 75 2022 3943 284 244 280 diff --git a/scripts/compare-llama-bench.py b/scripts/compare-llama-bench.py index ef7f19ecb..3892fd25c 100755 --- a/scripts/compare-llama-bench.py +++ b/scripts/compare-llama-bench.py @@ -1,5 +1,6 @@ #!/usr/bin/env python3 +import logging import argparse import heapq import sys @@ -11,9 +12,11 @@ try: import git from tabulate import tabulate except ImportError as e: - print("ERROR: the following Python libraries are required: GitPython, tabulate.") + print("the following Python libraries are required: GitPython, tabulate.") # noqa: NP100 raise e +logger = logging.getLogger("compare-llama-bench") + # Properties by which to differentiate results per commit: KEY_PROPERTIES = [ "cpu_info", "gpu_info", "n_gpu_layers", "main_gpu", "cuda", "opencl", "metal", "gpu_blas", @@ -94,8 +97,7 @@ parser.add_argument("-s", "--show", help=help_s) known_args, unknown_args = parser.parse_known_args() if unknown_args: - print(f"ERROR: Received unknown args: {unknown_args}.") - print() + logger.error(f"Received unknown args: {unknown_args}.") parser.print_help() sys.exit(1) @@ -108,8 +110,7 @@ if input_file is None: input_file = sqlite_files[0] if input_file is None: - print("ERROR: Cannot find a suitable input file, please provide one.") - print() + logger.error("Cannot find a suitable input file, please provide one.") parser.print_help() sys.exit(1) @@ -194,23 +195,19 @@ if known_args.baseline is not None: hexsha8_baseline = get_commit_hexsha8(known_args.baseline) name_baseline = known_args.baseline if hexsha8_baseline is None: - print(f"ERROR: cannot find data for baseline={known_args.baseline}.") + logger.error(f"cannot find data for baseline={known_args.baseline}.") sys.exit(1) # Otherwise, search for the most recent parent of master for which there is data: elif repo is not None: hexsha8_baseline = find_parent_in_data(repo.heads.master.commit) if hexsha8_baseline is None: - print("ERROR: No baseline was provided and did not find data for any master branch commits.") - print() + logger.error("No baseline was provided and did not find data for any master branch commits.") parser.print_help() sys.exit(1) else: - print( - "ERROR: No baseline was provided and the current working directory " - "is not part of a git repository from which a baseline could be inferred." - ) - print() + logger.error("No baseline was provided and the current working directory " + "is not part of a git repository from which a baseline could be inferred.") parser.print_help() sys.exit(1) @@ -227,7 +224,7 @@ if known_args.compare is not None: hexsha8_compare = get_commit_hexsha8(known_args.compare) name_compare = known_args.compare if hexsha8_compare is None: - print(f"ERROR: cannot find data for compare={known_args.compare}.") + logger.error(f"cannot find data for compare={known_args.compare}.") sys.exit(1) # Otherwise, search for the commit for llama-bench was most recently run # and that is not a parent of master: @@ -241,16 +238,12 @@ elif repo is not None: break if hexsha8_compare is None: - print("ERROR: No compare target was provided and did not find data for any non-master commits.") - print() + logger.error("No compare target was provided and did not find data for any non-master commits.") parser.print_help() sys.exit(1) else: - print( - "ERROR: No compare target was provided and the current working directory " - "is not part of a git repository from which a compare target could be inferred." - ) - print() + logger.error("No compare target was provided and the current working directory " + "is not part of a git repository from which a compare target could be inferred.\n") parser.print_help() sys.exit(1) @@ -284,8 +277,7 @@ if known_args.show is not None: if prop not in KEY_PROPERTIES[:-2]: # Last two values are n_prompt, n_gen. unknown_cols.append(prop) if unknown_cols: - print(f"ERROR: Unknown values for --show: {', '.join(unknown_cols)}") - print() + logger.error(f"Unknown values for --show: {', '.join(unknown_cols)}") parser.print_usage() sys.exit(1) rows_show = get_rows(show) @@ -369,7 +361,7 @@ if "gpu_info" in show: headers = [PRETTY_NAMES[p] for p in show] headers += ["Test", f"t/s {name_baseline}", f"t/s {name_compare}", "Speedup"] -print(tabulate( +logger.info(tabulate( table, headers=headers, floatfmt=".2f", diff --git a/scripts/gen-unicode-data.py b/scripts/gen-unicode-data.py new file mode 100644 index 000000000..d49cbf2a0 --- /dev/null +++ b/scripts/gen-unicode-data.py @@ -0,0 +1,66 @@ +import regex + + +def cpt_to_utf8_str(cpt): + if cpt <= 0xFF: + return bytes([cpt, 0, 0, 0]) + elif cpt <= 0xFFFF: + return bytes([cpt & 0xFF, cpt >> 8, 0, 0]) + elif cpt <= 0xFFFFFF: + return bytes([cpt & 0xFF, (cpt >> 8) & 0xFF, (cpt >> 16) & 0xFF, 0]) + else: + return bytes([cpt & 0xFF, (cpt >> 8) & 0xFF, (cpt >> 16) & 0xFF, cpt >> 24]) + + +def is_match(codepoint, regex_expr): + try: + res = regex.match(regex_expr, cpt_to_utf8_str(codepoint).decode('utf-32')) + return res is not None + except Exception: + return False + + +def get_matches(regex_expr): + unicode_ranges = [] + current_range = None + + for codepoint in range(0x110000): + if is_match(codepoint, regex_expr): + if current_range is None: + current_range = [codepoint, codepoint] + else: + current_range[1] = codepoint + elif current_range is not None: + unicode_ranges.append(tuple(current_range)) + current_range = None + + if current_range is not None: + unicode_ranges.append(tuple(current_range)) + + return unicode_ranges + + +def print_cat(cat, ranges): + print("const std::vector> unicode_ranges_{} = {{".format(cat)) + cnt = 0 + for start, end in ranges: + if cnt % 4 != 0: + print(" ", end="") + print("{{0x{:08X}, 0x{:08X}}},".format(start, end), end="") + if cnt % 4 == 3: + print("") + cnt += 1 + + if cnt % 4 != 0: + print("") + print("};") + print("") + + +print_cat("number", get_matches(r'\p{N}')) +print_cat("letter", get_matches(r'\p{L}')) +print_cat("whitespace", get_matches(r'\p{Z}')) +print_cat("accent_mark", get_matches(r'\p{M}')) +print_cat("punctuation", get_matches(r'\p{P}')) +print_cat("symbol", get_matches(r'\p{S}')) +print_cat("control", get_matches(r'\p{C}')) diff --git a/scripts/run-with-preset.py b/scripts/run-with-preset.py index a18252730..e986a3604 100755 --- a/scripts/run-with-preset.py +++ b/scripts/run-with-preset.py @@ -1,5 +1,6 @@ #!/usr/bin/env python3 +import logging import argparse import os import subprocess @@ -7,6 +8,8 @@ import sys import yaml +logger = logging.getLogger("run-with-preset") + CLI_ARGS_MAIN_PERPLEXITY = [ "batch-size", "cfg-negative-prompt", "cfg-scale", "chunks", "color", "ctx-size", "escape", "export", "file", "frequency-penalty", "grammar", "grammar-file", "hellaswag", @@ -56,6 +59,7 @@ parser.add_argument("-bin", "--binary", help="The binary to run.") parser.add_argument("yaml_files", nargs="*", help="Arbitrary number of YAML files from which to read preset values. " "If two files specify the same values the later one will be used.") +parser.add_argument("--verbose", action="store_true", help="increase output verbosity") known_args, unknown_args = parser.parse_known_args() @@ -63,6 +67,8 @@ if not known_args.yaml_files and not unknown_args: parser.print_help() sys.exit(0) +logging.basicConfig(level=logging.DEBUG if known_args.verbose else logging.INFO) + props = dict() for yaml_file in known_args.yaml_files: @@ -85,7 +91,7 @@ elif binary.lower().endswith("llama-bench"): elif binary.lower().endswith("server"): cli_args = CLI_ARGS_SERVER else: - print(f"Unknown binary: {binary}") + logger.error(f"Unknown binary: {binary}") sys.exit(1) command_list = [binary] @@ -121,11 +127,11 @@ for cli_arg in cli_args: num_unused = len(props) if num_unused > 10: - print(f"The preset file contained a total of {num_unused} unused properties.") + logger.info(f"The preset file contained a total of {num_unused} unused properties.") elif num_unused > 0: - print("The preset file contained the following unused properties:") + logger.info("The preset file contained the following unused properties:") for prop, value in props.items(): - print(f" {prop}: {value}") + logger.info(f" {prop}: {value}") command_list += unknown_args diff --git a/scripts/verify-checksum-models.py b/scripts/verify-checksum-models.py index dff4b4734..0b5b9aafa 100755 --- a/scripts/verify-checksum-models.py +++ b/scripts/verify-checksum-models.py @@ -1,8 +1,11 @@ #!/usr/bin/env python3 +import logging import os import hashlib +logger = logging.getLogger("verify-checksum-models") + def sha256sum(file): block_size = 16 * 1024 * 1024 # 16 MB block size @@ -27,7 +30,7 @@ hash_list_file = os.path.join(llama_path, "SHA256SUMS") # Check if the hash list file exists if not os.path.exists(hash_list_file): - print(f"Hash list file not found: {hash_list_file}") + logger.error(f"Hash list file not found: {hash_list_file}") exit(1) # Read the hash file content and split it into an array of lines @@ -46,7 +49,7 @@ for line in hash_list: file_path = os.path.join(llama_path, filename) # Informing user of the progress of the integrity check - print(f"Verifying the checksum of {file_path}") + logger.info(f"Verifying the checksum of {file_path}") # Check if the file exists if os.path.exists(file_path): @@ -73,9 +76,9 @@ for line in hash_list: # Print column headers for results table -print("\n" + "filename".ljust(40) + "valid checksum".center(20) + "file missing".center(20)) -print("-" * 80) +print("filename".ljust(40) + "valid checksum".center(20) + "file missing".center(20)) # noqa: NP100 +print("-" * 80) # noqa: NP100 # Output the results as a table for r in results: - print(f"{r['filename']:40} {r['valid checksum']:^20} {r['file missing']:^20}") + print(f"{r['filename']:40} {r['valid checksum']:^20} {r['file missing']:^20}") # noqa: NP100 diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index d23e7f771..cad703fce 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -74,13 +74,15 @@ llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-spm ARGS ${CMAKE llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-bpe ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf) llama_test(test-tokenizer-0 NAME test-tokenizer-0-phi-3 ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-phi-3.gguf) llama_test(test-tokenizer-0 NAME test-tokenizer-0-falcon ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) -llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-llm ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-llm.gguf) -llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-coder ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-coder.gguf) llama_test(test-tokenizer-0 NAME test-tokenizer-0-bert-bge ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bert-bge.gguf) # TODO: enable when fixed +# https://github.com/ggerganov/llama.cpp/pull/7036 #llama_test(test-tokenizer-0 NAME test-tokenizer-0-mpt ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf) +#llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-llm ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-llm.gguf) +#llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-coder ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-coder.gguf) llama_test(test-tokenizer-0 NAME test-tokenizer-0-starcoder ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf) llama_test(test-tokenizer-0 NAME test-tokenizer-0-gpt-2 ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-2.gguf) +llama_test(test-tokenizer-0 NAME test-tokenizer-0-refact ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf) # build test-tokenizer-1-bpe target once and add many tests add_executable(test-tokenizer-1-bpe test-tokenizer-1-bpe.cpp) diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 02daad24b..b27c1291e 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -1090,6 +1090,12 @@ struct test_soft_max : public test_case { return VARS_TO_STR5(type, ne, mask, scale, max_bias); } + // the 1024 test with bias occasionally fails: + // SOFT_MAX(type=f32,ne=[1024,16,1,1],mask=1,scale=1.000000,max_bias=8.000000): [SOFT_MAX] NMSE = 0.000000103 > 0.000000100 FAIL + virtual double max_nmse_err() override { + return 1e-6; + } + test_soft_max(ggml_type type = GGML_TYPE_F32, std::array ne = {10, 10, 10, 10}, bool mask = false, @@ -1101,7 +1107,7 @@ struct test_soft_max : public test_case { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); ggml_tensor * mask = nullptr; if (this->mask) { - mask = ggml_new_tensor_2d(ctx, type, ne[0], ne[1]); + mask = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, ne[0], ne[1]); } ggml_tensor * pos = nullptr; if (max_bias > 0.0f) { @@ -1475,6 +1481,34 @@ struct test_leaky_relu : public test_case { } }; +// GGML_OP_FLASH_ATTN_EXT +struct test_flash_attn_ext : public test_case { + const int64_t hs; // head size + const int64_t nh; // num heads + const int64_t kv; // kv size + const int64_t nb; // batch size + + std::string vars() override { + return VARS_TO_STR4(hs, nh, kv, nb); + } + + double max_nmse_err() override { + return 5e-4; + } + + test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8) + : hs(hs), nh(nh), kv(kv), nb(nb) {} + + ggml_tensor * build_graph(ggml_context * ctx) override { + ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs, nb, nh, 1); + ggml_tensor * k = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1); + ggml_tensor * v = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1); + ggml_tensor * mask = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), 1, 1); + ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, mask, 1.0f/sqrtf(hs)); + return out; + } +}; + enum llm_norm_type { LLM_NORM, LLM_NORM_RMS, @@ -1661,7 +1695,7 @@ struct test_llama : public test_llm { struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, hp.n_tokens); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hp.n_kv, hp.n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx, GGML_TYPE_F16, hp.n_kv, hp.n_tokens, 1); ggml_tensor * k_l = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 1638400); ggml_tensor * v_l = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 1638400); @@ -1783,7 +1817,7 @@ struct test_falcon : public test_llm { struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, hp.n_tokens); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hp.n_kv, hp.n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx, GGML_TYPE_F16, hp.n_kv, hp.n_tokens, 1); ggml_tensor * k_l = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 1638400); ggml_tensor * v_l = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 1638400); @@ -2095,7 +2129,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op for (float scale : {1.0f, 0.1f}) { for (int64_t ne0 : {16, 1024}) { for (int64_t ne1 : {16, 1024}) { - test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, scale, max_bias)); + test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, scale, max_bias)); test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, scale, max_bias)); } } @@ -2139,6 +2173,16 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op test_cases.emplace_back(new test_timestep_embedding()); test_cases.emplace_back(new test_leaky_relu()); + for (int hs : { 64, 80, 128, 256, }) { + for (int nh : { 32, }) { + for (int kv : { 512, 1024, }) { + for (int nb : { 1, 2, 4, 8, }) { + test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb)); + } + } + } + } + // these tests are disabled to save execution time, but they can be handy for debugging #if 0 test_cases.emplace_back(new test_llama(1)); diff --git a/tests/test-grammar-integration.cpp b/tests/test-grammar-integration.cpp index 2d8f228e3..1a4004e2a 100644 --- a/tests/test-grammar-integration.cpp +++ b/tests/test-grammar-integration.cpp @@ -10,15 +10,10 @@ #include "unicode.h" #include #include +#include -static void test_simple_grammar() { - // Test case for a simple grammar - const std::string grammar_str = R"""(root ::= expr -expr ::= term ("+" term)* -term ::= number -number ::= [0-9]+)"""; - - grammar_parser::parse_state parsed_grammar = grammar_parser::parse(grammar_str.c_str()); +static llama_grammar* build_grammar(const std::string & grammar_str) { + auto parsed_grammar = grammar_parser::parse(grammar_str.c_str()); // Ensure we parsed correctly assert(!parsed_grammar.rules.empty()); @@ -30,8 +25,10 @@ number ::= [0-9]+)"""; llama_grammar* grammar = llama_grammar_init( grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root")); - std::string input = "123+456"; + return grammar; +} +static bool match_string(const std::string & input, llama_grammar* grammar) { auto decoded = decode_utf8(input, {}); const auto & code_points = decoded.first; @@ -39,159 +36,67 @@ number ::= [0-9]+)"""; for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) { auto prev_stacks = grammar->stacks; llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks); - assert(!grammar->stacks.empty()); - } - - bool completed_grammar = false; - - for (const auto & stack : grammar->stacks) { - if (stack.empty()) { - completed_grammar = true; - break; + if (grammar->stacks.empty()) { + // no stacks means that the grammar failed to match at this point + return false; } } - assert(completed_grammar); + for (const auto & stack : grammar->stacks) { + if (stack.empty()) { + // An empty stack means that the grammar has been completed + return true; + } + } - // Clean up allocated memory - llama_grammar_free(grammar); + return false; } -static void test_complex_grammar() { - // Test case for a more complex grammar, with both failure strings and success strings - const std::string grammar_str = R"""(root ::= expression -expression ::= term ws (("+"|"-") ws term)* -term ::= factor ws (("*"|"/") ws factor)* -factor ::= number | variable | "(" expression ")" | function-call -number ::= [0-9]+ -variable ::= [a-zA-Z_][a-zA-Z0-9_]* -function-call ::= variable ws "(" (expression ("," ws expression)*)? ")" -ws ::= [ \t\n\r]?)"""; +static void test_grammar(const std::string & test_desc, const std::string & grammar_str, const std::vector & passing_strings, const std::vector & failing_strings) { + fprintf(stderr, "⚫ Testing %s. Grammar: %s\n", test_desc.c_str(), grammar_str.c_str()); + fflush(stderr); - grammar_parser::parse_state parsed_grammar = grammar_parser::parse(grammar_str.c_str()); - - // Ensure we parsed correctly - assert(!parsed_grammar.rules.empty()); - - // Ensure we have a root node - assert(!(parsed_grammar.symbol_ids.find("root") == parsed_grammar.symbol_ids.end())); - - std::vector grammar_rules(parsed_grammar.c_rules()); - llama_grammar* grammar = llama_grammar_init( - grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root")); + auto grammar = build_grammar(grammar_str); // Save the original grammar stacks so that we can reset after every new string we want to test auto original_stacks = grammar->stacks; - // Test a few strings - std::vector test_strings_pass = { - "42", - "1*2*3*4*5", - "x", - "x+10", - "x1+y2", - "(a+b)*(c-d)", - "func()", - "func(x,y+2)", - "a*(b+c)-d/e", - "f(g(x),h(y,z))", - "x + 10", - "x1 + y2", - "(a + b) * (c - d)", - "func()", - "func(x, y + 2)", - "a * (b + c) - d / e", - "f(g(x), h(y, z))", - "123+456", - "123*456*789-123/456+789*123", - "123+456*789-123/456+789*123-456/789+123*456-789/123+456*789-123/456+789*123-456" - }; - - std::vector test_strings_fail = { - "+", - "/ 3x", - "x + + y", - "a * / b", - "func(,)", - "func(x y)", - "(a + b", - "x + y)", - "a + b * (c - d", - "42 +", - "x +", - "x + 10 +", - "(a + b) * (c - d", - "func(", - "func(x, y + 2", - "a * (b + c) - d /", - "f(g(x), h(y, z)", - "123+456*789-123/456+789*123-456/789+123*456-789/123+456*789-123/456+789*123-456/", - }; + fprintf(stderr, " 🔵 Valid strings:\n"); // Passing strings - for (const auto & test_string : test_strings_pass) { - auto decoded = decode_utf8(test_string, {}); + for (const auto & test_string : passing_strings) { + fprintf(stderr, " \"%s\" ", test_string.c_str()); + fflush(stderr); - const auto & code_points = decoded.first; + bool matched = match_string(test_string, grammar); - int pos = 0; - for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) { - ++pos; - auto prev_stacks = grammar->stacks; - llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks); - - // Expect that each code point will not cause the grammar to fail - if (grammar->stacks.empty()) { - fprintf(stdout, "Error at position %d\n", pos); - fprintf(stderr, "Unexpected character '%s'\n", unicode_cpt_to_utf8(*it).c_str()); - fprintf(stderr, "Input string is %s:\n", test_string.c_str()); - } - assert(!grammar->stacks.empty()); + if (!matched) { + fprintf(stderr, "❌ (failed to match)\n"); + } else { + fprintf(stdout, "✅︎\n"); } - bool completed_grammar = false; - - for (const auto & stack : grammar->stacks) { - if (stack.empty()) { - completed_grammar = true; - break; - } - } - - assert(completed_grammar); + assert(matched); // Reset the grammar stacks grammar->stacks = original_stacks; } + fprintf(stderr, " 🟠 Invalid strings:\n"); + // Failing strings - for (const auto & test_string : test_strings_fail) { - auto decoded = decode_utf8(test_string, {}); + for (const auto & test_string : failing_strings) { + fprintf(stderr, " \"%s\" ", test_string.c_str()); + fflush(stderr); - const auto & code_points = decoded.first; - bool parse_failed = false; + bool matched = match_string(test_string, grammar); - for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) { - auto prev_stacks = grammar->stacks; - llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks); - if (grammar->stacks.empty()) { - parse_failed = true; - break; - } - assert(!grammar->stacks.empty()); + if (matched) { + fprintf(stderr, "❌ (incorrectly matched)\n"); + } else { + fprintf(stdout, "✅︎\n"); } - - bool completed_grammar = false; - - for (const auto & stack : grammar->stacks) { - if (stack.empty()) { - completed_grammar = true; - break; - } - } - - // Ensure that the grammar is not completed, or that each string failed to match as-expected - assert((!completed_grammar) || parse_failed); + assert(!matched); // Reset the grammar stacks grammar->stacks = original_stacks; @@ -201,7 +106,183 @@ ws ::= [ \t\n\r]?)"""; llama_grammar_free(grammar); } +static void test_simple_grammar() { + // Test case for a simple grammar + test_grammar( + "simple grammar", + R"""( + root ::= expr + expr ::= term ("+" term)* + term ::= number + number ::= [0-9]+)""", + // Passing strings + { + "42", + "1+2+3+4+5", + "123+456", + }, + // Failing strings + { + "+", + "/ 3", + "1+2+3+4+5+", + "12a45", + } + ); +} + +static void test_complex_grammar() { + // Test case for a more complex grammar, with both failure strings and success strings + test_grammar( + "medium complexity grammar", + // Grammar + R"""( + root ::= expression + expression ::= term ws (("+"|"-") ws term)* + term ::= factor ws (("*"|"/") ws factor)* + factor ::= number | variable | "(" expression ")" | function-call + number ::= [0-9]+ + variable ::= [a-zA-Z_][a-zA-Z0-9_]* + function-call ::= variable ws "(" (expression ("," ws expression)*)? ")" + ws ::= [ \t\n\r]?)""", + // Passing strings + { + "42", + "1*2*3*4*5", + "x", + "x+10", + "x1+y2", + "(a+b)*(c-d)", + "func()", + "func(x,y+2)", + "a*(b+c)-d/e", + "f(g(x),h(y,z))", + "x + 10", + "x1 + y2", + "(a + b) * (c - d)", + "func()", + "func(x, y + 2)", + "a * (b + c) - d / e", + "f(g(x), h(y, z))", + "123+456", + "123*456*789-123/456+789*123", + "123+456*789-123/456+789*123-456/789+123*456-789/123+456*789-123/456+789*123-456" + }, + // Failing strings + { + "+", + "/ 3x", + "x + + y", + "a * / b", + "func(,)", + "func(x y)", + "(a + b", + "x + y)", + "a + b * (c - d", + "42 +", + "x +", + "x + 10 +", + "(a + b) * (c - d", + "func(", + "func(x, y + 2", + "a * (b + c) - d /", + "f(g(x), h(y, z)", + "123+456*789-123/456+789*123-456/789+123*456-789/123+456*789-123/456+789*123-456/", + } + ); +} + +static void test_quantifiers() { + // A collection of tests to exercise * + and ? quantifiers + + test_grammar( + "* quantifier", + // Grammar + R"""(root ::= "a"*)""", + // Passing strings + { + "", + "a", + "aaaaa", + "aaaaaaaaaaaaaaaaaa", + "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa" + }, + // Failing strings + { + "b", + "ab", + "aab", + "ba", + "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaab" + } + ); + test_grammar( + "+ quantifier", + // Grammar + R"""(root ::= "a"+)""", + // Passing strings + { + "a", + "aaaaa", + "aaaaaaaaaaaaaaaaaa", + "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa" + }, + // Failing strings + { + "", + "b", + "ab", + "aab", + "ba", + "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaab" + } + ); + test_grammar( + "? quantifier", + // Grammar + R"""(root ::= "a"?)""", + // Passing strings + { + "", + "a" + }, + // Failing strings + { + "b", + "ab", + "aa", + "ba", + } + ); + test_grammar( + "mixed quantifiers", + // Grammar + R"""( + root ::= cons+ vowel* cons? (vowel cons)* + vowel ::= [aeiouy] + cons ::= [bcdfghjklmnpqrstvwxyz] + )""", + // Passing strings + { + "yes", + "no", + "noyes", + "crwth", + "four", + "bryyyy", + }, + // Failing strings + { + "yess", + "yesno", + "forty", + "catyyy", + } + ); +} + static void test_failure_missing_root() { + fprintf(stderr, "⚫ Testing missing root node:\n"); // Test case for a grammar that is missing a root rule const std::string grammar_str = R"""(rot ::= expr expr ::= term ("+" term)* @@ -215,29 +296,37 @@ number ::= [0-9]+)"""; // Ensure we do NOT have a root node assert(parsed_grammar.symbol_ids.find("root") == parsed_grammar.symbol_ids.end()); + fprintf(stderr, " ✅︎ Passed\n"); } static void test_failure_missing_reference() { + fprintf(stderr, "⚫ Testing missing reference node:\n"); + // Test case for a grammar that is missing a referenced rule - const std::string grammar_str = R"""(root ::= expr + const std::string grammar_str = +R"""(root ::= expr expr ::= term ("+" term)* term ::= numero number ::= [0-9]+)"""; - fprintf(stderr, "Expected error: "); + fprintf(stderr, " Expected error: "); grammar_parser::parse_state parsed_grammar = grammar_parser::parse(grammar_str.c_str()); // Ensure we did NOT parsed correctly assert(parsed_grammar.rules.empty()); - fprintf(stderr, "End of expected error. Test successful.\n"); + fprintf(stderr, " End of expected error.\n"); + fprintf(stderr, " ✅︎ Passed\n"); } int main() { + fprintf(stdout, "Running grammar integration tests...\n"); test_simple_grammar(); test_complex_grammar(); + test_quantifiers(); test_failure_missing_root(); test_failure_missing_reference(); + fprintf(stdout, "All tests passed.\n"); return 0; } diff --git a/tests/test-tokenizer-0.cpp b/tests/test-tokenizer-0.cpp index 5122757cf..d478f1041 100644 --- a/tests/test-tokenizer-0.cpp +++ b/tests/test-tokenizer-0.cpp @@ -55,8 +55,10 @@ // return _k_tests; //} -static std::map> read_tests(const std::string & fname_inp, const std::string & fname_out) { - std::map> tests; +using llama_tests = std::map>; + +static llama_tests read_tests(const std::string & fname_inp, const std::string & fname_out) { + llama_tests tests; std::ifstream ifs_inp(fname_inp); if (!ifs_inp) { @@ -175,12 +177,20 @@ int main(int argc, char **argv) { bool success = true; - const auto k_tests = read_tests(fname_inp, fname_out); + const auto k_tests = [&]() -> llama_tests { + if (!fname_text.empty()) { + return {}; + } - if (k_tests.empty()) { - fprintf(stderr, "%s : error: no tests found\n", __func__); - return 1; - } + const auto res = read_tests(fname_inp, fname_out); + + if (res.empty()) { + fprintf(stderr, "%s : error: no tests found\n", __func__); + exit(1); + } + + return res; + }(); const bool add_special = false; @@ -238,7 +248,17 @@ int main(int argc, char **argv) { fprintf(stderr, "%s : text size: %zu\n", __func__, text.size()); - const std::vector res = llama_tokenize(ctx, text, add_special); + std::vector res; + + { + const auto t_start = ggml_time_us(); + + res = llama_tokenize(ctx, text, add_special); + + const auto t_end = ggml_time_us(); + + fprintf(stderr, "%s : tokenized in %.3f ms (cpp)\n", __func__, (t_end - t_start) / 1000.0); + } fprintf(stderr, "%s : tokens: %zu\n", __func__, res.size()); @@ -252,7 +272,8 @@ int main(int argc, char **argv) { } for (const auto & tok : res) { - ofs << tok << " '" << string_strip(llama_detokenize_bpe(ctx, std::vector{tok})) << "'" << std::endl; + //ofs << tok << " '" << string_strip(llama_detokenize_bpe(ctx, std::vector{tok})) << "'" << std::endl; + ofs << tok << "\n"; } } diff --git a/tests/test-tokenizer-0.py b/tests/test-tokenizer-0.py new file mode 100644 index 000000000..8e7638e42 --- /dev/null +++ b/tests/test-tokenizer-0.py @@ -0,0 +1,46 @@ +import time +import argparse + +from transformers import AutoTokenizer + +parser = argparse.ArgumentParser() +parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file") +parser.add_argument("--fname-tok", help="path to a text file to tokenize", required=True) +args = parser.parse_args() + +dir_tokenizer = args.dir_tokenizer +fname_tok = args.fname_tok + +tokenizer = AutoTokenizer.from_pretrained(dir_tokenizer) + +print('tokenizing file: ', fname_tok) +fname_out = fname_tok + '.tok' +with open(fname_tok, 'r', encoding='utf-8') as f: + lines = f.readlines() + s = ''.join(lines) + t_start = time.time() + res = tokenizer.encode(s, add_special_tokens=False) + t_end = time.time() + print('\nmain : tokenized in', "{:.3f}".format(1000.0 * (t_end - t_start)), 'ms (py)') + with open(fname_out, 'w', encoding='utf-8') as f: + for x in res: + # LLaMA v3 for some reason strips the space for these tokens (and others) + # if x == 662: + # f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n') + # elif x == 1174: + # f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n') + # elif x == 2564: + # f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n') + # elif x == 758: + # f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n') + # elif x == 949: + # f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n') + # elif x == 5354: + # f.write(str(x) + ' \' ' + tokenizer.decode(x) + '\'\n') + # else: + # f.write(str(x) + ' \'' + tokenizer.decode(x) + '\'\n') + # f.write(str(x) + ' \'' + tokenizer.decode(x).strip() + '\'\n') + f.write(str(x) + '\n') + print('len(res): ', len(res)) + print('len(lines): ', len(lines)) +print('results written to: ', fname_out) diff --git a/tests/test-tokenizer-0.sh b/tests/test-tokenizer-0.sh new file mode 100755 index 000000000..2fb8632d8 --- /dev/null +++ b/tests/test-tokenizer-0.sh @@ -0,0 +1,34 @@ +#!/bin/bash +# +# Usage: +# +# test-tokenizer-0.sh +# + +if [ $# -ne 2 ]; then + printf "Usage: $0 \n" + exit 1 +fi + +name=$1 +input=$2 + +make -j tests/test-tokenizer-0 + +printf "Testing %s on %s ...\n" $name $input + +python3 ./tests/test-tokenizer-0.py ./models/tokenizers/$name --fname-tok $input > /tmp/test-tokenizer-0-$name-py.log 2>&1 +cat /tmp/test-tokenizer-0-$name-py.log | grep "tokenized in" + +./tests/test-tokenizer-0 ./models/ggml-vocab-$name.gguf $input > /tmp/test-tokenizer-0-$name-cpp.log 2>&1 +cat /tmp/test-tokenizer-0-$name-cpp.log | grep "tokenized in" + +diff $input.tok $input.tokcpp > /dev/null 2>&1 + +if [ $? -eq 0 ]; then + printf "Tokenization is correct!\n" +else + diff $input.tok $input.tokcpp | head -n 32 + + printf "Tokenization differs!\n" +fi diff --git a/unicode-data.cpp b/unicode-data.cpp index 1b3849628..07bf02c45 100644 --- a/unicode-data.cpp +++ b/unicode-data.cpp @@ -5,29 +5,47 @@ #include #include -const std::vector> unicode_ranges_digit = { -{0x00000030, 0x00000039}, {0x000000B2, 0x000000B3}, {0x000000B9, 0x000000B9}, {0x00000660, 0x00000669}, -{0x000006F0, 0x000006F9}, {0x000007C0, 0x000007C9}, {0x00000966, 0x0000096F}, {0x000009E6, 0x000009EF}, -{0x00000A66, 0x00000A6F}, {0x00000AE6, 0x00000AEF}, {0x00000B66, 0x00000B6F}, {0x00000BE6, 0x00000BEF}, -{0x00000C66, 0x00000C6F}, {0x00000CE6, 0x00000CEF}, {0x00000D66, 0x00000D6F}, {0x00000DE6, 0x00000DEF}, -{0x00000E50, 0x00000E59}, {0x00000ED0, 0x00000ED9}, {0x00000F20, 0x00000F29}, {0x00001040, 0x00001049}, -{0x00001090, 0x00001099}, {0x00001369, 0x00001371}, {0x000017E0, 0x000017E9}, {0x00001810, 0x00001819}, -{0x00001946, 0x0000194F}, {0x000019D0, 0x000019DA}, {0x00001A80, 0x00001A89}, {0x00001A90, 0x00001A99}, -{0x00001B50, 0x00001B59}, {0x00001BB0, 0x00001BB9}, {0x00001C40, 0x00001C49}, {0x00001C50, 0x00001C59}, -{0x00002070, 0x00002070}, {0x00002074, 0x00002079}, {0x00002080, 0x00002089}, {0x00002460, 0x00002468}, -{0x00002474, 0x0000247C}, {0x00002488, 0x00002490}, {0x000024EA, 0x000024EA}, {0x000024F5, 0x000024FD}, -{0x000024FF, 0x000024FF}, {0x00002776, 0x0000277E}, {0x00002780, 0x00002788}, {0x0000278A, 0x00002792}, -{0x0000A620, 0x0000A629}, {0x0000A8D0, 0x0000A8D9}, {0x0000A900, 0x0000A909}, {0x0000A9D0, 0x0000A9D9}, -{0x0000A9F0, 0x0000A9F9}, {0x0000AA50, 0x0000AA59}, {0x0000ABF0, 0x0000ABF9}, {0x0000FF10, 0x0000FF19}, -{0x000104A0, 0x000104A9}, {0x00010A40, 0x00010A43}, {0x00010D30, 0x00010D39}, {0x00010E60, 0x00010E68}, -{0x00011052, 0x0001105A}, {0x00011066, 0x0001106F}, {0x000110F0, 0x000110F9}, {0x00011136, 0x0001113F}, -{0x000111D0, 0x000111D9}, {0x000112F0, 0x000112F9}, {0x00011450, 0x00011459}, {0x000114D0, 0x000114D9}, -{0x00011650, 0x00011659}, {0x000116C0, 0x000116C9}, {0x00011730, 0x00011739}, {0x000118E0, 0x000118E9}, -{0x00011950, 0x00011959}, {0x00011C50, 0x00011C59}, {0x00011D50, 0x00011D59}, {0x00011DA0, 0x00011DA9}, -{0x00016A60, 0x00016A69}, {0x00016B50, 0x00016B59}, {0x0001D7CE, 0x0001D7FF}, {0x0001E140, 0x0001E149}, -{0x0001E2F0, 0x0001E2F9}, {0x0001E950, 0x0001E959}, {0x0001F100, 0x0001F10A}, {0x0001FBF0, 0x0001FBF9}, -{0x000000BC, 0x000000BE}, {0x00003007, 0x00003007}, -//{0x00002150, 0x0000218F}, // Number Forms +// generated with scripts/gen-unicode-data.py +// +// TODO: generate unicode_map_lowercase +// TODO: generate unicode_map_nfd + +const std::vector> unicode_ranges_number = { +{0x00000030, 0x00000039}, {0x000000B2, 0x000000B3}, {0x000000B9, 0x000000B9}, {0x000000BC, 0x000000BE}, +{0x00000660, 0x00000669}, {0x000006F0, 0x000006F9}, {0x000007C0, 0x000007C9}, {0x00000966, 0x0000096F}, +{0x000009E6, 0x000009EF}, {0x000009F4, 0x000009F9}, {0x00000A66, 0x00000A6F}, {0x00000AE6, 0x00000AEF}, +{0x00000B66, 0x00000B6F}, {0x00000B72, 0x00000B77}, {0x00000BE6, 0x00000BF2}, {0x00000C66, 0x00000C6F}, +{0x00000C78, 0x00000C7E}, {0x00000CE6, 0x00000CEF}, {0x00000D58, 0x00000D5E}, {0x00000D66, 0x00000D78}, +{0x00000DE6, 0x00000DEF}, {0x00000E50, 0x00000E59}, {0x00000ED0, 0x00000ED9}, {0x00000F20, 0x00000F33}, +{0x00001040, 0x00001049}, {0x00001090, 0x00001099}, {0x00001369, 0x0000137C}, {0x000016EE, 0x000016F0}, +{0x000017E0, 0x000017E9}, {0x000017F0, 0x000017F9}, {0x00001810, 0x00001819}, {0x00001946, 0x0000194F}, +{0x000019D0, 0x000019DA}, {0x00001A80, 0x00001A89}, {0x00001A90, 0x00001A99}, {0x00001B50, 0x00001B59}, +{0x00001BB0, 0x00001BB9}, {0x00001C40, 0x00001C49}, {0x00001C50, 0x00001C59}, {0x00002070, 0x00002070}, +{0x00002074, 0x00002079}, {0x00002080, 0x00002089}, {0x00002150, 0x00002182}, {0x00002185, 0x00002189}, +{0x00002460, 0x0000249B}, {0x000024EA, 0x000024FF}, {0x00002776, 0x00002793}, {0x00002CFD, 0x00002CFD}, +{0x00003007, 0x00003007}, {0x00003021, 0x00003029}, {0x00003038, 0x0000303A}, {0x00003192, 0x00003195}, +{0x00003220, 0x00003229}, {0x00003248, 0x0000324F}, {0x00003251, 0x0000325F}, {0x00003280, 0x00003289}, +{0x000032B1, 0x000032BF}, {0x0000A620, 0x0000A629}, {0x0000A6E6, 0x0000A6EF}, {0x0000A830, 0x0000A835}, +{0x0000A8D0, 0x0000A8D9}, {0x0000A900, 0x0000A909}, {0x0000A9D0, 0x0000A9D9}, {0x0000A9F0, 0x0000A9F9}, +{0x0000AA50, 0x0000AA59}, {0x0000ABF0, 0x0000ABF9}, {0x0000FF10, 0x0000FF19}, {0x00010107, 0x00010133}, +{0x00010140, 0x00010178}, {0x0001018A, 0x0001018B}, {0x000102E1, 0x000102FB}, {0x00010320, 0x00010323}, +{0x00010341, 0x00010341}, {0x0001034A, 0x0001034A}, {0x000103D1, 0x000103D5}, {0x000104A0, 0x000104A9}, +{0x00010858, 0x0001085F}, {0x00010879, 0x0001087F}, {0x000108A7, 0x000108AF}, {0x000108FB, 0x000108FF}, +{0x00010916, 0x0001091B}, {0x000109BC, 0x000109BD}, {0x000109C0, 0x000109CF}, {0x000109D2, 0x000109FF}, +{0x00010A40, 0x00010A48}, {0x00010A7D, 0x00010A7E}, {0x00010A9D, 0x00010A9F}, {0x00010AEB, 0x00010AEF}, +{0x00010B58, 0x00010B5F}, {0x00010B78, 0x00010B7F}, {0x00010BA9, 0x00010BAF}, {0x00010CFA, 0x00010CFF}, +{0x00010D30, 0x00010D39}, {0x00010E60, 0x00010E7E}, {0x00010F1D, 0x00010F26}, {0x00010F51, 0x00010F54}, +{0x00010FC5, 0x00010FCB}, {0x00011052, 0x0001106F}, {0x000110F0, 0x000110F9}, {0x00011136, 0x0001113F}, +{0x000111D0, 0x000111D9}, {0x000111E1, 0x000111F4}, {0x000112F0, 0x000112F9}, {0x00011450, 0x00011459}, +{0x000114D0, 0x000114D9}, {0x00011650, 0x00011659}, {0x000116C0, 0x000116C9}, {0x00011730, 0x0001173B}, +{0x000118E0, 0x000118F2}, {0x00011950, 0x00011959}, {0x00011C50, 0x00011C6C}, {0x00011D50, 0x00011D59}, +{0x00011DA0, 0x00011DA9}, {0x00011F50, 0x00011F59}, {0x00011FC0, 0x00011FD4}, {0x00012400, 0x0001246E}, +{0x00016A60, 0x00016A69}, {0x00016AC0, 0x00016AC9}, {0x00016B50, 0x00016B59}, {0x00016B5B, 0x00016B61}, +{0x00016E80, 0x00016E96}, {0x0001D2C0, 0x0001D2D3}, {0x0001D2E0, 0x0001D2F3}, {0x0001D360, 0x0001D378}, +{0x0001D7CE, 0x0001D7FF}, {0x0001E140, 0x0001E149}, {0x0001E2F0, 0x0001E2F9}, {0x0001E4F0, 0x0001E4F9}, +{0x0001E8C7, 0x0001E8CF}, {0x0001E950, 0x0001E959}, {0x0001EC71, 0x0001ECAB}, {0x0001ECAD, 0x0001ECAF}, +{0x0001ECB1, 0x0001ECB4}, {0x0001ED01, 0x0001ED2D}, {0x0001ED2F, 0x0001ED3D}, {0x0001F100, 0x0001F10C}, +{0x0001FBF0, 0x0001FBF9}, }; const std::vector> unicode_ranges_letter = { @@ -43,73 +61,73 @@ const std::vector> unicode_ranges_letter = { {0x00000710, 0x00000710}, {0x00000712, 0x0000072F}, {0x0000074D, 0x000007A5}, {0x000007B1, 0x000007B1}, {0x000007CA, 0x000007EA}, {0x000007F4, 0x000007F5}, {0x000007FA, 0x000007FA}, {0x00000800, 0x00000815}, {0x0000081A, 0x0000081A}, {0x00000824, 0x00000824}, {0x00000828, 0x00000828}, {0x00000840, 0x00000858}, -{0x00000860, 0x0000086A}, {0x000008A0, 0x000008B4}, {0x000008B6, 0x000008C7}, {0x00000904, 0x00000939}, -{0x0000093D, 0x0000093D}, {0x00000950, 0x00000950}, {0x00000958, 0x00000961}, {0x00000971, 0x00000980}, -{0x00000985, 0x0000098C}, {0x0000098F, 0x00000990}, {0x00000993, 0x000009A8}, {0x000009AA, 0x000009B0}, -{0x000009B2, 0x000009B2}, {0x000009B6, 0x000009B9}, {0x000009BD, 0x000009BD}, {0x000009CE, 0x000009CE}, -{0x000009DC, 0x000009DD}, {0x000009DF, 0x000009E1}, {0x000009F0, 0x000009F1}, {0x000009FC, 0x000009FC}, -{0x00000A05, 0x00000A0A}, {0x00000A0F, 0x00000A10}, {0x00000A13, 0x00000A28}, {0x00000A2A, 0x00000A30}, -{0x00000A32, 0x00000A33}, {0x00000A35, 0x00000A36}, {0x00000A38, 0x00000A39}, {0x00000A59, 0x00000A5C}, -{0x00000A5E, 0x00000A5E}, {0x00000A72, 0x00000A74}, {0x00000A85, 0x00000A8D}, {0x00000A8F, 0x00000A91}, -{0x00000A93, 0x00000AA8}, {0x00000AAA, 0x00000AB0}, {0x00000AB2, 0x00000AB3}, {0x00000AB5, 0x00000AB9}, -{0x00000ABD, 0x00000ABD}, {0x00000AD0, 0x00000AD0}, {0x00000AE0, 0x00000AE1}, {0x00000AF9, 0x00000AF9}, -{0x00000B05, 0x00000B0C}, {0x00000B0F, 0x00000B10}, {0x00000B13, 0x00000B28}, {0x00000B2A, 0x00000B30}, -{0x00000B32, 0x00000B33}, {0x00000B35, 0x00000B39}, {0x00000B3D, 0x00000B3D}, {0x00000B5C, 0x00000B5D}, -{0x00000B5F, 0x00000B61}, {0x00000B71, 0x00000B71}, {0x00000B83, 0x00000B83}, {0x00000B85, 0x00000B8A}, -{0x00000B8E, 0x00000B90}, {0x00000B92, 0x00000B95}, {0x00000B99, 0x00000B9A}, {0x00000B9C, 0x00000B9C}, -{0x00000B9E, 0x00000B9F}, {0x00000BA3, 0x00000BA4}, {0x00000BA8, 0x00000BAA}, {0x00000BAE, 0x00000BB9}, -{0x00000BD0, 0x00000BD0}, {0x00000C05, 0x00000C0C}, {0x00000C0E, 0x00000C10}, {0x00000C12, 0x00000C28}, -{0x00000C2A, 0x00000C39}, {0x00000C3D, 0x00000C3D}, {0x00000C58, 0x00000C5A}, {0x00000C60, 0x00000C61}, -{0x00000C80, 0x00000C80}, {0x00000C85, 0x00000C8C}, {0x00000C8E, 0x00000C90}, {0x00000C92, 0x00000CA8}, -{0x00000CAA, 0x00000CB3}, {0x00000CB5, 0x00000CB9}, {0x00000CBD, 0x00000CBD}, {0x00000CDE, 0x00000CDE}, -{0x00000CE0, 0x00000CE1}, {0x00000CF1, 0x00000CF2}, {0x00000D04, 0x00000D0C}, {0x00000D0E, 0x00000D10}, -{0x00000D12, 0x00000D3A}, {0x00000D3D, 0x00000D3D}, {0x00000D4E, 0x00000D4E}, {0x00000D54, 0x00000D56}, -{0x00000D5F, 0x00000D61}, {0x00000D7A, 0x00000D7F}, {0x00000D85, 0x00000D96}, {0x00000D9A, 0x00000DB1}, -{0x00000DB3, 0x00000DBB}, {0x00000DBD, 0x00000DBD}, {0x00000DC0, 0x00000DC6}, {0x00000E01, 0x00000E30}, -{0x00000E32, 0x00000E33}, {0x00000E40, 0x00000E46}, {0x00000E81, 0x00000E82}, {0x00000E84, 0x00000E84}, -{0x00000E86, 0x00000E8A}, {0x00000E8C, 0x00000EA3}, {0x00000EA5, 0x00000EA5}, {0x00000EA7, 0x00000EB0}, -{0x00000EB2, 0x00000EB3}, {0x00000EBD, 0x00000EBD}, {0x00000EC0, 0x00000EC4}, {0x00000EC6, 0x00000EC6}, -{0x00000EDC, 0x00000EDF}, {0x00000F00, 0x00000F00}, {0x00000F40, 0x00000F47}, {0x00000F49, 0x00000F6C}, -{0x00000F88, 0x00000F8C}, {0x00001000, 0x0000102A}, {0x0000103F, 0x0000103F}, {0x00001050, 0x00001055}, -{0x0000105A, 0x0000105D}, {0x00001061, 0x00001061}, {0x00001065, 0x00001066}, {0x0000106E, 0x00001070}, -{0x00001075, 0x00001081}, {0x0000108E, 0x0000108E}, {0x000010A0, 0x000010C5}, {0x000010C7, 0x000010C7}, -{0x000010CD, 0x000010CD}, {0x000010D0, 0x000010FA}, {0x000010FC, 0x00001248}, {0x0000124A, 0x0000124D}, -{0x00001250, 0x00001256}, {0x00001258, 0x00001258}, {0x0000125A, 0x0000125D}, {0x00001260, 0x00001288}, -{0x0000128A, 0x0000128D}, {0x00001290, 0x000012B0}, {0x000012B2, 0x000012B5}, {0x000012B8, 0x000012BE}, -{0x000012C0, 0x000012C0}, {0x000012C2, 0x000012C5}, {0x000012C8, 0x000012D6}, {0x000012D8, 0x00001310}, -{0x00001312, 0x00001315}, {0x00001318, 0x0000135A}, {0x00001380, 0x0000138F}, {0x000013A0, 0x000013F5}, -{0x000013F8, 0x000013FD}, {0x00001401, 0x0000166C}, {0x0000166F, 0x0000167F}, {0x00001681, 0x0000169A}, -{0x000016A0, 0x000016EA}, {0x000016F1, 0x000016F8}, {0x00001700, 0x0000170C}, {0x0000170E, 0x00001711}, -{0x00001720, 0x00001731}, {0x00001740, 0x00001751}, {0x00001760, 0x0000176C}, {0x0000176E, 0x00001770}, -{0x00001780, 0x000017B3}, {0x000017D7, 0x000017D7}, {0x000017DC, 0x000017DC}, {0x00001820, 0x00001878}, -{0x00001880, 0x00001884}, {0x00001887, 0x000018A8}, {0x000018AA, 0x000018AA}, {0x000018B0, 0x000018F5}, -{0x00001900, 0x0000191E}, {0x00001950, 0x0000196D}, {0x00001970, 0x00001974}, {0x00001980, 0x000019AB}, -{0x000019B0, 0x000019C9}, {0x00001A00, 0x00001A16}, {0x00001A20, 0x00001A54}, {0x00001AA7, 0x00001AA7}, -{0x00001B05, 0x00001B33}, {0x00001B45, 0x00001B4B}, {0x00001B83, 0x00001BA0}, {0x00001BAE, 0x00001BAF}, -{0x00001BBA, 0x00001BE5}, {0x00001C00, 0x00001C23}, {0x00001C4D, 0x00001C4F}, {0x00001C5A, 0x00001C7D}, -{0x00001C80, 0x00001C88}, {0x00001C90, 0x00001CBA}, {0x00001CBD, 0x00001CBF}, {0x00001CE9, 0x00001CEC}, -{0x00001CEE, 0x00001CF3}, {0x00001CF5, 0x00001CF6}, {0x00001CFA, 0x00001CFA}, {0x00001D00, 0x00001DBF}, -{0x00001E00, 0x00001F15}, {0x00001F18, 0x00001F1D}, {0x00001F20, 0x00001F45}, {0x00001F48, 0x00001F4D}, -{0x00001F50, 0x00001F57}, {0x00001F59, 0x00001F59}, {0x00001F5B, 0x00001F5B}, {0x00001F5D, 0x00001F5D}, -{0x00001F5F, 0x00001F7D}, {0x00001F80, 0x00001FB4}, {0x00001FB6, 0x00001FBC}, {0x00001FBE, 0x00001FBE}, -{0x00001FC2, 0x00001FC4}, {0x00001FC6, 0x00001FCC}, {0x00001FD0, 0x00001FD3}, {0x00001FD6, 0x00001FDB}, -{0x00001FE0, 0x00001FEC}, {0x00001FF2, 0x00001FF4}, {0x00001FF6, 0x00001FFC}, {0x00002071, 0x00002071}, -{0x0000207F, 0x0000207F}, {0x00002090, 0x0000209C}, {0x00002102, 0x00002102}, {0x00002107, 0x00002107}, -{0x0000210A, 0x00002113}, {0x00002115, 0x00002115}, {0x00002119, 0x0000211D}, {0x00002124, 0x00002124}, -{0x00002126, 0x00002126}, {0x00002128, 0x00002128}, {0x0000212A, 0x0000212D}, {0x0000212F, 0x00002139}, -{0x0000213C, 0x0000213F}, {0x00002145, 0x00002149}, {0x0000214E, 0x0000214E}, {0x00002183, 0x00002184}, -{0x00002C00, 0x00002C2E}, {0x00002C30, 0x00002C5E}, {0x00002C60, 0x00002CE4}, {0x00002CEB, 0x00002CEE}, -{0x00002CF2, 0x00002CF3}, {0x00002D00, 0x00002D25}, {0x00002D27, 0x00002D27}, {0x00002D2D, 0x00002D2D}, -{0x00002D30, 0x00002D67}, {0x00002D6F, 0x00002D6F}, {0x00002D80, 0x00002D96}, {0x00002DA0, 0x00002DA6}, -{0x00002DA8, 0x00002DAE}, {0x00002DB0, 0x00002DB6}, {0x00002DB8, 0x00002DBE}, {0x00002DC0, 0x00002DC6}, -{0x00002DC8, 0x00002DCE}, {0x00002DD0, 0x00002DD6}, {0x00002DD8, 0x00002DDE}, {0x00002E2F, 0x00002E2F}, -{0x00003005, 0x00003006}, {0x00003031, 0x00003035}, {0x0000303B, 0x0000303C}, {0x00003041, 0x00003096}, -{0x0000309D, 0x0000309F}, {0x000030A1, 0x000030FA}, {0x000030FC, 0x000030FF}, {0x00003105, 0x0000312F}, -{0x00003131, 0x0000318E}, {0x000031A0, 0x000031BF}, {0x000031F0, 0x000031FF}, {0x00003400, 0x00004DBF}, -{0x00004E00, 0x00009FFC}, {0x0000A000, 0x0000A48C}, {0x0000A4D0, 0x0000A4FD}, {0x0000A500, 0x0000A60C}, -{0x0000A610, 0x0000A61F}, {0x0000A62A, 0x0000A62B}, {0x0000A640, 0x0000A66E}, {0x0000A67F, 0x0000A69D}, -{0x0000A6A0, 0x0000A6E5}, {0x0000A717, 0x0000A71F}, {0x0000A722, 0x0000A788}, {0x0000A78B, 0x0000A7BF}, -{0x0000A7C2, 0x0000A7CA}, {0x0000A7F5, 0x0000A801}, {0x0000A803, 0x0000A805}, {0x0000A807, 0x0000A80A}, +{0x00000860, 0x0000086A}, {0x00000870, 0x00000887}, {0x00000889, 0x0000088E}, {0x000008A0, 0x000008C9}, +{0x00000904, 0x00000939}, {0x0000093D, 0x0000093D}, {0x00000950, 0x00000950}, {0x00000958, 0x00000961}, +{0x00000971, 0x00000980}, {0x00000985, 0x0000098C}, {0x0000098F, 0x00000990}, {0x00000993, 0x000009A8}, +{0x000009AA, 0x000009B0}, {0x000009B2, 0x000009B2}, {0x000009B6, 0x000009B9}, {0x000009BD, 0x000009BD}, +{0x000009CE, 0x000009CE}, {0x000009DC, 0x000009DD}, {0x000009DF, 0x000009E1}, {0x000009F0, 0x000009F1}, +{0x000009FC, 0x000009FC}, {0x00000A05, 0x00000A0A}, {0x00000A0F, 0x00000A10}, {0x00000A13, 0x00000A28}, +{0x00000A2A, 0x00000A30}, {0x00000A32, 0x00000A33}, {0x00000A35, 0x00000A36}, {0x00000A38, 0x00000A39}, +{0x00000A59, 0x00000A5C}, {0x00000A5E, 0x00000A5E}, {0x00000A72, 0x00000A74}, {0x00000A85, 0x00000A8D}, +{0x00000A8F, 0x00000A91}, {0x00000A93, 0x00000AA8}, {0x00000AAA, 0x00000AB0}, {0x00000AB2, 0x00000AB3}, +{0x00000AB5, 0x00000AB9}, {0x00000ABD, 0x00000ABD}, {0x00000AD0, 0x00000AD0}, {0x00000AE0, 0x00000AE1}, +{0x00000AF9, 0x00000AF9}, {0x00000B05, 0x00000B0C}, {0x00000B0F, 0x00000B10}, {0x00000B13, 0x00000B28}, +{0x00000B2A, 0x00000B30}, {0x00000B32, 0x00000B33}, {0x00000B35, 0x00000B39}, {0x00000B3D, 0x00000B3D}, +{0x00000B5C, 0x00000B5D}, {0x00000B5F, 0x00000B61}, {0x00000B71, 0x00000B71}, {0x00000B83, 0x00000B83}, +{0x00000B85, 0x00000B8A}, {0x00000B8E, 0x00000B90}, {0x00000B92, 0x00000B95}, {0x00000B99, 0x00000B9A}, +{0x00000B9C, 0x00000B9C}, {0x00000B9E, 0x00000B9F}, {0x00000BA3, 0x00000BA4}, {0x00000BA8, 0x00000BAA}, +{0x00000BAE, 0x00000BB9}, {0x00000BD0, 0x00000BD0}, {0x00000C05, 0x00000C0C}, {0x00000C0E, 0x00000C10}, +{0x00000C12, 0x00000C28}, {0x00000C2A, 0x00000C39}, {0x00000C3D, 0x00000C3D}, {0x00000C58, 0x00000C5A}, +{0x00000C5D, 0x00000C5D}, {0x00000C60, 0x00000C61}, {0x00000C80, 0x00000C80}, {0x00000C85, 0x00000C8C}, +{0x00000C8E, 0x00000C90}, {0x00000C92, 0x00000CA8}, {0x00000CAA, 0x00000CB3}, {0x00000CB5, 0x00000CB9}, +{0x00000CBD, 0x00000CBD}, {0x00000CDD, 0x00000CDE}, {0x00000CE0, 0x00000CE1}, {0x00000CF1, 0x00000CF2}, +{0x00000D04, 0x00000D0C}, {0x00000D0E, 0x00000D10}, {0x00000D12, 0x00000D3A}, {0x00000D3D, 0x00000D3D}, +{0x00000D4E, 0x00000D4E}, {0x00000D54, 0x00000D56}, {0x00000D5F, 0x00000D61}, {0x00000D7A, 0x00000D7F}, +{0x00000D85, 0x00000D96}, {0x00000D9A, 0x00000DB1}, {0x00000DB3, 0x00000DBB}, {0x00000DBD, 0x00000DBD}, +{0x00000DC0, 0x00000DC6}, {0x00000E01, 0x00000E30}, {0x00000E32, 0x00000E33}, {0x00000E40, 0x00000E46}, +{0x00000E81, 0x00000E82}, {0x00000E84, 0x00000E84}, {0x00000E86, 0x00000E8A}, {0x00000E8C, 0x00000EA3}, +{0x00000EA5, 0x00000EA5}, {0x00000EA7, 0x00000EB0}, {0x00000EB2, 0x00000EB3}, {0x00000EBD, 0x00000EBD}, +{0x00000EC0, 0x00000EC4}, {0x00000EC6, 0x00000EC6}, {0x00000EDC, 0x00000EDF}, {0x00000F00, 0x00000F00}, +{0x00000F40, 0x00000F47}, {0x00000F49, 0x00000F6C}, {0x00000F88, 0x00000F8C}, {0x00001000, 0x0000102A}, +{0x0000103F, 0x0000103F}, {0x00001050, 0x00001055}, {0x0000105A, 0x0000105D}, {0x00001061, 0x00001061}, +{0x00001065, 0x00001066}, {0x0000106E, 0x00001070}, {0x00001075, 0x00001081}, {0x0000108E, 0x0000108E}, +{0x000010A0, 0x000010C5}, {0x000010C7, 0x000010C7}, {0x000010CD, 0x000010CD}, {0x000010D0, 0x000010FA}, +{0x000010FC, 0x00001248}, {0x0000124A, 0x0000124D}, {0x00001250, 0x00001256}, {0x00001258, 0x00001258}, +{0x0000125A, 0x0000125D}, {0x00001260, 0x00001288}, {0x0000128A, 0x0000128D}, {0x00001290, 0x000012B0}, +{0x000012B2, 0x000012B5}, {0x000012B8, 0x000012BE}, {0x000012C0, 0x000012C0}, {0x000012C2, 0x000012C5}, +{0x000012C8, 0x000012D6}, {0x000012D8, 0x00001310}, {0x00001312, 0x00001315}, {0x00001318, 0x0000135A}, +{0x00001380, 0x0000138F}, {0x000013A0, 0x000013F5}, {0x000013F8, 0x000013FD}, {0x00001401, 0x0000166C}, +{0x0000166F, 0x0000167F}, {0x00001681, 0x0000169A}, {0x000016A0, 0x000016EA}, {0x000016F1, 0x000016F8}, +{0x00001700, 0x00001711}, {0x0000171F, 0x00001731}, {0x00001740, 0x00001751}, {0x00001760, 0x0000176C}, +{0x0000176E, 0x00001770}, {0x00001780, 0x000017B3}, {0x000017D7, 0x000017D7}, {0x000017DC, 0x000017DC}, +{0x00001820, 0x00001878}, {0x00001880, 0x00001884}, {0x00001887, 0x000018A8}, {0x000018AA, 0x000018AA}, +{0x000018B0, 0x000018F5}, {0x00001900, 0x0000191E}, {0x00001950, 0x0000196D}, {0x00001970, 0x00001974}, +{0x00001980, 0x000019AB}, {0x000019B0, 0x000019C9}, {0x00001A00, 0x00001A16}, {0x00001A20, 0x00001A54}, +{0x00001AA7, 0x00001AA7}, {0x00001B05, 0x00001B33}, {0x00001B45, 0x00001B4C}, {0x00001B83, 0x00001BA0}, +{0x00001BAE, 0x00001BAF}, {0x00001BBA, 0x00001BE5}, {0x00001C00, 0x00001C23}, {0x00001C4D, 0x00001C4F}, +{0x00001C5A, 0x00001C7D}, {0x00001C80, 0x00001C88}, {0x00001C90, 0x00001CBA}, {0x00001CBD, 0x00001CBF}, +{0x00001CE9, 0x00001CEC}, {0x00001CEE, 0x00001CF3}, {0x00001CF5, 0x00001CF6}, {0x00001CFA, 0x00001CFA}, +{0x00001D00, 0x00001DBF}, {0x00001E00, 0x00001F15}, {0x00001F18, 0x00001F1D}, {0x00001F20, 0x00001F45}, +{0x00001F48, 0x00001F4D}, {0x00001F50, 0x00001F57}, {0x00001F59, 0x00001F59}, {0x00001F5B, 0x00001F5B}, +{0x00001F5D, 0x00001F5D}, {0x00001F5F, 0x00001F7D}, {0x00001F80, 0x00001FB4}, {0x00001FB6, 0x00001FBC}, +{0x00001FBE, 0x00001FBE}, {0x00001FC2, 0x00001FC4}, {0x00001FC6, 0x00001FCC}, {0x00001FD0, 0x00001FD3}, +{0x00001FD6, 0x00001FDB}, {0x00001FE0, 0x00001FEC}, {0x00001FF2, 0x00001FF4}, {0x00001FF6, 0x00001FFC}, +{0x00002071, 0x00002071}, {0x0000207F, 0x0000207F}, {0x00002090, 0x0000209C}, {0x00002102, 0x00002102}, +{0x00002107, 0x00002107}, {0x0000210A, 0x00002113}, {0x00002115, 0x00002115}, {0x00002119, 0x0000211D}, +{0x00002124, 0x00002124}, {0x00002126, 0x00002126}, {0x00002128, 0x00002128}, {0x0000212A, 0x0000212D}, +{0x0000212F, 0x00002139}, {0x0000213C, 0x0000213F}, {0x00002145, 0x00002149}, {0x0000214E, 0x0000214E}, +{0x00002183, 0x00002184}, {0x00002C00, 0x00002CE4}, {0x00002CEB, 0x00002CEE}, {0x00002CF2, 0x00002CF3}, +{0x00002D00, 0x00002D25}, {0x00002D27, 0x00002D27}, {0x00002D2D, 0x00002D2D}, {0x00002D30, 0x00002D67}, +{0x00002D6F, 0x00002D6F}, {0x00002D80, 0x00002D96}, {0x00002DA0, 0x00002DA6}, {0x00002DA8, 0x00002DAE}, +{0x00002DB0, 0x00002DB6}, {0x00002DB8, 0x00002DBE}, {0x00002DC0, 0x00002DC6}, {0x00002DC8, 0x00002DCE}, +{0x00002DD0, 0x00002DD6}, {0x00002DD8, 0x00002DDE}, {0x00002E2F, 0x00002E2F}, {0x00003005, 0x00003006}, +{0x00003031, 0x00003035}, {0x0000303B, 0x0000303C}, {0x00003041, 0x00003096}, {0x0000309D, 0x0000309F}, +{0x000030A1, 0x000030FA}, {0x000030FC, 0x000030FF}, {0x00003105, 0x0000312F}, {0x00003131, 0x0000318E}, +{0x000031A0, 0x000031BF}, {0x000031F0, 0x000031FF}, {0x00003400, 0x00004DBF}, {0x00004E00, 0x0000A48C}, +{0x0000A4D0, 0x0000A4FD}, {0x0000A500, 0x0000A60C}, {0x0000A610, 0x0000A61F}, {0x0000A62A, 0x0000A62B}, +{0x0000A640, 0x0000A66E}, {0x0000A67F, 0x0000A69D}, {0x0000A6A0, 0x0000A6E5}, {0x0000A717, 0x0000A71F}, +{0x0000A722, 0x0000A788}, {0x0000A78B, 0x0000A7CA}, {0x0000A7D0, 0x0000A7D1}, {0x0000A7D3, 0x0000A7D3}, +{0x0000A7D5, 0x0000A7D9}, {0x0000A7F2, 0x0000A801}, {0x0000A803, 0x0000A805}, {0x0000A807, 0x0000A80A}, {0x0000A80C, 0x0000A822}, {0x0000A840, 0x0000A873}, {0x0000A882, 0x0000A8B3}, {0x0000A8F2, 0x0000A8F7}, {0x0000A8FB, 0x0000A8FB}, {0x0000A8FD, 0x0000A8FE}, {0x0000A90A, 0x0000A925}, {0x0000A930, 0x0000A946}, {0x0000A960, 0x0000A97C}, {0x0000A984, 0x0000A9B2}, {0x0000A9CF, 0x0000A9CF}, {0x0000A9E0, 0x0000A9E4}, @@ -131,51 +149,60 @@ const std::vector> unicode_ranges_letter = { {0x000102A0, 0x000102D0}, {0x00010300, 0x0001031F}, {0x0001032D, 0x00010340}, {0x00010342, 0x00010349}, {0x00010350, 0x00010375}, {0x00010380, 0x0001039D}, {0x000103A0, 0x000103C3}, {0x000103C8, 0x000103CF}, {0x00010400, 0x0001049D}, {0x000104B0, 0x000104D3}, {0x000104D8, 0x000104FB}, {0x00010500, 0x00010527}, -{0x00010530, 0x00010563}, {0x00010600, 0x00010736}, {0x00010740, 0x00010755}, {0x00010760, 0x00010767}, -{0x00010800, 0x00010805}, {0x00010808, 0x00010808}, {0x0001080A, 0x00010835}, {0x00010837, 0x00010838}, -{0x0001083C, 0x0001083C}, {0x0001083F, 0x00010855}, {0x00010860, 0x00010876}, {0x00010880, 0x0001089E}, -{0x000108E0, 0x000108F2}, {0x000108F4, 0x000108F5}, {0x00010900, 0x00010915}, {0x00010920, 0x00010939}, -{0x00010980, 0x000109B7}, {0x000109BE, 0x000109BF}, {0x00010A00, 0x00010A00}, {0x00010A10, 0x00010A13}, -{0x00010A15, 0x00010A17}, {0x00010A19, 0x00010A35}, {0x00010A60, 0x00010A7C}, {0x00010A80, 0x00010A9C}, -{0x00010AC0, 0x00010AC7}, {0x00010AC9, 0x00010AE4}, {0x00010B00, 0x00010B35}, {0x00010B40, 0x00010B55}, -{0x00010B60, 0x00010B72}, {0x00010B80, 0x00010B91}, {0x00010C00, 0x00010C48}, {0x00010C80, 0x00010CB2}, -{0x00010CC0, 0x00010CF2}, {0x00010D00, 0x00010D23}, {0x00010E80, 0x00010EA9}, {0x00010EB0, 0x00010EB1}, -{0x00010F00, 0x00010F1C}, {0x00010F27, 0x00010F27}, {0x00010F30, 0x00010F45}, {0x00010FB0, 0x00010FC4}, -{0x00010FE0, 0x00010FF6}, {0x00011003, 0x00011037}, {0x00011083, 0x000110AF}, {0x000110D0, 0x000110E8}, -{0x00011103, 0x00011126}, {0x00011144, 0x00011144}, {0x00011147, 0x00011147}, {0x00011150, 0x00011172}, -{0x00011176, 0x00011176}, {0x00011183, 0x000111B2}, {0x000111C1, 0x000111C4}, {0x000111DA, 0x000111DA}, -{0x000111DC, 0x000111DC}, {0x00011200, 0x00011211}, {0x00011213, 0x0001122B}, {0x00011280, 0x00011286}, -{0x00011288, 0x00011288}, {0x0001128A, 0x0001128D}, {0x0001128F, 0x0001129D}, {0x0001129F, 0x000112A8}, -{0x000112B0, 0x000112DE}, {0x00011305, 0x0001130C}, {0x0001130F, 0x00011310}, {0x00011313, 0x00011328}, -{0x0001132A, 0x00011330}, {0x00011332, 0x00011333}, {0x00011335, 0x00011339}, {0x0001133D, 0x0001133D}, -{0x00011350, 0x00011350}, {0x0001135D, 0x00011361}, {0x00011400, 0x00011434}, {0x00011447, 0x0001144A}, -{0x0001145F, 0x00011461}, {0x00011480, 0x000114AF}, {0x000114C4, 0x000114C5}, {0x000114C7, 0x000114C7}, -{0x00011580, 0x000115AE}, {0x000115D8, 0x000115DB}, {0x00011600, 0x0001162F}, {0x00011644, 0x00011644}, -{0x00011680, 0x000116AA}, {0x000116B8, 0x000116B8}, {0x00011700, 0x0001171A}, {0x00011800, 0x0001182B}, +{0x00010530, 0x00010563}, {0x00010570, 0x0001057A}, {0x0001057C, 0x0001058A}, {0x0001058C, 0x00010592}, +{0x00010594, 0x00010595}, {0x00010597, 0x000105A1}, {0x000105A3, 0x000105B1}, {0x000105B3, 0x000105B9}, +{0x000105BB, 0x000105BC}, {0x00010600, 0x00010736}, {0x00010740, 0x00010755}, {0x00010760, 0x00010767}, +{0x00010780, 0x00010785}, {0x00010787, 0x000107B0}, {0x000107B2, 0x000107BA}, {0x00010800, 0x00010805}, +{0x00010808, 0x00010808}, {0x0001080A, 0x00010835}, {0x00010837, 0x00010838}, {0x0001083C, 0x0001083C}, +{0x0001083F, 0x00010855}, {0x00010860, 0x00010876}, {0x00010880, 0x0001089E}, {0x000108E0, 0x000108F2}, +{0x000108F4, 0x000108F5}, {0x00010900, 0x00010915}, {0x00010920, 0x00010939}, {0x00010980, 0x000109B7}, +{0x000109BE, 0x000109BF}, {0x00010A00, 0x00010A00}, {0x00010A10, 0x00010A13}, {0x00010A15, 0x00010A17}, +{0x00010A19, 0x00010A35}, {0x00010A60, 0x00010A7C}, {0x00010A80, 0x00010A9C}, {0x00010AC0, 0x00010AC7}, +{0x00010AC9, 0x00010AE4}, {0x00010B00, 0x00010B35}, {0x00010B40, 0x00010B55}, {0x00010B60, 0x00010B72}, +{0x00010B80, 0x00010B91}, {0x00010C00, 0x00010C48}, {0x00010C80, 0x00010CB2}, {0x00010CC0, 0x00010CF2}, +{0x00010D00, 0x00010D23}, {0x00010E80, 0x00010EA9}, {0x00010EB0, 0x00010EB1}, {0x00010F00, 0x00010F1C}, +{0x00010F27, 0x00010F27}, {0x00010F30, 0x00010F45}, {0x00010F70, 0x00010F81}, {0x00010FB0, 0x00010FC4}, +{0x00010FE0, 0x00010FF6}, {0x00011003, 0x00011037}, {0x00011071, 0x00011072}, {0x00011075, 0x00011075}, +{0x00011083, 0x000110AF}, {0x000110D0, 0x000110E8}, {0x00011103, 0x00011126}, {0x00011144, 0x00011144}, +{0x00011147, 0x00011147}, {0x00011150, 0x00011172}, {0x00011176, 0x00011176}, {0x00011183, 0x000111B2}, +{0x000111C1, 0x000111C4}, {0x000111DA, 0x000111DA}, {0x000111DC, 0x000111DC}, {0x00011200, 0x00011211}, +{0x00011213, 0x0001122B}, {0x0001123F, 0x00011240}, {0x00011280, 0x00011286}, {0x00011288, 0x00011288}, +{0x0001128A, 0x0001128D}, {0x0001128F, 0x0001129D}, {0x0001129F, 0x000112A8}, {0x000112B0, 0x000112DE}, +{0x00011305, 0x0001130C}, {0x0001130F, 0x00011310}, {0x00011313, 0x00011328}, {0x0001132A, 0x00011330}, +{0x00011332, 0x00011333}, {0x00011335, 0x00011339}, {0x0001133D, 0x0001133D}, {0x00011350, 0x00011350}, +{0x0001135D, 0x00011361}, {0x00011400, 0x00011434}, {0x00011447, 0x0001144A}, {0x0001145F, 0x00011461}, +{0x00011480, 0x000114AF}, {0x000114C4, 0x000114C5}, {0x000114C7, 0x000114C7}, {0x00011580, 0x000115AE}, +{0x000115D8, 0x000115DB}, {0x00011600, 0x0001162F}, {0x00011644, 0x00011644}, {0x00011680, 0x000116AA}, +{0x000116B8, 0x000116B8}, {0x00011700, 0x0001171A}, {0x00011740, 0x00011746}, {0x00011800, 0x0001182B}, {0x000118A0, 0x000118DF}, {0x000118FF, 0x00011906}, {0x00011909, 0x00011909}, {0x0001190C, 0x00011913}, {0x00011915, 0x00011916}, {0x00011918, 0x0001192F}, {0x0001193F, 0x0001193F}, {0x00011941, 0x00011941}, {0x000119A0, 0x000119A7}, {0x000119AA, 0x000119D0}, {0x000119E1, 0x000119E1}, {0x000119E3, 0x000119E3}, {0x00011A00, 0x00011A00}, {0x00011A0B, 0x00011A32}, {0x00011A3A, 0x00011A3A}, {0x00011A50, 0x00011A50}, -{0x00011A5C, 0x00011A89}, {0x00011A9D, 0x00011A9D}, {0x00011AC0, 0x00011AF8}, {0x00011C00, 0x00011C08}, +{0x00011A5C, 0x00011A89}, {0x00011A9D, 0x00011A9D}, {0x00011AB0, 0x00011AF8}, {0x00011C00, 0x00011C08}, {0x00011C0A, 0x00011C2E}, {0x00011C40, 0x00011C40}, {0x00011C72, 0x00011C8F}, {0x00011D00, 0x00011D06}, {0x00011D08, 0x00011D09}, {0x00011D0B, 0x00011D30}, {0x00011D46, 0x00011D46}, {0x00011D60, 0x00011D65}, {0x00011D67, 0x00011D68}, {0x00011D6A, 0x00011D89}, {0x00011D98, 0x00011D98}, {0x00011EE0, 0x00011EF2}, -{0x00011FB0, 0x00011FB0}, {0x00012000, 0x00012399}, {0x00012480, 0x00012543}, {0x00013000, 0x0001342E}, -{0x00014400, 0x00014646}, {0x00016800, 0x00016A38}, {0x00016A40, 0x00016A5E}, {0x00016AD0, 0x00016AED}, -{0x00016B00, 0x00016B2F}, {0x00016B40, 0x00016B43}, {0x00016B63, 0x00016B77}, {0x00016B7D, 0x00016B8F}, -{0x00016E40, 0x00016E7F}, {0x00016F00, 0x00016F4A}, {0x00016F50, 0x00016F50}, {0x00016F93, 0x00016F9F}, -{0x00016FE0, 0x00016FE1}, {0x00016FE3, 0x00016FE3}, {0x00017000, 0x000187F7}, {0x00018800, 0x00018CD5}, -{0x00018D00, 0x00018D08}, {0x0001B000, 0x0001B11E}, {0x0001B150, 0x0001B152}, {0x0001B164, 0x0001B167}, -{0x0001B170, 0x0001B2FB}, {0x0001BC00, 0x0001BC6A}, {0x0001BC70, 0x0001BC7C}, {0x0001BC80, 0x0001BC88}, -{0x0001BC90, 0x0001BC99}, {0x0001D400, 0x0001D454}, {0x0001D456, 0x0001D49C}, {0x0001D49E, 0x0001D49F}, -{0x0001D4A2, 0x0001D4A2}, {0x0001D4A5, 0x0001D4A6}, {0x0001D4A9, 0x0001D4AC}, {0x0001D4AE, 0x0001D4B9}, -{0x0001D4BB, 0x0001D4BB}, {0x0001D4BD, 0x0001D4C3}, {0x0001D4C5, 0x0001D505}, {0x0001D507, 0x0001D50A}, -{0x0001D50D, 0x0001D514}, {0x0001D516, 0x0001D51C}, {0x0001D51E, 0x0001D539}, {0x0001D53B, 0x0001D53E}, -{0x0001D540, 0x0001D544}, {0x0001D546, 0x0001D546}, {0x0001D54A, 0x0001D550}, {0x0001D552, 0x0001D6A5}, -{0x0001D6A8, 0x0001D6C0}, {0x0001D6C2, 0x0001D6DA}, {0x0001D6DC, 0x0001D6FA}, {0x0001D6FC, 0x0001D714}, -{0x0001D716, 0x0001D734}, {0x0001D736, 0x0001D74E}, {0x0001D750, 0x0001D76E}, {0x0001D770, 0x0001D788}, -{0x0001D78A, 0x0001D7A8}, {0x0001D7AA, 0x0001D7C2}, {0x0001D7C4, 0x0001D7CB}, {0x0001E100, 0x0001E12C}, -{0x0001E137, 0x0001E13D}, {0x0001E14E, 0x0001E14E}, {0x0001E2C0, 0x0001E2EB}, {0x0001E800, 0x0001E8C4}, +{0x00011F02, 0x00011F02}, {0x00011F04, 0x00011F10}, {0x00011F12, 0x00011F33}, {0x00011FB0, 0x00011FB0}, +{0x00012000, 0x00012399}, {0x00012480, 0x00012543}, {0x00012F90, 0x00012FF0}, {0x00013000, 0x0001342F}, +{0x00013441, 0x00013446}, {0x00014400, 0x00014646}, {0x00016800, 0x00016A38}, {0x00016A40, 0x00016A5E}, +{0x00016A70, 0x00016ABE}, {0x00016AD0, 0x00016AED}, {0x00016B00, 0x00016B2F}, {0x00016B40, 0x00016B43}, +{0x00016B63, 0x00016B77}, {0x00016B7D, 0x00016B8F}, {0x00016E40, 0x00016E7F}, {0x00016F00, 0x00016F4A}, +{0x00016F50, 0x00016F50}, {0x00016F93, 0x00016F9F}, {0x00016FE0, 0x00016FE1}, {0x00016FE3, 0x00016FE3}, +{0x00017000, 0x000187F7}, {0x00018800, 0x00018CD5}, {0x00018D00, 0x00018D08}, {0x0001AFF0, 0x0001AFF3}, +{0x0001AFF5, 0x0001AFFB}, {0x0001AFFD, 0x0001AFFE}, {0x0001B000, 0x0001B122}, {0x0001B132, 0x0001B132}, +{0x0001B150, 0x0001B152}, {0x0001B155, 0x0001B155}, {0x0001B164, 0x0001B167}, {0x0001B170, 0x0001B2FB}, +{0x0001BC00, 0x0001BC6A}, {0x0001BC70, 0x0001BC7C}, {0x0001BC80, 0x0001BC88}, {0x0001BC90, 0x0001BC99}, +{0x0001D400, 0x0001D454}, {0x0001D456, 0x0001D49C}, {0x0001D49E, 0x0001D49F}, {0x0001D4A2, 0x0001D4A2}, +{0x0001D4A5, 0x0001D4A6}, {0x0001D4A9, 0x0001D4AC}, {0x0001D4AE, 0x0001D4B9}, {0x0001D4BB, 0x0001D4BB}, +{0x0001D4BD, 0x0001D4C3}, {0x0001D4C5, 0x0001D505}, {0x0001D507, 0x0001D50A}, {0x0001D50D, 0x0001D514}, +{0x0001D516, 0x0001D51C}, {0x0001D51E, 0x0001D539}, {0x0001D53B, 0x0001D53E}, {0x0001D540, 0x0001D544}, +{0x0001D546, 0x0001D546}, {0x0001D54A, 0x0001D550}, {0x0001D552, 0x0001D6A5}, {0x0001D6A8, 0x0001D6C0}, +{0x0001D6C2, 0x0001D6DA}, {0x0001D6DC, 0x0001D6FA}, {0x0001D6FC, 0x0001D714}, {0x0001D716, 0x0001D734}, +{0x0001D736, 0x0001D74E}, {0x0001D750, 0x0001D76E}, {0x0001D770, 0x0001D788}, {0x0001D78A, 0x0001D7A8}, +{0x0001D7AA, 0x0001D7C2}, {0x0001D7C4, 0x0001D7CB}, {0x0001DF00, 0x0001DF1E}, {0x0001DF25, 0x0001DF2A}, +{0x0001E030, 0x0001E06D}, {0x0001E100, 0x0001E12C}, {0x0001E137, 0x0001E13D}, {0x0001E14E, 0x0001E14E}, +{0x0001E290, 0x0001E2AD}, {0x0001E2C0, 0x0001E2EB}, {0x0001E4D0, 0x0001E4EB}, {0x0001E7E0, 0x0001E7E6}, +{0x0001E7E8, 0x0001E7EB}, {0x0001E7ED, 0x0001E7EE}, {0x0001E7F0, 0x0001E7FE}, {0x0001E800, 0x0001E8C4}, {0x0001E900, 0x0001E943}, {0x0001E94B, 0x0001E94B}, {0x0001EE00, 0x0001EE03}, {0x0001EE05, 0x0001EE1F}, {0x0001EE21, 0x0001EE22}, {0x0001EE24, 0x0001EE24}, {0x0001EE27, 0x0001EE27}, {0x0001EE29, 0x0001EE32}, {0x0001EE34, 0x0001EE37}, {0x0001EE39, 0x0001EE39}, {0x0001EE3B, 0x0001EE3B}, {0x0001EE42, 0x0001EE42}, @@ -184,29 +211,16 @@ const std::vector> unicode_ranges_letter = { {0x0001EE5B, 0x0001EE5B}, {0x0001EE5D, 0x0001EE5D}, {0x0001EE5F, 0x0001EE5F}, {0x0001EE61, 0x0001EE62}, {0x0001EE64, 0x0001EE64}, {0x0001EE67, 0x0001EE6A}, {0x0001EE6C, 0x0001EE72}, {0x0001EE74, 0x0001EE77}, {0x0001EE79, 0x0001EE7C}, {0x0001EE7E, 0x0001EE7E}, {0x0001EE80, 0x0001EE89}, {0x0001EE8B, 0x0001EE9B}, -{0x0001EEA1, 0x0001EEA3}, {0x0001EEA5, 0x0001EEA9}, {0x0001EEAB, 0x0001EEBB}, {0x00020000, 0x0002A6DD}, -{0x0002A700, 0x0002B734}, {0x0002B740, 0x0002B81D}, {0x0002B820, 0x0002CEA1}, {0x0002CEB0, 0x0002EBE0}, -{0x0002F800, 0x0002FA1D}, {0x00030000, 0x0003134A}, +{0x0001EEA1, 0x0001EEA3}, {0x0001EEA5, 0x0001EEA9}, {0x0001EEAB, 0x0001EEBB}, {0x00020000, 0x0002A6DF}, +{0x0002A700, 0x0002B739}, {0x0002B740, 0x0002B81D}, {0x0002B820, 0x0002CEA1}, {0x0002CEB0, 0x0002EBE0}, +{0x0002F800, 0x0002FA1D}, {0x00030000, 0x0003134A}, {0x00031350, 0x000323AF}, }; const std::vector> unicode_ranges_whitespace = { -{0x00000009, 0x0000000D}, {0x00000020, 0x00000020}, {0x00000085, 0x00000085}, {0x000000A0, 0x000000A0}, -{0x00001680, 0x00001680}, {0x00002000, 0x0000200A}, {0x00002028, 0x00002029}, {0x0000202F, 0x0000202F}, -{0x0000205F, 0x0000205F}, {0x00003000, 0x00003000}, +{0x00000020, 0x00000020}, {0x000000A0, 0x000000A0}, {0x00001680, 0x00001680}, {0x00002000, 0x0000200A}, +{0x00002028, 0x00002029}, {0x0000202F, 0x0000202F}, {0x0000205F, 0x0000205F}, {0x00003000, 0x00003000}, }; - -//const std::vector> unicode_ranges_separator_ = { -//{0x00000020, 0x00000020}, {0x000000A0, 0x000000A0}, {0x00001680, 0x00001680}, -//{0x00002000, 0x0000200A}, {0x00002028, 0x00002029}, {0x0000202F, 0x0000202F}, -//{0x0000205F, 0x0000205F}, {0x00003000, 0x00003000}, -//}; - -//const std::vector> unicode_ranges_whitespace_ = { -//{'\r', '\r'}, {'\n', '\n'}, {'\t', '\t'}, {'\f', '\f'}, {'\v', '\v'}, {' ', ' '}, -//}; - - const std::vector> unicode_ranges_accent_mark = { {0x00000300, 0x0000036F}, {0x00000483, 0x00000489}, {0x00000591, 0x000005BD}, {0x000005BF, 0x000005BF}, {0x000005C1, 0x000005C2}, {0x000005C4, 0x000005C5}, {0x000005C7, 0x000005C7}, {0x00000610, 0x0000061A}, @@ -214,72 +228,77 @@ const std::vector> unicode_ranges_accent_mark = { {0x000006E7, 0x000006E8}, {0x000006EA, 0x000006ED}, {0x00000711, 0x00000711}, {0x00000730, 0x0000074A}, {0x000007A6, 0x000007B0}, {0x000007EB, 0x000007F3}, {0x000007FD, 0x000007FD}, {0x00000816, 0x00000819}, {0x0000081B, 0x00000823}, {0x00000825, 0x00000827}, {0x00000829, 0x0000082D}, {0x00000859, 0x0000085B}, -{0x000008D3, 0x000008E1}, {0x000008E3, 0x00000903}, {0x0000093A, 0x0000093C}, {0x0000093E, 0x0000094F}, -{0x00000951, 0x00000957}, {0x00000962, 0x00000963}, {0x00000981, 0x00000983}, {0x000009BC, 0x000009BC}, -{0x000009BE, 0x000009C4}, {0x000009C7, 0x000009C8}, {0x000009CB, 0x000009CD}, {0x000009D7, 0x000009D7}, -{0x000009E2, 0x000009E3}, {0x000009FE, 0x000009FE}, {0x00000A01, 0x00000A03}, {0x00000A3C, 0x00000A3C}, -{0x00000A3E, 0x00000A42}, {0x00000A47, 0x00000A48}, {0x00000A4B, 0x00000A4D}, {0x00000A51, 0x00000A51}, -{0x00000A70, 0x00000A71}, {0x00000A75, 0x00000A75}, {0x00000A81, 0x00000A83}, {0x00000ABC, 0x00000ABC}, -{0x00000ABE, 0x00000AC5}, {0x00000AC7, 0x00000AC9}, {0x00000ACB, 0x00000ACD}, {0x00000AE2, 0x00000AE3}, -{0x00000AFA, 0x00000AFF}, {0x00000B01, 0x00000B03}, {0x00000B3C, 0x00000B3C}, {0x00000B3E, 0x00000B44}, -{0x00000B47, 0x00000B48}, {0x00000B4B, 0x00000B4D}, {0x00000B55, 0x00000B57}, {0x00000B62, 0x00000B63}, -{0x00000B82, 0x00000B82}, {0x00000BBE, 0x00000BC2}, {0x00000BC6, 0x00000BC8}, {0x00000BCA, 0x00000BCD}, -{0x00000BD7, 0x00000BD7}, {0x00000C00, 0x00000C04}, {0x00000C3E, 0x00000C44}, {0x00000C46, 0x00000C48}, -{0x00000C4A, 0x00000C4D}, {0x00000C55, 0x00000C56}, {0x00000C62, 0x00000C63}, {0x00000C81, 0x00000C83}, -{0x00000CBC, 0x00000CBC}, {0x00000CBE, 0x00000CC4}, {0x00000CC6, 0x00000CC8}, {0x00000CCA, 0x00000CCD}, -{0x00000CD5, 0x00000CD6}, {0x00000CE2, 0x00000CE3}, {0x00000D00, 0x00000D03}, {0x00000D3B, 0x00000D3C}, -{0x00000D3E, 0x00000D44}, {0x00000D46, 0x00000D48}, {0x00000D4A, 0x00000D4D}, {0x00000D57, 0x00000D57}, -{0x00000D62, 0x00000D63}, {0x00000D81, 0x00000D83}, {0x00000DCA, 0x00000DCA}, {0x00000DCF, 0x00000DD4}, -{0x00000DD6, 0x00000DD6}, {0x00000DD8, 0x00000DDF}, {0x00000DF2, 0x00000DF3}, {0x00000E31, 0x00000E31}, -{0x00000E34, 0x00000E3A}, {0x00000E47, 0x00000E4E}, {0x00000EB1, 0x00000EB1}, {0x00000EB4, 0x00000EBC}, -{0x00000EC8, 0x00000ECD}, {0x00000F18, 0x00000F19}, {0x00000F35, 0x00000F35}, {0x00000F37, 0x00000F37}, -{0x00000F39, 0x00000F39}, {0x00000F3E, 0x00000F3F}, {0x00000F71, 0x00000F84}, {0x00000F86, 0x00000F87}, -{0x00000F8D, 0x00000F97}, {0x00000F99, 0x00000FBC}, {0x00000FC6, 0x00000FC6}, {0x0000102B, 0x0000103E}, -{0x00001056, 0x00001059}, {0x0000105E, 0x00001060}, {0x00001062, 0x00001064}, {0x00001067, 0x0000106D}, -{0x00001071, 0x00001074}, {0x00001082, 0x0000108D}, {0x0000108F, 0x0000108F}, {0x0000109A, 0x0000109D}, -{0x0000135D, 0x0000135F}, {0x00001712, 0x00001714}, {0x00001732, 0x00001734}, {0x00001752, 0x00001753}, -{0x00001772, 0x00001773}, {0x000017B4, 0x000017D3}, {0x000017DD, 0x000017DD}, {0x0000180B, 0x0000180D}, +{0x00000898, 0x0000089F}, {0x000008CA, 0x000008E1}, {0x000008E3, 0x00000903}, {0x0000093A, 0x0000093C}, +{0x0000093E, 0x0000094F}, {0x00000951, 0x00000957}, {0x00000962, 0x00000963}, {0x00000981, 0x00000983}, +{0x000009BC, 0x000009BC}, {0x000009BE, 0x000009C4}, {0x000009C7, 0x000009C8}, {0x000009CB, 0x000009CD}, +{0x000009D7, 0x000009D7}, {0x000009E2, 0x000009E3}, {0x000009FE, 0x000009FE}, {0x00000A01, 0x00000A03}, +{0x00000A3C, 0x00000A3C}, {0x00000A3E, 0x00000A42}, {0x00000A47, 0x00000A48}, {0x00000A4B, 0x00000A4D}, +{0x00000A51, 0x00000A51}, {0x00000A70, 0x00000A71}, {0x00000A75, 0x00000A75}, {0x00000A81, 0x00000A83}, +{0x00000ABC, 0x00000ABC}, {0x00000ABE, 0x00000AC5}, {0x00000AC7, 0x00000AC9}, {0x00000ACB, 0x00000ACD}, +{0x00000AE2, 0x00000AE3}, {0x00000AFA, 0x00000AFF}, {0x00000B01, 0x00000B03}, {0x00000B3C, 0x00000B3C}, +{0x00000B3E, 0x00000B44}, {0x00000B47, 0x00000B48}, {0x00000B4B, 0x00000B4D}, {0x00000B55, 0x00000B57}, +{0x00000B62, 0x00000B63}, {0x00000B82, 0x00000B82}, {0x00000BBE, 0x00000BC2}, {0x00000BC6, 0x00000BC8}, +{0x00000BCA, 0x00000BCD}, {0x00000BD7, 0x00000BD7}, {0x00000C00, 0x00000C04}, {0x00000C3C, 0x00000C3C}, +{0x00000C3E, 0x00000C44}, {0x00000C46, 0x00000C48}, {0x00000C4A, 0x00000C4D}, {0x00000C55, 0x00000C56}, +{0x00000C62, 0x00000C63}, {0x00000C81, 0x00000C83}, {0x00000CBC, 0x00000CBC}, {0x00000CBE, 0x00000CC4}, +{0x00000CC6, 0x00000CC8}, {0x00000CCA, 0x00000CCD}, {0x00000CD5, 0x00000CD6}, {0x00000CE2, 0x00000CE3}, +{0x00000CF3, 0x00000CF3}, {0x00000D00, 0x00000D03}, {0x00000D3B, 0x00000D3C}, {0x00000D3E, 0x00000D44}, +{0x00000D46, 0x00000D48}, {0x00000D4A, 0x00000D4D}, {0x00000D57, 0x00000D57}, {0x00000D62, 0x00000D63}, +{0x00000D81, 0x00000D83}, {0x00000DCA, 0x00000DCA}, {0x00000DCF, 0x00000DD4}, {0x00000DD6, 0x00000DD6}, +{0x00000DD8, 0x00000DDF}, {0x00000DF2, 0x00000DF3}, {0x00000E31, 0x00000E31}, {0x00000E34, 0x00000E3A}, +{0x00000E47, 0x00000E4E}, {0x00000EB1, 0x00000EB1}, {0x00000EB4, 0x00000EBC}, {0x00000EC8, 0x00000ECE}, +{0x00000F18, 0x00000F19}, {0x00000F35, 0x00000F35}, {0x00000F37, 0x00000F37}, {0x00000F39, 0x00000F39}, +{0x00000F3E, 0x00000F3F}, {0x00000F71, 0x00000F84}, {0x00000F86, 0x00000F87}, {0x00000F8D, 0x00000F97}, +{0x00000F99, 0x00000FBC}, {0x00000FC6, 0x00000FC6}, {0x0000102B, 0x0000103E}, {0x00001056, 0x00001059}, +{0x0000105E, 0x00001060}, {0x00001062, 0x00001064}, {0x00001067, 0x0000106D}, {0x00001071, 0x00001074}, +{0x00001082, 0x0000108D}, {0x0000108F, 0x0000108F}, {0x0000109A, 0x0000109D}, {0x0000135D, 0x0000135F}, +{0x00001712, 0x00001715}, {0x00001732, 0x00001734}, {0x00001752, 0x00001753}, {0x00001772, 0x00001773}, +{0x000017B4, 0x000017D3}, {0x000017DD, 0x000017DD}, {0x0000180B, 0x0000180D}, {0x0000180F, 0x0000180F}, {0x00001885, 0x00001886}, {0x000018A9, 0x000018A9}, {0x00001920, 0x0000192B}, {0x00001930, 0x0000193B}, {0x00001A17, 0x00001A1B}, {0x00001A55, 0x00001A5E}, {0x00001A60, 0x00001A7C}, {0x00001A7F, 0x00001A7F}, -{0x00001AB0, 0x00001AC0}, {0x00001B00, 0x00001B04}, {0x00001B34, 0x00001B44}, {0x00001B6B, 0x00001B73}, +{0x00001AB0, 0x00001ACE}, {0x00001B00, 0x00001B04}, {0x00001B34, 0x00001B44}, {0x00001B6B, 0x00001B73}, {0x00001B80, 0x00001B82}, {0x00001BA1, 0x00001BAD}, {0x00001BE6, 0x00001BF3}, {0x00001C24, 0x00001C37}, {0x00001CD0, 0x00001CD2}, {0x00001CD4, 0x00001CE8}, {0x00001CED, 0x00001CED}, {0x00001CF4, 0x00001CF4}, -{0x00001CF7, 0x00001CF9}, {0x00001DC0, 0x00001DF9}, {0x00001DFB, 0x00001DFF}, {0x000020D0, 0x000020F0}, -{0x00002CEF, 0x00002CF1}, {0x00002D7F, 0x00002D7F}, {0x00002DE0, 0x00002DFF}, {0x0000302A, 0x0000302F}, -{0x00003099, 0x0000309A}, {0x0000A66F, 0x0000A672}, {0x0000A674, 0x0000A67D}, {0x0000A69E, 0x0000A69F}, -{0x0000A6F0, 0x0000A6F1}, {0x0000A802, 0x0000A802}, {0x0000A806, 0x0000A806}, {0x0000A80B, 0x0000A80B}, -{0x0000A823, 0x0000A827}, {0x0000A82C, 0x0000A82C}, {0x0000A880, 0x0000A881}, {0x0000A8B4, 0x0000A8C5}, -{0x0000A8E0, 0x0000A8F1}, {0x0000A8FF, 0x0000A8FF}, {0x0000A926, 0x0000A92D}, {0x0000A947, 0x0000A953}, -{0x0000A980, 0x0000A983}, {0x0000A9B3, 0x0000A9C0}, {0x0000A9E5, 0x0000A9E5}, {0x0000AA29, 0x0000AA36}, -{0x0000AA43, 0x0000AA43}, {0x0000AA4C, 0x0000AA4D}, {0x0000AA7B, 0x0000AA7D}, {0x0000AAB0, 0x0000AAB0}, -{0x0000AAB2, 0x0000AAB4}, {0x0000AAB7, 0x0000AAB8}, {0x0000AABE, 0x0000AABF}, {0x0000AAC1, 0x0000AAC1}, -{0x0000AAEB, 0x0000AAEF}, {0x0000AAF5, 0x0000AAF6}, {0x0000ABE3, 0x0000ABEA}, {0x0000ABEC, 0x0000ABED}, -{0x0000FB1E, 0x0000FB1E}, {0x0000FE00, 0x0000FE0F}, {0x0000FE20, 0x0000FE2F}, {0x000101FD, 0x000101FD}, -{0x000102E0, 0x000102E0}, {0x00010376, 0x0001037A}, {0x00010A01, 0x00010A03}, {0x00010A05, 0x00010A06}, -{0x00010A0C, 0x00010A0F}, {0x00010A38, 0x00010A3A}, {0x00010A3F, 0x00010A3F}, {0x00010AE5, 0x00010AE6}, -{0x00010D24, 0x00010D27}, {0x00010EAB, 0x00010EAC}, {0x00010F46, 0x00010F50}, {0x00011000, 0x00011002}, -{0x00011038, 0x00011046}, {0x0001107F, 0x00011082}, {0x000110B0, 0x000110BA}, {0x00011100, 0x00011102}, +{0x00001CF7, 0x00001CF9}, {0x00001DC0, 0x00001DFF}, {0x000020D0, 0x000020F0}, {0x00002CEF, 0x00002CF1}, +{0x00002D7F, 0x00002D7F}, {0x00002DE0, 0x00002DFF}, {0x0000302A, 0x0000302F}, {0x00003099, 0x0000309A}, +{0x0000A66F, 0x0000A672}, {0x0000A674, 0x0000A67D}, {0x0000A69E, 0x0000A69F}, {0x0000A6F0, 0x0000A6F1}, +{0x0000A802, 0x0000A802}, {0x0000A806, 0x0000A806}, {0x0000A80B, 0x0000A80B}, {0x0000A823, 0x0000A827}, +{0x0000A82C, 0x0000A82C}, {0x0000A880, 0x0000A881}, {0x0000A8B4, 0x0000A8C5}, {0x0000A8E0, 0x0000A8F1}, +{0x0000A8FF, 0x0000A8FF}, {0x0000A926, 0x0000A92D}, {0x0000A947, 0x0000A953}, {0x0000A980, 0x0000A983}, +{0x0000A9B3, 0x0000A9C0}, {0x0000A9E5, 0x0000A9E5}, {0x0000AA29, 0x0000AA36}, {0x0000AA43, 0x0000AA43}, +{0x0000AA4C, 0x0000AA4D}, {0x0000AA7B, 0x0000AA7D}, {0x0000AAB0, 0x0000AAB0}, {0x0000AAB2, 0x0000AAB4}, +{0x0000AAB7, 0x0000AAB8}, {0x0000AABE, 0x0000AABF}, {0x0000AAC1, 0x0000AAC1}, {0x0000AAEB, 0x0000AAEF}, +{0x0000AAF5, 0x0000AAF6}, {0x0000ABE3, 0x0000ABEA}, {0x0000ABEC, 0x0000ABED}, {0x0000FB1E, 0x0000FB1E}, +{0x0000FE00, 0x0000FE0F}, {0x0000FE20, 0x0000FE2F}, {0x000101FD, 0x000101FD}, {0x000102E0, 0x000102E0}, +{0x00010376, 0x0001037A}, {0x00010A01, 0x00010A03}, {0x00010A05, 0x00010A06}, {0x00010A0C, 0x00010A0F}, +{0x00010A38, 0x00010A3A}, {0x00010A3F, 0x00010A3F}, {0x00010AE5, 0x00010AE6}, {0x00010D24, 0x00010D27}, +{0x00010EAB, 0x00010EAC}, {0x00010EFD, 0x00010EFF}, {0x00010F46, 0x00010F50}, {0x00010F82, 0x00010F85}, +{0x00011000, 0x00011002}, {0x00011038, 0x00011046}, {0x00011070, 0x00011070}, {0x00011073, 0x00011074}, +{0x0001107F, 0x00011082}, {0x000110B0, 0x000110BA}, {0x000110C2, 0x000110C2}, {0x00011100, 0x00011102}, {0x00011127, 0x00011134}, {0x00011145, 0x00011146}, {0x00011173, 0x00011173}, {0x00011180, 0x00011182}, {0x000111B3, 0x000111C0}, {0x000111C9, 0x000111CC}, {0x000111CE, 0x000111CF}, {0x0001122C, 0x00011237}, -{0x0001123E, 0x0001123E}, {0x000112DF, 0x000112EA}, {0x00011300, 0x00011303}, {0x0001133B, 0x0001133C}, -{0x0001133E, 0x00011344}, {0x00011347, 0x00011348}, {0x0001134B, 0x0001134D}, {0x00011357, 0x00011357}, -{0x00011362, 0x00011363}, {0x00011366, 0x0001136C}, {0x00011370, 0x00011374}, {0x00011435, 0x00011446}, -{0x0001145E, 0x0001145E}, {0x000114B0, 0x000114C3}, {0x000115AF, 0x000115B5}, {0x000115B8, 0x000115C0}, -{0x000115DC, 0x000115DD}, {0x00011630, 0x00011640}, {0x000116AB, 0x000116B7}, {0x0001171D, 0x0001172B}, -{0x0001182C, 0x0001183A}, {0x00011930, 0x00011935}, {0x00011937, 0x00011938}, {0x0001193B, 0x0001193E}, -{0x00011940, 0x00011940}, {0x00011942, 0x00011943}, {0x000119D1, 0x000119D7}, {0x000119DA, 0x000119E0}, -{0x000119E4, 0x000119E4}, {0x00011A01, 0x00011A0A}, {0x00011A33, 0x00011A39}, {0x00011A3B, 0x00011A3E}, -{0x00011A47, 0x00011A47}, {0x00011A51, 0x00011A5B}, {0x00011A8A, 0x00011A99}, {0x00011C2F, 0x00011C36}, -{0x00011C38, 0x00011C3F}, {0x00011C92, 0x00011CA7}, {0x00011CA9, 0x00011CB6}, {0x00011D31, 0x00011D36}, -{0x00011D3A, 0x00011D3A}, {0x00011D3C, 0x00011D3D}, {0x00011D3F, 0x00011D45}, {0x00011D47, 0x00011D47}, -{0x00011D8A, 0x00011D8E}, {0x00011D90, 0x00011D91}, {0x00011D93, 0x00011D97}, {0x00011EF3, 0x00011EF6}, -{0x00016AF0, 0x00016AF4}, {0x00016B30, 0x00016B36}, {0x00016F4F, 0x00016F4F}, {0x00016F51, 0x00016F87}, -{0x00016F8F, 0x00016F92}, {0x00016FE4, 0x00016FE4}, {0x00016FF0, 0x00016FF1}, {0x0001BC9D, 0x0001BC9E}, -{0x0001D165, 0x0001D169}, {0x0001D16D, 0x0001D172}, {0x0001D17B, 0x0001D182}, {0x0001D185, 0x0001D18B}, -{0x0001D1AA, 0x0001D1AD}, {0x0001D242, 0x0001D244}, {0x0001DA00, 0x0001DA36}, {0x0001DA3B, 0x0001DA6C}, -{0x0001DA75, 0x0001DA75}, {0x0001DA84, 0x0001DA84}, {0x0001DA9B, 0x0001DA9F}, {0x0001DAA1, 0x0001DAAF}, -{0x0001E000, 0x0001E006}, {0x0001E008, 0x0001E018}, {0x0001E01B, 0x0001E021}, {0x0001E023, 0x0001E024}, -{0x0001E026, 0x0001E02A}, {0x0001E130, 0x0001E136}, {0x0001E2EC, 0x0001E2EF}, {0x0001E8D0, 0x0001E8D6}, +{0x0001123E, 0x0001123E}, {0x00011241, 0x00011241}, {0x000112DF, 0x000112EA}, {0x00011300, 0x00011303}, +{0x0001133B, 0x0001133C}, {0x0001133E, 0x00011344}, {0x00011347, 0x00011348}, {0x0001134B, 0x0001134D}, +{0x00011357, 0x00011357}, {0x00011362, 0x00011363}, {0x00011366, 0x0001136C}, {0x00011370, 0x00011374}, +{0x00011435, 0x00011446}, {0x0001145E, 0x0001145E}, {0x000114B0, 0x000114C3}, {0x000115AF, 0x000115B5}, +{0x000115B8, 0x000115C0}, {0x000115DC, 0x000115DD}, {0x00011630, 0x00011640}, {0x000116AB, 0x000116B7}, +{0x0001171D, 0x0001172B}, {0x0001182C, 0x0001183A}, {0x00011930, 0x00011935}, {0x00011937, 0x00011938}, +{0x0001193B, 0x0001193E}, {0x00011940, 0x00011940}, {0x00011942, 0x00011943}, {0x000119D1, 0x000119D7}, +{0x000119DA, 0x000119E0}, {0x000119E4, 0x000119E4}, {0x00011A01, 0x00011A0A}, {0x00011A33, 0x00011A39}, +{0x00011A3B, 0x00011A3E}, {0x00011A47, 0x00011A47}, {0x00011A51, 0x00011A5B}, {0x00011A8A, 0x00011A99}, +{0x00011C2F, 0x00011C36}, {0x00011C38, 0x00011C3F}, {0x00011C92, 0x00011CA7}, {0x00011CA9, 0x00011CB6}, +{0x00011D31, 0x00011D36}, {0x00011D3A, 0x00011D3A}, {0x00011D3C, 0x00011D3D}, {0x00011D3F, 0x00011D45}, +{0x00011D47, 0x00011D47}, {0x00011D8A, 0x00011D8E}, {0x00011D90, 0x00011D91}, {0x00011D93, 0x00011D97}, +{0x00011EF3, 0x00011EF6}, {0x00011F00, 0x00011F01}, {0x00011F03, 0x00011F03}, {0x00011F34, 0x00011F3A}, +{0x00011F3E, 0x00011F42}, {0x00013440, 0x00013440}, {0x00013447, 0x00013455}, {0x00016AF0, 0x00016AF4}, +{0x00016B30, 0x00016B36}, {0x00016F4F, 0x00016F4F}, {0x00016F51, 0x00016F87}, {0x00016F8F, 0x00016F92}, +{0x00016FE4, 0x00016FE4}, {0x00016FF0, 0x00016FF1}, {0x0001BC9D, 0x0001BC9E}, {0x0001CF00, 0x0001CF2D}, +{0x0001CF30, 0x0001CF46}, {0x0001D165, 0x0001D169}, {0x0001D16D, 0x0001D172}, {0x0001D17B, 0x0001D182}, +{0x0001D185, 0x0001D18B}, {0x0001D1AA, 0x0001D1AD}, {0x0001D242, 0x0001D244}, {0x0001DA00, 0x0001DA36}, +{0x0001DA3B, 0x0001DA6C}, {0x0001DA75, 0x0001DA75}, {0x0001DA84, 0x0001DA84}, {0x0001DA9B, 0x0001DA9F}, +{0x0001DAA1, 0x0001DAAF}, {0x0001E000, 0x0001E006}, {0x0001E008, 0x0001E018}, {0x0001E01B, 0x0001E021}, +{0x0001E023, 0x0001E024}, {0x0001E026, 0x0001E02A}, {0x0001E08F, 0x0001E08F}, {0x0001E130, 0x0001E136}, +{0x0001E2AE, 0x0001E2AE}, {0x0001E2EC, 0x0001E2EF}, {0x0001E4EC, 0x0001E4EF}, {0x0001E8D0, 0x0001E8D6}, {0x0001E944, 0x0001E94A}, {0x000E0100, 0x000E01EF}, }; @@ -290,7 +309,7 @@ const std::vector> unicode_ranges_punctuation = { {0x000000B6, 0x000000B7}, {0x000000BB, 0x000000BB}, {0x000000BF, 0x000000BF}, {0x0000037E, 0x0000037E}, {0x00000387, 0x00000387}, {0x0000055A, 0x0000055F}, {0x00000589, 0x0000058A}, {0x000005BE, 0x000005BE}, {0x000005C0, 0x000005C0}, {0x000005C3, 0x000005C3}, {0x000005C6, 0x000005C6}, {0x000005F3, 0x000005F4}, -{0x00000609, 0x0000060A}, {0x0000060C, 0x0000060D}, {0x0000061B, 0x0000061B}, {0x0000061E, 0x0000061F}, +{0x00000609, 0x0000060A}, {0x0000060C, 0x0000060D}, {0x0000061B, 0x0000061B}, {0x0000061D, 0x0000061F}, {0x0000066A, 0x0000066D}, {0x000006D4, 0x000006D4}, {0x00000700, 0x0000070D}, {0x000007F7, 0x000007F9}, {0x00000830, 0x0000083E}, {0x0000085E, 0x0000085E}, {0x00000964, 0x00000965}, {0x00000970, 0x00000970}, {0x000009FD, 0x000009FD}, {0x00000A76, 0x00000A76}, {0x00000AF0, 0x00000AF0}, {0x00000C77, 0x00000C77}, @@ -300,37 +319,38 @@ const std::vector> unicode_ranges_punctuation = { {0x00001360, 0x00001368}, {0x00001400, 0x00001400}, {0x0000166E, 0x0000166E}, {0x0000169B, 0x0000169C}, {0x000016EB, 0x000016ED}, {0x00001735, 0x00001736}, {0x000017D4, 0x000017D6}, {0x000017D8, 0x000017DA}, {0x00001800, 0x0000180A}, {0x00001944, 0x00001945}, {0x00001A1E, 0x00001A1F}, {0x00001AA0, 0x00001AA6}, -{0x00001AA8, 0x00001AAD}, {0x00001B5A, 0x00001B60}, {0x00001BFC, 0x00001BFF}, {0x00001C3B, 0x00001C3F}, -{0x00001C7E, 0x00001C7F}, {0x00001CC0, 0x00001CC7}, {0x00001CD3, 0x00001CD3}, {0x00002010, 0x00002027}, -{0x00002030, 0x00002043}, {0x00002045, 0x00002051}, {0x00002053, 0x0000205E}, {0x0000207D, 0x0000207E}, -{0x0000208D, 0x0000208E}, {0x00002308, 0x0000230B}, {0x00002329, 0x0000232A}, {0x00002768, 0x00002775}, -{0x000027C5, 0x000027C6}, {0x000027E6, 0x000027EF}, {0x00002983, 0x00002998}, {0x000029D8, 0x000029DB}, -{0x000029FC, 0x000029FD}, {0x00002CF9, 0x00002CFC}, {0x00002CFE, 0x00002CFF}, {0x00002D70, 0x00002D70}, -{0x00002E00, 0x00002E2E}, {0x00002E30, 0x00002E4F}, {0x00002E52, 0x00002E52}, {0x00003001, 0x00003003}, -{0x00003008, 0x00003011}, {0x00003014, 0x0000301F}, {0x00003030, 0x00003030}, {0x0000303D, 0x0000303D}, -{0x000030A0, 0x000030A0}, {0x000030FB, 0x000030FB}, {0x0000A4FE, 0x0000A4FF}, {0x0000A60D, 0x0000A60F}, -{0x0000A673, 0x0000A673}, {0x0000A67E, 0x0000A67E}, {0x0000A6F2, 0x0000A6F7}, {0x0000A874, 0x0000A877}, -{0x0000A8CE, 0x0000A8CF}, {0x0000A8F8, 0x0000A8FA}, {0x0000A8FC, 0x0000A8FC}, {0x0000A92E, 0x0000A92F}, -{0x0000A95F, 0x0000A95F}, {0x0000A9C1, 0x0000A9CD}, {0x0000A9DE, 0x0000A9DF}, {0x0000AA5C, 0x0000AA5F}, -{0x0000AADE, 0x0000AADF}, {0x0000AAF0, 0x0000AAF1}, {0x0000ABEB, 0x0000ABEB}, {0x0000FD3E, 0x0000FD3F}, -{0x0000FE10, 0x0000FE19}, {0x0000FE30, 0x0000FE52}, {0x0000FE54, 0x0000FE61}, {0x0000FE63, 0x0000FE63}, -{0x0000FE68, 0x0000FE68}, {0x0000FE6A, 0x0000FE6B}, {0x0000FF01, 0x0000FF03}, {0x0000FF05, 0x0000FF0A}, -{0x0000FF0C, 0x0000FF0F}, {0x0000FF1A, 0x0000FF1B}, {0x0000FF1F, 0x0000FF20}, {0x0000FF3B, 0x0000FF3D}, -{0x0000FF3F, 0x0000FF3F}, {0x0000FF5B, 0x0000FF5B}, {0x0000FF5D, 0x0000FF5D}, {0x0000FF5F, 0x0000FF65}, -{0x00010100, 0x00010102}, {0x0001039F, 0x0001039F}, {0x000103D0, 0x000103D0}, {0x0001056F, 0x0001056F}, -{0x00010857, 0x00010857}, {0x0001091F, 0x0001091F}, {0x0001093F, 0x0001093F}, {0x00010A50, 0x00010A58}, -{0x00010A7F, 0x00010A7F}, {0x00010AF0, 0x00010AF6}, {0x00010B39, 0x00010B3F}, {0x00010B99, 0x00010B9C}, -{0x00010EAD, 0x00010EAD}, {0x00010F55, 0x00010F59}, {0x00011047, 0x0001104D}, {0x000110BB, 0x000110BC}, -{0x000110BE, 0x000110C1}, {0x00011140, 0x00011143}, {0x00011174, 0x00011175}, {0x000111C5, 0x000111C8}, -{0x000111CD, 0x000111CD}, {0x000111DB, 0x000111DB}, {0x000111DD, 0x000111DF}, {0x00011238, 0x0001123D}, -{0x000112A9, 0x000112A9}, {0x0001144B, 0x0001144F}, {0x0001145A, 0x0001145B}, {0x0001145D, 0x0001145D}, -{0x000114C6, 0x000114C6}, {0x000115C1, 0x000115D7}, {0x00011641, 0x00011643}, {0x00011660, 0x0001166C}, -{0x0001173C, 0x0001173E}, {0x0001183B, 0x0001183B}, {0x00011944, 0x00011946}, {0x000119E2, 0x000119E2}, -{0x00011A3F, 0x00011A46}, {0x00011A9A, 0x00011A9C}, {0x00011A9E, 0x00011AA2}, {0x00011C41, 0x00011C45}, -{0x00011C70, 0x00011C71}, {0x00011EF7, 0x00011EF8}, {0x00011FFF, 0x00011FFF}, {0x00012470, 0x00012474}, -{0x00016A6E, 0x00016A6F}, {0x00016AF5, 0x00016AF5}, {0x00016B37, 0x00016B3B}, {0x00016B44, 0x00016B44}, -{0x00016E97, 0x00016E9A}, {0x00016FE2, 0x00016FE2}, {0x0001BC9F, 0x0001BC9F}, {0x0001DA87, 0x0001DA8B}, -{0x0001E95E, 0x0001E95F}, +{0x00001AA8, 0x00001AAD}, {0x00001B5A, 0x00001B60}, {0x00001B7D, 0x00001B7E}, {0x00001BFC, 0x00001BFF}, +{0x00001C3B, 0x00001C3F}, {0x00001C7E, 0x00001C7F}, {0x00001CC0, 0x00001CC7}, {0x00001CD3, 0x00001CD3}, +{0x00002010, 0x00002027}, {0x00002030, 0x00002043}, {0x00002045, 0x00002051}, {0x00002053, 0x0000205E}, +{0x0000207D, 0x0000207E}, {0x0000208D, 0x0000208E}, {0x00002308, 0x0000230B}, {0x00002329, 0x0000232A}, +{0x00002768, 0x00002775}, {0x000027C5, 0x000027C6}, {0x000027E6, 0x000027EF}, {0x00002983, 0x00002998}, +{0x000029D8, 0x000029DB}, {0x000029FC, 0x000029FD}, {0x00002CF9, 0x00002CFC}, {0x00002CFE, 0x00002CFF}, +{0x00002D70, 0x00002D70}, {0x00002E00, 0x00002E2E}, {0x00002E30, 0x00002E4F}, {0x00002E52, 0x00002E5D}, +{0x00003001, 0x00003003}, {0x00003008, 0x00003011}, {0x00003014, 0x0000301F}, {0x00003030, 0x00003030}, +{0x0000303D, 0x0000303D}, {0x000030A0, 0x000030A0}, {0x000030FB, 0x000030FB}, {0x0000A4FE, 0x0000A4FF}, +{0x0000A60D, 0x0000A60F}, {0x0000A673, 0x0000A673}, {0x0000A67E, 0x0000A67E}, {0x0000A6F2, 0x0000A6F7}, +{0x0000A874, 0x0000A877}, {0x0000A8CE, 0x0000A8CF}, {0x0000A8F8, 0x0000A8FA}, {0x0000A8FC, 0x0000A8FC}, +{0x0000A92E, 0x0000A92F}, {0x0000A95F, 0x0000A95F}, {0x0000A9C1, 0x0000A9CD}, {0x0000A9DE, 0x0000A9DF}, +{0x0000AA5C, 0x0000AA5F}, {0x0000AADE, 0x0000AADF}, {0x0000AAF0, 0x0000AAF1}, {0x0000ABEB, 0x0000ABEB}, +{0x0000FD3E, 0x0000FD3F}, {0x0000FE10, 0x0000FE19}, {0x0000FE30, 0x0000FE52}, {0x0000FE54, 0x0000FE61}, +{0x0000FE63, 0x0000FE63}, {0x0000FE68, 0x0000FE68}, {0x0000FE6A, 0x0000FE6B}, {0x0000FF01, 0x0000FF03}, +{0x0000FF05, 0x0000FF0A}, {0x0000FF0C, 0x0000FF0F}, {0x0000FF1A, 0x0000FF1B}, {0x0000FF1F, 0x0000FF20}, +{0x0000FF3B, 0x0000FF3D}, {0x0000FF3F, 0x0000FF3F}, {0x0000FF5B, 0x0000FF5B}, {0x0000FF5D, 0x0000FF5D}, +{0x0000FF5F, 0x0000FF65}, {0x00010100, 0x00010102}, {0x0001039F, 0x0001039F}, {0x000103D0, 0x000103D0}, +{0x0001056F, 0x0001056F}, {0x00010857, 0x00010857}, {0x0001091F, 0x0001091F}, {0x0001093F, 0x0001093F}, +{0x00010A50, 0x00010A58}, {0x00010A7F, 0x00010A7F}, {0x00010AF0, 0x00010AF6}, {0x00010B39, 0x00010B3F}, +{0x00010B99, 0x00010B9C}, {0x00010EAD, 0x00010EAD}, {0x00010F55, 0x00010F59}, {0x00010F86, 0x00010F89}, +{0x00011047, 0x0001104D}, {0x000110BB, 0x000110BC}, {0x000110BE, 0x000110C1}, {0x00011140, 0x00011143}, +{0x00011174, 0x00011175}, {0x000111C5, 0x000111C8}, {0x000111CD, 0x000111CD}, {0x000111DB, 0x000111DB}, +{0x000111DD, 0x000111DF}, {0x00011238, 0x0001123D}, {0x000112A9, 0x000112A9}, {0x0001144B, 0x0001144F}, +{0x0001145A, 0x0001145B}, {0x0001145D, 0x0001145D}, {0x000114C6, 0x000114C6}, {0x000115C1, 0x000115D7}, +{0x00011641, 0x00011643}, {0x00011660, 0x0001166C}, {0x000116B9, 0x000116B9}, {0x0001173C, 0x0001173E}, +{0x0001183B, 0x0001183B}, {0x00011944, 0x00011946}, {0x000119E2, 0x000119E2}, {0x00011A3F, 0x00011A46}, +{0x00011A9A, 0x00011A9C}, {0x00011A9E, 0x00011AA2}, {0x00011B00, 0x00011B09}, {0x00011C41, 0x00011C45}, +{0x00011C70, 0x00011C71}, {0x00011EF7, 0x00011EF8}, {0x00011F43, 0x00011F4F}, {0x00011FFF, 0x00011FFF}, +{0x00012470, 0x00012474}, {0x00012FF1, 0x00012FF2}, {0x00016A6E, 0x00016A6F}, {0x00016AF5, 0x00016AF5}, +{0x00016B37, 0x00016B3B}, {0x00016B44, 0x00016B44}, {0x00016E97, 0x00016E9A}, {0x00016FE2, 0x00016FE2}, +{0x0001BC9F, 0x0001BC9F}, {0x0001DA87, 0x0001DA8B}, {0x0001E95E, 0x0001E95F}, }; const std::vector> unicode_ranges_symbol = { @@ -342,40 +362,41 @@ const std::vector> unicode_ranges_symbol = { {0x00000375, 0x00000375}, {0x00000384, 0x00000385}, {0x000003F6, 0x000003F6}, {0x00000482, 0x00000482}, {0x0000058D, 0x0000058F}, {0x00000606, 0x00000608}, {0x0000060B, 0x0000060B}, {0x0000060E, 0x0000060F}, {0x000006DE, 0x000006DE}, {0x000006E9, 0x000006E9}, {0x000006FD, 0x000006FE}, {0x000007F6, 0x000007F6}, -{0x000007FE, 0x000007FF}, {0x000009F2, 0x000009F3}, {0x000009FA, 0x000009FB}, {0x00000AF1, 0x00000AF1}, -{0x00000B70, 0x00000B70}, {0x00000BF3, 0x00000BFA}, {0x00000C7F, 0x00000C7F}, {0x00000D4F, 0x00000D4F}, -{0x00000D79, 0x00000D79}, {0x00000E3F, 0x00000E3F}, {0x00000F01, 0x00000F03}, {0x00000F13, 0x00000F13}, -{0x00000F15, 0x00000F17}, {0x00000F1A, 0x00000F1F}, {0x00000F34, 0x00000F34}, {0x00000F36, 0x00000F36}, -{0x00000F38, 0x00000F38}, {0x00000FBE, 0x00000FC5}, {0x00000FC7, 0x00000FCC}, {0x00000FCE, 0x00000FCF}, -{0x00000FD5, 0x00000FD8}, {0x0000109E, 0x0000109F}, {0x00001390, 0x00001399}, {0x0000166D, 0x0000166D}, -{0x000017DB, 0x000017DB}, {0x00001940, 0x00001940}, {0x000019DE, 0x000019FF}, {0x00001B61, 0x00001B6A}, -{0x00001B74, 0x00001B7C}, {0x00001FBD, 0x00001FBD}, {0x00001FBF, 0x00001FC1}, {0x00001FCD, 0x00001FCF}, -{0x00001FDD, 0x00001FDF}, {0x00001FED, 0x00001FEF}, {0x00001FFD, 0x00001FFE}, {0x00002044, 0x00002044}, -{0x00002052, 0x00002052}, {0x0000207A, 0x0000207C}, {0x0000208A, 0x0000208C}, {0x000020A0, 0x000020BF}, -{0x00002100, 0x00002101}, {0x00002103, 0x00002106}, {0x00002108, 0x00002109}, {0x00002114, 0x00002114}, -{0x00002116, 0x00002118}, {0x0000211E, 0x00002123}, {0x00002125, 0x00002125}, {0x00002127, 0x00002127}, -{0x00002129, 0x00002129}, {0x0000212E, 0x0000212E}, {0x0000213A, 0x0000213B}, {0x00002140, 0x00002144}, -{0x0000214A, 0x0000214D}, {0x0000214F, 0x0000214F}, {0x0000218A, 0x0000218B}, {0x00002190, 0x00002307}, -{0x0000230C, 0x00002328}, {0x0000232B, 0x00002426}, {0x00002440, 0x0000244A}, {0x0000249C, 0x000024E9}, -{0x00002500, 0x00002767}, {0x00002794, 0x000027C4}, {0x000027C7, 0x000027E5}, {0x000027F0, 0x00002982}, -{0x00002999, 0x000029D7}, {0x000029DC, 0x000029FB}, {0x000029FE, 0x00002B73}, {0x00002B76, 0x00002B95}, -{0x00002B97, 0x00002BFF}, {0x00002CE5, 0x00002CEA}, {0x00002E50, 0x00002E51}, {0x00002E80, 0x00002E99}, -{0x00002E9B, 0x00002EF3}, {0x00002F00, 0x00002FD5}, {0x00002FF0, 0x00002FFB}, {0x00003004, 0x00003004}, -{0x00003012, 0x00003013}, {0x00003020, 0x00003020}, {0x00003036, 0x00003037}, {0x0000303E, 0x0000303F}, -{0x0000309B, 0x0000309C}, {0x00003190, 0x00003191}, {0x00003196, 0x0000319F}, {0x000031C0, 0x000031E3}, -{0x00003200, 0x0000321E}, {0x0000322A, 0x00003247}, {0x00003250, 0x00003250}, {0x00003260, 0x0000327F}, -{0x0000328A, 0x000032B0}, {0x000032C0, 0x000033FF}, {0x00004DC0, 0x00004DFF}, {0x0000A490, 0x0000A4C6}, -{0x0000A700, 0x0000A716}, {0x0000A720, 0x0000A721}, {0x0000A789, 0x0000A78A}, {0x0000A828, 0x0000A82B}, -{0x0000A836, 0x0000A839}, {0x0000AA77, 0x0000AA79}, {0x0000AB5B, 0x0000AB5B}, {0x0000AB6A, 0x0000AB6B}, -{0x0000FB29, 0x0000FB29}, {0x0000FBB2, 0x0000FBC1}, {0x0000FDFC, 0x0000FDFD}, {0x0000FE62, 0x0000FE62}, -{0x0000FE64, 0x0000FE66}, {0x0000FE69, 0x0000FE69}, {0x0000FF04, 0x0000FF04}, {0x0000FF0B, 0x0000FF0B}, -{0x0000FF1C, 0x0000FF1E}, {0x0000FF3E, 0x0000FF3E}, {0x0000FF40, 0x0000FF40}, {0x0000FF5C, 0x0000FF5C}, -{0x0000FF5E, 0x0000FF5E}, {0x0000FFE0, 0x0000FFE6}, {0x0000FFE8, 0x0000FFEE}, {0x0000FFFC, 0x0000FFFD}, -{0x00010137, 0x0001013F}, {0x00010179, 0x00010189}, {0x0001018C, 0x0001018E}, {0x00010190, 0x0001019C}, -{0x000101A0, 0x000101A0}, {0x000101D0, 0x000101FC}, {0x00010877, 0x00010878}, {0x00010AC8, 0x00010AC8}, -{0x0001173F, 0x0001173F}, {0x00011FD5, 0x00011FF1}, {0x00016B3C, 0x00016B3F}, {0x00016B45, 0x00016B45}, -{0x0001BC9C, 0x0001BC9C}, {0x0001D000, 0x0001D0F5}, {0x0001D100, 0x0001D126}, {0x0001D129, 0x0001D164}, -{0x0001D16A, 0x0001D16C}, {0x0001D183, 0x0001D184}, {0x0001D18C, 0x0001D1A9}, {0x0001D1AE, 0x0001D1E8}, +{0x000007FE, 0x000007FF}, {0x00000888, 0x00000888}, {0x000009F2, 0x000009F3}, {0x000009FA, 0x000009FB}, +{0x00000AF1, 0x00000AF1}, {0x00000B70, 0x00000B70}, {0x00000BF3, 0x00000BFA}, {0x00000C7F, 0x00000C7F}, +{0x00000D4F, 0x00000D4F}, {0x00000D79, 0x00000D79}, {0x00000E3F, 0x00000E3F}, {0x00000F01, 0x00000F03}, +{0x00000F13, 0x00000F13}, {0x00000F15, 0x00000F17}, {0x00000F1A, 0x00000F1F}, {0x00000F34, 0x00000F34}, +{0x00000F36, 0x00000F36}, {0x00000F38, 0x00000F38}, {0x00000FBE, 0x00000FC5}, {0x00000FC7, 0x00000FCC}, +{0x00000FCE, 0x00000FCF}, {0x00000FD5, 0x00000FD8}, {0x0000109E, 0x0000109F}, {0x00001390, 0x00001399}, +{0x0000166D, 0x0000166D}, {0x000017DB, 0x000017DB}, {0x00001940, 0x00001940}, {0x000019DE, 0x000019FF}, +{0x00001B61, 0x00001B6A}, {0x00001B74, 0x00001B7C}, {0x00001FBD, 0x00001FBD}, {0x00001FBF, 0x00001FC1}, +{0x00001FCD, 0x00001FCF}, {0x00001FDD, 0x00001FDF}, {0x00001FED, 0x00001FEF}, {0x00001FFD, 0x00001FFE}, +{0x00002044, 0x00002044}, {0x00002052, 0x00002052}, {0x0000207A, 0x0000207C}, {0x0000208A, 0x0000208C}, +{0x000020A0, 0x000020C0}, {0x00002100, 0x00002101}, {0x00002103, 0x00002106}, {0x00002108, 0x00002109}, +{0x00002114, 0x00002114}, {0x00002116, 0x00002118}, {0x0000211E, 0x00002123}, {0x00002125, 0x00002125}, +{0x00002127, 0x00002127}, {0x00002129, 0x00002129}, {0x0000212E, 0x0000212E}, {0x0000213A, 0x0000213B}, +{0x00002140, 0x00002144}, {0x0000214A, 0x0000214D}, {0x0000214F, 0x0000214F}, {0x0000218A, 0x0000218B}, +{0x00002190, 0x00002307}, {0x0000230C, 0x00002328}, {0x0000232B, 0x00002426}, {0x00002440, 0x0000244A}, +{0x0000249C, 0x000024E9}, {0x00002500, 0x00002767}, {0x00002794, 0x000027C4}, {0x000027C7, 0x000027E5}, +{0x000027F0, 0x00002982}, {0x00002999, 0x000029D7}, {0x000029DC, 0x000029FB}, {0x000029FE, 0x00002B73}, +{0x00002B76, 0x00002B95}, {0x00002B97, 0x00002BFF}, {0x00002CE5, 0x00002CEA}, {0x00002E50, 0x00002E51}, +{0x00002E80, 0x00002E99}, {0x00002E9B, 0x00002EF3}, {0x00002F00, 0x00002FD5}, {0x00002FF0, 0x00002FFB}, +{0x00003004, 0x00003004}, {0x00003012, 0x00003013}, {0x00003020, 0x00003020}, {0x00003036, 0x00003037}, +{0x0000303E, 0x0000303F}, {0x0000309B, 0x0000309C}, {0x00003190, 0x00003191}, {0x00003196, 0x0000319F}, +{0x000031C0, 0x000031E3}, {0x00003200, 0x0000321E}, {0x0000322A, 0x00003247}, {0x00003250, 0x00003250}, +{0x00003260, 0x0000327F}, {0x0000328A, 0x000032B0}, {0x000032C0, 0x000033FF}, {0x00004DC0, 0x00004DFF}, +{0x0000A490, 0x0000A4C6}, {0x0000A700, 0x0000A716}, {0x0000A720, 0x0000A721}, {0x0000A789, 0x0000A78A}, +{0x0000A828, 0x0000A82B}, {0x0000A836, 0x0000A839}, {0x0000AA77, 0x0000AA79}, {0x0000AB5B, 0x0000AB5B}, +{0x0000AB6A, 0x0000AB6B}, {0x0000FB29, 0x0000FB29}, {0x0000FBB2, 0x0000FBC2}, {0x0000FD40, 0x0000FD4F}, +{0x0000FDCF, 0x0000FDCF}, {0x0000FDFC, 0x0000FDFF}, {0x0000FE62, 0x0000FE62}, {0x0000FE64, 0x0000FE66}, +{0x0000FE69, 0x0000FE69}, {0x0000FF04, 0x0000FF04}, {0x0000FF0B, 0x0000FF0B}, {0x0000FF1C, 0x0000FF1E}, +{0x0000FF3E, 0x0000FF3E}, {0x0000FF40, 0x0000FF40}, {0x0000FF5C, 0x0000FF5C}, {0x0000FF5E, 0x0000FF5E}, +{0x0000FFE0, 0x0000FFE6}, {0x0000FFE8, 0x0000FFEE}, {0x0000FFFC, 0x0000FFFD}, {0x00010137, 0x0001013F}, +{0x00010179, 0x00010189}, {0x0001018C, 0x0001018E}, {0x00010190, 0x0001019C}, {0x000101A0, 0x000101A0}, +{0x000101D0, 0x000101FC}, {0x00010877, 0x00010878}, {0x00010AC8, 0x00010AC8}, {0x0001173F, 0x0001173F}, +{0x00011FD5, 0x00011FF1}, {0x00016B3C, 0x00016B3F}, {0x00016B45, 0x00016B45}, {0x0001BC9C, 0x0001BC9C}, +{0x0001CF50, 0x0001CFC3}, {0x0001D000, 0x0001D0F5}, {0x0001D100, 0x0001D126}, {0x0001D129, 0x0001D164}, +{0x0001D16A, 0x0001D16C}, {0x0001D183, 0x0001D184}, {0x0001D18C, 0x0001D1A9}, {0x0001D1AE, 0x0001D1EA}, {0x0001D200, 0x0001D241}, {0x0001D245, 0x0001D245}, {0x0001D300, 0x0001D356}, {0x0001D6C1, 0x0001D6C1}, {0x0001D6DB, 0x0001D6DB}, {0x0001D6FB, 0x0001D6FB}, {0x0001D715, 0x0001D715}, {0x0001D735, 0x0001D735}, {0x0001D74F, 0x0001D74F}, {0x0001D76F, 0x0001D76F}, {0x0001D789, 0x0001D789}, {0x0001D7A9, 0x0001D7A9}, @@ -385,104 +406,100 @@ const std::vector> unicode_ranges_symbol = { {0x0001F000, 0x0001F02B}, {0x0001F030, 0x0001F093}, {0x0001F0A0, 0x0001F0AE}, {0x0001F0B1, 0x0001F0BF}, {0x0001F0C1, 0x0001F0CF}, {0x0001F0D1, 0x0001F0F5}, {0x0001F10D, 0x0001F1AD}, {0x0001F1E6, 0x0001F202}, {0x0001F210, 0x0001F23B}, {0x0001F240, 0x0001F248}, {0x0001F250, 0x0001F251}, {0x0001F260, 0x0001F265}, -{0x0001F300, 0x0001F6D7}, {0x0001F6E0, 0x0001F6EC}, {0x0001F6F0, 0x0001F6FC}, {0x0001F700, 0x0001F773}, -{0x0001F780, 0x0001F7D8}, {0x0001F7E0, 0x0001F7EB}, {0x0001F800, 0x0001F80B}, {0x0001F810, 0x0001F847}, -{0x0001F850, 0x0001F859}, {0x0001F860, 0x0001F887}, {0x0001F890, 0x0001F8AD}, {0x0001F8B0, 0x0001F8B1}, -{0x0001F900, 0x0001F978}, {0x0001F97A, 0x0001F9CB}, {0x0001F9CD, 0x0001FA53}, {0x0001FA60, 0x0001FA6D}, -{0x0001FA70, 0x0001FA74}, {0x0001FA78, 0x0001FA7A}, {0x0001FA80, 0x0001FA86}, {0x0001FA90, 0x0001FAA8}, -{0x0001FAB0, 0x0001FAB6}, {0x0001FAC0, 0x0001FAC2}, {0x0001FAD0, 0x0001FAD6}, {0x0001FB00, 0x0001FB92}, -{0x0001FB94, 0x0001FBCA}, +{0x0001F300, 0x0001F6D7}, {0x0001F6DC, 0x0001F6EC}, {0x0001F6F0, 0x0001F6FC}, {0x0001F700, 0x0001F776}, +{0x0001F77B, 0x0001F7D9}, {0x0001F7E0, 0x0001F7EB}, {0x0001F7F0, 0x0001F7F0}, {0x0001F800, 0x0001F80B}, +{0x0001F810, 0x0001F847}, {0x0001F850, 0x0001F859}, {0x0001F860, 0x0001F887}, {0x0001F890, 0x0001F8AD}, +{0x0001F8B0, 0x0001F8B1}, {0x0001F900, 0x0001FA53}, {0x0001FA60, 0x0001FA6D}, {0x0001FA70, 0x0001FA7C}, +{0x0001FA80, 0x0001FA88}, {0x0001FA90, 0x0001FABD}, {0x0001FABF, 0x0001FAC5}, {0x0001FACE, 0x0001FADB}, +{0x0001FAE0, 0x0001FAE8}, {0x0001FAF0, 0x0001FAF8}, {0x0001FB00, 0x0001FB92}, {0x0001FB94, 0x0001FBCA}, }; const std::vector> unicode_ranges_control = { -//{0x00000009, 0x0000000D}, {0x00000085, 0x00000085}, -{0x0000001C, 0x0000001F}, -{0x00000000, 0x00000008}, {0x0000000E, 0x0000001B}, {0x0000007F, 0x00000084}, {0x00000086, 0x0000009F}, -{0x000000AD, 0x000000AD}, {0x00000378, 0x00000379}, {0x00000380, 0x00000383}, {0x0000038B, 0x0000038B}, -{0x0000038D, 0x0000038D}, {0x000003A2, 0x000003A2}, {0x00000530, 0x00000530}, {0x00000557, 0x00000558}, -{0x0000058B, 0x0000058C}, {0x00000590, 0x00000590}, {0x000005C8, 0x000005CF}, {0x000005EB, 0x000005EE}, -{0x000005F5, 0x00000605}, {0x0000061C, 0x0000061D}, {0x000006DD, 0x000006DD}, {0x0000070E, 0x0000070F}, -{0x0000074B, 0x0000074C}, {0x000007B2, 0x000007BF}, {0x000007FB, 0x000007FC}, {0x0000082E, 0x0000082F}, -{0x0000083F, 0x0000083F}, {0x0000085C, 0x0000085D}, {0x0000085F, 0x0000085F}, {0x0000086B, 0x0000089F}, -{0x000008B5, 0x000008B5}, {0x000008C8, 0x000008D2}, {0x000008E2, 0x000008E2}, {0x00000984, 0x00000984}, -{0x0000098D, 0x0000098E}, {0x00000991, 0x00000992}, {0x000009A9, 0x000009A9}, {0x000009B1, 0x000009B1}, -{0x000009B3, 0x000009B5}, {0x000009BA, 0x000009BB}, {0x000009C5, 0x000009C6}, {0x000009C9, 0x000009CA}, -{0x000009CF, 0x000009D6}, {0x000009D8, 0x000009DB}, {0x000009DE, 0x000009DE}, {0x000009E4, 0x000009E5}, -{0x000009FF, 0x00000A00}, {0x00000A04, 0x00000A04}, {0x00000A0B, 0x00000A0E}, {0x00000A11, 0x00000A12}, -{0x00000A29, 0x00000A29}, {0x00000A31, 0x00000A31}, {0x00000A34, 0x00000A34}, {0x00000A37, 0x00000A37}, -{0x00000A3A, 0x00000A3B}, {0x00000A3D, 0x00000A3D}, {0x00000A43, 0x00000A46}, {0x00000A49, 0x00000A4A}, -{0x00000A4E, 0x00000A50}, {0x00000A52, 0x00000A58}, {0x00000A5D, 0x00000A5D}, {0x00000A5F, 0x00000A65}, -{0x00000A77, 0x00000A80}, {0x00000A84, 0x00000A84}, {0x00000A8E, 0x00000A8E}, {0x00000A92, 0x00000A92}, -{0x00000AA9, 0x00000AA9}, {0x00000AB1, 0x00000AB1}, {0x00000AB4, 0x00000AB4}, {0x00000ABA, 0x00000ABB}, -{0x00000AC6, 0x00000AC6}, {0x00000ACA, 0x00000ACA}, {0x00000ACE, 0x00000ACF}, {0x00000AD1, 0x00000ADF}, -{0x00000AE4, 0x00000AE5}, {0x00000AF2, 0x00000AF8}, {0x00000B00, 0x00000B00}, {0x00000B04, 0x00000B04}, -{0x00000B0D, 0x00000B0E}, {0x00000B11, 0x00000B12}, {0x00000B29, 0x00000B29}, {0x00000B31, 0x00000B31}, -{0x00000B34, 0x00000B34}, {0x00000B3A, 0x00000B3B}, {0x00000B45, 0x00000B46}, {0x00000B49, 0x00000B4A}, -{0x00000B4E, 0x00000B54}, {0x00000B58, 0x00000B5B}, {0x00000B5E, 0x00000B5E}, {0x00000B64, 0x00000B65}, -{0x00000B78, 0x00000B81}, {0x00000B84, 0x00000B84}, {0x00000B8B, 0x00000B8D}, {0x00000B91, 0x00000B91}, -{0x00000B96, 0x00000B98}, {0x00000B9B, 0x00000B9B}, {0x00000B9D, 0x00000B9D}, {0x00000BA0, 0x00000BA2}, -{0x00000BA5, 0x00000BA7}, {0x00000BAB, 0x00000BAD}, {0x00000BBA, 0x00000BBD}, {0x00000BC3, 0x00000BC5}, -{0x00000BC9, 0x00000BC9}, {0x00000BCE, 0x00000BCF}, {0x00000BD1, 0x00000BD6}, {0x00000BD8, 0x00000BE5}, -{0x00000BFB, 0x00000BFF}, {0x00000C0D, 0x00000C0D}, {0x00000C11, 0x00000C11}, {0x00000C29, 0x00000C29}, -{0x00000C3A, 0x00000C3C}, {0x00000C45, 0x00000C45}, {0x00000C49, 0x00000C49}, {0x00000C4E, 0x00000C54}, -{0x00000C57, 0x00000C57}, {0x00000C5B, 0x00000C5F}, {0x00000C64, 0x00000C65}, {0x00000C70, 0x00000C76}, -{0x00000C8D, 0x00000C8D}, {0x00000C91, 0x00000C91}, {0x00000CA9, 0x00000CA9}, {0x00000CB4, 0x00000CB4}, -{0x00000CBA, 0x00000CBB}, {0x00000CC5, 0x00000CC5}, {0x00000CC9, 0x00000CC9}, {0x00000CCE, 0x00000CD4}, -{0x00000CD7, 0x00000CDD}, {0x00000CDF, 0x00000CDF}, {0x00000CE4, 0x00000CE5}, {0x00000CF0, 0x00000CF0}, -{0x00000CF3, 0x00000CFF}, {0x00000D0D, 0x00000D0D}, {0x00000D11, 0x00000D11}, {0x00000D45, 0x00000D45}, -{0x00000D49, 0x00000D49}, {0x00000D50, 0x00000D53}, {0x00000D64, 0x00000D65}, {0x00000D80, 0x00000D80}, -{0x00000D84, 0x00000D84}, {0x00000D97, 0x00000D99}, {0x00000DB2, 0x00000DB2}, {0x00000DBC, 0x00000DBC}, -{0x00000DBE, 0x00000DBF}, {0x00000DC7, 0x00000DC9}, {0x00000DCB, 0x00000DCE}, {0x00000DD5, 0x00000DD5}, -{0x00000DD7, 0x00000DD7}, {0x00000DE0, 0x00000DE5}, {0x00000DF0, 0x00000DF1}, {0x00000DF5, 0x00000E00}, -{0x00000E3B, 0x00000E3E}, {0x00000E5C, 0x00000E80}, {0x00000E83, 0x00000E83}, {0x00000E85, 0x00000E85}, -{0x00000E8B, 0x00000E8B}, {0x00000EA4, 0x00000EA4}, {0x00000EA6, 0x00000EA6}, {0x00000EBE, 0x00000EBF}, -{0x00000EC5, 0x00000EC5}, {0x00000EC7, 0x00000EC7}, {0x00000ECE, 0x00000ECF}, {0x00000EDA, 0x00000EDB}, -{0x00000EE0, 0x00000EFF}, {0x00000F48, 0x00000F48}, {0x00000F6D, 0x00000F70}, {0x00000F98, 0x00000F98}, -{0x00000FBD, 0x00000FBD}, {0x00000FCD, 0x00000FCD}, {0x00000FDB, 0x00000FFF}, {0x000010C6, 0x000010C6}, -{0x000010C8, 0x000010CC}, {0x000010CE, 0x000010CF}, {0x00001249, 0x00001249}, {0x0000124E, 0x0000124F}, -{0x00001257, 0x00001257}, {0x00001259, 0x00001259}, {0x0000125E, 0x0000125F}, {0x00001289, 0x00001289}, -{0x0000128E, 0x0000128F}, {0x000012B1, 0x000012B1}, {0x000012B6, 0x000012B7}, {0x000012BF, 0x000012BF}, -{0x000012C1, 0x000012C1}, {0x000012C6, 0x000012C7}, {0x000012D7, 0x000012D7}, {0x00001311, 0x00001311}, -{0x00001316, 0x00001317}, {0x0000135B, 0x0000135C}, {0x0000137D, 0x0000137F}, {0x0000139A, 0x0000139F}, -{0x000013F6, 0x000013F7}, {0x000013FE, 0x000013FF}, {0x0000169D, 0x0000169F}, {0x000016F9, 0x000016FF}, -{0x0000170D, 0x0000170D}, {0x00001715, 0x0000171F}, {0x00001737, 0x0000173F}, {0x00001754, 0x0000175F}, -{0x0000176D, 0x0000176D}, {0x00001771, 0x00001771}, {0x00001774, 0x0000177F}, {0x000017DE, 0x000017DF}, -{0x000017EA, 0x000017EF}, {0x000017FA, 0x000017FF}, {0x0000180E, 0x0000180F}, {0x0000181A, 0x0000181F}, -{0x00001879, 0x0000187F}, {0x000018AB, 0x000018AF}, {0x000018F6, 0x000018FF}, {0x0000191F, 0x0000191F}, -{0x0000192C, 0x0000192F}, {0x0000193C, 0x0000193F}, {0x00001941, 0x00001943}, {0x0000196E, 0x0000196F}, -{0x00001975, 0x0000197F}, {0x000019AC, 0x000019AF}, {0x000019CA, 0x000019CF}, {0x000019DB, 0x000019DD}, -{0x00001A1C, 0x00001A1D}, {0x00001A5F, 0x00001A5F}, {0x00001A7D, 0x00001A7E}, {0x00001A8A, 0x00001A8F}, -{0x00001A9A, 0x00001A9F}, {0x00001AAE, 0x00001AAF}, {0x00001AC1, 0x00001AFF}, {0x00001B4C, 0x00001B4F}, -{0x00001B7D, 0x00001B7F}, {0x00001BF4, 0x00001BFB}, {0x00001C38, 0x00001C3A}, {0x00001C4A, 0x00001C4C}, -{0x00001C89, 0x00001C8F}, {0x00001CBB, 0x00001CBC}, {0x00001CC8, 0x00001CCF}, {0x00001CFB, 0x00001CFF}, -{0x00001DFA, 0x00001DFA}, {0x00001F16, 0x00001F17}, {0x00001F1E, 0x00001F1F}, {0x00001F46, 0x00001F47}, +{0x00000000, 0x0000001F}, {0x0000007F, 0x0000009F}, {0x000000AD, 0x000000AD}, {0x00000378, 0x00000379}, +{0x00000380, 0x00000383}, {0x0000038B, 0x0000038B}, {0x0000038D, 0x0000038D}, {0x000003A2, 0x000003A2}, +{0x00000530, 0x00000530}, {0x00000557, 0x00000558}, {0x0000058B, 0x0000058C}, {0x00000590, 0x00000590}, +{0x000005C8, 0x000005CF}, {0x000005EB, 0x000005EE}, {0x000005F5, 0x00000605}, {0x0000061C, 0x0000061C}, +{0x000006DD, 0x000006DD}, {0x0000070E, 0x0000070F}, {0x0000074B, 0x0000074C}, {0x000007B2, 0x000007BF}, +{0x000007FB, 0x000007FC}, {0x0000082E, 0x0000082F}, {0x0000083F, 0x0000083F}, {0x0000085C, 0x0000085D}, +{0x0000085F, 0x0000085F}, {0x0000086B, 0x0000086F}, {0x0000088F, 0x00000897}, {0x000008E2, 0x000008E2}, +{0x00000984, 0x00000984}, {0x0000098D, 0x0000098E}, {0x00000991, 0x00000992}, {0x000009A9, 0x000009A9}, +{0x000009B1, 0x000009B1}, {0x000009B3, 0x000009B5}, {0x000009BA, 0x000009BB}, {0x000009C5, 0x000009C6}, +{0x000009C9, 0x000009CA}, {0x000009CF, 0x000009D6}, {0x000009D8, 0x000009DB}, {0x000009DE, 0x000009DE}, +{0x000009E4, 0x000009E5}, {0x000009FF, 0x00000A00}, {0x00000A04, 0x00000A04}, {0x00000A0B, 0x00000A0E}, +{0x00000A11, 0x00000A12}, {0x00000A29, 0x00000A29}, {0x00000A31, 0x00000A31}, {0x00000A34, 0x00000A34}, +{0x00000A37, 0x00000A37}, {0x00000A3A, 0x00000A3B}, {0x00000A3D, 0x00000A3D}, {0x00000A43, 0x00000A46}, +{0x00000A49, 0x00000A4A}, {0x00000A4E, 0x00000A50}, {0x00000A52, 0x00000A58}, {0x00000A5D, 0x00000A5D}, +{0x00000A5F, 0x00000A65}, {0x00000A77, 0x00000A80}, {0x00000A84, 0x00000A84}, {0x00000A8E, 0x00000A8E}, +{0x00000A92, 0x00000A92}, {0x00000AA9, 0x00000AA9}, {0x00000AB1, 0x00000AB1}, {0x00000AB4, 0x00000AB4}, +{0x00000ABA, 0x00000ABB}, {0x00000AC6, 0x00000AC6}, {0x00000ACA, 0x00000ACA}, {0x00000ACE, 0x00000ACF}, +{0x00000AD1, 0x00000ADF}, {0x00000AE4, 0x00000AE5}, {0x00000AF2, 0x00000AF8}, {0x00000B00, 0x00000B00}, +{0x00000B04, 0x00000B04}, {0x00000B0D, 0x00000B0E}, {0x00000B11, 0x00000B12}, {0x00000B29, 0x00000B29}, +{0x00000B31, 0x00000B31}, {0x00000B34, 0x00000B34}, {0x00000B3A, 0x00000B3B}, {0x00000B45, 0x00000B46}, +{0x00000B49, 0x00000B4A}, {0x00000B4E, 0x00000B54}, {0x00000B58, 0x00000B5B}, {0x00000B5E, 0x00000B5E}, +{0x00000B64, 0x00000B65}, {0x00000B78, 0x00000B81}, {0x00000B84, 0x00000B84}, {0x00000B8B, 0x00000B8D}, +{0x00000B91, 0x00000B91}, {0x00000B96, 0x00000B98}, {0x00000B9B, 0x00000B9B}, {0x00000B9D, 0x00000B9D}, +{0x00000BA0, 0x00000BA2}, {0x00000BA5, 0x00000BA7}, {0x00000BAB, 0x00000BAD}, {0x00000BBA, 0x00000BBD}, +{0x00000BC3, 0x00000BC5}, {0x00000BC9, 0x00000BC9}, {0x00000BCE, 0x00000BCF}, {0x00000BD1, 0x00000BD6}, +{0x00000BD8, 0x00000BE5}, {0x00000BFB, 0x00000BFF}, {0x00000C0D, 0x00000C0D}, {0x00000C11, 0x00000C11}, +{0x00000C29, 0x00000C29}, {0x00000C3A, 0x00000C3B}, {0x00000C45, 0x00000C45}, {0x00000C49, 0x00000C49}, +{0x00000C4E, 0x00000C54}, {0x00000C57, 0x00000C57}, {0x00000C5B, 0x00000C5C}, {0x00000C5E, 0x00000C5F}, +{0x00000C64, 0x00000C65}, {0x00000C70, 0x00000C76}, {0x00000C8D, 0x00000C8D}, {0x00000C91, 0x00000C91}, +{0x00000CA9, 0x00000CA9}, {0x00000CB4, 0x00000CB4}, {0x00000CBA, 0x00000CBB}, {0x00000CC5, 0x00000CC5}, +{0x00000CC9, 0x00000CC9}, {0x00000CCE, 0x00000CD4}, {0x00000CD7, 0x00000CDC}, {0x00000CDF, 0x00000CDF}, +{0x00000CE4, 0x00000CE5}, {0x00000CF0, 0x00000CF0}, {0x00000CF4, 0x00000CFF}, {0x00000D0D, 0x00000D0D}, +{0x00000D11, 0x00000D11}, {0x00000D45, 0x00000D45}, {0x00000D49, 0x00000D49}, {0x00000D50, 0x00000D53}, +{0x00000D64, 0x00000D65}, {0x00000D80, 0x00000D80}, {0x00000D84, 0x00000D84}, {0x00000D97, 0x00000D99}, +{0x00000DB2, 0x00000DB2}, {0x00000DBC, 0x00000DBC}, {0x00000DBE, 0x00000DBF}, {0x00000DC7, 0x00000DC9}, +{0x00000DCB, 0x00000DCE}, {0x00000DD5, 0x00000DD5}, {0x00000DD7, 0x00000DD7}, {0x00000DE0, 0x00000DE5}, +{0x00000DF0, 0x00000DF1}, {0x00000DF5, 0x00000E00}, {0x00000E3B, 0x00000E3E}, {0x00000E5C, 0x00000E80}, +{0x00000E83, 0x00000E83}, {0x00000E85, 0x00000E85}, {0x00000E8B, 0x00000E8B}, {0x00000EA4, 0x00000EA4}, +{0x00000EA6, 0x00000EA6}, {0x00000EBE, 0x00000EBF}, {0x00000EC5, 0x00000EC5}, {0x00000EC7, 0x00000EC7}, +{0x00000ECF, 0x00000ECF}, {0x00000EDA, 0x00000EDB}, {0x00000EE0, 0x00000EFF}, {0x00000F48, 0x00000F48}, +{0x00000F6D, 0x00000F70}, {0x00000F98, 0x00000F98}, {0x00000FBD, 0x00000FBD}, {0x00000FCD, 0x00000FCD}, +{0x00000FDB, 0x00000FFF}, {0x000010C6, 0x000010C6}, {0x000010C8, 0x000010CC}, {0x000010CE, 0x000010CF}, +{0x00001249, 0x00001249}, {0x0000124E, 0x0000124F}, {0x00001257, 0x00001257}, {0x00001259, 0x00001259}, +{0x0000125E, 0x0000125F}, {0x00001289, 0x00001289}, {0x0000128E, 0x0000128F}, {0x000012B1, 0x000012B1}, +{0x000012B6, 0x000012B7}, {0x000012BF, 0x000012BF}, {0x000012C1, 0x000012C1}, {0x000012C6, 0x000012C7}, +{0x000012D7, 0x000012D7}, {0x00001311, 0x00001311}, {0x00001316, 0x00001317}, {0x0000135B, 0x0000135C}, +{0x0000137D, 0x0000137F}, {0x0000139A, 0x0000139F}, {0x000013F6, 0x000013F7}, {0x000013FE, 0x000013FF}, +{0x0000169D, 0x0000169F}, {0x000016F9, 0x000016FF}, {0x00001716, 0x0000171E}, {0x00001737, 0x0000173F}, +{0x00001754, 0x0000175F}, {0x0000176D, 0x0000176D}, {0x00001771, 0x00001771}, {0x00001774, 0x0000177F}, +{0x000017DE, 0x000017DF}, {0x000017EA, 0x000017EF}, {0x000017FA, 0x000017FF}, {0x0000180E, 0x0000180E}, +{0x0000181A, 0x0000181F}, {0x00001879, 0x0000187F}, {0x000018AB, 0x000018AF}, {0x000018F6, 0x000018FF}, +{0x0000191F, 0x0000191F}, {0x0000192C, 0x0000192F}, {0x0000193C, 0x0000193F}, {0x00001941, 0x00001943}, +{0x0000196E, 0x0000196F}, {0x00001975, 0x0000197F}, {0x000019AC, 0x000019AF}, {0x000019CA, 0x000019CF}, +{0x000019DB, 0x000019DD}, {0x00001A1C, 0x00001A1D}, {0x00001A5F, 0x00001A5F}, {0x00001A7D, 0x00001A7E}, +{0x00001A8A, 0x00001A8F}, {0x00001A9A, 0x00001A9F}, {0x00001AAE, 0x00001AAF}, {0x00001ACF, 0x00001AFF}, +{0x00001B4D, 0x00001B4F}, {0x00001B7F, 0x00001B7F}, {0x00001BF4, 0x00001BFB}, {0x00001C38, 0x00001C3A}, +{0x00001C4A, 0x00001C4C}, {0x00001C89, 0x00001C8F}, {0x00001CBB, 0x00001CBC}, {0x00001CC8, 0x00001CCF}, +{0x00001CFB, 0x00001CFF}, {0x00001F16, 0x00001F17}, {0x00001F1E, 0x00001F1F}, {0x00001F46, 0x00001F47}, {0x00001F4E, 0x00001F4F}, {0x00001F58, 0x00001F58}, {0x00001F5A, 0x00001F5A}, {0x00001F5C, 0x00001F5C}, {0x00001F5E, 0x00001F5E}, {0x00001F7E, 0x00001F7F}, {0x00001FB5, 0x00001FB5}, {0x00001FC5, 0x00001FC5}, {0x00001FD4, 0x00001FD5}, {0x00001FDC, 0x00001FDC}, {0x00001FF0, 0x00001FF1}, {0x00001FF5, 0x00001FF5}, {0x00001FFF, 0x00001FFF}, {0x0000200B, 0x0000200F}, {0x0000202A, 0x0000202E}, {0x00002060, 0x0000206F}, -{0x00002072, 0x00002073}, {0x0000208F, 0x0000208F}, {0x0000209D, 0x0000209F}, {0x000020C0, 0x000020CF}, +{0x00002072, 0x00002073}, {0x0000208F, 0x0000208F}, {0x0000209D, 0x0000209F}, {0x000020C1, 0x000020CF}, {0x000020F1, 0x000020FF}, {0x0000218C, 0x0000218F}, {0x00002427, 0x0000243F}, {0x0000244B, 0x0000245F}, -{0x00002B74, 0x00002B75}, {0x00002B96, 0x00002B96}, {0x00002C2F, 0x00002C2F}, {0x00002C5F, 0x00002C5F}, -{0x00002CF4, 0x00002CF8}, {0x00002D26, 0x00002D26}, {0x00002D28, 0x00002D2C}, {0x00002D2E, 0x00002D2F}, -{0x00002D68, 0x00002D6E}, {0x00002D71, 0x00002D7E}, {0x00002D97, 0x00002D9F}, {0x00002DA7, 0x00002DA7}, -{0x00002DAF, 0x00002DAF}, {0x00002DB7, 0x00002DB7}, {0x00002DBF, 0x00002DBF}, {0x00002DC7, 0x00002DC7}, -{0x00002DCF, 0x00002DCF}, {0x00002DD7, 0x00002DD7}, {0x00002DDF, 0x00002DDF}, {0x00002E53, 0x00002E7F}, -{0x00002E9A, 0x00002E9A}, {0x00002EF4, 0x00002EFF}, {0x00002FD6, 0x00002FEF}, {0x00002FFC, 0x00002FFF}, -{0x00003040, 0x00003040}, {0x00003097, 0x00003098}, {0x00003100, 0x00003104}, {0x00003130, 0x00003130}, -{0x0000318F, 0x0000318F}, {0x000031E4, 0x000031EF}, {0x0000321F, 0x0000321F}, {0x00009FFD, 0x00009FFF}, -{0x0000A48D, 0x0000A48F}, {0x0000A4C7, 0x0000A4CF}, {0x0000A62C, 0x0000A63F}, {0x0000A6F8, 0x0000A6FF}, -{0x0000A7C0, 0x0000A7C1}, {0x0000A7CB, 0x0000A7F4}, {0x0000A82D, 0x0000A82F}, {0x0000A83A, 0x0000A83F}, -{0x0000A878, 0x0000A87F}, {0x0000A8C6, 0x0000A8CD}, {0x0000A8DA, 0x0000A8DF}, {0x0000A954, 0x0000A95E}, -{0x0000A97D, 0x0000A97F}, {0x0000A9CE, 0x0000A9CE}, {0x0000A9DA, 0x0000A9DD}, {0x0000A9FF, 0x0000A9FF}, -{0x0000AA37, 0x0000AA3F}, {0x0000AA4E, 0x0000AA4F}, {0x0000AA5A, 0x0000AA5B}, {0x0000AAC3, 0x0000AADA}, -{0x0000AAF7, 0x0000AB00}, {0x0000AB07, 0x0000AB08}, {0x0000AB0F, 0x0000AB10}, {0x0000AB17, 0x0000AB1F}, -{0x0000AB27, 0x0000AB27}, {0x0000AB2F, 0x0000AB2F}, {0x0000AB6C, 0x0000AB6F}, {0x0000ABEE, 0x0000ABEF}, -{0x0000ABFA, 0x0000ABFF}, {0x0000D7A4, 0x0000D7AF}, {0x0000D7C7, 0x0000D7CA}, {0x0000D7FC, 0x0000F8FF}, +{0x00002B74, 0x00002B75}, {0x00002B96, 0x00002B96}, {0x00002CF4, 0x00002CF8}, {0x00002D26, 0x00002D26}, +{0x00002D28, 0x00002D2C}, {0x00002D2E, 0x00002D2F}, {0x00002D68, 0x00002D6E}, {0x00002D71, 0x00002D7E}, +{0x00002D97, 0x00002D9F}, {0x00002DA7, 0x00002DA7}, {0x00002DAF, 0x00002DAF}, {0x00002DB7, 0x00002DB7}, +{0x00002DBF, 0x00002DBF}, {0x00002DC7, 0x00002DC7}, {0x00002DCF, 0x00002DCF}, {0x00002DD7, 0x00002DD7}, +{0x00002DDF, 0x00002DDF}, {0x00002E5E, 0x00002E7F}, {0x00002E9A, 0x00002E9A}, {0x00002EF4, 0x00002EFF}, +{0x00002FD6, 0x00002FEF}, {0x00002FFC, 0x00002FFF}, {0x00003040, 0x00003040}, {0x00003097, 0x00003098}, +{0x00003100, 0x00003104}, {0x00003130, 0x00003130}, {0x0000318F, 0x0000318F}, {0x000031E4, 0x000031EF}, +{0x0000321F, 0x0000321F}, {0x0000A48D, 0x0000A48F}, {0x0000A4C7, 0x0000A4CF}, {0x0000A62C, 0x0000A63F}, +{0x0000A6F8, 0x0000A6FF}, {0x0000A7CB, 0x0000A7CF}, {0x0000A7D2, 0x0000A7D2}, {0x0000A7D4, 0x0000A7D4}, +{0x0000A7DA, 0x0000A7F1}, {0x0000A82D, 0x0000A82F}, {0x0000A83A, 0x0000A83F}, {0x0000A878, 0x0000A87F}, +{0x0000A8C6, 0x0000A8CD}, {0x0000A8DA, 0x0000A8DF}, {0x0000A954, 0x0000A95E}, {0x0000A97D, 0x0000A97F}, +{0x0000A9CE, 0x0000A9CE}, {0x0000A9DA, 0x0000A9DD}, {0x0000A9FF, 0x0000A9FF}, {0x0000AA37, 0x0000AA3F}, +{0x0000AA4E, 0x0000AA4F}, {0x0000AA5A, 0x0000AA5B}, {0x0000AAC3, 0x0000AADA}, {0x0000AAF7, 0x0000AB00}, +{0x0000AB07, 0x0000AB08}, {0x0000AB0F, 0x0000AB10}, {0x0000AB17, 0x0000AB1F}, {0x0000AB27, 0x0000AB27}, +{0x0000AB2F, 0x0000AB2F}, {0x0000AB6C, 0x0000AB6F}, {0x0000ABEE, 0x0000ABEF}, {0x0000ABFA, 0x0000ABFF}, +{0x0000D7A4, 0x0000D7AF}, {0x0000D7C7, 0x0000D7CA}, {0x0000D7FC, 0x0000D7FF}, {0x0000E000, 0x0000F8FF}, {0x0000FA6E, 0x0000FA6F}, {0x0000FADA, 0x0000FAFF}, {0x0000FB07, 0x0000FB12}, {0x0000FB18, 0x0000FB1C}, {0x0000FB37, 0x0000FB37}, {0x0000FB3D, 0x0000FB3D}, {0x0000FB3F, 0x0000FB3F}, {0x0000FB42, 0x0000FB42}, -{0x0000FB45, 0x0000FB45}, {0x0000FBC2, 0x0000FBD2}, {0x0000FD40, 0x0000FD4F}, {0x0000FD90, 0x0000FD91}, -{0x0000FDC8, 0x0000FDEF}, {0x0000FDFE, 0x0000FDFF}, {0x0000FE1A, 0x0000FE1F}, {0x0000FE53, 0x0000FE53}, -{0x0000FE67, 0x0000FE67}, {0x0000FE6C, 0x0000FE6F}, {0x0000FE75, 0x0000FE75}, {0x0000FEFD, 0x0000FF00}, +{0x0000FB45, 0x0000FB45}, {0x0000FBC3, 0x0000FBD2}, {0x0000FD90, 0x0000FD91}, {0x0000FDC8, 0x0000FDCE}, +{0x0000FDD0, 0x0000FDEF}, {0x0000FE1A, 0x0000FE1F}, {0x0000FE53, 0x0000FE53}, {0x0000FE67, 0x0000FE67}, +{0x0000FE6C, 0x0000FE6F}, {0x0000FE75, 0x0000FE75}, {0x0000FEFD, 0x0000FEFE}, {0x0000FF00, 0x0000FF00}, {0x0000FFBF, 0x0000FFC1}, {0x0000FFC8, 0x0000FFC9}, {0x0000FFD0, 0x0000FFD1}, {0x0000FFD8, 0x0000FFD9}, {0x0000FFDD, 0x0000FFDF}, {0x0000FFE7, 0x0000FFE7}, {0x0000FFEF, 0x0000FFFB}, {0x0000FFFE, 0x0000FFFF}, {0x0001000C, 0x0001000C}, {0x00010027, 0x00010027}, {0x0001003B, 0x0001003B}, {0x0001003E, 0x0001003E}, @@ -492,82 +509,91 @@ const std::vector> unicode_ranges_control = { {0x00010324, 0x0001032C}, {0x0001034B, 0x0001034F}, {0x0001037B, 0x0001037F}, {0x0001039E, 0x0001039E}, {0x000103C4, 0x000103C7}, {0x000103D6, 0x000103FF}, {0x0001049E, 0x0001049F}, {0x000104AA, 0x000104AF}, {0x000104D4, 0x000104D7}, {0x000104FC, 0x000104FF}, {0x00010528, 0x0001052F}, {0x00010564, 0x0001056E}, -{0x00010570, 0x000105FF}, {0x00010737, 0x0001073F}, {0x00010756, 0x0001075F}, {0x00010768, 0x000107FF}, -{0x00010806, 0x00010807}, {0x00010809, 0x00010809}, {0x00010836, 0x00010836}, {0x00010839, 0x0001083B}, -{0x0001083D, 0x0001083E}, {0x00010856, 0x00010856}, {0x0001089F, 0x000108A6}, {0x000108B0, 0x000108DF}, -{0x000108F3, 0x000108F3}, {0x000108F6, 0x000108FA}, {0x0001091C, 0x0001091E}, {0x0001093A, 0x0001093E}, -{0x00010940, 0x0001097F}, {0x000109B8, 0x000109BB}, {0x000109D0, 0x000109D1}, {0x00010A04, 0x00010A04}, -{0x00010A07, 0x00010A0B}, {0x00010A14, 0x00010A14}, {0x00010A18, 0x00010A18}, {0x00010A36, 0x00010A37}, -{0x00010A3B, 0x00010A3E}, {0x00010A49, 0x00010A4F}, {0x00010A59, 0x00010A5F}, {0x00010AA0, 0x00010ABF}, -{0x00010AE7, 0x00010AEA}, {0x00010AF7, 0x00010AFF}, {0x00010B36, 0x00010B38}, {0x00010B56, 0x00010B57}, -{0x00010B73, 0x00010B77}, {0x00010B92, 0x00010B98}, {0x00010B9D, 0x00010BA8}, {0x00010BB0, 0x00010BFF}, -{0x00010C49, 0x00010C7F}, {0x00010CB3, 0x00010CBF}, {0x00010CF3, 0x00010CF9}, {0x00010D28, 0x00010D2F}, -{0x00010D3A, 0x00010E5F}, {0x00010E7F, 0x00010E7F}, {0x00010EAA, 0x00010EAA}, {0x00010EAE, 0x00010EAF}, -{0x00010EB2, 0x00010EFF}, {0x00010F28, 0x00010F2F}, {0x00010F5A, 0x00010FAF}, {0x00010FCC, 0x00010FDF}, -{0x00010FF7, 0x00010FFF}, {0x0001104E, 0x00011051}, {0x00011070, 0x0001107E}, {0x000110BD, 0x000110BD}, -{0x000110C2, 0x000110CF}, {0x000110E9, 0x000110EF}, {0x000110FA, 0x000110FF}, {0x00011135, 0x00011135}, -{0x00011148, 0x0001114F}, {0x00011177, 0x0001117F}, {0x000111E0, 0x000111E0}, {0x000111F5, 0x000111FF}, -{0x00011212, 0x00011212}, {0x0001123F, 0x0001127F}, {0x00011287, 0x00011287}, {0x00011289, 0x00011289}, -{0x0001128E, 0x0001128E}, {0x0001129E, 0x0001129E}, {0x000112AA, 0x000112AF}, {0x000112EB, 0x000112EF}, -{0x000112FA, 0x000112FF}, {0x00011304, 0x00011304}, {0x0001130D, 0x0001130E}, {0x00011311, 0x00011312}, -{0x00011329, 0x00011329}, {0x00011331, 0x00011331}, {0x00011334, 0x00011334}, {0x0001133A, 0x0001133A}, -{0x00011345, 0x00011346}, {0x00011349, 0x0001134A}, {0x0001134E, 0x0001134F}, {0x00011351, 0x00011356}, -{0x00011358, 0x0001135C}, {0x00011364, 0x00011365}, {0x0001136D, 0x0001136F}, {0x00011375, 0x000113FF}, -{0x0001145C, 0x0001145C}, {0x00011462, 0x0001147F}, {0x000114C8, 0x000114CF}, {0x000114DA, 0x0001157F}, -{0x000115B6, 0x000115B7}, {0x000115DE, 0x000115FF}, {0x00011645, 0x0001164F}, {0x0001165A, 0x0001165F}, -{0x0001166D, 0x0001167F}, {0x000116B9, 0x000116BF}, {0x000116CA, 0x000116FF}, {0x0001171B, 0x0001171C}, -{0x0001172C, 0x0001172F}, {0x00011740, 0x000117FF}, {0x0001183C, 0x0001189F}, {0x000118F3, 0x000118FE}, -{0x00011907, 0x00011908}, {0x0001190A, 0x0001190B}, {0x00011914, 0x00011914}, {0x00011917, 0x00011917}, -{0x00011936, 0x00011936}, {0x00011939, 0x0001193A}, {0x00011947, 0x0001194F}, {0x0001195A, 0x0001199F}, -{0x000119A8, 0x000119A9}, {0x000119D8, 0x000119D9}, {0x000119E5, 0x000119FF}, {0x00011A48, 0x00011A4F}, -{0x00011AA3, 0x00011ABF}, {0x00011AF9, 0x00011BFF}, {0x00011C09, 0x00011C09}, {0x00011C37, 0x00011C37}, +{0x0001057B, 0x0001057B}, {0x0001058B, 0x0001058B}, {0x00010593, 0x00010593}, {0x00010596, 0x00010596}, +{0x000105A2, 0x000105A2}, {0x000105B2, 0x000105B2}, {0x000105BA, 0x000105BA}, {0x000105BD, 0x000105FF}, +{0x00010737, 0x0001073F}, {0x00010756, 0x0001075F}, {0x00010768, 0x0001077F}, {0x00010786, 0x00010786}, +{0x000107B1, 0x000107B1}, {0x000107BB, 0x000107FF}, {0x00010806, 0x00010807}, {0x00010809, 0x00010809}, +{0x00010836, 0x00010836}, {0x00010839, 0x0001083B}, {0x0001083D, 0x0001083E}, {0x00010856, 0x00010856}, +{0x0001089F, 0x000108A6}, {0x000108B0, 0x000108DF}, {0x000108F3, 0x000108F3}, {0x000108F6, 0x000108FA}, +{0x0001091C, 0x0001091E}, {0x0001093A, 0x0001093E}, {0x00010940, 0x0001097F}, {0x000109B8, 0x000109BB}, +{0x000109D0, 0x000109D1}, {0x00010A04, 0x00010A04}, {0x00010A07, 0x00010A0B}, {0x00010A14, 0x00010A14}, +{0x00010A18, 0x00010A18}, {0x00010A36, 0x00010A37}, {0x00010A3B, 0x00010A3E}, {0x00010A49, 0x00010A4F}, +{0x00010A59, 0x00010A5F}, {0x00010AA0, 0x00010ABF}, {0x00010AE7, 0x00010AEA}, {0x00010AF7, 0x00010AFF}, +{0x00010B36, 0x00010B38}, {0x00010B56, 0x00010B57}, {0x00010B73, 0x00010B77}, {0x00010B92, 0x00010B98}, +{0x00010B9D, 0x00010BA8}, {0x00010BB0, 0x00010BFF}, {0x00010C49, 0x00010C7F}, {0x00010CB3, 0x00010CBF}, +{0x00010CF3, 0x00010CF9}, {0x00010D28, 0x00010D2F}, {0x00010D3A, 0x00010E5F}, {0x00010E7F, 0x00010E7F}, +{0x00010EAA, 0x00010EAA}, {0x00010EAE, 0x00010EAF}, {0x00010EB2, 0x00010EFC}, {0x00010F28, 0x00010F2F}, +{0x00010F5A, 0x00010F6F}, {0x00010F8A, 0x00010FAF}, {0x00010FCC, 0x00010FDF}, {0x00010FF7, 0x00010FFF}, +{0x0001104E, 0x00011051}, {0x00011076, 0x0001107E}, {0x000110BD, 0x000110BD}, {0x000110C3, 0x000110CF}, +{0x000110E9, 0x000110EF}, {0x000110FA, 0x000110FF}, {0x00011135, 0x00011135}, {0x00011148, 0x0001114F}, +{0x00011177, 0x0001117F}, {0x000111E0, 0x000111E0}, {0x000111F5, 0x000111FF}, {0x00011212, 0x00011212}, +{0x00011242, 0x0001127F}, {0x00011287, 0x00011287}, {0x00011289, 0x00011289}, {0x0001128E, 0x0001128E}, +{0x0001129E, 0x0001129E}, {0x000112AA, 0x000112AF}, {0x000112EB, 0x000112EF}, {0x000112FA, 0x000112FF}, +{0x00011304, 0x00011304}, {0x0001130D, 0x0001130E}, {0x00011311, 0x00011312}, {0x00011329, 0x00011329}, +{0x00011331, 0x00011331}, {0x00011334, 0x00011334}, {0x0001133A, 0x0001133A}, {0x00011345, 0x00011346}, +{0x00011349, 0x0001134A}, {0x0001134E, 0x0001134F}, {0x00011351, 0x00011356}, {0x00011358, 0x0001135C}, +{0x00011364, 0x00011365}, {0x0001136D, 0x0001136F}, {0x00011375, 0x000113FF}, {0x0001145C, 0x0001145C}, +{0x00011462, 0x0001147F}, {0x000114C8, 0x000114CF}, {0x000114DA, 0x0001157F}, {0x000115B6, 0x000115B7}, +{0x000115DE, 0x000115FF}, {0x00011645, 0x0001164F}, {0x0001165A, 0x0001165F}, {0x0001166D, 0x0001167F}, +{0x000116BA, 0x000116BF}, {0x000116CA, 0x000116FF}, {0x0001171B, 0x0001171C}, {0x0001172C, 0x0001172F}, +{0x00011747, 0x000117FF}, {0x0001183C, 0x0001189F}, {0x000118F3, 0x000118FE}, {0x00011907, 0x00011908}, +{0x0001190A, 0x0001190B}, {0x00011914, 0x00011914}, {0x00011917, 0x00011917}, {0x00011936, 0x00011936}, +{0x00011939, 0x0001193A}, {0x00011947, 0x0001194F}, {0x0001195A, 0x0001199F}, {0x000119A8, 0x000119A9}, +{0x000119D8, 0x000119D9}, {0x000119E5, 0x000119FF}, {0x00011A48, 0x00011A4F}, {0x00011AA3, 0x00011AAF}, +{0x00011AF9, 0x00011AFF}, {0x00011B0A, 0x00011BFF}, {0x00011C09, 0x00011C09}, {0x00011C37, 0x00011C37}, {0x00011C46, 0x00011C4F}, {0x00011C6D, 0x00011C6F}, {0x00011C90, 0x00011C91}, {0x00011CA8, 0x00011CA8}, {0x00011CB7, 0x00011CFF}, {0x00011D07, 0x00011D07}, {0x00011D0A, 0x00011D0A}, {0x00011D37, 0x00011D39}, {0x00011D3B, 0x00011D3B}, {0x00011D3E, 0x00011D3E}, {0x00011D48, 0x00011D4F}, {0x00011D5A, 0x00011D5F}, {0x00011D66, 0x00011D66}, {0x00011D69, 0x00011D69}, {0x00011D8F, 0x00011D8F}, {0x00011D92, 0x00011D92}, -{0x00011D99, 0x00011D9F}, {0x00011DAA, 0x00011EDF}, {0x00011EF9, 0x00011FAF}, {0x00011FB1, 0x00011FBF}, -{0x00011FF2, 0x00011FFE}, {0x0001239A, 0x000123FF}, {0x0001246F, 0x0001246F}, {0x00012475, 0x0001247F}, -{0x00012544, 0x00012FFF}, {0x0001342F, 0x000143FF}, {0x00014647, 0x000167FF}, {0x00016A39, 0x00016A3F}, -{0x00016A5F, 0x00016A5F}, {0x00016A6A, 0x00016A6D}, {0x00016A70, 0x00016ACF}, {0x00016AEE, 0x00016AEF}, -{0x00016AF6, 0x00016AFF}, {0x00016B46, 0x00016B4F}, {0x00016B5A, 0x00016B5A}, {0x00016B62, 0x00016B62}, -{0x00016B78, 0x00016B7C}, {0x00016B90, 0x00016E3F}, {0x00016E9B, 0x00016EFF}, {0x00016F4B, 0x00016F4E}, -{0x00016F88, 0x00016F8E}, {0x00016FA0, 0x00016FDF}, {0x00016FE5, 0x00016FEF}, {0x00016FF2, 0x00016FFF}, -{0x000187F8, 0x000187FF}, {0x00018CD6, 0x00018CFF}, {0x00018D09, 0x0001AFFF}, {0x0001B11F, 0x0001B14F}, -{0x0001B153, 0x0001B163}, {0x0001B168, 0x0001B16F}, {0x0001B2FC, 0x0001BBFF}, {0x0001BC6B, 0x0001BC6F}, -{0x0001BC7D, 0x0001BC7F}, {0x0001BC89, 0x0001BC8F}, {0x0001BC9A, 0x0001BC9B}, {0x0001BCA0, 0x0001CFFF}, -{0x0001D0F6, 0x0001D0FF}, {0x0001D127, 0x0001D128}, {0x0001D173, 0x0001D17A}, {0x0001D1E9, 0x0001D1FF}, -{0x0001D246, 0x0001D2DF}, {0x0001D2F4, 0x0001D2FF}, {0x0001D357, 0x0001D35F}, {0x0001D379, 0x0001D3FF}, -{0x0001D455, 0x0001D455}, {0x0001D49D, 0x0001D49D}, {0x0001D4A0, 0x0001D4A1}, {0x0001D4A3, 0x0001D4A4}, -{0x0001D4A7, 0x0001D4A8}, {0x0001D4AD, 0x0001D4AD}, {0x0001D4BA, 0x0001D4BA}, {0x0001D4BC, 0x0001D4BC}, -{0x0001D4C4, 0x0001D4C4}, {0x0001D506, 0x0001D506}, {0x0001D50B, 0x0001D50C}, {0x0001D515, 0x0001D515}, -{0x0001D51D, 0x0001D51D}, {0x0001D53A, 0x0001D53A}, {0x0001D53F, 0x0001D53F}, {0x0001D545, 0x0001D545}, -{0x0001D547, 0x0001D549}, {0x0001D551, 0x0001D551}, {0x0001D6A6, 0x0001D6A7}, {0x0001D7CC, 0x0001D7CD}, -{0x0001DA8C, 0x0001DA9A}, {0x0001DAA0, 0x0001DAA0}, {0x0001DAB0, 0x0001DFFF}, {0x0001E007, 0x0001E007}, -{0x0001E019, 0x0001E01A}, {0x0001E022, 0x0001E022}, {0x0001E025, 0x0001E025}, {0x0001E02B, 0x0001E0FF}, -{0x0001E12D, 0x0001E12F}, {0x0001E13E, 0x0001E13F}, {0x0001E14A, 0x0001E14D}, {0x0001E150, 0x0001E2BF}, -{0x0001E2FA, 0x0001E2FE}, {0x0001E300, 0x0001E7FF}, {0x0001E8C5, 0x0001E8C6}, {0x0001E8D7, 0x0001E8FF}, -{0x0001E94C, 0x0001E94F}, {0x0001E95A, 0x0001E95D}, {0x0001E960, 0x0001EC70}, {0x0001ECB5, 0x0001ED00}, -{0x0001ED3E, 0x0001EDFF}, {0x0001EE04, 0x0001EE04}, {0x0001EE20, 0x0001EE20}, {0x0001EE23, 0x0001EE23}, -{0x0001EE25, 0x0001EE26}, {0x0001EE28, 0x0001EE28}, {0x0001EE33, 0x0001EE33}, {0x0001EE38, 0x0001EE38}, -{0x0001EE3A, 0x0001EE3A}, {0x0001EE3C, 0x0001EE41}, {0x0001EE43, 0x0001EE46}, {0x0001EE48, 0x0001EE48}, -{0x0001EE4A, 0x0001EE4A}, {0x0001EE4C, 0x0001EE4C}, {0x0001EE50, 0x0001EE50}, {0x0001EE53, 0x0001EE53}, -{0x0001EE55, 0x0001EE56}, {0x0001EE58, 0x0001EE58}, {0x0001EE5A, 0x0001EE5A}, {0x0001EE5C, 0x0001EE5C}, -{0x0001EE5E, 0x0001EE5E}, {0x0001EE60, 0x0001EE60}, {0x0001EE63, 0x0001EE63}, {0x0001EE65, 0x0001EE66}, -{0x0001EE6B, 0x0001EE6B}, {0x0001EE73, 0x0001EE73}, {0x0001EE78, 0x0001EE78}, {0x0001EE7D, 0x0001EE7D}, -{0x0001EE7F, 0x0001EE7F}, {0x0001EE8A, 0x0001EE8A}, {0x0001EE9C, 0x0001EEA0}, {0x0001EEA4, 0x0001EEA4}, -{0x0001EEAA, 0x0001EEAA}, {0x0001EEBC, 0x0001EEEF}, {0x0001EEF2, 0x0001EFFF}, {0x0001F02C, 0x0001F02F}, -{0x0001F094, 0x0001F09F}, {0x0001F0AF, 0x0001F0B0}, {0x0001F0C0, 0x0001F0C0}, {0x0001F0D0, 0x0001F0D0}, -{0x0001F0F6, 0x0001F0FF}, {0x0001F1AE, 0x0001F1E5}, {0x0001F203, 0x0001F20F}, {0x0001F23C, 0x0001F23F}, -{0x0001F249, 0x0001F24F}, {0x0001F252, 0x0001F25F}, {0x0001F266, 0x0001F2FF}, {0x0001F6D8, 0x0001F6DF}, -{0x0001F6ED, 0x0001F6EF}, {0x0001F6FD, 0x0001F6FF}, {0x0001F774, 0x0001F77F}, {0x0001F7D9, 0x0001F7DF}, -{0x0001F7EC, 0x0001F7FF}, {0x0001F80C, 0x0001F80F}, {0x0001F848, 0x0001F84F}, {0x0001F85A, 0x0001F85F}, -{0x0001F888, 0x0001F88F}, {0x0001F8AE, 0x0001F8AF}, {0x0001F8B2, 0x0001F8FF}, {0x0001F979, 0x0001F979}, -{0x0001F9CC, 0x0001F9CC}, {0x0001FA54, 0x0001FA5F}, {0x0001FA6E, 0x0001FA6F}, {0x0001FA75, 0x0001FA77}, -{0x0001FA7B, 0x0001FA7F}, {0x0001FA87, 0x0001FA8F}, {0x0001FAA9, 0x0001FAAF}, {0x0001FAB7, 0x0001FABF}, -{0x0001FAC3, 0x0001FACF}, {0x0001FAD7, 0x0001FAFF}, {0x0001FB93, 0x0001FB93}, {0x0001FBCB, 0x0001FBEF}, -{0x0001FBFA, 0x0001FFFF}, {0x0002A6DE, 0x0002A6FF}, {0x0002B735, 0x0002B73F}, {0x0002B81E, 0x0002B81F}, -{0x0002CEA2, 0x0002CEAF}, {0x0002EBE1, 0x0002F7FF}, {0x0002FA1E, 0x0002FFFF}, {0x0003134B, 0x000E00FF}, -{0x000E01F0, 0x0010FFFF}, +{0x00011D99, 0x00011D9F}, {0x00011DAA, 0x00011EDF}, {0x00011EF9, 0x00011EFF}, {0x00011F11, 0x00011F11}, +{0x00011F3B, 0x00011F3D}, {0x00011F5A, 0x00011FAF}, {0x00011FB1, 0x00011FBF}, {0x00011FF2, 0x00011FFE}, +{0x0001239A, 0x000123FF}, {0x0001246F, 0x0001246F}, {0x00012475, 0x0001247F}, {0x00012544, 0x00012F8F}, +{0x00012FF3, 0x00012FFF}, {0x00013430, 0x0001343F}, {0x00013456, 0x000143FF}, {0x00014647, 0x000167FF}, +{0x00016A39, 0x00016A3F}, {0x00016A5F, 0x00016A5F}, {0x00016A6A, 0x00016A6D}, {0x00016ABF, 0x00016ABF}, +{0x00016ACA, 0x00016ACF}, {0x00016AEE, 0x00016AEF}, {0x00016AF6, 0x00016AFF}, {0x00016B46, 0x00016B4F}, +{0x00016B5A, 0x00016B5A}, {0x00016B62, 0x00016B62}, {0x00016B78, 0x00016B7C}, {0x00016B90, 0x00016E3F}, +{0x00016E9B, 0x00016EFF}, {0x00016F4B, 0x00016F4E}, {0x00016F88, 0x00016F8E}, {0x00016FA0, 0x00016FDF}, +{0x00016FE5, 0x00016FEF}, {0x00016FF2, 0x00016FFF}, {0x000187F8, 0x000187FF}, {0x00018CD6, 0x00018CFF}, +{0x00018D09, 0x0001AFEF}, {0x0001AFF4, 0x0001AFF4}, {0x0001AFFC, 0x0001AFFC}, {0x0001AFFF, 0x0001AFFF}, +{0x0001B123, 0x0001B131}, {0x0001B133, 0x0001B14F}, {0x0001B153, 0x0001B154}, {0x0001B156, 0x0001B163}, +{0x0001B168, 0x0001B16F}, {0x0001B2FC, 0x0001BBFF}, {0x0001BC6B, 0x0001BC6F}, {0x0001BC7D, 0x0001BC7F}, +{0x0001BC89, 0x0001BC8F}, {0x0001BC9A, 0x0001BC9B}, {0x0001BCA0, 0x0001CEFF}, {0x0001CF2E, 0x0001CF2F}, +{0x0001CF47, 0x0001CF4F}, {0x0001CFC4, 0x0001CFFF}, {0x0001D0F6, 0x0001D0FF}, {0x0001D127, 0x0001D128}, +{0x0001D173, 0x0001D17A}, {0x0001D1EB, 0x0001D1FF}, {0x0001D246, 0x0001D2BF}, {0x0001D2D4, 0x0001D2DF}, +{0x0001D2F4, 0x0001D2FF}, {0x0001D357, 0x0001D35F}, {0x0001D379, 0x0001D3FF}, {0x0001D455, 0x0001D455}, +{0x0001D49D, 0x0001D49D}, {0x0001D4A0, 0x0001D4A1}, {0x0001D4A3, 0x0001D4A4}, {0x0001D4A7, 0x0001D4A8}, +{0x0001D4AD, 0x0001D4AD}, {0x0001D4BA, 0x0001D4BA}, {0x0001D4BC, 0x0001D4BC}, {0x0001D4C4, 0x0001D4C4}, +{0x0001D506, 0x0001D506}, {0x0001D50B, 0x0001D50C}, {0x0001D515, 0x0001D515}, {0x0001D51D, 0x0001D51D}, +{0x0001D53A, 0x0001D53A}, {0x0001D53F, 0x0001D53F}, {0x0001D545, 0x0001D545}, {0x0001D547, 0x0001D549}, +{0x0001D551, 0x0001D551}, {0x0001D6A6, 0x0001D6A7}, {0x0001D7CC, 0x0001D7CD}, {0x0001DA8C, 0x0001DA9A}, +{0x0001DAA0, 0x0001DAA0}, {0x0001DAB0, 0x0001DEFF}, {0x0001DF1F, 0x0001DF24}, {0x0001DF2B, 0x0001DFFF}, +{0x0001E007, 0x0001E007}, {0x0001E019, 0x0001E01A}, {0x0001E022, 0x0001E022}, {0x0001E025, 0x0001E025}, +{0x0001E02B, 0x0001E02F}, {0x0001E06E, 0x0001E08E}, {0x0001E090, 0x0001E0FF}, {0x0001E12D, 0x0001E12F}, +{0x0001E13E, 0x0001E13F}, {0x0001E14A, 0x0001E14D}, {0x0001E150, 0x0001E28F}, {0x0001E2AF, 0x0001E2BF}, +{0x0001E2FA, 0x0001E2FE}, {0x0001E300, 0x0001E4CF}, {0x0001E4FA, 0x0001E7DF}, {0x0001E7E7, 0x0001E7E7}, +{0x0001E7EC, 0x0001E7EC}, {0x0001E7EF, 0x0001E7EF}, {0x0001E7FF, 0x0001E7FF}, {0x0001E8C5, 0x0001E8C6}, +{0x0001E8D7, 0x0001E8FF}, {0x0001E94C, 0x0001E94F}, {0x0001E95A, 0x0001E95D}, {0x0001E960, 0x0001EC70}, +{0x0001ECB5, 0x0001ED00}, {0x0001ED3E, 0x0001EDFF}, {0x0001EE04, 0x0001EE04}, {0x0001EE20, 0x0001EE20}, +{0x0001EE23, 0x0001EE23}, {0x0001EE25, 0x0001EE26}, {0x0001EE28, 0x0001EE28}, {0x0001EE33, 0x0001EE33}, +{0x0001EE38, 0x0001EE38}, {0x0001EE3A, 0x0001EE3A}, {0x0001EE3C, 0x0001EE41}, {0x0001EE43, 0x0001EE46}, +{0x0001EE48, 0x0001EE48}, {0x0001EE4A, 0x0001EE4A}, {0x0001EE4C, 0x0001EE4C}, {0x0001EE50, 0x0001EE50}, +{0x0001EE53, 0x0001EE53}, {0x0001EE55, 0x0001EE56}, {0x0001EE58, 0x0001EE58}, {0x0001EE5A, 0x0001EE5A}, +{0x0001EE5C, 0x0001EE5C}, {0x0001EE5E, 0x0001EE5E}, {0x0001EE60, 0x0001EE60}, {0x0001EE63, 0x0001EE63}, +{0x0001EE65, 0x0001EE66}, {0x0001EE6B, 0x0001EE6B}, {0x0001EE73, 0x0001EE73}, {0x0001EE78, 0x0001EE78}, +{0x0001EE7D, 0x0001EE7D}, {0x0001EE7F, 0x0001EE7F}, {0x0001EE8A, 0x0001EE8A}, {0x0001EE9C, 0x0001EEA0}, +{0x0001EEA4, 0x0001EEA4}, {0x0001EEAA, 0x0001EEAA}, {0x0001EEBC, 0x0001EEEF}, {0x0001EEF2, 0x0001EFFF}, +{0x0001F02C, 0x0001F02F}, {0x0001F094, 0x0001F09F}, {0x0001F0AF, 0x0001F0B0}, {0x0001F0C0, 0x0001F0C0}, +{0x0001F0D0, 0x0001F0D0}, {0x0001F0F6, 0x0001F0FF}, {0x0001F1AE, 0x0001F1E5}, {0x0001F203, 0x0001F20F}, +{0x0001F23C, 0x0001F23F}, {0x0001F249, 0x0001F24F}, {0x0001F252, 0x0001F25F}, {0x0001F266, 0x0001F2FF}, +{0x0001F6D8, 0x0001F6DB}, {0x0001F6ED, 0x0001F6EF}, {0x0001F6FD, 0x0001F6FF}, {0x0001F777, 0x0001F77A}, +{0x0001F7DA, 0x0001F7DF}, {0x0001F7EC, 0x0001F7EF}, {0x0001F7F1, 0x0001F7FF}, {0x0001F80C, 0x0001F80F}, +{0x0001F848, 0x0001F84F}, {0x0001F85A, 0x0001F85F}, {0x0001F888, 0x0001F88F}, {0x0001F8AE, 0x0001F8AF}, +{0x0001F8B2, 0x0001F8FF}, {0x0001FA54, 0x0001FA5F}, {0x0001FA6E, 0x0001FA6F}, {0x0001FA7D, 0x0001FA7F}, +{0x0001FA89, 0x0001FA8F}, {0x0001FABE, 0x0001FABE}, {0x0001FAC6, 0x0001FACD}, {0x0001FADC, 0x0001FADF}, +{0x0001FAE9, 0x0001FAEF}, {0x0001FAF9, 0x0001FAFF}, {0x0001FB93, 0x0001FB93}, {0x0001FBCB, 0x0001FBEF}, +{0x0001FBFA, 0x0001FFFF}, {0x0002A6E0, 0x0002A6FF}, {0x0002B73A, 0x0002B73F}, {0x0002B81E, 0x0002B81F}, +{0x0002CEA2, 0x0002CEAF}, {0x0002EBE1, 0x0002F7FF}, {0x0002FA1E, 0x0002FFFF}, {0x0003134B, 0x0003134F}, +{0x000323B0, 0x000E00FF}, {0x000E01F0, 0x0010FFFF}, }; const std::multimap unicode_map_nfd = { diff --git a/unicode-data.h b/unicode-data.h index cb9dd8aa5..3cf84117c 100644 --- a/unicode-data.h +++ b/unicode-data.h @@ -5,7 +5,7 @@ #include #include -extern const std::vector> unicode_ranges_digit; +extern const std::vector> unicode_ranges_number; extern const std::vector> unicode_ranges_letter; extern const std::vector> unicode_ranges_whitespace; extern const std::vector> unicode_ranges_accent_mark; diff --git a/unicode.cpp b/unicode.cpp index 50a2018ea..d43f2a014 100644 --- a/unicode.cpp +++ b/unicode.cpp @@ -110,9 +110,9 @@ static uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) static std::unordered_map unicode_cpt_type_map() { std::unordered_map cpt_types; - for (auto p : unicode_ranges_digit) { + for (auto p : unicode_ranges_number) { for (auto i = p.first; i <= p.second; ++ i) { - cpt_types[i] = CODEPOINT_TYPE_DIGIT; + cpt_types[i] = CODEPOINT_TYPE_NUMBER; } } for (auto p : unicode_ranges_letter) { @@ -291,18 +291,18 @@ static std::vector unicode_regex_split_custom_gpt2(const std::string & t continue; } // regex: ?\p{N}+ - if (cpt2_type == CODEPOINT_TYPE_DIGIT) { + if (cpt2_type == CODEPOINT_TYPE_NUMBER) { pos += (cpt == ' '); - while (cpt2_type == CODEPOINT_TYPE_DIGIT) { + while (cpt2_type == CODEPOINT_TYPE_NUMBER) { cpt2_type = _get_cpt_type(++pos); } _add_token(pos); continue; } // regex: ?[^\s\p{L}\p{N}]+ - if (cpt2_type != CODEPOINT_TYPE_WHITESPACE && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_DIGIT && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) { + if (cpt2_type != CODEPOINT_TYPE_WHITESPACE && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) { pos += (cpt == ' '); - while (cpt2_type != CODEPOINT_TYPE_WHITESPACE && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_DIGIT && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) { + while (cpt2_type != CODEPOINT_TYPE_WHITESPACE && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) { cpt2_type = _get_cpt_type(++pos); } _add_token(pos); @@ -398,7 +398,7 @@ static std::vector unicode_regex_split_custom_llama3(const std::string & } // regex: [^\r\n\p{L}\p{N}]?\p{L}+ //####FIXME: the first \p{L} is correct? - if (cpt != '\r' && cpt != '\n' && /*cpt_type != CODEPOINT_TYPE_LETTER &&*/ cpt_type != CODEPOINT_TYPE_DIGIT) { + if (cpt != '\r' && cpt != '\n' && /*cpt_type != CODEPOINT_TYPE_LETTER &&*/ cpt_type != CODEPOINT_TYPE_NUMBER) { if (cpt_type == CODEPOINT_TYPE_LETTER || _get_cpt_type(pos+1) == CODEPOINT_TYPE_LETTER) { // one or more letters pos++; while (_get_cpt_type(pos) == CODEPOINT_TYPE_LETTER) { @@ -410,9 +410,9 @@ static std::vector unicode_regex_split_custom_llama3(const std::string & } // regex: \p{N}{1,3} - if (cpt_type == CODEPOINT_TYPE_DIGIT) { + if (cpt_type == CODEPOINT_TYPE_NUMBER) { size_t ini = pos; - while (_get_cpt_type(pos) == CODEPOINT_TYPE_DIGIT) { + while (_get_cpt_type(pos) == CODEPOINT_TYPE_NUMBER) { if (++pos - ini >= 3 ) { _add_token(pos); ini = pos; @@ -424,9 +424,9 @@ static std::vector unicode_regex_split_custom_llama3(const std::string & // regex: ?[^\s\p{L}\p{N}]+[\r\n]* int cpt2_type = (cpt == ' ' ? _get_cpt_type(pos+1) : cpt_type); - if (cpt2_type != CODEPOINT_TYPE_WHITESPACE && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_DIGIT && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) { + if (cpt2_type != CODEPOINT_TYPE_WHITESPACE && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) { pos += (cpt == ' '); - while (cpt2_type != CODEPOINT_TYPE_WHITESPACE && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_DIGIT && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) { + while (cpt2_type != CODEPOINT_TYPE_WHITESPACE && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) { cpt2_type = _get_cpt_type(++pos); } char32_t cpt2 = _get_cpt(pos); @@ -639,19 +639,19 @@ char32_t unicode_tolower(char32_t cp) { std::vector unicode_regex_split(const std::string & text, const std::vector & regex_exprs) { // unicode categories static const std::map k_ucat_enum = { - { "\\p{N}", CODEPOINT_TYPE_DIGIT }, + { "\\p{N}", CODEPOINT_TYPE_NUMBER }, { "\\p{L}", CODEPOINT_TYPE_LETTER }, { "\\p{P}", CODEPOINT_TYPE_PUNCTUATION }, }; static const std::map k_ucat_cpt = { - { CODEPOINT_TYPE_DIGIT, 0xD1 }, + { CODEPOINT_TYPE_NUMBER, 0xD1 }, { CODEPOINT_TYPE_LETTER, 0xD2 }, { CODEPOINT_TYPE_PUNCTUATION, 0xD3 }, }; static const std::map k_ucat_map = { - { CODEPOINT_TYPE_DIGIT, "\x30-\x39" }, // 0-9 + { CODEPOINT_TYPE_NUMBER, "\x30-\x39" }, // 0-9 { CODEPOINT_TYPE_LETTER, "\x41-\x5A\x61-\x7A" }, // A-Za-z { CODEPOINT_TYPE_PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\} }; diff --git a/unicode.h b/unicode.h index ce2bcef5a..e9026dc81 100644 --- a/unicode.h +++ b/unicode.h @@ -5,7 +5,7 @@ #include #define CODEPOINT_TYPE_UNIDENTIFIED 0 -#define CODEPOINT_TYPE_DIGIT 1 +#define CODEPOINT_TYPE_NUMBER 1 #define CODEPOINT_TYPE_LETTER 2 #define CODEPOINT_TYPE_WHITESPACE 3 #define CODEPOINT_TYPE_ACCENT_MARK 4