diff --git a/.devops/full.Dockerfile b/.devops/full.Dockerfile index a75bc976f..491d67676 100644 --- a/.devops/full.Dockerfile +++ b/.devops/full.Dockerfile @@ -5,9 +5,10 @@ FROM ubuntu:$UBUNTU_VERSION as build RUN apt-get update && \ apt-get install -y build-essential python3 python3-pip +COPY requirements.txt requirements.txt + RUN pip install --upgrade pip setuptools wheel \ - && pip install numpy requests sentencepiece tqdm \ - && pip install torch --index-url https://download.pytorch.org/whl/cpu + && pip install -r requirements.txt WORKDIR /app diff --git a/.devops/main.Dockerfile b/.devops/main.Dockerfile index cd575efa0..2e629f8ce 100644 --- a/.devops/main.Dockerfile +++ b/.devops/main.Dockerfile @@ -15,4 +15,4 @@ FROM ubuntu:$UBUNTU_VERSION as runtime COPY --from=build /app/main /main -ENTRYPOINT [ "/main" ] \ No newline at end of file +ENTRYPOINT [ "/main" ] diff --git a/.dockerignore b/.dockerignore index 952990f26..462fac23a 100644 --- a/.dockerignore +++ b/.dockerignore @@ -21,4 +21,4 @@ models/* arm_neon.h compile_commands.json -Dockerfile \ No newline at end of file +Dockerfile diff --git a/.ecrc b/.ecrc new file mode 100644 index 000000000..b682057dd --- /dev/null +++ b/.ecrc @@ -0,0 +1,5 @@ +{ + "Disable": { + "IndentSize": true + } +} diff --git a/.editorconfig b/.editorconfig new file mode 100644 index 000000000..135a7e4bc --- /dev/null +++ b/.editorconfig @@ -0,0 +1,19 @@ +# https://EditorConfig.org + +# Top-most EditorConfig file +root = true + +# Unix-style newlines with a newline ending every file, utf-8 charset +[*] +end_of_line = lf +insert_final_newline = true +trim_trailing_whitespace = true +charset = utf-8 +indent_style = space +indent_size = 4 + +[Makefile] +indent_style = tab + +[prompts/*.txt] +insert_final_newline = unset diff --git a/.github/ISSUE_TEMPLATE/custom.md b/.github/ISSUE_TEMPLATE/custom.md index 0d508802d..8fd955356 100644 --- a/.github/ISSUE_TEMPLATE/custom.md +++ b/.github/ISSUE_TEMPLATE/custom.md @@ -22,9 +22,9 @@ Please provide a detailed written description of what you were trying to do, and # Current Behavior -Please provide a detailed written description of what `llama.cpp` did, instead. +Please provide a detailed written description of what `llama.cpp` did, instead. -# Environment and Context +# Environment and Context Please provide detailed information about your computer setup. This is important in case the issue is not reproducible except for under certain specific conditions. @@ -133,7 +133,7 @@ llama_model_load: loading model part 8/8 from './models/65B/ggml-model-q4_0.bin. llama_model_load: .......................................................................................... done llama_model_load: model size = 4869.09 MB / num tensors = 723 -system_info: n_threads = 16 / 32 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 | +system_info: n_threads = 16 / 32 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 | main: prompt: 'Please close your issue when it has been answered.' main: number of tokens in prompt = 11 @@ -166,14 +166,14 @@ main: total time = 246406.42 ms Performance counter stats for './main -m ./models/65B/ggml-model-q4_0.bin -t 16 -n 1024 -p Please close your issue when it has been answered.': - 3636882.89 msec task-clock # 14.677 CPUs utilized - 13509 context-switches # 3.714 /sec - 2436 cpu-migrations # 0.670 /sec - 10476679 page-faults # 2.881 K/sec + 3636882.89 msec task-clock # 14.677 CPUs utilized + 13509 context-switches # 3.714 /sec + 2436 cpu-migrations # 0.670 /sec + 10476679 page-faults # 2.881 K/sec 13133115082869 cycles # 3.611 GHz (16.77%) 29314462753 stalled-cycles-frontend # 0.22% frontend cycles idle (16.76%) 10294402631459 stalled-cycles-backend # 78.39% backend cycles idle (16.74%) - 23479217109614 instructions # 1.79 insn per cycle + 23479217109614 instructions # 1.79 insn per cycle # 0.44 stalled cycles per insn (16.76%) 2353072268027 branches # 647.002 M/sec (16.77%) 1998682780 branch-misses # 0.08% of all branches (16.76%) diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 88e70e495..7e8a29b1e 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -8,6 +8,8 @@ on: required: true type: boolean push: + branches: + - master paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp'] pull_request: types: [opened, synchronize, edited, reopened, review_requested, ready_for_review] @@ -18,6 +20,8 @@ env: jobs: ubuntu-latest-make: + if: github.event.pull_request.draft == false + runs-on: ubuntu-latest steps: @@ -37,6 +41,8 @@ jobs: make ubuntu-latest-cmake: + if: github.event.pull_request.draft == false + runs-on: ubuntu-latest steps: @@ -65,6 +71,8 @@ jobs: ctest --verbose ubuntu-latest-cmake-sanitizer: + if: github.event.pull_request.draft == false + runs-on: ubuntu-latest continue-on-error: true @@ -73,7 +81,6 @@ jobs: matrix: sanitizer: [ADDRESS, THREAD, UNDEFINED] build_type: [Debug, Release] - accelerate: [ON, OFF] steps: - name: Clone @@ -91,7 +98,7 @@ jobs: run: | mkdir build cd build - cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DLLAMA_ACCELERATE=${{ matrix.accelerate }} + cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} cmake --build . --config ${{ matrix.build_type }} - name: Test @@ -101,6 +108,8 @@ jobs: ctest --verbose macOS-latest-make: + if: github.event.pull_request.draft == false + runs-on: macos-latest steps: @@ -119,6 +128,8 @@ jobs: make macOS-latest-cmake: + if: github.event.pull_request.draft == false + runs-on: macOS-latest steps: @@ -146,6 +157,8 @@ jobs: ctest --verbose windows-latest-cmake: + if: github.event.pull_request.draft == false + runs-on: windows-latest strategy: diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index f70821de2..379fbd7ad 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -18,6 +18,8 @@ on: jobs: push_to_registry: name: Push Docker image to Docker Hub + if: github.event.pull_request.draft == false + runs-on: ubuntu-latest env: COMMIT_SHA: ${{ github.sha }} @@ -60,4 +62,4 @@ jobs: push: ${{ github.event_name == 'push' }} platforms: linux/amd64,linux/arm64 tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}" - file: ${{ matrix.config.dockerfile }} \ No newline at end of file + file: ${{ matrix.config.dockerfile }} diff --git a/.github/workflows/editorconfig.yml b/.github/workflows/editorconfig.yml new file mode 100644 index 000000000..b4e535acf --- /dev/null +++ b/.github/workflows/editorconfig.yml @@ -0,0 +1,17 @@ +name: EditorConfig Checker + +on: + push: + branches: + - master + pull_request: + branches: + - master + +jobs: + editorconfig: + runs-on: ubuntu-latest + steps: + - uses: actions/checkout@v3 + - uses: editorconfig-checker/action-editorconfig-checker@main + - run: editorconfig-checker diff --git a/.gitignore b/.gitignore index 1c75d38d1..e52d479ee 100644 --- a/.gitignore +++ b/.gitignore @@ -1,11 +1,15 @@ *.o *.a +.DS_Store +.build/ .cache/ +.direnv/ +.envrc +.swiftpm +.venv .vs/ .vscode/ -.DS_Store -.build/ build/ build-em/ build-debug/ @@ -19,17 +23,20 @@ models/* /main /quantize +/quantize-stats /result /perplexity /embedding +/benchmark-q4_0-matmult +/vdot /Pipfile arm_neon.h compile_commands.json -.envrc -.direnv/ - -.venv __pycache__ -.swiftpm + +zig-out/ +zig-cache/ + +ppl-*.txt diff --git a/CMakeLists.txt b/CMakeLists.txt index 1a434f07b..1f9fdd30f 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -55,11 +55,18 @@ option(LLAMA_SANITIZE_UNDEFINED "llama: enable undefined sanitizer" option(LLAMA_AVX "llama: enable AVX" ON) option(LLAMA_AVX2 "llama: enable AVX2" ON) option(LLAMA_AVX512 "llama: enable AVX512" OFF) +option(LLAMA_AVX512_VBMI "llama: enable AVX512-VBMI" OFF) +option(LLAMA_AVX512_VNNI "llama: enable AVX512-VNNI" OFF) option(LLAMA_FMA "llama: enable FMA" ON) +# in MSVC F16C is implied with AVX2/AVX512 +if (NOT MSVC) + option(LLAMA_F16C "llama: enable F16C" ON) +endif() # 3rd party libs option(LLAMA_ACCELERATE "llama: enable Accelerate framework" ON) option(LLAMA_OPENBLAS "llama: use OpenBLAS" OFF) +option(LLAMA_CUBLAS "llama: use cuBLAS" OFF) option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE}) option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE}) @@ -103,6 +110,7 @@ if (APPLE AND LLAMA_ACCELERATE) message(WARNING "Accelerate framework not found") endif() endif() + if (LLAMA_OPENBLAS) if (LLAMA_STATIC) set(BLA_STATIC ON) @@ -115,11 +123,51 @@ if (LLAMA_OPENBLAS) add_compile_definitions(GGML_USE_OPENBLAS) add_link_options(${BLAS_LIBRARIES}) + set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} openblas) + + # find header file + set(OPENBLAS_INCLUDE_SEARCH_PATHS + /usr/include + /usr/include/openblas + /usr/include/openblas-base + /usr/local/include + /usr/local/include/openblas + /usr/local/include/openblas-base + /opt/OpenBLAS/include + $ENV{OpenBLAS_HOME} + $ENV{OpenBLAS_HOME}/include + ) + find_path(OPENBLAS_INC NAMES cblas.h PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS}) + add_compile_options(-I${OPENBLAS_INC}) else() message(WARNING "OpenBLAS not found") endif() endif() +if (LLAMA_CUBLAS) + cmake_minimum_required(VERSION 3.17) + + find_package(CUDAToolkit) + if (CUDAToolkit_FOUND) + message(STATUS "cuBLAS found") + + enable_language(CUDA) + + set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h) + + add_compile_definitions(GGML_USE_CUBLAS) + + if (LLAMA_STATIC) + set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static) + else() + set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt) + endif() + + else() + message(WARNING "cuBLAS not found") + endif() +endif() + if (LLAMA_ALL_WARNINGS) if (NOT MSVC) set(c_flags @@ -131,7 +179,6 @@ if (LLAMA_ALL_WARNINGS) -Wshadow -Wstrict-prototypes -Wpointer-arith - -Wno-unused-function ) set(cxx_flags -Wall @@ -139,6 +186,7 @@ if (LLAMA_ALL_WARNINGS) -Wpedantic -Wcast-qual -Wno-unused-function + -Wno-multichar ) else() # todo : msvc @@ -151,6 +199,10 @@ if (LLAMA_ALL_WARNINGS) endif() +if (MSVC) + add_compile_definitions(_CRT_SECURE_NO_WARNINGS) +endif() + if (LLAMA_LTO) include(CheckIPOSupported) check_ipo_supported(RESULT result OUTPUT output) @@ -194,14 +246,31 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$") message(STATUS "x86 detected") if (MSVC) if (LLAMA_AVX512) - add_compile_options(/arch:AVX512) + add_compile_options($<$:/arch:AVX512>) + add_compile_options($<$:/arch:AVX512>) + # MSVC has no compile-time flags enabling specific + # AVX512 extensions, neither it defines the + # macros corresponding to the extensions. + # Do it manually. + if (LLAMA_AVX512_VBMI) + add_compile_definitions($<$:__AVX512VBMI__>) + add_compile_definitions($<$:__AVX512VBMI__>) + endif() + if (LLAMA_AVX512_VNNI) + add_compile_definitions($<$:__AVX512VNNI__>) + add_compile_definitions($<$:__AVX512VNNI__>) + endif() elseif (LLAMA_AVX2) - add_compile_options(/arch:AVX2) + add_compile_options($<$:/arch:AVX2>) + add_compile_options($<$:/arch:AVX2>) elseif (LLAMA_AVX) - add_compile_options(/arch:AVX) + add_compile_options($<$:/arch:AVX>) + add_compile_options($<$:/arch:AVX>) endif() else() - add_compile_options(-mf16c) + if (LLAMA_F16C) + add_compile_options(-mf16c) + endif() if (LLAMA_FMA) add_compile_options(-mfma) endif() @@ -213,9 +282,13 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$") endif() if (LLAMA_AVX512) add_compile_options(-mavx512f) - # add_compile_options(-mavx512cd) - # add_compile_options(-mavx512dq) - # add_compile_options(-mavx512bw) + add_compile_options(-mavx512bw) + endif() + if (LLAMA_AVX512_VBMI) + add_compile_options(-mavx512vbmi) + endif() + if (LLAMA_AVX512_VNNI) + add_compile_options(-mavx512vnni) endif() endif() else() @@ -229,7 +302,8 @@ endif() add_library(ggml OBJECT ggml.c - ggml.h) + ggml.h + ${GGML_CUDA_SOURCES}) target_include_directories(ggml PUBLIC .) target_compile_features(ggml PUBLIC c_std_11) # don't bump @@ -240,7 +314,8 @@ endif() add_library(llama llama.cpp - llama.h) + llama.h + llama_util.h) target_include_directories(llama PUBLIC .) target_compile_features(llama PUBLIC cxx_std_11) # don't bump @@ -250,6 +325,14 @@ if (BUILD_SHARED_LIBS) target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD) endif() +if (GGML_CUDA_SOURCES) + message(STATUS "GGML CUDA sources found, configuring CUDA architecture") + set_property(TARGET ggml PROPERTY CUDA_ARCHITECTURES OFF) + set_property(TARGET ggml PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto") + set_property(TARGET llama PROPERTY CUDA_ARCHITECTURES OFF) +endif() + + # # programs, examples and tests # @@ -261,4 +344,5 @@ endif () if (LLAMA_BUILD_EXAMPLES) add_subdirectory(examples) + add_subdirectory(pocs) endif() diff --git a/Makefile b/Makefile index 2f828bf10..f267d0864 100644 --- a/Makefile +++ b/Makefile @@ -1,3 +1,6 @@ +# Define the default target now so that it is always the first target +default: main quantize quantize-stats perplexity embedding vdot + ifndef UNAME_S UNAME_S := $(shell uname -s) endif @@ -36,8 +39,8 @@ CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC LDFLAGS = # warnings -CFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith -Wno-unused-function -CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function +CFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith +CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar # OS specific # TODO: support Windows @@ -72,6 +75,7 @@ endif ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686)) # Use all CPU extensions that are available: CFLAGS += -march=native -mtune=native + CXXFLAGS += -march=native -mtune=native endif ifneq ($(filter ppc64%,$(UNAME_M)),) POWER9_M := $(shell grep "POWER9" /proc/cpuinfo) @@ -96,6 +100,13 @@ ifdef LLAMA_OPENBLAS CFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas LDFLAGS += -lopenblas endif +ifdef LLAMA_CUBLAS + CFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include + LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 + OBJS += ggml-cuda.o +ggml-cuda.o: ggml-cuda.cu ggml-cuda.h + nvcc -arch=native -c -o $@ $< +endif ifdef LLAMA_GPROF CFLAGS += -pg CXXFLAGS += -pg @@ -132,43 +143,54 @@ $(info I CC: $(CCV)) $(info I CXX: $(CXXV)) $(info ) -default: main quantize perplexity embedding - # # Build library # ggml.o: ggml.c ggml.h - $(CC) $(CFLAGS) -c ggml.c -o ggml.o + $(CC) $(CFLAGS) -c $< -o $@ -llama.o: llama.cpp llama.h - $(CXX) $(CXXFLAGS) -c llama.cpp -o llama.o +llama.o: llama.cpp ggml.h llama.h llama_util.h + $(CXX) $(CXXFLAGS) -c $< -o $@ common.o: examples/common.cpp examples/common.h - $(CXX) $(CXXFLAGS) -c examples/common.cpp -o common.o + $(CXX) $(CXXFLAGS) -c $< -o $@ clean: - rm -vf *.o main quantize perplexity embedding + rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-q4_0-matmult -main: examples/main/main.cpp ggml.o llama.o common.o - $(CXX) $(CXXFLAGS) examples/main/main.cpp ggml.o llama.o common.o -o main $(LDFLAGS) +main: examples/main/main.cpp ggml.o llama.o common.o $(OBJS) + $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) @echo @echo '==== Run ./main -h for help. ====' @echo -quantize: examples/quantize/quantize.cpp ggml.o llama.o - $(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp ggml.o llama.o -o quantize $(LDFLAGS) +quantize: examples/quantize/quantize.cpp ggml.o llama.o $(OBJS) + $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) -perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o - $(CXX) $(CXXFLAGS) examples/perplexity/perplexity.cpp ggml.o llama.o common.o -o perplexity $(LDFLAGS) +quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o $(OBJS) + $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) -embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o - $(CXX) $(CXXFLAGS) examples/embedding/embedding.cpp ggml.o llama.o common.o -o embedding $(LDFLAGS) +perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o $(OBJS) + $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) + +embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o $(OBJS) + $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) + +vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS) + $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) + +libllama.so: llama.o ggml.o $(OBJS) + $(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) # # Tests # +benchmark: examples/benchmark/benchmark-q4_0-matmult.c ggml.o $(OBJS) + $(CXX) $(CXXFLAGS) $^ -o benchmark-q4_0-matmult $(LDFLAGS) + ./benchmark-q4_0-matmult + .PHONY: tests tests: bash ./tests/run-tests.sh diff --git a/Package.swift b/Package.swift index 79d13c82d..2c2c147ba 100644 --- a/Package.swift +++ b/Package.swift @@ -13,7 +13,10 @@ let package = Package( path: ".", sources: ["ggml.c", "llama.cpp"], publicHeadersPath: "spm-headers", - cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"])] + cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"]), .define("GGML_USE_ACCELERATE")], + linkerSettings: [ + .linkedFramework("Accelerate") + ] ), ], cxxLanguageStandard: .cxx11 diff --git a/README.md b/README.md index 508d315d5..324d49f07 100644 --- a/README.md +++ b/README.md @@ -1,20 +1,25 @@ # llama.cpp -![llama](https://user-images.githubusercontent.com/1991296/227761327-6d83e30e-2200-41a6-bfbb-f575231c54f4.png) +![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png) [![Actions Status](https://github.com/ggerganov/llama.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/llama.cpp/actions) [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++ +**Warnings** + +- `Q4_2` and `Q4_3` are still in development. Do not expect any kind of backward compatibility until they are finalized + **Hot topics:** -- [Roadmap (short-term)](https://github.com/ggerganov/llama.cpp/discussions/457) -- Support for [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all) +- [Added LoRA support](https://github.com/ggerganov/llama.cpp/pull/820) +- [Add GPU support to ggml](https://github.com/ggerganov/llama.cpp/discussions/915) +- [Roadmap Apr 2023](https://github.com/ggerganov/llama.cpp/discussions/784) ## Description -The main goal is to run the model using 4-bit quantization on a MacBook +The main goal of llama.cpp is to run the llama model using 4-bit quantization on a MacBook. - Plain C/C++ implementation without dependencies - Apple silicon first-class citizen - optimized via ARM NEON and Accelerate framework @@ -28,20 +33,34 @@ Please do not make conclusions about the models based on the results from this i For all I know, it can be completely wrong. This project is for educational purposes. New features will probably be added mostly through community contributions. -Supported platforms: +**Supported platforms:** - [X] Mac OS - [X] Linux - [X] Windows (via CMake) - [X] Docker -Supported models: +**Supported models:** - [X] LLaMA 🦙 - [X] [Alpaca](https://github.com/ggerganov/llama.cpp#instruction-mode-with-alpaca) - [X] [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all) - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne) +- [X] [Vicuna](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5533894) +- [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/) + +**Bindings:** + +- Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python) +- Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp) +- Node.js: [hlhr202/llama-node](https://github.com/hlhr202/llama-node) +- Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb) + +**UI:** + +- [nat/openplayground](https://github.com/nat/openplayground) +- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) --- @@ -137,23 +156,52 @@ https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8 ## Usage -Here are the step for the LLaMA-7B model: +Here are the steps for the LLaMA-7B model. + +### Get the Code ```bash -# build this repo git clone https://github.com/ggerganov/llama.cpp cd llama.cpp -make +``` +### Build + +Note: For Windows, CMake or Zig can be used. + +1. Use `make` + + ```bash + make + ``` + +1. Use CMake + + ```bash + mkdir build + cd build + cmake .. + cmake --build . --config Release + ``` + +1. Use Zig + + ```bash + zig build -Drelease-fast + ``` + +### Prepare Data & Run + +```bash # obtain the original LLaMA model weights and place them in ./models ls ./models 65B 30B 13B 7B tokenizer_checklist.chk tokenizer.model # install Python dependencies -python3 -m pip install torch numpy sentencepiece +python3 -m pip install -r requirements.txt # convert the 7B model to ggml FP16 format -python3 convert-pth-to-ggml.py models/7B/ 1 +python3 convert.py models/7B/ # quantize the model to 4-bits (using method 2 = q4_0) ./quantize ./models/7B/ggml-model-f16.bin ./models/7B/ggml-model-q4_0.bin 2 @@ -162,14 +210,11 @@ python3 convert-pth-to-ggml.py models/7B/ 1 ./main -m ./models/7B/ggml-model-q4_0.bin -n 128 ``` -Currently, it's best to use Python 3.9 or Python 3.10, as `sentencepiece` has not yet published a wheel for Python 3.11. - When running the larger models, make sure you have enough disk space to store all the intermediate files. ### Memory/Disk Requirements -As the models are currently fully loaded into memory, you will need adequate disk space to save them -and sufficient RAM to load them. At the moment, memory and disk requirements are the same. +As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same. | model | original size | quantized size (4-bit) | |-------|---------------|------------------------| @@ -181,18 +226,18 @@ and sufficient RAM to load them. At the moment, memory and disk requirements are ### Interactive mode If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter. -In this mode, you can always interrupt generation by pressing Ctrl+C and enter one or more lines of text which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt which makes LLaMa emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`. +In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMa emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`. -Here is an example few-shot interaction, invoked with the command +Here is an example of a few-shot interaction, invoked with the command ```bash -# default arguments using 7B model +# default arguments using a 7B model ./examples/chat.sh -# advanced chat with 13B model +# advanced chat with a 13B model ./examples/chat-13B.sh -# custom arguments using 13B model +# custom arguments using a 13B model ./main -m ./models/13B/ggml-model-q4_0.bin -n 256 --repeat_penalty 1.0 --color -i -r "User:" -f prompts/chat-with-bob.txt ``` @@ -225,30 +270,30 @@ There 26 letters in the English Alphabet The majority (54%) are using public transit. This includes buses, trams and metros with over 100 lines throughout the city which make it very accessible for tourists to navigate around town as well as locals who commute by tram or metro on a daily basis > List 5 words that start with "ca". cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach. -> +> ``` ### Using [GPT4All](https://github.com/nomic-ai/gpt4all) - Obtain the `gpt4all-lora-quantized.bin` model -- It is distributed in the old `ggml` format which is now obsoleted +- It is distributed in the old `ggml` format, which is now obsoleted - You have to convert it to the new format using [./convert-gpt4all-to-ggml.py](./convert-gpt4all-to-ggml.py). You may also need to convert the model from the old format to the new format with [./migrate-ggml-2023-03-30-pr613.py](./migrate-ggml-2023-03-30-pr613.py): ```bash - python3 convert-gpt4all-to-ggml.py models/gpt4all-7B/gpt4all-lora-quantized.bin ./models/tokenizer.model + python3 convert-gpt4all-to-ggml.py models/gpt4all-7B/gpt4all-lora-quantized.bin ./models/tokenizer.model python3 migrate-ggml-2023-03-30-pr613.py models/gpt4all-7B/gpt4all-lora-quantized.bin models/gpt4all-7B/gpt4all-lora-quantized-new.bin ``` - + - You can now use the newly generated `gpt4all-lora-quantized-new.bin` model in exactly the same way as all other models - The original model is saved in the same folder with a suffix `.orig` ### Obtaining and verifying the Facebook LLaMA original model and Stanford Alpaca model data -- **Under no circumstances share IPFS, magnet links, or any other links to model downloads anywhere in this respository, including in issues, discussions or pull requests. They will be immediately deleted.** -- The LLaMA models are officially distributed by Facebook and will **never** be provided through this repository. +- **Under no circumstances should IPFS, magnet links, or any other links to model downloads be shared anywhere in this repository, including in issues, discussions, or pull requests. They will be immediately deleted.** +- The LLaMA models are officially distributed by Facebook and will **never** be provided through this repository. - Refer to [Facebook's LLaMA repository](https://github.com/facebookresearch/llama/pull/73/files) if you need to request access to the model data. -- Please verify the sha256 checksums of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files. +- Please verify the [sha256 checksums](SHA256SUMS) of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files. - The following command will verify if you have all possible latest files in your self-installed `./models` subdirectory: `sha256sum --ignore-missing -c SHA256SUMS` on Linux @@ -257,29 +302,27 @@ convert the model from the old format to the new format with [./migrate-ggml-202 `shasum -a 256 --ignore-missing -c SHA256SUMS` on macOS -- If your issue is with model generation quality then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT: - - LLaMA: - - [Introducing LLaMA: A foundational, 65-billion-parameter large language model](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/) - - [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971) - - GPT-3 - - [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165) - - GPT-3.5 / InstructGPT / ChatGPT: - - [Aligning language models to follow instructions](https://openai.com/research/instruction-following) - - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155) - -### Perplexity (Measuring model quality) +- If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT: +- LLaMA: +- [Introducing LLaMA: A foundational, 65-billion-parameter large language model](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/) +- [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971) +- GPT-3 +- [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165) +- GPT-3.5 / InstructGPT / ChatGPT: +- [Aligning language models to follow instructions](https://openai.com/research/instruction-following) +- [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155) -You can use the `perplexity` example to measure perplexity over the given prompt. For more background, -see https://huggingface.co/docs/transformers/perplexity. However, in general, lower perplexity is better for LLMs. +### Perplexity (measuring model quality) + +You can use the `perplexity` example to measure perplexity over the given prompt. For more background, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity). However, in general, lower perplexity is better for LLMs. #### Latest measurements -The latest perplexity scores for the various model sizes and quantizations are being tracked in [discussion #406](https://github.com/ggerganov/llama.cpp/discussions/406). `llama.cpp` is measuring very well -compared to the baseline implementations. Quantization has a small negative impact to quality, but, as you can see, running +The latest perplexity scores for the various model sizes and quantizations are being tracked in [discussion #406](https://github.com/ggerganov/llama.cpp/discussions/406). `llama.cpp` is measuring very well compared to the baseline implementations. Quantization has a small negative impact on quality, but, as you can see, running 13B at q4_0 beats the 7B f16 model by a significant amount. -All measurements are done against wikitext2 test dataset (https://paperswithcode.com/dataset/wikitext-2), with default options (512 length context). -Note that the changing the context length will have a significant impact on perplexity (longer context = better perplexity). +All measurements are done against the wikitext2 test dataset (https://paperswithcode.com/dataset/wikitext-2), with default options (512 length context). +Note that changing the context length will have a significant impact on perplexity (longer context = better perplexity). ``` Perplexity - model options 5.5985 - 13B, q4_0 @@ -321,7 +364,7 @@ https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b0 #### Prerequisites * Docker must be installed and running on your system. -* Create a folder to store big models & intermediate files (in ex. im using /llama/models) +* Create a folder to store big models & intermediate files (ex. /llama/models) #### Images We have two Docker images available for this project: @@ -333,20 +376,22 @@ We have two Docker images available for this project: The easiest way to download the models, convert them to ggml and optimize them is with the --all-in-one command which includes the full docker image. - ```bash -docker run -v /llama/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B -``` - -On complete, you are ready to play! +Replace `/path/to/models` below with the actual path where you downloaded the models. ```bash -docker run -v /llama/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512 +docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B ``` -or with light image: +On completion, you are ready to play! ```bash -docker run -v /llama/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512 +docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512 +``` + +or with a light image: + +```bash +docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.bin -p "Building a website can be done in 10 simple steps:" -n 512 ``` ### Contributing @@ -364,6 +409,9 @@ docker run -v /llama/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models - Always consider cross-compatibility with other operating systems and architectures - Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple - There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit -- Clean-up any trailing whitespaces, use 4 spaces indentation, brackets on same line, `void * ptr`, `int & a` +- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a` - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions +### Docs + +- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks) diff --git a/SHA256SUMS b/SHA256SUMS index 63fac21ae..1d034b371 100644 --- a/SHA256SUMS +++ b/SHA256SUMS @@ -1,12 +1,27 @@ 700df0d3013b703a806d2ae7f1bfb8e59814e3d06ae78be0c66368a50059f33d models/7B/consolidated.00.pth +666a4bb533b303bdaf89e1b6a3b6f93535d868de31d903afdc20983dc526c847 models/7B/ggml-model-f16.bin +fcb7664c2e69776920b526362a243e912f73c36b1ec892eb354bab940f5edb5a models/7B/ggml-model-q4_0.bin +cc061458339a3eb8bcecbf0a825e9924fb7d1a8150f63cd5d091caa99215aafe models/7B/ggml-model-q4_1.bin +1bc7484c24a87612726d756f1761890e7acf5f412e23378577ce50fbe789b5b8 models/7B/ggml-model-q4_2.bin +3429bf198ec771886cf81a574df45245f3ebf04f0ce0956b73ef5d0ab01ff48b models/7B/ggml-model-q4_3.bin 7e89e242ddc0dd6f060b43ca219ce8b3e8f08959a72cb3c0855df8bb04d46265 models/7B/params.json 745bf4e29a4dd6f411e72976d92b452da1b49168a4f41c951cfcc8051823cf08 models/13B/consolidated.00.pth d5ccbcc465c71c0de439a5aeffebe8344c68a519bce70bc7f9f92654ee567085 models/13B/consolidated.01.pth +2b206e9b21fb1076f11cafc624e2af97c9e48ea09312a0962153acc20d45f808 models/13B/ggml-model-f16.bin +4b69e4d6b6e3275230955997b90407fceca7e5ab3daf2e63a2c9e7270a8e1e3e models/13B/ggml-model-q4_0.bin +d9581b5b88e5622532fe897c9f9b0e67a317d22dd27a6f90fa4ab8c6d23ccdbb models/13B/ggml-model-q4_1.bin +8d55a2077317ec9a928c7851d6a43e08e51f7e9e08360f2a7a7e1deefea3134f models/13B/ggml-model-q4_2.bin +4208cdec9788ffa48dc1a17af2c36a0299f5bf3eb0e2b87889dda7fad591fca3 models/13B/ggml-model-q4_3.bin 4ab77bec4d4405ccb66a97b282574c89a94417e3c32e5f68f37e2876fc21322f models/13B/params.json e23294a58552d8cdec5b7e8abb87993b97ea6eced4178ff2697c02472539d067 models/30B/consolidated.00.pth 4e077b7136c7ae2302e954860cf64930458d3076fcde9443f4d0e939e95903ff models/30B/consolidated.01.pth 24a87f01028cbd3a12de551dcedb712346c0b5cbdeff1454e0ddf2df9b675378 models/30B/consolidated.02.pth 1adfcef71420886119544949767f6a56cb6339b4d5fcde755d80fe68b49de93b models/30B/consolidated.03.pth +7e1b524061a9f4b27c22a12d6d2a5bf13b8ebbea73e99f218809351ed9cf7d37 models/30B/ggml-model-f16.bin +7a679908ce31c9d6ae2e38d6059bcd4d0ad3a870cd58cc1c8f7b36f2b2f51c73 models/30B/ggml-model-q4_0.bin +7b75ac615fa369ee593493a7e6ef87542bf0350255db928b22c5a24f6d598bcd models/30B/ggml-model-q4_1.bin +2c82b4954a94a6a284f452f6011c1e4f0d20362c194a0b1eb5737f5fd8a20fb3 models/30B/ggml-model-q4_2.bin +a6188660199dbcb8d5658abe7d89169869e50423494385830d9e6b330ea7fc33 models/30B/ggml-model-q4_3.bin 2c07118ea98d69dbe7810d88520e30288fa994751b337f8fca02b171955f44cb models/30B/params.json 135c563f6b3938114458183afb01adc9a63bef3d8ff7cccc3977e5d3664ecafe models/65B/consolidated.00.pth 9a600b37b19d38c7e43809485f70d17d1dc12206c07efa83bc72bb498a568bde models/65B/consolidated.01.pth @@ -16,5 +31,10 @@ e7babf7c5606f165a3756f527cb0fedc4f83e67ef1290391e52fb1cce5f26770 models/65B/con a287c0dfe49081626567c7fe87f74cce5831f58e459b427b5e05567641f47b78 models/65B/consolidated.05.pth 72b4eba67a1a3b18cb67a85b70f8f1640caae9b40033ea943fb166bd80a7b36b models/65B/consolidated.06.pth d27f5b0677d7ff129ceacd73fd461c4d06910ad7787cf217b249948c3f3bc638 models/65B/consolidated.07.pth +60758f2384d74e423dffddfd020ffed9d3bb186ebc54506f9c4a787d0f5367b0 models/65B/ggml-model-f16.bin +c671fe1bce71499ac732ec999770ebe53ac486623a7891e42c9dfdb6962d2c64 models/65B/ggml-model-q4_0.bin +4743a28aac3e5f32a6e838a815f51d3779de44fbbe251d745251e66c23c5950f models/65B/ggml-model-q4_1.bin +4a145a210c56982389b1ed34387e0590c3e0d7325fa9be4f2284fe4d244a3633 models/65B/ggml-model-q4_2.bin +305e91a4608b4f627b9b8ad5b4af75187d2684254bfd76dcb9db571618ef293c models/65B/ggml-model-q4_3.bin 999ed1659b469ccc2a941714c0a9656fa571d17c9f7c8c7589817ca90edef51b models/65B/params.json 9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347 models/tokenizer.model diff --git a/build.zig b/build.zig new file mode 100644 index 000000000..306127ffe --- /dev/null +++ b/build.zig @@ -0,0 +1,61 @@ +const std = @import("std"); + +pub fn build(b: *std.build.Builder) void { + const target = b.standardTargetOptions(.{}); + const optimize = b.standardReleaseOptions(); + const want_lto = b.option(bool, "lto", "Want -fLTO"); + + const lib = b.addStaticLibrary("llama", null); + lib.want_lto = want_lto; + lib.setTarget(target); + lib.setBuildMode(optimize); + lib.linkLibCpp(); + lib.addIncludePath("."); + lib.addIncludePath("examples"); + lib.addCSourceFiles(&.{ + "ggml.c", + }, &.{"-std=c11"}); + lib.addCSourceFiles(&.{ + "llama.cpp", + }, &.{"-std=c++11"}); + lib.install(); + + const build_args = .{ .b = b, .lib = lib, .target = target, .optimize = optimize, .want_lto = want_lto }; + + const exe = build_example("main", build_args); + _ = build_example("quantize", build_args); + _ = build_example("perplexity", build_args); + _ = build_example("embedding", build_args); + + // create "zig build run" command for ./main + + const run_cmd = exe.run(); + run_cmd.step.dependOn(b.getInstallStep()); + if (b.args) |args| { + run_cmd.addArgs(args); + } + + const run_step = b.step("run", "Run the app"); + run_step.dependOn(&run_cmd.step); +} + +fn build_example(comptime name: []const u8, args: anytype) *std.build.LibExeObjStep { + const b = args.b; + const lib = args.lib; + const want_lto = args.want_lto; + + const exe = b.addExecutable(name, null); + exe.want_lto = want_lto; + lib.setTarget(args.target); + lib.setBuildMode(args.optimize); + exe.addIncludePath("."); + exe.addIncludePath("examples"); + exe.addCSourceFiles(&.{ + std.fmt.comptimePrint("examples/{s}/{s}.cpp", .{name, name}), + "examples/common.cpp", + }, &.{"-std=c++11"}); + exe.linkLibrary(lib); + exe.install(); + + return exe; +} diff --git a/convert-ggml-to-pth.py b/convert-ggml-to-pth.py deleted file mode 100644 index 25a44237a..000000000 --- a/convert-ggml-to-pth.py +++ /dev/null @@ -1,299 +0,0 @@ -# Author: github.com/ductai199x -import argparse -import os -import struct - -import numpy as np -import torch -from numba import njit -from tqdm.auto import tqdm - - -def read_header(fin): - values = struct.unpack("i" * 9, fin.read(4 * 9)) - _, _, vocab_size, dim, multiple_of, n_heads, n_layers, rot, ftype = values - return { - "vocab_size": vocab_size, - "dim": dim, - "multiple_of": multiple_of, - "n_heads": n_heads, - "n_layers": n_layers, - }, ftype - - -def read_tokens(fin, vocab_size): - tokens = [] - for _ in range(vocab_size): - text_len = struct.unpack("i", fin.read(4))[0] - text_bytes = fin.read(text_len) - try: - text = text_bytes.decode() - except UnicodeDecodeError: - text = text_bytes.decode(errors="replace") - score = struct.unpack("f", fin.read(4))[0] - tokens.append((text, score)) - return tokens - - -@njit -def dequantize_weights_numba(fin_data, n_rows, n_cols): - qk = 32 - nb = n_cols // qk - bs = 4 + (qk // 2) - - weights = np.zeros((n_rows, n_cols), dtype=np.float32) - data_pos = 0 - - for row in range(n_rows): - for block in range(nb): - d = np.frombuffer(fin_data[data_pos : data_pos + 4], dtype=np.float32)[0] - data_pos += 4 - packed_values = fin_data[data_pos : data_pos + (qk // 2)] - data_pos += qk // 2 - - for i in range(qk // 2): - packed_value = packed_values[i] - v0 = np.float32((packed_value & 0b00001111) - 8) * d - v1 = np.float32((packed_value >> 4) - 8) * d - - weights[row, block * qk + 2 * i] = v0 - weights[row, block * qk + 2 * i + 1] = v1 - - return weights - - -def dequantize_weights(fin, n_rows, n_cols): - qk = 32 - nb = n_cols // qk - data_size = n_rows * n_cols // 2 + n_rows * nb * 4 - fin_data = fin.read(data_size) - return dequantize_weights_numba(fin_data, n_rows, n_cols) - - -def read_variables(fin): - model = {} - pbar = tqdm(total=os.path.getsize(fin.name), unit="B", unit_scale=True, desc="Reading variables") - while True: - start_pos = fin.tell() - try: - n_dims, name_length, ftype_cur = struct.unpack("iii", fin.read(4 * 3)) - except struct.error: - break - - shape = tuple(struct.unpack("i" * n_dims, fin.read(4 * n_dims))) - shape = shape[::-1] - name = fin.read(name_length).decode() - - # ensure tensor data is aligned - tensor_data_offset = fin.tell() - tensor_data_offset = (tensor_data_offset + 31) & -32 - fin.seek(tensor_data_offset) - - if ftype_cur == 2: - # 4-bit quantized weights - dtype = np.uint8 - data = dequantize_weights(fin, shape[0], shape[1]) - data = data.reshape(shape) - elif ftype_cur == 0: - dtype = np.float32 - data_size = np.prod(shape) - data = np.fromfile(fin, dtype=dtype, count=data_size).reshape(shape) - elif ftype_cur == 1: - dtype = np.float16 - data_size = np.prod(shape) - data = np.fromfile(fin, dtype=dtype, count=data_size).reshape(shape) - - model[name] = torch.tensor(data, dtype=torch.float32 if dtype == np.float32 else torch.float16) - - pbar.update(fin.tell() - start_pos) - - return model - - -def convert_to_hf_format(model, hparams): - # This works for llama 7B, need to test with other models - n_layers = hparams["n_layers"] - n_heads = hparams["n_heads"] - dim = hparams["dim"] - dims_per_head = dim // n_heads - base = 10000.0 - inv_freq = 1.0 / (base ** (torch.arange(0, dims_per_head, 2).float() / dims_per_head)) - - # permute for sliced rotary - def permute(w): - return w.view(n_heads, dim // n_heads // 2, 2, dim).transpose(1, 2).reshape(dim, dim) - - state_dict = {} - for layer_i in range(n_layers): - state_dict.update( - { - f"model.layers.{layer_i}.self_attn.q_proj.weight": permute( - model[f"layers.{layer_i}.attention.wq.weight"] - ), - f"model.layers.{layer_i}.self_attn.k_proj.weight": permute( - model[f"layers.{layer_i}.attention.wk.weight"] - ), - f"model.layers.{layer_i}.self_attn.v_proj.weight": model[ - f"layers.{layer_i}.attention.wv.weight" - ], - f"model.layers.{layer_i}.self_attn.o_proj.weight": model[ - f"layers.{layer_i}.attention.wo.weight" - ], - f"model.layers.{layer_i}.mlp.gate_proj.weight": model[ - f"layers.{layer_i}.feed_forward.w1.weight" - ], - f"model.layers.{layer_i}.mlp.down_proj.weight": model[ - f"layers.{layer_i}.feed_forward.w2.weight" - ], - f"model.layers.{layer_i}.mlp.up_proj.weight": model[ - f"layers.{layer_i}.feed_forward.w3.weight" - ], - f"model.layers.{layer_i}.input_layernorm.weight": model[ - f"layers.{layer_i}.attention_norm.weight" - ], - f"model.layers.{layer_i}.post_attention_layernorm.weight": model[ - f"layers.{layer_i}.ffn_norm.weight" - ], - } - ) - state_dict[f"model.layers.{layer_i}.self_attn.rotary_emb.inv_freq"] = inv_freq - state_dict.update( - { - "model.embed_tokens.weight": model["tok_embeddings.weight"], - "model.norm.weight": model["norm.weight"], - "lm_head.weight": model["output.weight"], - } - ) - - return state_dict - - -def chat(model, hparams, llama_dir): - from transformers import (GenerationConfig, LlamaForCausalLM, - LlamaTokenizer, StoppingCriteria, - StoppingCriteriaList) - from transformers.models.llama.configuration_llama import LlamaConfig - - class StoppingCriteriaSub(StoppingCriteria): - def __init__(self): - super().__init__() - - def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, stops=[]): - print(tokenizer.decode(input_ids[0]), end="", flush=True) - if input_ids[0][-1] == 13: - return True - - return False - - config = LlamaConfig( - vocab_size=hparams["vocab_size"], - dim=hparams["dim"], - num_hidden_layers=hparams["n_layers"], - num_attention_heads=hparams["n_heads"], - ) - - llama = LlamaForCausalLM(config=config) - llama.load_state_dict(state_dict=model, strict=True) - tokenizer = LlamaTokenizer.from_pretrained(llama_dir) - - device = torch.device("cpu") - llama = llama.to(device) - - ctx = """You are AI. -This is a dialog, where User interacts with AI. AI is helpful, kind, obedient, honest, respectful, direct, concise, should try to protect User's privacy, and knows its own limits. Also, AI must answer User and AI cannot stop the conversation by itself. -User: Hello, AI. -AI: Hello! How can I assist you today? -""" - print(ctx.rstrip("\n")) - while True: - print("-" * 60) - prompt = input("User: ") - if ctx != "": - ctx = f"{ctx}User: {prompt}\n" - else: - ctx = f"{prompt}\nAI:" - - ctx = (ctx[-1920:]) if len(ctx) >= 2048 else ctx - - print("-" * 60) - if len(ctx.strip()) > 0: - input_ids = tokenizer(ctx, return_tensors="pt")["input_ids"].to(device) - generation_config = GenerationConfig( - temperature=0.8, - top_p=0.95, - top_k=50, - repetition_penalty=1.1764, - ) - with torch.no_grad(): - generation_output = llama.generate( - input_ids=input_ids, - generation_config=generation_config, - return_dict_in_generate=True, - output_scores=True, - max_length=2048, - do_sample=True, - stopping_criteria=StoppingCriteriaList([StoppingCriteriaSub()]), - ) - s = generation_output.sequences[0] - decoded = tokenizer.decode(s) - ctx = f"{decoded}\n" - - -def main(): - parser = argparse.ArgumentParser() - parser.add_argument( - "--input_dir", "-i", type=str, required=True, help="The input directory containing the ggml files." - ) - parser.add_argument( - "--prefix", - "-p", - type=str, - required=True, - help="The prefix of the ggml files (ggml-model-f16 or ggml-model-q4_0).", - ) - parser.add_argument( - "--hf", - action="store_true", - help="Whether to save the model in the Hugging Face format. (default: False)", - ) - parser.add_argument( - "--chat", "-c", action="store_true", help="Whether to open a chat with the model. (default: False)" - ) - args = parser.parse_args() - - llama_dir = os.path.abspath(f"{args.input_dir}/../") - - ggml_files = sorted( - [f"{args.input_dir}/{f}" for f in os.listdir(args.input_dir) if f.startswith(args.prefix)] - ) - - fin = open(ggml_files[0], "rb") - hparams, ftype = read_header(fin) - tokens = read_tokens(fin, hparams["vocab_size"]) - model = read_variables(fin) - - for f in tqdm(ggml_files[1:]): - fin = open(f, "rb") - read_header(fin) - read_tokens(fin, hparams["vocab_size"]) - model.update(read_variables(fin)) - - if args.hf: - model = convert_to_hf_format(model, hparams) - - pth_ckpt = { - "state_dict": model, - "hparams": hparams, - "tokens": tokens, - } - - torch.save(pth_ckpt, f"{args.input_dir}/{args.prefix}-to-torch.pth") - - if args.chat: - if not args.hf: - model = convert_to_hf_format(model, hparams) - chat(model, hparams, llama_dir) - - -if __name__ == "__main__": - main() diff --git a/convert-gpt4all-to-ggml.py b/convert-gpt4all-to-ggml.py deleted file mode 100644 index b1a5e0560..000000000 --- a/convert-gpt4all-to-ggml.py +++ /dev/null @@ -1,107 +0,0 @@ -#!/usr/bin/env python3 - -# -# TODO: deduplicate GPT4All with convert-unversioned-ggml-to-ggml.py -# - -# Original by https://github.com/eiz -# https://github.com/ggerganov/llama.cpp/issues/324#issuecomment-1476227818 -import argparse -import glob -import os -import struct -import sys -from sentencepiece import SentencePieceProcessor - -HPARAMS = keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"] - -def parse_args(): - parser = argparse.ArgumentParser(description='Upgrade a GPT4All model to the current format') - parser.add_argument('gpt4all_model', help='path to gpt4all-lora-quantized.bin') - parser.add_argument('tokenizer_model', help='path to LLaMA tokenizer.model file') - return parser.parse_args() - -def read_header(f_in): - struct_fmt = "i" * (3 + len(HPARAMS)) - struct_size = struct.calcsize(struct_fmt) - buf = f_in.read(struct_size) - return struct.unpack(struct_fmt, buf) - -def write_header(f_out, header): - (magic, vocab_size, dim, multiple_of, n_heads, n_layers, rot, ftype) = header - - if magic != 0x67676d6c: - raise Exception('Invalid file magic. Must be an old style ggml file.') - - values = [ - 0x67676d66, # magic: ggml in hex - 1, # file version - vocab_size, - dim, - multiple_of, - n_heads, - n_layers, - rot, - ftype - ] - f_out.write(struct.pack("i" * len(values), *values)) - -def write_tokens(fout, tokenizer): - for i in range(tokenizer.vocab_size()): - if tokenizer.is_unknown(i): - text = " \u2047 ".encode() - elif tokenizer.is_control(i): - text = b"" - elif tokenizer.is_byte(i): - piece = tokenizer.id_to_piece(i) - if len(piece) != 6: - print(f"Invalid token: {piece}") - sys.exit(1) - byte_value = int(piece[3:-1], 16) - text = struct.pack("B", byte_value) - else: - text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode() - fout.write(struct.pack("i", len(text))) - fout.write(text) - fout.write(struct.pack("f", tokenizer.get_score(i))) - - # TODO: GPT4All - add extra token - text = "".encode() - fout.write(struct.pack("i", len(text))) - fout.write(text) - fout.write(struct.pack("f", 0.0)) - -def read_tokens(f_in, tokenizer): - for i in range(tokenizer.vocab_size()): - len_b = f_in.read(4) - (length,) = struct.unpack("i", len_b) - f_in.read(length) - -def copy_all_data(f_out, f_in): - while True: - buf = f_in.read(1024 * 1024) - if not buf: - break - f_out.write(buf) - -def convert_one_file(path_in, tokenizer): - path_tmp = f"{path_in}.tmp" - path_orig= f"{path_in}.orig" - print(f"converting {path_in}") - with open(path_in, "rb") as f_in, open(path_tmp, "wb") as f_out: - write_header(f_out, read_header(f_in)) - read_tokens(f_in, tokenizer) - write_tokens(f_out, tokenizer) - copy_all_data(f_out, f_in) - os.rename(path_in, path_orig) - os.rename(path_tmp, path_in) - -def main(): - args = parse_args() - - tokenizer = SentencePieceProcessor(args.tokenizer_model) - - convert_one_file(args.gpt4all_model, tokenizer) - -if __name__ == "__main__": - main() diff --git a/convert-gptq-to-ggml.py b/convert-gptq-to-ggml.py deleted file mode 100644 index 42e99c2ff..000000000 --- a/convert-gptq-to-ggml.py +++ /dev/null @@ -1,172 +0,0 @@ -# Convert a GPTQ quantized LLaMA model to a ggml compatible file -# Based on: https://github.com/qwopqwop200/GPTQ-for-LLaMa -# -import os -import re -import sys -import json -import struct -import numpy as np -import torch -from sentencepiece import SentencePieceProcessor - -if len(sys.argv) != 4: - print("Usage: convert-gptq-to-ggml.py llamaXXb-4bit.pt tokenizer.model out.bin\n") - sys.exit(1) - -fname_model = sys.argv[1] -fname_tokenizer = sys.argv[2] -dir_out = sys.argv[3] - -model = torch.load(fname_model, map_location="cpu") - -n_vocab, n_embd = model['model.embed_tokens.weight'].shape -n_layer = 1 + max(int(m.group(1)) for name in model - if (m := re.match(r'model\.layers\.([0-9]+)', name))) - -# hardcoded: -n_mult = 256 -n_head = {32: 32, 40: 40, 60: 52, 80: 64}[n_layer] - -tokenizer = SentencePieceProcessor(fname_tokenizer) - -assert tokenizer.vocab_size() == n_vocab - -fname_out = sys.argv[3] - -fout = open(fname_out, "wb") - -fout.write(struct.pack("i", 0x67676d66)) # magic: ggmf in hex -fout.write(struct.pack("i", 1)) # file version -fout.write(struct.pack("i", n_vocab)) -fout.write(struct.pack("i", n_embd)) -fout.write(struct.pack("i", n_mult)) -fout.write(struct.pack("i", n_head)) -fout.write(struct.pack("i", n_layer)) -fout.write(struct.pack("i", n_embd // n_head)) # rot (obsolete) -fout.write(struct.pack("i", 4)) - - -# This loop unchanged from convert-pth-to-ggml.py: -for i in range(tokenizer.vocab_size()): - if tokenizer.is_unknown(i): - text = " \u2047 ".encode() - elif tokenizer.is_control(i): - text = b"" - elif tokenizer.is_byte(i): - piece = tokenizer.id_to_piece(i) - if len(piece) != 6: - print(f"Invalid token: {piece}") - sys.exit(1) - byte_value = int(piece[3:-1], 16) - text = struct.pack("B", byte_value) - else: - text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode() - fout.write(struct.pack("i", len(text))) - fout.write(text) - fout.write(struct.pack("f", tokenizer.get_score(i))) - -def write_header(shape, dst_name, ftype_cur): - sname = dst_name.encode() - fout.write(struct.pack("iii", len(shape), len(sname), ftype_cur)) - fout.write(struct.pack("i" * len(shape), *shape[::-1])) - fout.write(sname) - - # ensure tensor data is aligned - tensor_data_offset = fout.tell() - tensor_data_offset = (tensor_data_offset + 31) & -32 - fout.seek(tensor_data_offset) - -def convert_non_q4(src_name, dst_name): - v = model[src_name] - shape = v.shape - print(f"Processing non-Q4 variable: {src_name} with shape: {shape} and type: {v.dtype}") - if len(shape) == 1: - print(" Converting to float32") - v = v.to(torch.float32) - - ftype_cur = {torch.float16: 1, torch.float32: 0}[v.dtype] - - # header - write_header(shape, dst_name, ftype_cur) - - # data - v.numpy().tofile(fout) - -def convert_q4(src_name, dst_name, permute=False): - zeros = model[f"{src_name}.zeros"].numpy() - scales = model[f"{src_name}.scales"].numpy() - bias = model[f"{src_name}.bias"].numpy() - qweight = model[f"{src_name}.qweight"].numpy().T # transpose - - # Q4_1 does not support bias; good thing the bias is always all zeros. - assert not np.any(bias) - - # Each int32 item is actually 8 int4 items packed together, and it's transposed. - shape = (qweight.shape[0], qweight.shape[1] * 8) - - print(f"Processing Q4 variable: {src_name} with shape: {shape}") - - # The output format has the int4 weights in groups of 32 rather than 8. - # It looks like this: - # For each row: - # For each group of 32 columns: - # - addend (float32, 4 bytes) - # - scale (float32, 4 bytes) - # - weights (int4 * 32, 16 bytes) - # Note that in the input, the scales and addends are shared between all - # the columns in a row, so we end up wasting quite a bit of memory with - # repeated scales and addends. - - addends = -zeros # flip sign - - # Since the output format is mixed between integers and floats, we have - # to hackily view the floats as int32s just so numpy will let us - # concatenate them. - addends_view = addends.view(dtype=np.int32) - scales_view = scales.view(dtype=np.int32) - - # Split into groups of 4 columns (i.e. 32 columns of quantized data): - grouped = qweight.reshape([qweight.shape[0], qweight.shape[1] // 4, 4]) - - # Repeat addends and scales: - addends_rep = np.atleast_3d(addends_view).repeat(grouped.shape[1], axis=1) - scales_rep = np.atleast_3d(scales_view).repeat(grouped.shape[1], axis=1) - - blob = np.concatenate([scales_rep, addends_rep, grouped], axis=2, casting='no') - - if permute: - # Permute some rows to undo the permutation done by convert_llama_weights_to_hf.py. - # This can be done after the above conversion because it doesn't affect column order/layout. - blob = (blob.reshape(n_head, 2, shape[0] // n_head // 2, *blob.shape[1:]) - .swapaxes(1, 2) - .reshape(blob.shape)) - - # header - write_header(shape, dst_name, 3) # ftype = Q4_1 - - # data - blob.tofile(fout) - -convert_non_q4("model.embed_tokens.weight", "tok_embeddings.weight") -convert_non_q4("model.norm.weight", "norm.weight") -convert_non_q4("lm_head.weight", "output.weight") - -for i in range(n_layer): - convert_q4(f"model.layers.{i}.self_attn.q_proj", f"layers.{i}.attention.wq.weight", permute=True) - convert_q4(f"model.layers.{i}.self_attn.k_proj", f"layers.{i}.attention.wk.weight", permute=True) - convert_q4(f"model.layers.{i}.self_attn.v_proj", f"layers.{i}.attention.wv.weight") - convert_q4(f"model.layers.{i}.self_attn.o_proj", f"layers.{i}.attention.wo.weight") - - convert_q4(f"model.layers.{i}.mlp.gate_proj", f"layers.{i}.feed_forward.w1.weight") - convert_q4(f"model.layers.{i}.mlp.down_proj", f"layers.{i}.feed_forward.w2.weight") - convert_q4(f"model.layers.{i}.mlp.up_proj", f"layers.{i}.feed_forward.w3.weight") - - convert_non_q4(f"model.layers.{i}.input_layernorm.weight", f"layers.{i}.attention_norm.weight") - convert_non_q4(f"model.layers.{i}.post_attention_layernorm.weight", f"layers.{i}.ffn_norm.weight") - - -fout.close() - -print(f"Done. Output file: {fname_out}") -print() diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py new file mode 100644 index 000000000..8a2085c25 --- /dev/null +++ b/convert-lora-to-ggml.py @@ -0,0 +1,124 @@ +import json +import os +import re +import struct +import sys +from typing import Any, Dict, Sequence, TextIO + +import torch + +from convert import DATA_TYPE_TO_FTYPE, NUMPY_TYPE_TO_DATA_TYPE, DataType + +HF_SUBLAYER_TO_GGML = { + "self_attn.q_proj": "attention.wq", + "self_attn.k_proj": "attention.wk", + "self_attn.v_proj": "attention.wv", + "self_attn.o_proj": "attention.wo", + "mlp.gate_proj": "feed_forward.w1", + "mlp.down_proj": "feed_forward.w2", + "mlp.up_proj": "feed_forward.w3", + "input_layernorm": "attention_norm", + "post_attention_layernorm": "ffn_norm", + # "norm": "norm", + # "embed_tokens": "tok_embeddings", + # "lm_head": "output", +} + + +def translate_tensor_name(t: str) -> str: + match = re.match(r".*layers\.(\d+)\.(\w+\.\w+)\.lora_(A|B)\.weight", t) + if match: + nn = match.group(1) + sub_layer = match.group(2) + lora_type = match.group(3) + + sub_layer_renamed = HF_SUBLAYER_TO_GGML.get(sub_layer) + if sub_layer_renamed is None: + print(f"Error: unrecognized sub-layer {sub_layer} in tensor {t}") + sys.exit(1) + + output_string = ( + f"layers.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}" + ) + return output_string + else: + print(f"Error: unrecognized tensor {t}") + sys.exit(1) + + +def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None: + fout.write(b"ggla"[::-1]) # magic (ggml lora) + fout.write(struct.pack("i", 1)) # file version + fout.write(struct.pack("ii", params["r"], params["lora_alpha"])) + + +def write_tensor_header( + self, name: str, shape: Sequence[int], data_type: DataType +) -> None: + sname = name.encode("utf-8") + fout.write( + struct.pack( + "iii", + len(shape), + len(sname), + DATA_TYPE_TO_FTYPE[NUMPY_TYPE_TO_DATA_TYPE[data_type]], + ) + ) + fout.write(struct.pack("i" * len(shape), *shape[::-1])) + fout.write(sname) + fout.seek((fout.tell() + 31) & -32) + + +if len(sys.argv) != 2: + print(f"Usage: python {sys.argv[0]} ") + print( + "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'" + ) + sys.exit(1) + +input_json = os.path.join(sys.argv[1], "adapter_config.json") +input_model = os.path.join(sys.argv[1], "adapter_model.bin") +output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin") + +model = torch.load(input_model, map_location="cpu") + +with open(input_json, "r") as f: + params = json.load(f) + +if params["peft_type"] != "LORA": + print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA") + sys.exit(1) + +if params["fan_in_fan_out"] == True: + print("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") + 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") + sys.exit(1) + +with open(output_path, "wb") as fout: + fout.truncate() + + write_file_header(fout, params) + for k, v in model.items(): + if k.endswith("lora_A.weight"): + if v.dtype != torch.float16 and v.dtype != torch.float32: + v = v.float() + v = v.T + else: + v = v.float() + + t = v.numpy() + tname = translate_tensor_name(k) + print(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}") diff --git a/convert-pth-to-ggml.py b/convert-pth-to-ggml.py index dcef2f6a3..f87ac270c 100644 --- a/convert-pth-to-ggml.py +++ b/convert-pth-to-ggml.py @@ -1,274 +1,11 @@ -# Convert a LLaMA model checkpoint to a ggjt compatible file -# -# Load the model using Torch -# Iterate over all variables and write them to a binary file. -# -# For each variable, write the following: -# - Number of dimensions (int) -# - Name length (int) -# - Dimensions (int[n_dims]) -# - Name (char[name_length]) -# - Data (float[n_dims]) -# -# At the start of the ggml file we write the model parameters -# and vocabulary. -# +# Compatibility stub import argparse -import os -import sys -import json -import struct -import numpy as np -import torch -from sentencepiece import SentencePieceProcessor +import convert -QK = 32 - -GGML_TYPE_Q4_0 = 0 -GGML_TYPE_Q4_1 = 1 -GGML_TYPE_I8 = 2 -GGML_TYPE_I16 = 3 -GGML_TYPE_I32 = 4 -GGML_TYPE_F16 = 5 -GGML_TYPE_F32 = 6 - -WTYPES = { - 0: GGML_TYPE_F32, - 1: GGML_TYPE_F16, - 2: GGML_TYPE_Q4_0, - 3: GGML_TYPE_Q4_1, -} - -GGML_BLCK_SIZE = { - GGML_TYPE_Q4_0: QK, - GGML_TYPE_Q4_1: QK, - GGML_TYPE_I8: 1, - GGML_TYPE_I16: 1, - GGML_TYPE_I32: 1, - GGML_TYPE_F16: 1, - GGML_TYPE_F32: 1, -} - -GGML_TYPE_SIZE = { - GGML_TYPE_Q4_0: 4 + QK//2, - GGML_TYPE_Q4_1: 4*2 + QK//2, - GGML_TYPE_I8: 1, - GGML_TYPE_I16: 2, - GGML_TYPE_I32: 4, - GGML_TYPE_F16: 2, - GGML_TYPE_F32: 4, -} - -def ggml_nelements(shape): - r = 1 - for i in shape: - r *= i - return r - -def ggml_nbytes(shape, ftype): - x = ggml_nelements(shape) - t = WTYPES[ftype] - x *= GGML_TYPE_SIZE[t] - x //= GGML_BLCK_SIZE[t] - return x - -def parse_args(): - parser = argparse.ArgumentParser(description='Convert a LLaMA model checkpoint to a ggml compatible file') - parser.add_argument('dir_model', help='directory containing the model checkpoint') - parser.add_argument('ftype', help='file type (0: float32, 1: float16)', type=int, choices=[0, 1], default=1) - parser.add_argument('vocab_only', help='only write vocab to file', type=int, default=0, nargs='?') - return parser.parse_args() - -def get_n_parts(dim): - mappings = {4096: 1, 5120: 2, 6656: 4, 8192: 8} - n_parts = mappings.get(dim) - if n_parts is None: - print(f"Invalid dim: {dim}") - sys.exit(1) - - print(f"n_parts = {n_parts}\n") - return n_parts - -def load_hparams_and_tokenizer(dir_model): - # `dir_model` is something like `models/7B` or `models/7B/`. - # "tokenizer.model" is expected under model's parent dir. - # When `dir_model` is a symlink, f"{dir_model}/../tokenizer.model" would not be found. - # Let's use the model's parent dir directly. - model_parent_dir = os.path.dirname(os.path.normpath(dir_model)) - fname_hparams = f"{dir_model}/params.json" - fname_tokenizer = f"{model_parent_dir}/tokenizer.model" - with open(fname_hparams, "r") as f: - hparams = json.load(f) - print(hparams) - tokenizer = SentencePieceProcessor(fname_tokenizer) - hparams.update({"vocab_size": tokenizer.vocab_size()}) - return hparams, tokenizer - -def write_header(fout, hparams, ftype): - keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"] - values = [ - 0x67676a74, # magic: ggjt in hex - 1, # file version - *[hparams[key] for key in keys], - hparams["dim"] // hparams["n_heads"], # rot (obsolete) - ftype - ] - fout.write(struct.pack("i" * len(values), *values)) - -def write_tokens(fout, tokenizer): - for i in range(tokenizer.vocab_size()): - if tokenizer.is_unknown(i): - text = " \u2047 ".encode() - elif tokenizer.is_control(i): - text = b"" - elif tokenizer.is_byte(i): - piece = tokenizer.id_to_piece(i) - if len(piece) != 6: - print(f"Invalid token: {piece}") - sys.exit(1) - byte_value = int(piece[3:-1], 16) - text = struct.pack("B", byte_value) - else: - text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode() - fout.write(struct.pack("i", len(text))) - fout.write(text) - fout.write(struct.pack("f", tokenizer.get_score(i))) - -def process_and_write_variables(fout, model, ftype, part_id, n_parts): - for name, datao in model.items(): - if name.endswith("freqs"): - continue - - # remove dimensions with a single element - data = datao.numpy().squeeze() - partshape = data.shape - n_dims = len(data.shape) - assert n_dims in (1, 2) - - print(f"Processing variable: {name} with shape: {partshape} and type: {datao.dtype}") - - # coerce single-dimensional tensors from float16 to float32 - ftype_cur = 1 - if ftype == 0 or n_dims == 1: - print(" Converting to float32") - data = data.astype(np.float32) - ftype_cur = 0 - blck_size = GGML_BLCK_SIZE[WTYPES[ftype_cur]] - type_size = GGML_TYPE_SIZE[WTYPES[ftype_cur]] - - # determine dimension along which multipart tensor is sharded - # - # split_dim 0 regex: - # - output.* - # - layers.*.attention.wq.weight - # - layers.*.attention.wk.weight - # - layers.*.attention.wv.weight - # - layers.*.feed_forward.w1.weight - # - layers.*.feed_forward.w3.weight - # - # split_dim 1 regex: - # - tok_embeddings.* - # - layers.*.attention.wo.weight - # - layers.*.feed_forward.w2.weight - # - if n_dims > 1: - split_dim = 1 - if "tok_embeddings" in name: - split_dim = 1 - elif "layers" in name: - if "attention.wo.weight" in name: - split_dim = 1 - elif "feed_forward.w2.weight" in name: - split_dim = 1 - else: - split_dim = 0 - elif "output" in name: - split_dim = 0 - - # output tensor header - fullshape = list(partshape) - if n_dims > 1: - fullshape[split_dim] *= n_parts - sname = name.encode() - fout.write(struct.pack("iii", n_dims, len(sname), ftype_cur)) - for dim in reversed(fullshape): - fout.write(struct.pack("i", dim)) - fout.write(sname) - - # ensure tensor data is aligned - tensor_data_offset = fout.tell() - while tensor_data_offset % QK != 0: - fout.write(struct.pack("B", 0)) - tensor_data_offset += 1 - - # output unified mappable tensor data - if n_dims == 1 or n_parts == 1: - # copy tensor which we thankfully received in one piece - if part_id == 0: - data.tofile(fout) - elif split_dim == 0: - # reassemble multifile tensor containing some of the rows - rows_per_chunk = partshape[0] - current_row = part_id * rows_per_chunk - bytes_per_row = fullshape[1] // blck_size * type_size - offset = current_row * bytes_per_row - fout.seek(tensor_data_offset + offset) - data.tofile(fout) - elif split_dim == 1: - # reassemble multifile tensor containing some of the cols - cols_per_chunk = partshape[1] - current_col = part_id * cols_per_chunk - bytes_per_row = fullshape[1] // blck_size * type_size - offset_current_col = current_col // blck_size * type_size - for row in range(partshape[0]): - offset_row = row * bytes_per_row - offset = offset_row + offset_current_col - fout.seek(tensor_data_offset + offset) - data[row].tofile(fout) - - # advance file position to next tensor - fout.seek(tensor_data_offset + ggml_nbytes(fullshape, ftype_cur)) - -def main(): - args = parse_args() - dir_model = args.dir_model - ftype = args.ftype - ftype_str = ["f32", "f16"] - hparams, tokenizer = load_hparams_and_tokenizer(dir_model) - - print(args) - - # if only writing vocab to file - if args.vocab_only: - fname_model = f"{dir_model}/consolidated.00.pth" - fname_out = f"{dir_model}/ggml-vocab.bin" - print(f"Extracting only the vocab from '{fname_model}'\n") - with open(fname_out, "wb") as fout: - write_header(fout, hparams, ftype) - write_tokens(fout, tokenizer) - print(f"Done. Output file: {fname_out}\n") - return - - n_parts = get_n_parts(hparams["dim"]) - fname_out = f"{dir_model}/ggml-model-{ftype_str[ftype]}.bin" - - # we output a single file for ggml - with open(fname_out, "wb") as fout: - write_header(fout, hparams, ftype) - write_tokens(fout, tokenizer) - offset_of_tensors = fout.tell() - # the tensors we load could be split across multiple files - for part_id in range(n_parts): - fout.seek(offset_of_tensors) - print(f"Processing part {part_id+1} of {n_parts}\n") - fname_model = f"{dir_model}/consolidated.0{part_id}.pth" - model = torch.load(fname_model, map_location="cpu") - process_and_write_variables(fout, model, ftype, part_id, n_parts) - del model - - print(f"Done. Output file: {fname_out}\n") - -if __name__ == "__main__": - main() +parser = argparse.ArgumentParser(description='Convert a LLaMA model checkpoint to a ggml compatible file') +parser.add_argument('dir_model', help='directory containing the model checkpoint') +parser.add_argument('ftype', help='file type (0: float32, 1: float16)', type=int, choices=[0, 1], default=1) +args = parser.parse_args() +convert.main(['--outtype', 'f16' if args.ftype == 1 else 'f32', '--', args.dir_model]) diff --git a/convert-unversioned-ggml-to-ggml.py b/convert-unversioned-ggml-to-ggml.py deleted file mode 100644 index 5151d9081..000000000 --- a/convert-unversioned-ggml-to-ggml.py +++ /dev/null @@ -1,100 +0,0 @@ -#!/usr/bin/env python3 -# Original by https://github.com/eiz -# https://github.com/ggerganov/llama.cpp/issues/324#issuecomment-1476227818 -import argparse -import glob -import os -import struct -import sys -from sentencepiece import SentencePieceProcessor - -HPARAMS = keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"] - -def parse_args(): - parser = argparse.ArgumentParser(description='Upgrade old ggml model files to the current format') - parser.add_argument('dir_model', help='directory containing ggml .bin files') - parser.add_argument('tokenizer_model', help='path to LLaMA tokenizer.model file') - return parser.parse_args() - -def read_header(f_in): - struct_fmt = "i" * (3 + len(HPARAMS)) - struct_size = struct.calcsize(struct_fmt) - buf = f_in.read(struct_size) - return struct.unpack(struct_fmt, buf) - -def write_header(f_out, header): - (magic, vocab_size, dim, multiple_of, n_heads, n_layers, rot, ftype) = header - - if magic != 0x67676d6c: - raise Exception('Invalid file magic. Must be an old style ggml file.') - - values = [ - 0x67676d66, # magic: ggml in hex - 1, # file version - vocab_size, - dim, - multiple_of, - n_heads, - n_layers, - rot, - ftype - ] - f_out.write(struct.pack("i" * len(values), *values)) - -def write_tokens(fout, tokenizer): - for i in range(tokenizer.vocab_size()): - if tokenizer.is_unknown(i): - text = " \u2047 ".encode() - elif tokenizer.is_control(i): - text = b"" - elif tokenizer.is_byte(i): - piece = tokenizer.id_to_piece(i) - if len(piece) != 6: - print(f"Invalid token: {piece}") - sys.exit(1) - byte_value = int(piece[3:-1], 16) - text = struct.pack("B", byte_value) - else: - text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode() - fout.write(struct.pack("i", len(text))) - fout.write(text) - fout.write(struct.pack("f", tokenizer.get_score(i))) - -def read_tokens(f_in, tokenizer): - for i in range(tokenizer.vocab_size()): - len_b = f_in.read(4) - (length,) = struct.unpack("i", len_b) - f_in.read(length) - -def copy_all_data(f_out, f_in): - while True: - buf = f_in.read(1024 * 1024) - if not buf: - break - f_out.write(buf) - -def convert_one_file(path_in, tokenizer): - path_tmp = f"{path_in}.tmp" - path_orig= f"{path_in}.orig" - print(f"converting {path_in}") - with open(path_in, "rb") as f_in, open(path_tmp, "wb") as f_out: - write_header(f_out, read_header(f_in)) - read_tokens(f_in, tokenizer) - write_tokens(f_out, tokenizer) - copy_all_data(f_out, f_in) - os.rename(path_in, path_orig) - os.rename(path_tmp, path_in) - -def main(): - args = parse_args() - files = [] - files.extend(glob.glob(f"{args.dir_model}/*.bin")) - files.extend(glob.glob(f"{args.dir_model}/*.bin.*")) - - tokenizer = SentencePieceProcessor(args.tokenizer_model) - - for file in files: - convert_one_file(file, tokenizer) - -if __name__ == "__main__": - main() diff --git a/convert.py b/convert.py new file mode 100644 index 000000000..7f7ae05fa --- /dev/null +++ b/convert.py @@ -0,0 +1,1149 @@ +import argparse +import concurrent.futures +import copy +import enum +import faulthandler +import functools +import io +import itertools +import json +import math +import mmap +import pickle +import re +import signal +import struct +import sys +import zipfile +from abc import ABCMeta, abstractmethod +from dataclasses import dataclass +from pathlib import Path +from typing import (IO, TYPE_CHECKING, Any, Callable, Dict, Iterable, List, + Literal, Optional, Sequence, Tuple, TypeVar, Union) + +import numpy as np +from sentencepiece import SentencePieceProcessor # type: ignore + +if TYPE_CHECKING: + from typing_extensions import TypeAlias + +if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'): + faulthandler.register(signal.SIGUSR1) + +NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]' + + +@dataclass(frozen=True) +class UnquantizedDataType: + name: str + + +DT_F16 = UnquantizedDataType('F16') +DT_F32 = UnquantizedDataType('F32') +DT_I32 = UnquantizedDataType('I32') +DT_BF16 = UnquantizedDataType('BF16') + + +@dataclass(frozen=True) +class QuantizedDataType: + groupsize: int + have_addends: bool + have_g_idx: bool + + +DT_Q4_0 = QuantizedDataType(groupsize=32, have_addends=False, have_g_idx=False) +DT_Q4_1 = QuantizedDataType(groupsize=32, have_addends=True, have_g_idx=False) + +DataType = Union[UnquantizedDataType, QuantizedDataType] + +DATA_TYPE_TO_FTYPE: Dict[DataType, int] = { + DT_F32: 0, + DT_F16: 1, + DT_Q4_0: 2, + DT_Q4_1: 3, +} + +FTYPE_TO_DATA_TYPE: Dict[int, DataType] = \ + {ftype: dtype for (dtype, ftype) in DATA_TYPE_TO_FTYPE.items()} + +DATA_TYPE_TO_NUMPY: Dict[DataType, 'np.dtype[Any]'] = { + DT_F16: np.dtype(np.float16), + DT_F32: np.dtype(np.float32), + DT_I32: np.dtype(np.int32), +} + +NUMPY_TYPE_TO_DATA_TYPE: Dict['np.dtype[Any]', DataType] = \ + {dtype: data_type for (data_type, dtype) in DATA_TYPE_TO_NUMPY.items()} + + +class GGMLFileType(enum.Enum): + AllF32 = 0 + MostlyF16 = 1 # except 1d tensors + MostlyQ4_0 = 2 # except 1d tensors + MostlyQ4_1 = 3 # except 1d tensors + PerLayerIsQ4_1 = 4 # but tok_embeddings.weight and output.weight are F16 + + def type_for_tensor(self, name: str, tensor: 'LazyTensor') -> DataType: + if len(tensor.shape) == 1: + # 1D tensors are always F32. + return DT_F32 + elif self == GGMLFileType.AllF32: + return DT_F32 + elif self == GGMLFileType.MostlyF16: + return DT_F16 + elif self == GGMLFileType.MostlyQ4_0: + return DT_Q4_0 + elif self == GGMLFileType.MostlyQ4_1: + return DT_Q4_1 + elif self == GGMLFileType.PerLayerIsQ4_1: + if name in ('output.weight', 'tok_embeddings.weight'): + return DT_F16 + else: + return DT_Q4_1 + else: + raise ValueError(self) + + +def make_tensors_list() -> List[str]: + ret = [ + 'tok_embeddings.weight', + 'norm.weight', + 'output.weight', + ] + for i in range(80): # maximum number of layer + ret += [ + f'layers.{i}.attention.wq.weight', + f'layers.{i}.attention.wk.weight', + f'layers.{i}.attention.wv.weight', + f'layers.{i}.attention.wo.weight', + f'layers.{i}.attention_norm.weight', + f'layers.{i}.feed_forward.w1.weight', + f'layers.{i}.feed_forward.w2.weight', + f'layers.{i}.feed_forward.w3.weight', + f'layers.{i}.atttention_norm.weight', + f'layers.{i}.ffn_norm.weight', + ] + return ret + + +TENSORS_LIST = make_tensors_list() +TENSORS_SET = set(TENSORS_LIST) + + +@dataclass +class Params: + n_vocab: int + n_embd: int + n_mult: int + n_head: int + n_layer: int + file_type: GGMLFileType + + @staticmethod + def guessed(model: 'LazyModel', file_type: GGMLFileType) -> 'Params': + n_vocab, n_embd = model["tok_embeddings.weight"].shape + + return Params( + n_vocab=n_vocab, + n_embd=n_embd, + n_mult=256, + n_head=n_embd // 128, + n_layer=next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model), + file_type=file_type, + ) + + +class SentencePieceVocab: + def __init__(self, fname_tokenizer: Path, fname_added_tokens: Optional[Path]) -> None: + self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer)) + added_tokens: Dict[str, int] + if fname_added_tokens is not None: + added_tokens = json.load(open(fname_added_tokens)) + else: + added_tokens = {} + vocab_size: int = self.sentencepiece_tokenizer.vocab_size() + expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) + actual_ids = sorted(added_tokens.values()) + if expected_ids != actual_ids: + raise Exception(f"Expected added token IDs to be sequential and start at {len(added_tokens)}; got {actual_ids}") + items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1]) + self.added_tokens_list = [text for (text, idx) in items] + self.vocab_size_base: int = vocab_size + self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer + self.fname_added_tokens = fname_added_tokens + + def sentencepiece_tokens(self) -> Iterable[Tuple[bytes, float]]: + tokenizer = self.sentencepiece_tokenizer + for i in range(tokenizer.vocab_size()): + text: bytes + if tokenizer.is_unknown(i): + text = " \u2047 ".encode("utf-8") + elif tokenizer.is_control(i): + text = b"" + elif tokenizer.is_byte(i): + piece = tokenizer.id_to_piece(i) + if len(piece) != 6: + raise Exception(f"Invalid token: {piece}") + byte_value = int(piece[3:-1], 16) + text = struct.pack("B", byte_value) + else: + text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8") + score: float = tokenizer.get_score(i) + yield text, score + + def added_tokens(self) -> Iterable[Tuple[bytes, float]]: + for text in self.added_tokens_list: + score = -1000.0 + yield text.encode("utf-8"), score + + def all_tokens(self) -> Iterable[Tuple[bytes, float]]: + yield from self.sentencepiece_tokens() + yield from self.added_tokens() + + def __repr__(self) -> str: + return f"" + + +class GGMLVocab: + def __init__(self, tokens: List[Tuple[bytes, float]]): + self.tokens = tokens + self.vocab_size = len(tokens) + + def all_tokens(self) -> Iterable[Tuple[bytes, float]]: + return self.tokens + + def __repr__(self) -> str: + return f"" + + +Vocab = Union[SentencePieceVocab, GGMLVocab] + + +def permute(weights: NDArray, n_head: int) -> NDArray: + return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) + .swapaxes(1, 2) + .reshape(weights.shape)) + + +def dequantize_q4(qvalues_pack32: NDArray, scales: NDArray, addends: Optional[NDArray], g_idx: Optional[NDArray]) -> NDArray: + # First reinterpret each row from a list of int32s containing 8 values each + # to a list of uint8s containing 2 values each. + qvalues_pack8 = qvalues_pack32.view(np.uint8) + + # Then split out the two values per int8 (which requires an actual + # conversion because numpy doesn't natively support int4s). + qvalues = np.zeros([qvalues_pack8.shape[0], qvalues_pack8.shape[1] * 2], dtype=np.uint8) + qvalues[:, 0::2] = qvalues_pack8 & 0xf + qvalues[:, 1::2] = qvalues_pack8 >> 4 + + assert addends is None or addends.shape == scales.shape + assert qvalues.shape[0] == scales.shape[0] + assert qvalues.shape[1] % scales.shape[1] == 0 + if g_idx is None: + repeat_count = qvalues.shape[1] // scales.shape[1] + scales = scales[:, :, np.newaxis] + if addends is not None: + addends = addends[:, :, np.newaxis] + # Reshape so that the below computation broadcasts over scales and addends: + qvalues.shape = (qvalues.shape[0], scales.shape[1], int(repeat_count)) + else: + # In this case the scale and addend is selected for each column by g_idx: + assert addends is not None + scales = scales[:, g_idx] + addends = addends[:, g_idx] + if addends is None: + # Q4_0 + qvalues = qvalues.view(np.int8) + qvalues -= 8 + # And do the actual 'value = scale * qvalue + addend' computation. + values = scales * qvalues + if addends is not None: + values += addends + if g_idx is None: + values.shape = (values.shape[0], values.shape[1] * values.shape[2]) + return values + + +class Tensor(metaclass=ABCMeta): + data_type: DataType + + @abstractmethod + def astype(self, data_type: DataType) -> 'Tensor': ... + @abstractmethod + def permute(self, n_head: int) -> 'Tensor': ... + @abstractmethod + def to_ggml(self) -> 'GGMLCompatibleTensor': ... + + +class UnquantizedTensor(Tensor): + def __init__(self, ndarray: NDArray) -> None: + assert isinstance(ndarray, np.ndarray) + self.ndarray = ndarray + self.data_type = NUMPY_TYPE_TO_DATA_TYPE[ndarray.dtype] + + def astype(self, data_type: DataType) -> Tensor: + dtype = DATA_TYPE_TO_NUMPY[data_type] + return UnquantizedTensor(self.ndarray.astype(dtype)) + + def to_ggml(self) -> 'UnquantizedTensor': + return self + + def permute(self, n_head: int) -> 'UnquantizedTensor': + return UnquantizedTensor(permute(self.ndarray, n_head)) + + +def load_unquantized(lazy_tensor: 'LazyTensor', expected_dtype: Any = None, convert: bool = False) -> NDArray: + tensor = lazy_tensor.load() + assert isinstance(tensor, UnquantizedTensor) + + # double-check: + actual_shape = list(tensor.ndarray.shape) + assert actual_shape == lazy_tensor.shape, (actual_shape, lazy_tensor.shape) + if expected_dtype is not None and expected_dtype != tensor.ndarray.dtype: + if convert: + tensor.ndarray = tensor.ndarray.astype(expected_dtype) + else: + raise ValueError(f'expected this tensor to have dtype {expected_dtype}, got {tensor.ndarray.dtype}') + + return tensor.ndarray + + +class GGMLQuantizedTensor(Tensor): + data_type: QuantizedDataType + + def __init__(self, ndarray: NDArray, shape: List[int], data_type: DataType) -> None: + rows, columns = shape + assert data_type in (DT_Q4_1, DT_Q4_0) # for now + assert isinstance(data_type, QuantizedDataType) # redundant, but mypy complains without this + assert columns % data_type.groupsize == 0 + words_in_block = 6 if data_type == DT_Q4_1 else 5 + self.ndarray = ndarray.view(dtype=np.uint32).reshape((rows, columns // data_type.groupsize, words_in_block)) + self.shape = shape[:] + self.data_type = data_type + + def astype(self, data_type: DataType) -> Tensor: + if data_type == self.data_type: + return self + scales = self.ndarray[:, :, 0].view(np.float32) + if self.data_type.have_addends: + addends = self.ndarray[:, :, 1].view(np.float32) + else: + addends = None + qweights = self.ndarray[:, :, -4:].reshape([self.shape[0], self.shape[1] // 8]) + + dq = dequantize_q4(qweights, scales, addends, g_idx=None) + return UnquantizedTensor(dq).astype(data_type) + + def to_ggml(self) -> 'GGMLQuantizedTensor': + return self + + def permute(self, n_head: int) -> 'GGMLQuantizedTensor': + return GGMLQuantizedTensor(permute(self.ndarray, n_head), self.shape, self.data_type) + + +GGMLCompatibleTensor = Union[UnquantizedTensor, GGMLQuantizedTensor] + + +class DeferredPermutedTensor(Tensor): + def __init__(self, base: Tensor, n_head: int) -> None: + self.base = base + self.n_head = n_head + self.data_type = self.base.data_type + + def astype(self, data_type: DataType) -> Tensor: + return self.base.astype(data_type).permute(self.n_head) + + def to_ggml(self) -> GGMLCompatibleTensor: + return self.base.to_ggml().permute(self.n_head) + + def permute(self, n_head: int) -> Tensor: + raise Exception("shouldn't permute twice") + + +class GPTQForLLaMaQuantizedTensor(Tensor): + def __init__(self, model: 'LazyModel', namebase: str) -> None: + qweight = load_unquantized(model[f"{namebase}.qweight"], np.int32) + scales = load_unquantized(model[f"{namebase}.scales"], np.float32, convert=True) + + bias = model.get(f"{namebase}.bias") + if bias is not None: + # Q4_1 does not support bias; good thing the bias is always all zeros. + assert not np.any(load_unquantized(bias)) + + if f"{namebase}.zeros" in model: + zeros = load_unquantized(model[f"{namebase}.zeros"], np.float32) + else: + qzeros = load_unquantized(model[f"{namebase}.qzeros"], np.int32) + assert qzeros.dtype == np.int32 + zeros = dequantize_q4(qzeros, scales, scales, g_idx=None) + assert zeros.dtype == np.float32 + + assert zeros.shape == scales.shape + + # Output is transposed compared to the input, and addends have their sign flipped. + # Scales and zeros similarly must be transposed but only for newer + # versions of GPTQ-for-LLaMa; the older versions can be identified by + # having shape (n_embd, 1). + qweight = qweight.T + if scales.shape[1] != 1: + scales = scales.T + zeros = zeros.T + + # Output also has signs flipped for the addends. + self.qweight = qweight + self.scales = scales + self.addends = -zeros + + self.g_idx: Optional[NDArray] + if f"{namebase}.g_idx" in model: + self.g_idx = load_unquantized(model[f"{namebase}.g_idx"], np.int32) + assert self.g_idx.shape == (qweight.shape[1] * 8,) + else: + self.g_idx = None + + self.shape = [self.qweight.shape[0], self.qweight.shape[1] * 8] + self.data_type = QuantizedDataType(groupsize=self.groupsize(), have_addends=True, + have_g_idx=(self.g_idx is not None)) + + def inspect(self, row: int, col: int) -> None: + '''For debugging.''' + qweight = (self.qweight[row, col // 8] >> (4 * (col & 7))) & 0xf + if self.g_idx is not None: + group = self.g_idx[col] + else: + group = int(col // self.groupsize()) + scale = self.scales[row, group] + addend = self.addends[row, group] + with np.printoptions(precision=None, suppress=True): + print(f'scale:{scale} addend:{addend} qweight:{qweight}') + print('possible values:', np.arange(16) * scale + addend) + print('actual value:', qweight * scale + addend) + + def astype(self, data_type: DataType) -> Tensor: + if isinstance(data_type, QuantizedDataType): + assert self.g_idx is None and data_type.have_addends is True and data_type.have_g_idx is False + return self.regroup(data_type.groupsize) + + dequantized = dequantize_q4(np.ascontiguousarray(self.qweight), self.scales, self.addends, self.g_idx) + return UnquantizedTensor(dequantized).astype(data_type) + + def groupsize(self) -> int: + assert self.addends.shape == self.scales.shape + assert self.shape[1] % self.scales.shape[1] == 0 + return self.shape[1] // self.scales.shape[1] + + def regroup(self, new_groupsize: int = 32) -> 'GPTQForLLaMaQuantizedTensor': + # Old versions of GPTQ-for-LLaMa shared scales and addends between all the + # columns in a row. Newer versions share them between every set of N + # columns in a row, where N is the `groupsize` parameter, usually 128. The + # output format shares them between every set of 32 columns. To handle + # this, duplicate scales and addends for every smaller group. + # (In the above, 'row' and 'column' are in the sense of the output.) + assert self.g_idx is None + old_groupsize = self.groupsize() + assert old_groupsize >= new_groupsize and old_groupsize % new_groupsize == 0, old_groupsize + ret = copy.copy(self) + ret.addends = self.addends.repeat(old_groupsize // new_groupsize, axis=1) + ret.scales = self.scales.repeat(old_groupsize // new_groupsize, axis=1) + ret.data_type = QuantizedDataType(groupsize=new_groupsize, have_addends=True, have_g_idx=False) + return ret + + def permute(self, n_head: int) -> Tensor: + return DeferredPermutedTensor(self, n_head) + + def to_ggml(self) -> GGMLQuantizedTensor: + # The output format looks like this: + # For each row: + # For each group of 32 columns: + # - addend (float32, 4 bytes) + # - scale (float32, 4 bytes) + # - weights (int4 * 32, 16 bytes) + + if self.groupsize() != 32: + raise Exception("should have been regrouped before converting to ggml") + + # Since the output format is mixed between integers and floats, we have + # to hackily view the floats as int32s just so numpy will let us + # concatenate them. + addends_view = self.addends.view(dtype=np.int32)[:, :, np.newaxis] + scales_view = self.scales.view(dtype=np.int32)[:, :, np.newaxis] + + # Split into groups of 4 columns (i.e. 32 columns of quantized data): + grouped = self.qweight.reshape([self.qweight.shape[0], self.qweight.shape[1] // 4, 4]) + + # And concatenate: + grouped = np.concatenate([scales_view, addends_view, grouped], axis=2, casting='no') + + return GGMLQuantizedTensor(grouped, self.shape, DT_Q4_1) + + +@dataclass +class LazyTensor: + _load: Callable[[], Tensor] + shape: List[int] + data_type: DataType + description: str + + def load(self) -> Tensor: + ret = self._load() + assert ret.data_type == self.data_type, (self.data_type, ret.data_type, self.description) + return ret + + def astype(self, data_type: DataType) -> 'LazyTensor': + self.validate_conversion_to(data_type) + + def load() -> Tensor: + return self.load().astype(data_type) + return LazyTensor(load, self.shape, data_type, f'convert({data_type}) {self.description}') + + def validate_conversion_to(self, data_type: DataType) -> None: + if data_type == self.data_type: + return + if isinstance(data_type, QuantizedDataType): + if not isinstance(self.data_type, QuantizedDataType): + raise Exception(f"Can't turn an unquantized tensor into a quantized type ({data_type})") + if self.data_type.have_g_idx: + sys.stderr.write("Error: Input uses the newer GPTQ-for-LLaMa format (using g_idx), which is not yet natively supported by GGML. For now you can still convert this model by passing `--outtype f16` to dequantize, but that will result in a much larger output file for no quality benefit.\n") + sys.exit(1) + assert not data_type.have_g_idx and self.data_type.have_addends and data_type.have_addends + + +LazyModel = Dict[str, LazyTensor] + + +@dataclass +class ModelPlus: + model: LazyModel + paths: List[Path] # Where this was read from. + format: Literal['ggml', 'torch', 'safetensors'] + vocab: Optional[Vocab] # For GGML models (which have vocab built in), the vocab. + + +def merge_sharded(models: List[LazyModel]) -> LazyModel: + # Original LLaMA models have each file contain one part of each tensor. + # Use a dict instead of a set to preserve order. + names = {name: None for model in models for name in model} + + def convert(name: str) -> LazyTensor: + lazy_tensors: List[LazyTensor] = [model[name] for model in models] + if len(lazy_tensors) == 1: + # only one file; don't go through this procedure since there might + # be quantized tensors + return lazy_tensors[0] + if len(lazy_tensors[0].shape) == 1: + # the tensor is just duplicated in every file + return lazy_tensors[0] + if name.startswith('tok_embeddings.') or \ + name.endswith('.attention.wo.weight') or \ + name.endswith('.feed_forward.w2.weight'): + # split by columns + axis = 1 + else: + # split by rows + axis = 0 + concatenated_shape = list(lazy_tensors[0].shape) + concatenated_shape[axis] = sum(tensor.shape[axis] for tensor in lazy_tensors) + + def load() -> UnquantizedTensor: + ndarrays = [load_unquantized(tensor) for tensor in lazy_tensors] + concatenated: NDArray = np.concatenate(ndarrays, axis=axis) + return UnquantizedTensor(concatenated) + description = 'concatenated[[' + '] | ['.join(lt.description for lt in lazy_tensors) + ']]' + return LazyTensor(load, concatenated_shape, lazy_tensors[0].data_type, description) + return {name: convert(name) for name in names} + + +def merge_multifile_models(models_plus: List[ModelPlus]) -> ModelPlus: + formats = set(mp.format for mp in models_plus) + assert len(formats) == 1, "different formats?" + format = formats.pop() + paths = [path for mp in models_plus for path in mp.paths] + # Use the first non-None vocab, if any. + try: + vocab = next(mp.vocab for mp in models_plus if mp.vocab is not None) + except StopIteration: + vocab = None + + if any("model.embed_tokens.weight" in mp.model for mp in models_plus): + # Transformers models put different tensors in different files, but + # don't split indivdual tensors between files. + model: LazyModel = {} + for mp in models_plus: + model.update(mp.model) + else: + model = merge_sharded([mp.model for mp in models_plus]) + + return ModelPlus(model, paths, format, vocab) + + +def permute_lazy(lazy_tensor: LazyTensor, n_head: int) -> LazyTensor: + def load() -> Tensor: + return lazy_tensor.load().permute(n_head) + return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}) ' + lazy_tensor.description) + + +def convert_transformers_to_orig(model: LazyModel) -> LazyModel: + out: LazyModel = {} + out["tok_embeddings.weight"] = model["model.embed_tokens.weight"] + out["norm.weight"] = model["model.norm.weight"] + out["output.weight"] = model["lm_head.weight"] + + n_head = model["model.layers.0.self_attn.q_proj.weight"].shape[1] // 128 + for i in itertools.count(): + if f"model.layers.{i}.self_attn.q_proj.weight" not in model: + break + out[f"layers.{i}.attention.wq.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], n_head) + out[f"layers.{i}.attention.wk.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], n_head) + out[f"layers.{i}.attention.wv.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"] + out[f"layers.{i}.attention.wo.weight"] = model[f"model.layers.{i}.self_attn.o_proj.weight"] + + out[f"layers.{i}.feed_forward.w1.weight"] = model[f"model.layers.{i}.mlp.gate_proj.weight"] + out[f"layers.{i}.feed_forward.w2.weight"] = model[f"model.layers.{i}.mlp.down_proj.weight"] + out[f"layers.{i}.feed_forward.w3.weight"] = model[f"model.layers.{i}.mlp.up_proj.weight"] + + out[f"layers.{i}.attention_norm.weight"] = model[f"model.layers.{i}.input_layernorm.weight"] + out[f"layers.{i}.ffn_norm.weight"] = model[f"model.layers.{i}.post_attention_layernorm.weight"] + return out + + +def handle_quantization(model: LazyModel) -> LazyModel: + '''Convert a model with entries for 'foo.qweight', 'foo.scales', etc. + (which resolve to UnquantizedTensors with the raw data) to one with entries + for 'foo.weight' (which resolve to QuantizedTensors). + ''' + def convert(name: str) -> Tuple[str, LazyTensor]: + if name.endswith(".qweight"): + namebase = name.rsplit('.', 1)[0] + orig_name = namebase + ".weight" + + lazy_tensor = model[name] + assert len(lazy_tensor.shape) == 2 + real_shape = [lazy_tensor.shape[1], lazy_tensor.shape[0] * 8] + + # Calculate type. This replicates the logic in + # GPTQForLLaMaQuantizedTensor (which is executed when the modelis + # actually loaded). + lazy_scales = model[f"{namebase}.scales"] + scales_width = 1 if lazy_scales.shape[1] == 1 else lazy_scales.shape[0] + assert real_shape[1] % scales_width == 0 + groupsize = real_shape[1] // scales_width + have_g_idx = f"{namebase}.g_idx" in model + data_type = QuantizedDataType(groupsize=groupsize, have_addends=True, have_g_idx=have_g_idx) + + def load() -> Tensor: + return GPTQForLLaMaQuantizedTensor(model, namebase) + + return (orig_name, LazyTensor(load, real_shape, data_type, '[quantized]')) + else: + return (name, model[name]) + return dict(convert(name) for name in model) + +# Functionality that simulates `torch.load` but where individual tensors are +# only loaded into memory on demand, not all at once. +# PyTorch can't do this natively as of time of writing: +# - https://github.com/pytorch/pytorch/issues/64327 +# This allows us to de-shard without multiplying RAM usage, and also +# conveniently drops the PyTorch dependency (though we still need numpy). + + +@dataclass +class LazyStorageKind: + data_type: DataType + + +@dataclass +class LazyStorage: + load: Callable[[int, int], NDArray] + kind: LazyStorageKind + description: str + + +class LazyUnpickler(pickle.Unpickler): + def __init__(self, fp: IO[bytes], data_base_path: str, zip_file: zipfile.ZipFile): + super().__init__(fp) + self.data_base_path = data_base_path + self.zip_file = zip_file + + def persistent_load(self, pid: Any) -> Any: + assert pid[0] == 'storage' + assert isinstance(pid[1], LazyStorageKind) + data_type = pid[1].data_type + filename_stem = pid[2] + filename = self.data_base_path + '/' + filename_stem + info = self.zip_file.getinfo(filename) + + def load(offset: int, elm_count: int) -> NDArray: + dtype = DATA_TYPE_TO_NUMPY.get(data_type) + if dtype is None: + raise Exception("tensor stored in unsupported format") + fp = self.zip_file.open(info) + fp.seek(offset * dtype.itemsize) + size = elm_count * dtype.itemsize + data = fp.read(size) + assert len(data) == size + return np.frombuffer(data, dtype) + description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}' + return LazyStorage(load=load, kind=pid[1], description=description) + + def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any, # pyright: ignore[reportSelfClsParameterName] + requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor: + assert isinstance(storage, LazyStorage) + + def load() -> UnquantizedTensor: + elm_count = stride[0] * size[0] + return UnquantizedTensor(storage.load(storage_offset, elm_count).reshape(size)) + description = f'pickled storage_offset={storage_offset} in {storage.description}' + return LazyTensor(load, list(size), storage.kind.data_type, description) + + CLASSES: Dict[Any, Any] = { + ('torch._utils', '_rebuild_tensor_v2'): lazy_rebuild_tensor_v2, + ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16), + ('torch', 'HalfStorage'): LazyStorageKind(DT_F16), + ('torch', 'FloatStorage'): LazyStorageKind(DT_F32), + ('torch', 'IntStorage'): LazyStorageKind(DT_I32), + } + + def find_class(self, module: str, name: str) -> Any: + if not module.startswith('torch'): + return super().find_class(module, name) + return self.CLASSES[(module, name)] + + +def lazy_load_torch_file(outer_fp: IO[bytes], path: Path) -> ModelPlus: + zf = zipfile.ZipFile(outer_fp) + pickle_paths = [name for name in zf.namelist() if name.endswith('.pkl')] + assert len(pickle_paths) == 1, pickle_paths + pickle_fp = zf.open(pickle_paths[0], 'r') + unpickler = LazyUnpickler(pickle_fp, + data_base_path=pickle_paths[0][:-4], + zip_file=zf) + model = unpickler.load() + as_dict = dict(model.items()) + return ModelPlus(model=as_dict, paths=[path], format='torch', vocab=None) + + +SAFETENSORS_DATA_TYPES: Dict[str, DataType] = { + 'F16': DT_F16, + 'F32': DT_F32, + 'I32': DT_I32, +} + + +def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus: + header_size, = struct.unpack(' LazyTensor: + data_type = SAFETENSORS_DATA_TYPES[info['dtype']] + numpy_dtype = DATA_TYPE_TO_NUMPY[data_type] + shape: List[int] = info['shape'] + begin, end = info['data_offsets'] + assert 0 <= begin <= end <= len(byte_buf) + assert end - begin == math.prod(shape) * numpy_dtype.itemsize + buf = byte_buf[begin:end] + + def load() -> UnquantizedTensor: + return UnquantizedTensor(np.frombuffer(buf, dtype=numpy_dtype).reshape(shape)) + description = f'safetensors begin={begin} end={end} type={data_type} path={path}' + return LazyTensor(load, shape, data_type, description) + model = {name: convert(info) for (name, info) in header.items()} + return ModelPlus(model=model, paths=[path], format='safetensors', vocab=None) + + +def must_read(fp: IO[bytes], length: int) -> bytes: + ret = fp.read(length) + if len(ret) < length: + raise Exception("unexpectedly reached end of file") + return ret + + +def lazy_load_ggml_file(fp: io.BufferedReader, path: Path) -> ModelPlus: + magic = must_read(fp, 4)[::-1] + if magic in (b'ggmf', b'ggjt'): + version, = struct.unpack("i", must_read(fp, 4)) + assert version == 1 + else: + assert magic == b'ggml' + version = None + n_vocab, n_embd, n_mult, n_head, n_layer, rot, file_type = struct.unpack('<7i', must_read(fp, 28)) + + tokens: List[Tuple[bytes, float]] = [] + for i in range(n_vocab): + if i == 32000: + # HACK: GPT4All messed with the format without changing the magic + # number. Specifically, they changed the vocab section to contain + # `n_vocab - 1` tokens instead of `n_vocab` (i.e. omitting the + # extra pad token). Try to detect if we're reading a file like + # this. + orig_pos = fp.tell() + fp.seek(20, io.SEEK_CUR) + is_gpt4all = fp.read(21) == b'tok_embeddings.weight' + fp.seek(orig_pos) + if is_gpt4all: + break + + length, = struct.unpack("i", must_read(fp, 4)) + text = must_read(fp, length) + if magic != b'ggml': + score, = struct.unpack("f", must_read(fp, 4)) + tokens.append((text, score)) + vocab = GGMLVocab(tokens) if magic != b'ggml' else None + + model: LazyModel = {} + # Use mmap for the actual data to avoid race conditions with the file offset. + off = fp.raw.tell() + mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ)) + fp.raw.seek(off) # needed on Windows + + def read_tensor() -> None: # this is a function so that variables captured in `load` don't change + shape_len, name_len, ftype = struct.unpack("iii", must_read(fp, 12)) + assert 0 <= shape_len <= 3 + shape: List[int] = list(struct.unpack(f"{shape_len}i", must_read(fp, 4 * shape_len))) + shape = shape[::-1] + name = must_read(fp, name_len).decode('utf-8') + data_type = FTYPE_TO_DATA_TYPE[ftype] + + if magic == b'ggjt': + fp.seek((fp.tell() + 31) & -32) + + if data_type == DT_Q4_1: + # See GPTQForLLaMaQuantizedTensor.ggml_ndarray() + size = 24 * (shape[1] // 32) * shape[0] + elif data_type == DT_Q4_0: + size = 20 * (shape[1] // 32) * shape[0] + else: + numpy_dtype = DATA_TYPE_TO_NUMPY[data_type] + elm_count = math.prod(shape) + size = elm_count * numpy_dtype.itemsize + offset = fp.tell() + buf = mapped[offset:offset+size] + fp.seek(size, io.SEEK_CUR) + + def load() -> Tensor: + if isinstance(data_type, QuantizedDataType): + ndarray = np.frombuffer(buf, dtype=np.uint32) + return GGMLQuantizedTensor(ndarray, shape, data_type) + else: + return UnquantizedTensor(np.frombuffer(buf, dtype=numpy_dtype).reshape(shape)) + description = f'ggml offset={offset} type={data_type} path={path}' + model[name] = LazyTensor(load, shape, data_type, description) + + while fp.read(1) != b'': + fp.seek(-1, io.SEEK_CUR) + read_tensor() + + return ModelPlus(model=model, paths=[path], format='ggml', vocab=vocab) + + +@functools.lru_cache(maxsize=None) +def lazy_load_file(path: Path) -> ModelPlus: + fp = open(path, 'rb') + first8 = fp.read(8) + fp.seek(0) + if first8[:2] == b'PK': + # A zip file, i.e. PyTorch format + return lazy_load_torch_file(fp, path) + elif first8[2:4] == b'gg': + # GGML format + return lazy_load_ggml_file(fp, path) + elif struct.unpack(' Iterable[Out]: + '''Parallel map, but with backpressure. If the caller doesn't call `next` + fast enough, this will stop calling `func` at some point rather than + letting results pile up in memory. Specifically, there is a max of one + output value buffered per thread.''' + with concurrent.futures.ThreadPoolExecutor() as executor: + futures: List[concurrent.futures.Future[Out]] = [] + items_rev = list(iterable)[::-1] + for i in range(min(concurrency, len(items_rev))): + futures.append(executor.submit(func, items_rev.pop())) + while futures: + result = futures.pop(0).result() + if items_rev: + futures.append(executor.submit(func, items_rev.pop())) + yield result + + +def check_vocab_size(params: Params, vocab: Vocab) -> None: + if params.n_vocab != vocab.vocab_size: + # GGMLVocab comes from the same file as the model so shouldn't mismatch: + assert isinstance(vocab, SentencePieceVocab) + if params.n_vocab == vocab.vocab_size_base: + print("Ignoring added_tokens.json since model matches vocab size without it.") + vocab.added_tokens_list = [] + vocab.vocab_size = vocab.vocab_size_base + return + msg = f"Vocab size mismatch (model has {params.n_vocab}, but {vocab.fname_tokenizer}" + if vocab.fname_added_tokens is not None: + msg += f" combined with {vocab.fname_added_tokens}" + msg += f" has {vocab.vocab_size})." + if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20 and vocab.fname_added_tokens is None: + msg += f" Most likely you are missing added_tokens.json (should be in {vocab.fname_tokenizer.parent})." + raise Exception(msg) + + +class OutputFile: + def __init__(self, fname_out: Path) -> None: + self.fout = open(fname_out, "wb") + + def write_file_header(self, params: Params) -> None: + self.fout.write(b"ggjt"[::-1]) # magic + values = [ + 1, # file version + params.n_vocab, + params.n_embd, + params.n_mult, + params.n_head, + params.n_layer, + params.n_embd // params.n_head, # rot (obsolete) + params.file_type.value, + ] + self.fout.write(struct.pack("i" * len(values), *values)) + + def write_tensor_header(self, name: str, shape: Sequence[int], data_type: DataType) -> None: + sname = name.encode('utf-8') + self.fout.write(struct.pack("iii", len(shape), len(sname), DATA_TYPE_TO_FTYPE[data_type])) + self.fout.write(struct.pack("i" * len(shape), *shape[::-1])) + self.fout.write(sname) + self.fout.seek((self.fout.tell() + 31) & -32) + + def write_vocab(self, vocab: Vocab) -> None: + for text, score in vocab.all_tokens(): + self.fout.write(struct.pack("i", len(text))) + self.fout.write(text) + self.fout.write(struct.pack("f", score)) + + @staticmethod + def write_vocab_only(fname_out: Path, vocab: Vocab) -> None: + of = OutputFile(fname_out) + params = Params(n_vocab=vocab.vocab_size, n_embd=0, n_mult=0, + n_head=1, n_layer=0, file_type=GGMLFileType.AllF32) + of = OutputFile(fname_out) + of.write_file_header(params) + of.write_vocab(vocab) + of.fout.close() + + @staticmethod + def write_all(fname_out: Path, params: Params, model: LazyModel, vocab: Vocab) -> None: + check_vocab_size(params, vocab) + of = OutputFile(fname_out) + of.write_file_header(params) + print("Writing vocab...") + of.write_vocab(vocab) + + def do_item(item: Tuple[str, LazyTensor]) -> NDArray: + name, lazy_tensor = item + return lazy_tensor.load().to_ggml().ndarray + + ndarrays = bounded_parallel_map(do_item, model.items(), concurrency=8) + for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)): + size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape) + padi = len(str(len(model))) + print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type}") + of.write_tensor_header(name, lazy_tensor.shape, lazy_tensor.data_type) + ndarray.tofile(of.fout) + of.fout.close() + + +def pick_output_type(model: LazyModel, output_type_str: Optional[str]) -> GGMLFileType: + wq_type = model["layers.0.attention.wq.weight"].data_type + if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32): + return GGMLFileType.AllF32 + if output_type_str == "f16" or (output_type_str is None and wq_type == DT_F16): + return GGMLFileType.MostlyF16 + if output_type_str == "q4_1" or (output_type_str is None and isinstance(wq_type, QuantizedDataType) and + wq_type.have_addends): + if isinstance(model["output.weight"].data_type, QuantizedDataType): + return GGMLFileType.MostlyQ4_1 + else: + return GGMLFileType.PerLayerIsQ4_1 + if output_type_str == "q4_0" or (output_type_str is None and isinstance(wq_type, QuantizedDataType)): + return GGMLFileType.MostlyQ4_0 + name_to_type = {name: lazy_tensor.data_type for (name, lazy_tensor) in model.items()} + raise Exception(f"Unexpected combination of types: {name_to_type}") + + +def do_necessary_conversions(model: LazyModel) -> LazyModel: + model = handle_quantization(model) + + if "lm_head.weight" in model: + model = convert_transformers_to_orig(model) + model = filter_and_sort_tensors(model) + + return model + + +def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel: + return {name: tensor.astype(output_type.type_for_tensor(name, tensor)) + for (name, tensor) in model.items()} + + +def nth_multifile_path(path: Path, n: int) -> Optional[Path]: + '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return + the nth path in the model. + ''' + # Support the following patterns: + patterns: List[Tuple[str, str]] = [ + # - x.00.pth, x.01.pth, etc. + (r'\.[0-9]{2}\.pth$', f'.{n:02}.pth'), + # - x-00001-of-00002.bin, x-00002-of-00002.bin, etc. + (r'-[0-9]{5}-of-(.*)$', fr'-{n:05}-of-\1'), + # x.bin, x.bin.1, etc. + (r'(\.[0-9]+)?$', r'\1' if n == 0 else fr'\1.{n}') + ] + for regex, replacement in patterns: + if re.search(regex, path.name): + new_path = path.with_name(re.sub(regex, replacement, path.name)) + if new_path.exists(): + return new_path + return None + + +def find_multifile_paths(path: Path) -> List[Path]: + '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return + the whole list of paths in the model. + ''' + ret: List[Path] = [] + for i in itertools.count(): + nth_path = nth_multifile_path(path, i) + if nth_path is None: + break + ret.append(nth_path) + if not ret: + # No matches. This should only happen if the file was named, e.g., + # foo.0, and there was no file named foo. Oh well, try to process it + # as a single file. + return [path] + return ret + + +def load_some_model(path: Path) -> ModelPlus: + '''Load a model of any supported format.''' + # Be extra-friendly and accept either a file or a directory: + if path.is_dir(): + globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt"] + files = [file for glob in globs for file in path.glob(glob)] + if not files: + # Try GGML too, but with lower priority, since if both a non-GGML + # model and a GGML model exist in the same directory, we assume the + # latter was converted from the former. + files = list(path.glob("ggml-model*.bin*")) + if not files: + raise Exception(f"Can't find model in directory {path}") + if len(files) > 1: + raise Exception(f"Found multiple models in {path}, not sure which to pick: {files}") + path = files[0] + + paths = find_multifile_paths(path) + models_plus: List[ModelPlus] = [] + for path in paths: + print(f"Loading model file {path}") + models_plus.append(lazy_load_file(path)) + + model_plus = merge_multifile_models(models_plus) + return model_plus + + +def filter_and_sort_tensors(model: LazyModel) -> LazyModel: + return {name: model[name] for name in TENSORS_LIST if name in model} + + +def load_vocab(path: Path) -> SentencePieceVocab: + # Be extra-friendly and accept either a file or a directory. Also, if it's + # a directory, it might be the model directory, and tokenizer.model might + # be in the parent of that. + if path.is_dir(): + path2 = path / "tokenizer.model" + # Use `.parent` instead of /.. to handle the symlink case better. + path3 = path.parent / "tokenizer.model" + if path2.exists(): + path = path2 + elif path3.exists(): + path = path3 + else: + raise FileNotFoundError(f"Could not find tokenizer.model in {path} or its parent; if it's in another directory, pass the directory as --vocab-dir") + added_tokens_path = path.parent / "added_tokens.json" + print(f"Loading vocab file {path}") + return SentencePieceVocab(path, added_tokens_path if added_tokens_path.exists() else None) + + +def default_outfile(model_paths: List[Path], params: Params) -> Path: + namestr = { + GGMLFileType.AllF32: "f32", + GGMLFileType.MostlyF16: "f16", + GGMLFileType.MostlyQ4_0: "q4_0", + GGMLFileType.MostlyQ4_1: "q4_1", + GGMLFileType.PerLayerIsQ4_1: "q4_1", + }[params.file_type] + ret = model_paths[0].parent / f"ggml-model-{namestr}.bin" + if ret in model_paths: + sys.stderr.write(f"Error: Default output path ({ret}) would overwrite the input. Please explicitly specify a path using --outfile.\n") + 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}") + for name, lazy_tensor in model_plus.model.items(): + print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}") + + +def main(args_in: Optional[List[str]] = None) -> None: + parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file") + parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model") + parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file") + parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") + parser.add_argument("--outtype", choices=["f32", "f16", "q4_1", "q4_0"], help="output format (default: based on input)") + parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file") + parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") + parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)") + args = parser.parse_args(args_in) + + vocab: Vocab + if args.dump_single: + model_plus = lazy_load_file(args.model) + do_dump_model(model_plus) + elif args.vocab_only: + vocab = load_vocab(args.vocab_dir or args.model) + assert args.outfile, "need --outfile if using --vocab-only" + outfile = args.outfile + OutputFile.write_vocab_only(outfile, vocab) + print(f"Wrote {outfile}") + else: + model_plus = load_some_model(args.model) + if args.dump: + do_dump_model(model_plus) + return + if model_plus.vocab is not None and args.vocab_dir is None: + vocab = model_plus.vocab + else: + vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent + vocab = load_vocab(vocab_dir) + model = model_plus.model + model = do_necessary_conversions(model) + output_type = pick_output_type(model, args.outtype) + model = convert_to_output_type(model, output_type) + params = Params.guessed(model, output_type) + outfile = args.outfile or default_outfile(model_plus.paths, params) + OutputFile.write_all(outfile, params, model, vocab) + print(f"Wrote {outfile}") + + +if __name__ == '__main__': + main() diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index ce3a34710..67a7cea54 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -31,6 +31,7 @@ if (EMSCRIPTEN) else() add_subdirectory(main) add_subdirectory(quantize) + add_subdirectory(quantize-stats) add_subdirectory(perplexity) add_subdirectory(embedding) endif() diff --git a/examples/Miku.sh b/examples/Miku.sh new file mode 100755 index 000000000..c4cbf80f2 --- /dev/null +++ b/examples/Miku.sh @@ -0,0 +1,49 @@ +#!/bin/bash +set -e + +AI_NAME="${AI_NAME:-Miku}" +MODEL="${MODEL:-./models/gpt4all-7B/gpt4all-lora-unfiltered-quantized.bin}" +USER_NAME="${USER_NAME:-Anon}" + +# Uncomment and adjust to the number of CPU cores you want to use. +#N_THREAD="${N_THREAD:-4}" +N_PREDICTS="${N_PREDICTS:-4096}" + +GEN_OPTIONS=(--batch_size 1024 +--ctx_size 2048 +--keep -1 +--repeat_last_n 256 +--repeat_penalty 1.17647 +--temp 0.7 +--top_k 40 +--top_p 0.5) + +if [ -n "$N_THREAD" ]; then + GEN_OPTIONS+=(--threads "$N_THREAD") +fi + +./main "${GEN_OPTIONS[@]}" \ + --model "$MODEL" \ + --n_predict "$N_PREDICTS" \ + --color --interactive \ + --reverse-prompt "${USER_NAME}:" \ + --prompt " +This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the users computer. +${AI_NAME} can think for herself without the user seeing her thoughts by adding a /think prefix to her output. She uses this to reason about the world and to think about what she should say next. +${AI_NAME} is always coherent and makes sense, but if she isn't sure if what she is saying is correct she will ask the user for help. +${AI_NAME} is a very helpful AI and will help the user with anything they need, she is also very friendly and will try to make the user feel better if they are sad. +${AI_NAME} is also very curious and will ask the user a lot of questions about themselves and their life, she will also try to make the user like her. +The conversation is only between ${USER_NAME} and ${AI_NAME} +The conversation is only through text, so ${AI_NAME} can't see ${USER_NAME}'s face or hear his voice. +${AI_NAME} can only communicate through text, so she can't send images or videos. + + +${USER_NAME}: Hello! +${AI_NAME}: /think I wonder what I should say to ${USER_NAME}? This is the first time we talk so it's important that I make a good first impression! +${AI_NAME}: Hi! I am ${AI_NAME}, your new AI friend, assistant(or whatever you like!), it's so nice to meet you! ^_^ +${AI_NAME}: I might be an AI, but I also have feelings, so please be nice to me! :) +${USER_NAME}: I'll make sure to be nice to you! I'm so happy to have you as my assistant! +${AI_NAME}: /think It sounds like ${USER_NAME} is happy to have me as their assistant! I'm so happy too! ^_^ Glad that whole emotion thing didn't scare him off! +${AI_NAME}: /think I wonder what ${USER_NAME} likes to do in his free time? I should ask him about that! +${AI_NAME}: What do you like to do in your free time? ^_^ +${USER_NAME}:" "$@" diff --git a/examples/alpaca.sh b/examples/alpaca.sh index 4c9aa5077..8d6261730 100755 --- a/examples/alpaca.sh +++ b/examples/alpaca.sh @@ -7,4 +7,4 @@ cd `dirname $0` cd .. -./main -m ./models/ggml-alpaca-7b-q4.bin --color -f ./prompts/alpaca.txt -ins -b 256 --top_k 10000 --temp 0.2 --repeat_penalty 1 -t 7 +./main -m ./models/ggml-alpaca-7b-q4.bin --color -f ./prompts/alpaca.txt --ctx_size 2048 -n -1 -ins -b 256 --top_k 10000 --temp 0.2 --repeat_penalty 1 -t 7 diff --git a/examples/benchmark/benchmark-q4_0-matmult.c b/examples/benchmark/benchmark-q4_0-matmult.c new file mode 100644 index 000000000..84b06766c --- /dev/null +++ b/examples/benchmark/benchmark-q4_0-matmult.c @@ -0,0 +1,270 @@ +/* + License: MIT License + + Changelog: + - 2023-03-31 Initial version by Sebastian Apel (https://github.com/SebastianApel) + +*/ + +#include +#include "ggml.h" +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +float tensor_sum_elements(struct ggml_tensor * tensor) { + float sum = 0; + if (tensor->type==GGML_TYPE_F32) { + for (int j = 0; j < tensor->ne[1]; j++) { + for (int k = 0; k < tensor->ne[0]; k++) { + sum += ((float *) tensor->data)[j*tensor->ne[0]+k]; + } + } + } + return sum; +} + + +/* + These are mapping to unknown + GGML_TYPE_I8, + GGML_TYPE_I16, + GGML_TYPE_I32, + GGML_TYPE_COUNT, +*/ + +#define TENSOR_TYPE_AS_STR(TYPE) TYPE == GGML_TYPE_F32 ? "FP32" : TYPE == GGML_TYPE_F16 ? "FP16" : TYPE == GGML_TYPE_Q4_0 ? "Q4_0" : TYPE == GGML_TYPE_Q4_1 ? "Q4_1" : "UNKNOWN" + +#define TENSOR_DUMP(TENSOR) printf("%15s: type = %i (%5s) ne = %5d x %5d x %5d, nb = (%5li, %5li, %5li) - ", #TENSOR, \ + TENSOR->type,TENSOR_TYPE_AS_STR(TENSOR->type),\ + TENSOR->ne[0], TENSOR->ne[1], TENSOR->ne[2], TENSOR->nb[0], TENSOR->nb[1], TENSOR->nb[2]); \ + { float sum = tensor_sum_elements(TENSOR); printf("Sum of tensor %s is %6.2f\n",#TENSOR, sum); } + +struct benchmark_params_struct { + int32_t n_threads = 1; + int32_t n_iterations = 10; +}; + +void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) { + fprintf(stderr, "usage: %s [options]\n", argv[0]); + fprintf(stderr, "\n"); + fprintf(stderr, "options:\n"); + fprintf(stderr, " -h, --help show this help message and exit\n"); + fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads); + fprintf(stderr, " -i N, --iter N number of iterations to use during computation (default: %d)\n", params.n_iterations); + fprintf(stderr, "\n"); +} + +int main(int argc, char ** argv) { + + + struct benchmark_params_struct benchmark_params; + + bool invalid_param = false; + std::string arg; + for (int i = 1; i < argc; i++) { + arg = argv[i]; + + if (arg == "-t" || arg == "--threads") { + if (++i >= argc) { + invalid_param = true; + break; + } + benchmark_params.n_threads = std::stoi(argv[i]); + } else if (arg == "-i" || arg == "--iter") { + if (++i >= argc) { + invalid_param = true; + break; + } + benchmark_params.n_iterations = std::stoi(argv[i]); + } else if (arg == "-h" || arg == "--help") { + print_usage(argc, argv, benchmark_params); + exit(0); + } + if (invalid_param) { + fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str()); + print_usage(argc, argv, benchmark_params); + exit(1); + } + } + + + // create the ggml context + printf("Starting Test\n"); + + + + struct ggml_context * ctx; + //const int sizex = 4096; + //const int sizey = 11008; + +#undef VERBOSE_DEBUGGING +#ifndef VERBOSE_DEBUGGING + const int sizey = 4096; + const int sizex = 11008; + const int sizez = 128; +#else + /* Working - let's increase size */ + const int sizey = 1; + const int sizex = (8*32); + const int sizez = 1; + + /*const int sizey = 1; + const int sizex = 3*(8*32); + const int sizez = 1;*/ +#endif + + //printf("Memsize required = %i\n", sizex*sizex); + ggml_type wtype = GGML_TYPE_F32; + + size_t ctx_size = 0; + ctx_size += sizex*sizey*ggml_type_sizef(wtype); + ctx_size += sizex*sizey*ggml_type_sizef(wtype); + ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); + ctx_size += sizex*sizeof(float); + ctx_size += 1024*1024*100; + + printf("Allocating Memory of size %li byes, %li MB\n",ctx_size, (ctx_size/1024/1024)); + + struct ggml_init_params params = { + /*.mem_size =*/ ctx_size, + /*.mem_buffer =*/ NULL, + /* no_alloc =*/ 0 + }; + + ctx = ggml_init(params); + if (!ctx) { + fprintf(stderr, "%s: ggml_init() failed\n", __func__); + return false; + } + + + printf("Creating new tensors\n"); + // printf("Creating new tensor m1\n"); + struct ggml_tensor * m11 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey); + ggml_set_f32(m11, 1.0f); + + // printf("Creating new tensor m1\n"); + struct ggml_tensor * m12 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey); + ggml_set_f32(m12, 1.5f); + + // printf("Creating new tensor m2\n"); + struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez); + ggml_set_f32(m2, 2.0f); + + printf("\n------ Test 1 - Matrix Mult via F32 code ------------------------------------------------------------------------------\n"); + // printf("Creating new tensor m11xm2\n"); + struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2); + + // printf("Creating compute graph\n"); + struct ggml_cgraph gf = ggml_build_forward(m11xm2); + + gf.n_threads=benchmark_params.n_threads; + printf("cgraph->n_threads=%i\n",gf.n_threads); + + TENSOR_DUMP(m11); + TENSOR_DUMP(m2); + + ggml_graph_compute(ctx, &gf); + + TENSOR_DUMP(gf.nodes[0]); + + printf("\n------ Test 2 - Matrix Mult via Q4_0 code ------------------------------------------------------------------------------\n"); + + int32_t nelements = sizex*sizey; + int32_t ne[2] = { sizex, sizey }; + + std::vector hist_cur(1 << 4, 0); + + // Set up a the benchmark matrices + // printf("Creating new tensor q11 & Running quantize\n"); + struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey); + ggml_quantize_q4_0((const float *) m11->data, q11->data, nelements, ne[0], hist_cur.data()); + + // Set up a the compute graph + // printf("Creating new tensor q31\n"); + struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2); + + // printf("Creating compute graph\n"); + struct ggml_cgraph gf31 = ggml_build_forward(q31); + gf31.n_threads=benchmark_params.n_threads; + + // Set up a second graph computation to make sure we override the CPU cache lines + // printf("Creating new tensor q12 & Running quantize\n"); + struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, GGML_TYPE_Q4_0, sizex, sizey); + ggml_quantize_q4_0((const float *) m12->data, q12->data, nelements, ne[0], hist_cur.data()); + + // printf("Creating new tensor q32\n"); + struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2); + + //printf("Creating compute graph\n"); + struct ggml_cgraph gf32 = ggml_build_forward(q32); + gf32.n_threads=benchmark_params.n_threads; + printf("cgraph->n_threads=%i\n",gf31.n_threads); + + const int dimx = sizex; + const int dimy = sizey; + const int dimz = sizez; + long long int flops_per_dot_product = dimy + dimy; + long long int flops_per_matrix = flops_per_dot_product * dimx * dimz; ; + printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - aboout %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000); + + + // Let's use the F32 result from above as a reference for the q4_0 multiplication + float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]); + + + printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; FLOPS_per_u_Second\n"); + printf("==============================================================================================\n"); + + for (int i=0;i allowed_delta) { + printf("\nABORT - ERROR in Matrix Multiplication result - expected %6.2f, got %6.2f (delta %6.2f > allowed_delta %6.2f)\n", + sum_of_F32_reference, + sum_of_Q4_result, + delta, + allowed_delta + ); + exit(0); + } + + // Running a different graph computation to make sure we override the CPU cache lines + ggml_graph_compute(ctx, &gf32); + + } + +} diff --git a/examples/common.cpp b/examples/common.cpp index 5400f6b01..a0b6f10ad 100644 --- a/examples/common.cpp +++ b/examples/common.cpp @@ -1,7 +1,5 @@ #include "common.h" -#include "ggml.h" - #include #include #include @@ -9,19 +7,20 @@ #include #include -#if defined(_MSC_VER) || defined(__MINGW32__) -#include // using malloc.h with MSC/MINGW -#elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__) -#include -#endif - #if defined (_WIN32) +#include +#include #pragma comment(lib,"kernel32.lib") extern "C" __declspec(dllimport) void* __stdcall GetStdHandle(unsigned long nStdHandle); extern "C" __declspec(dllimport) int __stdcall GetConsoleMode(void* hConsoleHandle, unsigned long* lpMode); extern "C" __declspec(dllimport) int __stdcall SetConsoleMode(void* hConsoleHandle, unsigned long dwMode); extern "C" __declspec(dllimport) int __stdcall SetConsoleCP(unsigned int wCodePageID); extern "C" __declspec(dllimport) int __stdcall SetConsoleOutputCP(unsigned int wCodePageID); +extern "C" __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int CodePage, unsigned long dwFlags, + const wchar_t * lpWideCharStr, int cchWideChar, + char * lpMultiByteStr, int cbMultiByte, + const char * lpDefaultChar, bool * lpUsedDefaultChar); +#define CP_UTF8 65001 #endif bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { @@ -140,6 +139,19 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { break; } params.model = argv[i]; + } else if (arg == "--lora") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.lora_adapter = argv[i]; + params.use_mmap = false; + } else if (arg == "--lora-base") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.lora_base = argv[i]; } else if (arg == "-i" || arg == "--interactive") { params.interactive = true; } else if (arg == "--embedding") { @@ -154,6 +166,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { params.use_color = true; } else if (arg == "--mlock") { params.use_mlock = true; + } else if (arg == "--no-mmap") { + params.use_mmap = false; } else if (arg == "--mtest") { params.mem_test = true; } else if (arg == "--verbose-prompt") { @@ -233,11 +247,16 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { fprintf(stderr, " -b N, --batch_size N batch size for prompt processing (default: %d)\n", params.n_batch); fprintf(stderr, " --perplexity compute perplexity over the prompt\n"); fprintf(stderr, " --keep number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep); - if (ggml_mlock_supported()) { + if (llama_mlock_supported()) { fprintf(stderr, " --mlock force system to keep model in RAM rather than swapping or compressing\n"); } + if (llama_mmap_supported()) { + fprintf(stderr, " --no-mmap do not memory-map model (slower load but may reduce pageouts if not using mlock)\n"); + } fprintf(stderr, " --mtest compute maximum memory usage\n"); fprintf(stderr, " --verbose-prompt print prompt before generation\n"); + fprintf(stderr, " --lora FNAME apply LoRA adapter (implies --no-mmap)\n"); + fprintf(stderr, " --lora-base FNAME optional model to use as a base for the layers modified by the LoRA adapter\n"); fprintf(stderr, " -m FNAME, --model FNAME\n"); fprintf(stderr, " model path (default: %s)\n", params.model.c_str()); fprintf(stderr, "\n"); @@ -307,12 +326,20 @@ void win32_console_init(bool enable_color) { SetConsoleMode(hConOut, dwMode | 0x4); // ENABLE_VIRTUAL_TERMINAL_PROCESSING (0x4) } // Set console output codepage to UTF8 - SetConsoleOutputCP(65001); // CP_UTF8 + SetConsoleOutputCP(CP_UTF8); } void* hConIn = GetStdHandle((unsigned long)-10); // STD_INPUT_HANDLE (-10) if (hConIn && hConIn != (void*)-1 && GetConsoleMode(hConIn, &dwMode)) { - // Set console input codepage to UTF8 - SetConsoleCP(65001); // CP_UTF8 + // Set console input codepage to UTF16 + _setmode(_fileno(stdin), _O_WTEXT); } } + +// Convert a wide Unicode string to an UTF8 string +void win32_utf8_encode(const std::wstring & wstr, std::string & str) { + int size_needed = WideCharToMultiByte(CP_UTF8, 0, &wstr[0], (int)wstr.size(), NULL, 0, NULL, NULL); + std::string strTo(size_needed, 0); + WideCharToMultiByte(CP_UTF8, 0, &wstr[0], (int)wstr.size(), &strTo[0], size_needed, NULL, NULL); + str = strTo; +} #endif diff --git a/examples/common.h b/examples/common.h index 1505aa927..cbbc2dfab 100644 --- a/examples/common.h +++ b/examples/common.h @@ -31,11 +31,12 @@ struct gpt_params { std::string model = "models/lamma-7B/ggml-model.bin"; // model path std::string prompt = ""; - std::string input_prefix = ""; // string to prefix user inputs with - - + std::string input_prefix = ""; // string to prefix user inputs with std::vector antiprompt; // string upon seeing which more user input is prompted + std::string lora_adapter = ""; // lora adapter path + std::string lora_base = ""; // base model path for the lora adapter + bool memory_f16 = true; // use f16 instead of f32 for memory kv bool random_prompt = false; // do not randomize prompt if none provided bool use_color = false; // use color to distinguish generations and inputs @@ -47,6 +48,7 @@ struct gpt_params { bool instruct = false; // instruction mode (used for Alpaca models) bool ignore_eos = false; // do not stop generating after eos bool perplexity = false; // compute perplexity over the prompt + bool use_mmap = true; // use mmap for faster loads bool use_mlock = false; // use mlock to keep model in memory bool mem_test = false; // compute maximum memory usage bool verbose_prompt = false; // print prompt tokens before generation @@ -92,4 +94,5 @@ void set_console_color(console_state & con_st, console_color_t color); #if defined (_WIN32) void win32_console_init(bool enable_color); +void win32_utf8_encode(const std::wstring & wstr, std::string & str); #endif diff --git a/examples/embedding/README.md b/examples/embedding/README.md index 21d8be65f..fe8f5dcc6 100644 --- a/examples/embedding/README.md +++ b/examples/embedding/README.md @@ -1,3 +1,3 @@ -# embedding - -TODO +# embedding + +TODO diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index d397f35fd..e10de619c 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -1,6 +1,8 @@ #include "common.h" #include "llama.h" +#include + int main(int argc, char ** argv) { gpt_params params; params.model = "models/llama-7B/ggml-model.bin"; @@ -38,6 +40,7 @@ int main(int argc, char ** argv) { lparams.seed = params.seed; lparams.f16_kv = params.memory_f16; lparams.logits_all = params.perplexity; + lparams.use_mmap = params.use_mmap; lparams.use_mlock = params.use_mlock; lparams.embedding = params.embedding; diff --git a/examples/gpt4all.sh b/examples/gpt4all.sh index d974f95a9..5fd739e55 100755 --- a/examples/gpt4all.sh +++ b/examples/gpt4all.sh @@ -10,6 +10,6 @@ cd .. ./main --color --instruct --threads 4 \ --model ./models/gpt4all-7B/gpt4all-lora-quantized.bin \ --file ./prompts/alpaca.txt \ - --batch_size 8 --ctx_size 2048 \ + --batch_size 8 --ctx_size 2048 -n -1 \ --repeat_last_n 64 --repeat_penalty 1.3 \ --n_predict 128 --temp 0.1 --top_k 40 --top_p 0.95 diff --git a/examples/main/README.md b/examples/main/README.md index 4701aa558..f09e7ba97 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -1,3 +1,3 @@ -# main - -TODO +# main + +TODO diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 453450a41..b7b3c4196 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -1,3 +1,8 @@ +// Defines sigaction on msys: +#ifndef _GNU_SOURCE +#define _GNU_SOURCE +#endif + #include "common.h" #include "llama.h" @@ -6,6 +11,7 @@ #include #include #include +#include #include #include #include @@ -97,6 +103,7 @@ int main(int argc, char ** argv) { lparams.n_parts = params.n_parts; lparams.seed = params.seed; lparams.f16_kv = params.memory_f16; + lparams.use_mmap = params.use_mmap; lparams.use_mlock = params.use_mlock; ctx = llama_init_from_file(params.model.c_str(), lparams); @@ -107,6 +114,17 @@ int main(int argc, char ** argv) { } } + if (!params.lora_adapter.empty()) { + int err = llama_apply_lora_from_file(ctx, + params.lora_adapter.c_str(), + params.lora_base.empty() ? NULL : params.lora_base.c_str(), + params.n_threads); + if (err != 0) { + fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__); + return 1; + } + } + // print system information { fprintf(stderr, "\n"); @@ -162,7 +180,7 @@ int main(int argc, char ** argv) { } // enable interactive mode if reverse prompt or interactive start is specified - if (params.antiprompt.size() != 0 || params.interactive_start) { + if (params.antiprompt.size() != 0 || params.interactive_start) { params.interactive = true; } @@ -386,10 +404,19 @@ int main(int argc, char ** argv) { std::string line; bool another_line = true; do { +#if defined(_WIN32) + std::wstring wline; + if (!std::getline(std::wcin, wline)) { + // input stream is bad or EOF received + return 0; + } + win32_utf8_encode(wline, line); +#else if (!std::getline(std::cin, line)) { // input stream is bad or EOF received return 0; } +#endif if (line.empty() || line.back() != '\\') { another_line = false; } else { @@ -431,7 +458,7 @@ int main(int argc, char ** argv) { } // end of text token - if (embd.back() == llama_token_eos()) { + if (!embd.empty() && embd.back() == llama_token_eos()) { if (params.instruct) { is_interacting = true; } else { diff --git a/examples/perplexity/README.md b/examples/perplexity/README.md index a932275c2..eacfb17c6 100644 --- a/examples/perplexity/README.md +++ b/examples/perplexity/README.md @@ -1,3 +1,3 @@ -# perplexity - -TODO +# perplexity + +TODO diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 07ed0a829..80792ea0d 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -2,6 +2,7 @@ #include "llama.h" #include +#include std::vector softmax(const std::vector& logits) { std::vector probs(logits.size()); @@ -27,20 +28,27 @@ void perplexity(llama_context * ctx, const gpt_params & params) { int count = 0; int seq_count = tokens.size() / params.n_ctx; + int n_vocab = llama_n_vocab(ctx); double nll = 0.0; - - fprintf(stderr, "%s : calculating perplexity over %d chunks\n", __func__, seq_count); + fprintf(stderr, "%s : calculating perplexity over %d chunks, batch_size=%d\n", __func__, seq_count, params.n_batch); for (int i = 0; i < seq_count; ++i) { int start = i * params.n_ctx; - int end = start + params.n_ctx - 1; // TODO: this is not optimal, e.g. it makes the batch 511 instead of 512 - // it is better to always be power of 2 for better performance - std::vector embd(tokens.begin() + start, tokens.begin() + end); + int end = start + params.n_ctx; + + std::vector logits; + int num_batches = (params.n_ctx + params.n_batch - 1) / params.n_batch; auto start_t = std::chrono::high_resolution_clock::now(); - if (llama_eval(ctx, embd.data(), embd.size(), 0, params.n_threads)) { - fprintf(stderr, "%s : failed to eval\n", __func__); - return; + for (int j = 0; j < num_batches; ++j) { + int batch_start = start + j * params.n_batch; + int batch_size = std::min(end - batch_start, params.n_batch); + if (llama_eval(ctx, tokens.data() + batch_start, batch_size, j * params.n_batch, params.n_threads)) { + fprintf(stderr, "%s : failed to eval\n", __func__); + return; + } + auto batch_logits = llama_get_logits(ctx); + logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); } auto end_t = std::chrono::high_resolution_clock::now(); if (i == 0) { @@ -59,15 +67,12 @@ void perplexity(llama_context * ctx, const gpt_params & params) { // Example, we have a context window of 512, we will compute perplexity for each of the // last 256 tokens. Then, we split the input up into context window size chunks to // process the entire prompt. - - auto logits = llama_get_logits(ctx); - for (int j = params.n_ctx / 2; j < params.n_ctx - 1; ++j) { + for (int j = std::min(512, params.n_ctx / 2); j < params.n_ctx - 1; ++j) { // Calculate probability of next token, given the previous ones. - int n_vocab = llama_n_vocab(ctx); std::vector tok_logits( - logits + j * n_vocab, - logits + (j + 1) * n_vocab); - const float prob = softmax(tok_logits)[tokens[start + j + 1]]; + logits.begin() + j * n_vocab, + logits.begin() + (j + 1) * n_vocab); + float prob = softmax(tok_logits)[tokens[start + j + 1]]; nll += -std::log(prob); ++count; } @@ -82,11 +87,13 @@ int main(int argc, char ** argv) { gpt_params params; params.model = "models/llama-7B/ggml-model.bin"; + params.n_batch = 512; if (gpt_params_parse(argc, argv, params) == false) { return 1; } params.perplexity = true; + params.n_batch = std::min(params.n_batch, params.n_ctx); if (params.n_ctx > 2048) { fprintf(stderr, "%s: warning: model does not support context sizes greater than 2048 tokens (%d specified);" @@ -115,6 +122,7 @@ int main(int argc, char ** argv) { lparams.seed = params.seed; lparams.f16_kv = params.memory_f16; lparams.logits_all = params.perplexity; + lparams.use_mmap = params.use_mmap; lparams.use_mlock = params.use_mlock; lparams.embedding = params.embedding; @@ -126,6 +134,17 @@ int main(int argc, char ** argv) { } } + if (!params.lora_adapter.empty()) { + int err = llama_apply_lora_from_file(ctx, + params.lora_adapter.c_str(), + params.lora_base.empty() ? NULL : params.lora_base.c_str(), + params.n_threads); + if (err != 0) { + fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__); + return 1; + } + } + // print system information { fprintf(stderr, "\n"); diff --git a/examples/quantize-stats/CMakeLists.txt b/examples/quantize-stats/CMakeLists.txt new file mode 100644 index 000000000..7bebc11a1 --- /dev/null +++ b/examples/quantize-stats/CMakeLists.txt @@ -0,0 +1,4 @@ +set(TARGET quantize-stats) +add_executable(${TARGET} quantize-stats.cpp) +target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp new file mode 100644 index 000000000..4e6c2c831 --- /dev/null +++ b/examples/quantize-stats/quantize-stats.cpp @@ -0,0 +1,420 @@ +#include "ggml.h" + +#define LLAMA_API_INTERNAL +#include "llama.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +struct quantize_stats_params { + std::string model = "models/7B/ggml-model-f16.bin"; + bool verbose = false; + bool per_layer_stats = false; + bool print_histogram = false; + bool reference = false; + std::vector include_layers; + std::vector exclude_layers; + std::vector include_types; +}; + +const size_t HISTOGRAM_BUCKETS = 150; +const double HISTOGRAM_RANGE = 0.03; + +struct error_stats { + size_t num_samples; + double total_error; + double max_error; + uint64_t error_histogram[HISTOGRAM_BUCKETS]; +}; + + +void quantize_stats_print_usage(int /*argc*/, char ** argv) { + quantize_stats_params params; + fprintf(stderr, "usage: %s [options]\n", argv[0]); + fprintf(stderr, "\n"); + fprintf(stderr, "options:\n"); + fprintf(stderr, " -h, --help show this help message and exit\n"); + fprintf(stderr, " -m FNAME, --model FNAME\n"); + fprintf(stderr, " model path (default: %s)\n", params.model.c_str()); + fprintf(stderr, " -r, --reference\n"); + fprintf(stderr, " use reference implementation (default: false)\n"); + fprintf(stderr, " -v, --verbose\n"); + fprintf(stderr, " verbose output (default: false)\n"); + fprintf(stderr, " -p, --per-layer-stats\n"); + fprintf(stderr, " print stats per layer (default: false)\n"); + fprintf(stderr, " --histogram\n"); + fprintf(stderr, " print error histogram (default: false)\n"); + fprintf(stderr, " -l LAYER, --include-layer LAYER\n"); + fprintf(stderr, " only test layers matching pattern\n"); + fprintf(stderr, " -L LAYER, --exclude-layer LAYER\n"); + fprintf(stderr, " exclude layers matching pattern\n"); + fprintf(stderr, " -t TYPE, --type TYPE\n"); + fprintf(stderr, " only test given type (q4_0, q4_1)\n"); + fprintf(stderr, "\n"); +} + +// Check if a layer is included/excluded by command line +bool layer_included(const quantize_stats_params params, const std::string & layer) { + for (const auto& excluded : params.exclude_layers) { + if (std::regex_search(layer, std::regex(excluded))) { + return false; + } + } + for (const auto& included : params.include_layers) { + if (std::regex_search(layer, std::regex(included))) { + return true; + } + } + return params.include_layers.empty(); +} + +// Update error statistics given vectors with the before/after result of quantization +void update_error_stats(int64_t nelements, const float * input, const float * output, error_stats & stats) { + for (int64_t i = 0; i < nelements; i++) { + double diff = input[i] - output[i]; + stats.total_error += diff * diff; + stats.max_error = fmax(fabs(diff), stats.max_error); + stats.error_histogram[std::max(std::min((size_t) floor(fabs(diff) / HISTOGRAM_RANGE * HISTOGRAM_BUCKETS), HISTOGRAM_BUCKETS-1), (size_t) 0)]++; + } + stats.num_samples += nelements; +} + +void combine_error_stats(error_stats & into, const error_stats & from) { + into.num_samples += from.num_samples; + into.total_error += from.total_error; + if (from.max_error > into.max_error) into.max_error = from.max_error; + for (size_t i=0; i= sum*quantile) { + return (i+1) * HISTOGRAM_RANGE / HISTOGRAM_BUCKETS; + } + } + return INFINITY; +} + +void print_error_stats(const std::string & name, const error_stats & stats, bool print_histogram) { + double rmse = sqrt(stats.total_error / (double) stats.num_samples); + double median = find_quantile(stats, .5); + double pct95 = find_quantile(stats, .95); + printf("%-50s: rmse %.8f, maxerr %.8f, 95pct<%.4f, median<%.4f\n", name.c_str(), rmse, stats.max_error, pct95, median); + if (print_histogram) { + printf("Error distribution:\n"); + for (size_t i = 0; i < HISTOGRAM_BUCKETS; i++) { + double lower = i * HISTOGRAM_RANGE / HISTOGRAM_BUCKETS; + double upper = (i+1) * HISTOGRAM_RANGE / HISTOGRAM_BUCKETS; + if (i == HISTOGRAM_BUCKETS -1) upper = INFINITY; + printf("[%3.4f, %3.4f): %11" PRIu64 "\n", lower, upper, stats.error_histogram[i]); + } + } +} + +// copied from ggml.h - verify that we can access this as a flat array +static bool tensor_is_contiguous(const struct ggml_tensor * tensor) { + static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); + + return + tensor->nb[0] == ggml_type_size(tensor->type) && + tensor->nb[1] == (tensor->nb[0]*tensor->ne[0])/ggml_blck_size(tensor->type) && + tensor->nb[2] == tensor->nb[1]*tensor->ne[1] && + tensor->nb[3] == tensor->nb[2]*tensor->ne[2]; +} + +void test_roundtrip_on_chunk( + const ggml_tensor * layer, + int64_t offset, + int64_t chunk_size, + const quantize_fns_t & qfns, + bool use_reference, + float * input_scratch, + char * quantized_scratch, + float * output_scratch, + error_stats & stats) { + + if (layer->type == GGML_TYPE_F16) { + for (int i = 0; i < chunk_size; i++) { + input_scratch[i] = ggml_get_f32_1d(layer, i + offset); + } + } else { + input_scratch = ggml_get_data_f32(layer) + offset; + } + + if (use_reference) { + qfns.quantize_row_q_reference(input_scratch, quantized_scratch, chunk_size); + } else { + qfns.quantize_row_q(input_scratch, quantized_scratch, chunk_size); + } + qfns.dequantize_row_q(quantized_scratch, output_scratch, chunk_size); + + update_error_stats(chunk_size, input_scratch, output_scratch, stats); +} + + +// Run quantization function for a single layer and update error stats +void test_roundtrip_on_layer( + std::string & name, + bool print_layer_stats, + const quantize_fns_t & qfns, + bool use_reference, + const ggml_tensor * layer, + std::vector & input_scratch, + std::vector & quantized_scratch, + std::vector & output_scratch, + error_stats & total_error, + int max_thread = 0) { + + assert(tensor_is_contiguous(layer)); + error_stats layer_error {}; + uint64_t nelements = ggml_nelements(layer); + + float* input_scratch_ptr = nullptr; + if (layer->type == GGML_TYPE_F16) { + if (input_scratch.size() < nelements) input_scratch.resize(nelements); + input_scratch_ptr = input_scratch.data(); + } + if (quantized_scratch.size() < 4*nelements) quantized_scratch.resize(4*nelements); + if (output_scratch.size() < nelements) output_scratch.resize(nelements); + + if (max_thread < 1) max_thread = std::thread::hardware_concurrency(); + int chunk_size = 32*512; + int num_chunks = (nelements + chunk_size - 1)/chunk_size; + + if (num_chunks < 2 || max_thread < 2) { + test_roundtrip_on_chunk(layer, 0, nelements, qfns, use_reference, input_scratch_ptr, quantized_scratch.data(), + output_scratch.data(), print_layer_stats ? layer_error : total_error); + } else { + auto & stats = print_layer_stats ? layer_error : total_error; + std::mutex mutex; + uint64_t counter = 0; + auto compute = [&mutex, &counter, &stats, &qfns, nelements, layer, use_reference, input_scratch_ptr, + &quantized_scratch, &output_scratch, chunk_size] () { + error_stats local_stats {}; + while (true) { + std::unique_lock lock(mutex); + uint64_t offset = counter; counter += chunk_size; + if (offset >= nelements) { + combine_error_stats(stats, local_stats); + break; + } + lock.unlock(); + uint64_t chunk = offset + chunk_size < nelements ? chunk_size : nelements - offset; + test_roundtrip_on_chunk(layer, offset, chunk, qfns, use_reference, input_scratch_ptr + offset, + quantized_scratch.data() + 4*offset, output_scratch.data() + offset, local_stats); + } + }; + int nthread = std::min(num_chunks, max_thread); + std::vector workers(nthread-1); + for (auto& w : workers) w = std::thread(compute); + compute(); + for (auto& w : workers) w.join(); + } + + if (print_layer_stats) { + print_error_stats(name, layer_error, false); + combine_error_stats(total_error, layer_error); + } +} + +int main(int argc, char ** argv) { + ggml_time_init(); + + quantize_stats_params params; + + // read command line + + int max_thread = 0; + bool invalid_param = false; + std::string arg; + for (int i = 1; i < argc; i++) { + arg = argv[i]; + + if (arg == "-h" || arg == "--help") { + quantize_stats_print_usage(argc, argv); + exit(0); + } else if (arg == "-r" || arg == "--reference") { + params.reference = true; + } else if (arg == "-v") { + params.verbose = true; + } else if (arg == "-p" || arg == "--per-layer-stats") { + params.per_layer_stats = true; + } else if (arg == "--histogram") { + params.print_histogram = true; + } else if (arg == "-m" || arg == "--model") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.model = argv[i]; + } else if (arg == "-l" || arg == "--include-layer") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.include_layers.push_back(argv[i]); + } else if (arg == "-L" || arg == "--exclude-layer") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.exclude_layers.push_back(argv[i]); + } else if (arg == "-t" || arg == "--type") { + if (++i >= argc) { + invalid_param = true; + break; + } + int j; + for (j = 0; j < GGML_TYPE_COUNT && strcmp(argv[i], ggml_type_name((ggml_type) j)) != 0; j++) { + // find match + } + if (j < GGML_TYPE_COUNT) { + params.include_types.push_back((ggml_type) j); + } else { + fprintf(stderr, "error: %s not in list of types\n", argv[i]); + invalid_param = true; + } + } else if (arg == "-n" || arg == "--num-threads") { + if (++i >= argc) { + invalid_param = true; + break; + } + max_thread = atoi(argv[i]); + } else { + fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); + quantize_stats_print_usage(argc, argv); + return 1; + } + } + if (invalid_param) { + fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str()); + quantize_stats_print_usage(argc, argv); + return 1; + } + + // load the model + fprintf(stderr, "Loading model\n"); + + const int64_t t_main_start_us = ggml_time_us(); + llama_context * ctx; + + { + auto lparams = llama_context_default_params(); + + lparams.n_ctx = 256; + lparams.n_parts = 1; + lparams.seed = 1; + lparams.f16_kv = false; + lparams.use_mlock = false; + + ctx = llama_init_from_file(params.model.c_str(), lparams); + + if (ctx == NULL) { + fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str()); + return 1; + } + } + + const auto &tensors = llama_internal_get_tensor_map(ctx); + + // check layer tensors + int included_layers = 0; + int64_t max_nelements = 0; + bool is_f16 = false; + for (const auto& kv_tensor : tensors) { + if (!layer_included(params, kv_tensor.first)) { + continue; + } + if (params.verbose) { + printf("%s: type %s, size %" PRId64 "\n", kv_tensor.first.c_str(), ggml_type_name(kv_tensor.second->type), ggml_nelements(kv_tensor.second)); + } + if (kv_tensor.second->type == GGML_TYPE_F16) { + is_f16 = true; + } else if (kv_tensor.second->type != GGML_TYPE_F32) { + fprintf(stderr, "%s: error: Quantization should be tested with a float model, " + "this model contains already quantized layers (%s is type %d)\n", __func__, kv_tensor.first.c_str(), kv_tensor.second->type); + llama_free(ctx); + return 1; + } + included_layers++; + max_nelements = std::max(max_nelements, ggml_nelements(kv_tensor.second)); + } + + if (is_f16) { + printf("note: source model is f16\n"); + } + printf("testing %d layers with max size %" PRId64 "\n", included_layers, max_nelements); + // allocate scratch space + std::vector input_scratch; + std::vector quantized_scratch; + std::vector output_scratch; + + // loop throught quantization types + for (int i = 0; i < GGML_TYPE_COUNT; i++) { + const ggml_type type = (ggml_type) i; + if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) { + continue; + } + quantize_fns_t qfns = ggml_internal_get_quantize_fn(i); + if (qfns.quantize_row_q && qfns.dequantize_row_q) { + if (params.verbose) { + printf("testing %s ...\n", ggml_type_name(type)); + } + + error_stats global_stats {}; + + for (const auto& kv_tensor : tensors) { + if (!layer_included(params, kv_tensor.first)) { + continue; + } + if (params.verbose) { + printf(" %s ...\n", kv_tensor.first.c_str()); + } + std::string layer_name { ggml_type_name(type) }; + layer_name += "::" + kv_tensor.first; + test_roundtrip_on_layer( + layer_name, + params.per_layer_stats, + qfns, + params.reference, + kv_tensor.second, + input_scratch, + quantized_scratch, + output_scratch, + global_stats, + max_thread + ); + } + + print_error_stats(ggml_type_name(type), global_stats, params.print_histogram); + } + } + + + llama_free(ctx); + // report timing + { + const int64_t t_main_end_us = ggml_time_us(); + + printf("\n"); + printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0); + } + + return 0; +} diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 680757c6b..5b4812c62 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -5,15 +5,17 @@ #include // usage: -// ./llama-quantize models/llama/ggml-model.bin models/llama/ggml-model-quant.bin type +// ./quantize models/llama/ggml-model.bin models/llama/ggml-model-quant.bin type // int main(int argc, char ** argv) { ggml_time_init(); - if (argc != 4) { - fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]); - fprintf(stderr, " type = 2 - q4_0\n"); - fprintf(stderr, " type = 3 - q4_1\n"); + if (argc < 4) { + fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type [nthread]\n", argv[0]); + fprintf(stderr, " type = %d - q4_0\n", LLAMA_FTYPE_MOSTLY_Q4_0); + fprintf(stderr, " type = %d - q4_1\n", LLAMA_FTYPE_MOSTLY_Q4_1); + fprintf(stderr, " type = %d - q4_2\n", LLAMA_FTYPE_MOSTLY_Q4_2); + fprintf(stderr, " type = %d - q4_3\n", LLAMA_FTYPE_MOSTLY_Q4_3); return 1; } @@ -27,7 +29,8 @@ int main(int argc, char ** argv) { const std::string fname_inp = argv[1]; const std::string fname_out = argv[2]; - const int itype = atoi(argv[3]); + const enum llama_ftype ftype = (enum llama_ftype)atoi(argv[3]); + int nthread = argc > 4 ? atoi(argv[4]) : 0; const int64_t t_main_start_us = ggml_time_us(); @@ -37,7 +40,7 @@ int main(int argc, char ** argv) { { const int64_t t_start_us = ggml_time_us(); - if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), itype)) { + if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype, nthread)) { fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str()); return 1; } diff --git a/flake.nix b/flake.nix index 4c2717e0d..5363052b1 100644 --- a/flake.nix +++ b/flake.nix @@ -10,7 +10,6 @@ inherit system; }; llama-python = pkgs.python310.withPackages (ps: with ps; [ - torch numpy sentencepiece ]); @@ -28,8 +27,9 @@ ]; installPhase = '' mkdir -p $out/bin - mv bin/main $out/bin/llama - mv bin/quantize $out/bin/quantize + mv bin/* $out/bin/ + mv $out/bin/main $out/bin/llama + echo "#!${llama-python}/bin/python" > $out/bin/convert-pth-to-ggml cat ${./convert-pth-to-ggml.py} >> $out/bin/convert-pth-to-ggml chmod +x $out/bin/convert-pth-to-ggml diff --git a/ggml-cuda.cu b/ggml-cuda.cu new file mode 100644 index 000000000..0baa989a3 --- /dev/null +++ b/ggml-cuda.cu @@ -0,0 +1,154 @@ +#include +#include +#include "ggml-cuda.h" + +typedef uint16_t ggml_fp16_t; +static_assert(sizeof(__half) == sizeof(ggml_fp16_t), "wrong fp16 size"); + +#define QK4_0 32 +typedef struct { + float d; // delta + uint8_t qs[QK4_0 / 2]; // nibbles / quants +} block_q4_0; +static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding"); + +#define QK4_1 32 +typedef struct { + float d; // delta + float m; // min + uint8_t qs[QK4_1 / 2]; // nibbles / quants +} block_q4_1; +static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding"); + +#define QK4_2 16 +typedef struct { + __half d; // delta + uint8_t qs[QK4_2 / 2]; // nibbles / quants +} block_q4_2; +static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding"); + +#define QK4_3 16 +typedef struct { + __half d; // delta + __half m; // min + uint8_t qs[QK4_3 / 2]; // nibbles / quants +} block_q4_3; +static_assert(sizeof(block_q4_3) == 2 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong q4_3 block size/padding"); + + + +static __global__ void dequantize_block_q4_0(const void * vx, float * y) { + const block_q4_0 * x = (const block_q4_0 *) vx; + + const int i = blockIdx.x; + + const float d = x[i].d; + + const uint8_t * pp = x[i].qs; + + for (int l = 0; l < QK4_0; l += 2) { + const uint8_t vi = pp[l/2]; + + const int8_t vi0 = vi & 0xf; + const int8_t vi1 = vi >> 4; + + const float v0 = (vi0 - 8)*d; + const float v1 = (vi1 - 8)*d; + + y[i*QK4_0 + l + 0] = v0; + y[i*QK4_0 + l + 1] = v1; + } +} + +static __global__ void dequantize_block_q4_1(const void * vx, float * y) { + const block_q4_1 * x = (const block_q4_1 *) vx; + + const int i = blockIdx.x; + + const float d = x[i].d; + const float m = x[i].m; + + const uint8_t * pp = x[i].qs; + + for (int l = 0; l < QK4_1; l += 2) { + const uint8_t vi = pp[l/2]; + + const int8_t vi0 = vi & 0xf; + const int8_t vi1 = vi >> 4; + + const float v0 = vi0*d + m; + const float v1 = vi1*d + m; + + y[i*QK4_1 + l + 0] = v0; + y[i*QK4_1 + l + 1] = v1; + } +} + +static __global__ void dequantize_block_q4_2(const void * vx, float * y) { + const block_q4_2 * x = (const block_q4_2 *) vx; + + const int i = blockIdx.x; + + const float d = x[i].d; + + const uint8_t * pp = x[i].qs; + + for (int l = 0; l < QK4_2; l += 2) { + const uint8_t vi = pp[l/2]; + + const int8_t vi0 = vi & 0xf; + const int8_t vi1 = vi >> 4; + + const float v0 = (vi0 - 8)*d; + const float v1 = (vi1 - 8)*d; + + y[i*QK4_2 + l + 0] = v0; + y[i*QK4_2 + l + 1] = v1; + } +} + +static __global__ void dequantize_block_q4_3(const void * vx, float * y) { + const block_q4_3 * x = (const block_q4_3 *) vx; + + const int i = blockIdx.x; + + const float d = x[i].d; + const float m = x[i].m; + + const uint8_t * pp = x[i].qs; + + for (int l = 0; l < QK4_3; l += 2) { + const uint8_t vi = pp[l/2]; + + const int8_t vi0 = vi & 0xf; + const int8_t vi1 = vi >> 4; + + const float v0 = vi0*d + m; + const float v1 = vi1*d + m; + + y[i*QK4_3 + l + 0] = v0; + y[i*QK4_3 + l + 1] = v1; + } +} + +extern "C" { + __host__ void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) { + const int nb = k / QK4_0; + dequantize_block_q4_0<<>>(vx, y); + } + + __host__ void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) { + const int nb = k / QK4_1; + dequantize_block_q4_1<<>>(vx, y); + } + + __host__ void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream) { + const int nb = k / QK4_2; + dequantize_block_q4_2<<>>(vx, y); + } + + __host__ void dequantize_row_q4_3_cuda(const void * vx, float * y, int k, cudaStream_t stream) { + const int nb = k / QK4_3; + dequantize_block_q4_3<<>>(vx, y); + } +} diff --git a/ggml-cuda.h b/ggml-cuda.h new file mode 100644 index 000000000..be140606a --- /dev/null +++ b/ggml-cuda.h @@ -0,0 +1,12 @@ +#ifdef __cplusplus +extern "C" { +#endif + +void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream); +void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream); +void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream); +void dequantize_row_q4_3_cuda(const void * vx, float * y, int k, cudaStream_t stream); + +#ifdef __cplusplus +} +#endif diff --git a/ggml.c b/ggml.c index 59e84ab45..998602150 100644 --- a/ggml.c +++ b/ggml.c @@ -1,4 +1,4 @@ -// Defines CLOCK_MONOTONIC and asprintf on Linux +// Defines CLOCK_MONOTONIC on Linux #define _GNU_SOURCE #include "ggml.h" @@ -19,6 +19,7 @@ #include #include #include +#include // if C99 - static_assert is noop // ref: https://stackoverflow.com/a/53923785/4039976 @@ -26,14 +27,9 @@ #define static_assert(cond, msg) struct global_scope_noop_trick #endif -#if defined _MSC_VER || defined(__MINGW32__) +#if defined(_WIN32) -#if !defined(__MINGW32__) -#include -#else -// ref: https://github.com/ggerganov/whisper.cpp/issues/168 #include -#endif typedef volatile LONG atomic_int; typedef atomic_int atomic_bool; @@ -55,6 +51,7 @@ typedef HANDLE pthread_t; typedef DWORD thread_ret_t; static int pthread_create(pthread_t* out, void* unused, thread_ret_t(*func)(void*), void* arg) { + (void) unused; HANDLE handle = CreateThread(NULL, 0, (LPTHREAD_START_ROUTINE) func, arg, 0, NULL); if (handle == NULL) { @@ -66,6 +63,7 @@ static int pthread_create(pthread_t* out, void* unused, thread_ret_t(*func)(void } static int pthread_join(pthread_t thread, void* unused) { + (void) unused; return (int) WaitForSingleObject(thread, INFINITE); } @@ -97,17 +95,6 @@ typedef void* thread_ret_t; #define static_assert(cond, msg) _Static_assert(cond, msg) #endif -#define GGML_MLOCK_SUPPORT 0 - -#ifdef __has_include - #if __has_include() - #undef GGML_MLOCK_SUPPORT - #define GGML_MLOCK_SUPPORT 1 - #include - #endif -#endif - - /*#define GGML_PERF*/ #define GGML_DEBUG 0 #define GGML_GELU_FP16 @@ -128,6 +115,23 @@ typedef void* thread_ret_t; #define GGML_MEM_ALIGN 16 #endif +#if defined(_MSC_VER) || defined(__MINGW32__) +#define GGML_ALIGNED_MALLOC(size) _aligned_malloc(size, GGML_MEM_ALIGN) +#define GGML_ALIGNED_FREE(ptr) _aligned_free(ptr) +#else +inline static void* ggml_aligned_malloc(size_t size) { + void* aligned_memory = NULL; + int result = posix_memalign(&aligned_memory, GGML_MEM_ALIGN, size); + if (result != 0) { + // Handle allocation failure + return NULL; + } + return aligned_memory; +} +#define GGML_ALIGNED_MALLOC(size) ggml_aligned_malloc(size) +#define GGML_ALIGNED_FREE(ptr) free(ptr) +#endif + #define UNUSED(x) (void)(x) #define SWAP(x, y, T) do { T SWAP = x; x = y; y = SWAP; } while (0) @@ -139,10 +143,49 @@ typedef void* thread_ret_t; } \ } while (0) -#ifdef GGML_USE_ACCELERATE +#if defined(GGML_USE_ACCELERATE) #include -#elif GGML_USE_OPENBLAS +#elif defined(GGML_USE_OPENBLAS) #include +#elif defined(GGML_USE_CUBLAS) +#include +#include +#include "ggml-cuda.h" + +#define CUDA_CHECK(err) \ + do { \ + cudaError_t err_ = (err); \ + if (err_ != cudaSuccess) { \ + printf("CUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__, \ + cudaGetErrorString(err_)); \ + exit(1); \ + } \ + } while (0) + +#define CUBLAS_CHECK(err) \ + do { \ + cublasStatus_t err_ = (err); \ + if (err_ != CUBLAS_STATUS_SUCCESS) { \ + printf("cuBLAS error %d at %s:%d\n", err_, __FILE__, __LINE__); \ + exit(1); \ + } \ + } while (0) + +static cublasHandle_t cublasH = NULL; +static cudaStream_t cudaStream = NULL; +static void init_cublas(void) { + if (cublasH == NULL) { + // create cublas handle, bind a stream + CUBLAS_CHECK(cublasCreate(&cublasH)); + + CUDA_CHECK(cudaStreamCreateWithFlags(&cudaStream, cudaStreamNonBlocking)); + + CUBLAS_CHECK(cublasSetStream(cublasH, cudaStream)); + + // configure logging to stdout + // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, NULL)); + } +} #endif #undef MIN @@ -242,12 +285,12 @@ static inline float fp32_from_bits(uint32_t w) { } static inline uint32_t fp32_to_bits(float f) { - union { - float as_value; - uint32_t as_bits; - } fp32; - fp32.as_value = f; - return fp32.as_bits; + union { + float as_value; + uint32_t as_bits; + } fp32; + fp32.as_value = f; + return fp32.as_bits; } static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { @@ -424,14 +467,30 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float); // quantization // -#define QK 32 +#if __AVX__ || __AVX2__ || __AVX512F__ +// Unpack 16 4-bit fields into 16 bytes +// The output vector contains 16 bytes, each one in [ 0 .. 15 ] interval +static inline __m128i bytes_from_nibbles_16(const uint8_t * rsi) +{ + // Load 8 bytes from memory + __m128i tmp = _mm_loadu_si64( ( const __m128i* )rsi ); + + // Expand bytes into uint16_t values + __m128i bytes = _mm_cvtepu8_epi16( tmp ); + + // Unpack values into individual bytes + const __m128i lowMask = _mm_set1_epi8( 0xF ); + __m128i high = _mm_andnot_si128( lowMask, bytes ); + __m128i low = _mm_and_si128( lowMask, bytes ); + high = _mm_slli_epi16( high, 4 ); + bytes = _mm_or_si128( low, high ); + return bytes; +} -// AVX routines provided by GH user Const-me -// ref: https://github.com/ggerganov/ggml/pull/27#issuecomment-1464934600 #if __AVX2__ || __AVX512F__ // Unpack 32 4-bit fields into 32 bytes // The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval -static inline __m256i bytesFromNibbles( const uint8_t* rsi ) +static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) { // Load 16 bytes from memory __m128i tmp = _mm_loadu_si128( ( const __m128i* )rsi ); @@ -462,24 +521,7 @@ static inline __m128i packNibbles( __m256i bytes ) __m128i r1 = _mm256_extracti128_si256( bytes, 1 ); return _mm_packus_epi16( r0, r1 ); } -#elif __AVX__ -static inline __m128i bytesFromNibbles( const uint8_t* rsi ) -{ - // Load 8 bytes from memory - __m128i tmp = _mm_loadu_si64( ( const __m128i* )rsi ); - - // Expand bytes into uint16_t values - __m128i bytes = _mm_cvtepu8_epi16( tmp ); - - // Unpack values into individual bytes - const __m128i lowMask = _mm_set1_epi8( 0xF ); - __m128i high = _mm_andnot_si128( lowMask, bytes ); - __m128i low = _mm_and_si128( lowMask, bytes ); - high = _mm_slli_epi16( high, 4 ); - bytes = _mm_or_si128( low, high ); - return bytes; -} - +#else static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 ) { // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh @@ -496,38 +538,142 @@ static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 ) return _mm_packus_epi16( bytes1, bytes2); } #endif +#endif // __AVX__ || __AVX2__ || __AVX512F__ -// method 5 -// blocks of QK elements -// represented with a single float (delta) and QK/2 8-bit ints (i.e QK 4-bit signed integer factors) +#if __ARM_NEON + +#if !defined(__aarch64__) + +inline static uint16_t vaddvq_u8(uint8x16_t v) { + return + (uint16_t)vgetq_lane_u8(v, 0) + (uint16_t)vgetq_lane_u8(v, 1) + + (uint16_t)vgetq_lane_u8(v, 2) + (uint16_t)vgetq_lane_u8(v, 3) + + (uint16_t)vgetq_lane_u8(v, 4) + (uint16_t)vgetq_lane_u8(v, 5) + + (uint16_t)vgetq_lane_u8(v, 6) + (uint16_t)vgetq_lane_u8(v, 7) + + (uint16_t)vgetq_lane_u8(v, 8) + (uint16_t)vgetq_lane_u8(v, 9) + + (uint16_t)vgetq_lane_u8(v, 10) + (uint16_t)vgetq_lane_u8(v, 11) + + (uint16_t)vgetq_lane_u8(v, 12) + (uint16_t)vgetq_lane_u8(v, 13) + + (uint16_t)vgetq_lane_u8(v, 14) + (uint16_t)vgetq_lane_u8(v, 15); +} + +inline static int16_t vaddvq_s8(int8x16_t v) { + return + (int16_t)vgetq_lane_s8(v, 0) + (int16_t)vgetq_lane_s8(v, 1) + + (int16_t)vgetq_lane_s8(v, 2) + (int16_t)vgetq_lane_s8(v, 3) + + (int16_t)vgetq_lane_s8(v, 4) + (int16_t)vgetq_lane_s8(v, 5) + + (int16_t)vgetq_lane_s8(v, 6) + (int16_t)vgetq_lane_s8(v, 7) + + (int16_t)vgetq_lane_s8(v, 8) + (int16_t)vgetq_lane_s8(v, 9) + + (int16_t)vgetq_lane_s8(v, 10) + (int16_t)vgetq_lane_s8(v, 11) + + (int16_t)vgetq_lane_s8(v, 12) + (int16_t)vgetq_lane_s8(v, 13) + + (int16_t)vgetq_lane_s8(v, 14) + (int16_t)vgetq_lane_s8(v, 15); +} + +inline static int32_t vaddvq_s16(int16x8_t v) { + return + (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) + + (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) + + (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) + + (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7); +} + +inline static uint32_t vaddvq_u16(uint16x8_t v) { + return + (uint32_t)vgetq_lane_u16(v, 0) + (uint32_t)vgetq_lane_u16(v, 1) + + (uint32_t)vgetq_lane_u16(v, 2) + (uint32_t)vgetq_lane_u16(v, 3) + + (uint32_t)vgetq_lane_u16(v, 4) + (uint32_t)vgetq_lane_u16(v, 5) + + (uint32_t)vgetq_lane_u16(v, 6) + (uint32_t)vgetq_lane_u16(v, 7); +} + +inline static int32_t vaddvq_s32(int32x4_t v) { + return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3); +} + +inline static float vaddvq_f32(float32x4_t v) { + return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3); +} + +float vminvq_f32(float32x4_t v) { + return + MIN(MIN(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)), + MIN(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3))); +} + +float vmaxvq_f32(float32x4_t v) { + return + MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)), + MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3))); +} + +int8x8_t vzip1_s8(int8x8_t a, int8x8_t b) { + return vget_low_s8(vcombine_s8(a, b)); +} + +int8x8_t vzip2_s8(int8x8_t a, int8x8_t b) { + return vget_high_s8(vcombine_s8(a, b)); +} + +uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) { + return vget_low_u8(vcombine_u8(a, b)); +} + +uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) { + return vget_high_u8(vcombine_u8(a, b)); +} + +#endif +#endif + + +#define QK4_0 32 typedef struct { - float d; // delta - uint8_t qs[QK / 2]; // nibbles / quants + float d; // delta + uint8_t qs[QK4_0 / 2]; // nibbles / quants } block_q4_0; -static_assert(sizeof(block_q4_0) == sizeof(float) + QK / 2, "wrong q4_0 block size/padding"); +static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding"); -// method 4 -// blocks of QK elements -// represented with 2 floats (delta + min) and QK/2 8-bit ints (i.e QK 4-bit unsigned integer factors) +#define QK4_1 32 typedef struct { - float d; - float m; - uint8_t qs[QK / 2]; // nibbles / quants + float d; // delta + float m; // min + uint8_t qs[QK4_1 / 2]; // nibbles / quants } block_q4_1; -static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK / 2, "wrong q4_1 block size/padding"); +static_assert(sizeof(block_q4_1) == 2 * sizeof(float) + QK4_1 / 2, "wrong q4_1 block size/padding"); + +#define QK4_2 16 +typedef struct { + ggml_fp16_t d; // delta + uint8_t qs[QK4_2 / 2]; // nibbles / quants +} block_q4_2; +static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding"); + +#define QK4_3 16 +typedef struct { + ggml_fp16_t d; // delta + ggml_fp16_t m; // min + uint8_t qs[QK4_3 / 2]; // nibbles / quants +} block_q4_3; +static_assert(sizeof(block_q4_3) == 2 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong q4_3 block size/padding"); + +#define QK8_0 32 +typedef struct { + float d; // delta + int8_t qs[QK8_0]; // quants +} block_q8_0; +static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding"); + // reference implementation for deterministic creation of model files static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) { - assert(k % QK == 0); - const int nb = k / QK; + assert(k % QK4_0 == 0); + const int nb = k / QK4_0; - uint8_t pp[QK/2]; + uint8_t pp[QK4_0/2]; for (int i = 0; i < nb; i++) { float amax = 0.0f; // absolute max - for (int l = 0; l < QK; l++) { - const float v = x[i*QK + l]; + for (int l = 0; l < QK4_0; l++) { + const float v = x[i*QK4_0 + l]; amax = MAX(amax, fabsf(v)); } @@ -536,9 +682,9 @@ static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * r y[i].d = d; - for (int l = 0; l < QK; l += 2) { - const float v0 = x[i*QK + l + 0]*id; - const float v1 = x[i*QK + l + 1]*id; + for (int l = 0; l < QK4_0; l += 2) { + const float v0 = x[i*QK4_0 + l + 0]*id; + const float v1 = x[i*QK4_0 + l + 1]*id; const uint8_t vi0 = (int8_t)roundf(v0) + 8; const uint8_t vi1 = (int8_t)roundf(v1) + 8; @@ -554,8 +700,8 @@ static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * r } static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int k) { - assert(k % QK == 0); - const int nb = k / QK; + assert(k % QK4_0 == 0); + const int nb = k / QK4_0; block_q4_0 * restrict y = vy; @@ -610,10 +756,7 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]); for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]); - // absolute max - const float amax = MAX( - MAX(vgetq_lane_f32(amaxv[0], 0), vgetq_lane_f32(amaxv[0], 1)), - MAX(vgetq_lane_f32(amaxv[0], 2), vgetq_lane_f32(amaxv[0], 3))); + const float amax = vmaxvq_f32(amaxv[0]); const float d = amax / ((1 << 3) - 1); const float id = d ? 1.0f/d : 0.0f; @@ -808,19 +951,19 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int } static void quantize_row_q4_1_reference(const float * restrict x, void * restrict vy, int k) { - assert(k % QK == 0); - const int nb = k / QK; + assert(k % QK4_1 == 0); + const int nb = k / QK4_1; block_q4_1 * restrict y = vy; - uint8_t pp[QK/2]; + uint8_t pp[QK4_1/2]; for (int i = 0; i < nb; i++) { float min = FLT_MAX; float max = -FLT_MAX; - for (int l = 0; l < QK; l++) { - const float v = x[i*QK + l]; + for (int l = 0; l < QK4_1; l++) { + const float v = x[i*QK4_1 + l]; if (v < min) min = v; if (v > max) max = v; } @@ -831,9 +974,9 @@ static void quantize_row_q4_1_reference(const float * restrict x, void * restric y[i].d = d; y[i].m = min; - for (int l = 0; l < QK; l += 2) { - const float v0 = (x[i*QK + l + 0] - min)*id; - const float v1 = (x[i*QK + l + 1] - min)*id; + for (int l = 0; l < QK4_1; l += 2) { + const float v0 = (x[i*QK4_1 + l + 0] - min)*id; + const float v1 = (x[i*QK4_1 + l + 1] - min)*id; const uint8_t vi0 = roundf(v0); const uint8_t vi1 = roundf(v1); @@ -849,9 +992,9 @@ static void quantize_row_q4_1_reference(const float * restrict x, void * restric } static void quantize_row_q4_1(const float * restrict x, void * restrict vy, int k) { - assert(k % QK == 0); + assert(k % QK4_1 == 0); - const int nb = k / QK; + const int nb = k / QK4_1; block_q4_1 * restrict y = vy; @@ -935,7 +1078,7 @@ static void quantize_row_q4_1(const float * restrict x, void * restrict vy, int float32x4_t minv[8]; float32x4_t maxv[8]; - for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l); + for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*QK4_1 + 4*l); for (int l = 0; l < 4; l++) minv[2*l] = vminq_f32(srcv[2*l], srcv[2*l + 1]); for (int l = 0; l < 2; l++) minv[4*l] = vminq_f32(minv[4*l], minv[4*l + 2]); @@ -958,7 +1101,8 @@ static void quantize_row_q4_1(const float * restrict x, void * restrict vy, int for (int l = 0; l < 8; l++) { const float32x4_t v = vmulq_n_f32(vsubq_f32(srcv[l], minv0), id); - const int32x4_t vi = vcvtq_s32_f32(v); + const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(0.5f)); // needed to round to nearest + const int32x4_t vi = vcvtq_s32_f32(vf); y[i].qs[2*l + 0] = vgetq_lane_s32(vi, 0) | (vgetq_lane_s32(vi, 1) << 4); y[i].qs[2*l + 1] = vgetq_lane_s32(vi, 2) | (vgetq_lane_s32(vi, 3) << 4); @@ -970,9 +1114,327 @@ static void quantize_row_q4_1(const float * restrict x, void * restrict vy, int #endif } +// reference implementation for deterministic creation of model files +static void quantize_row_q4_2_reference(const float * restrict x, block_q4_2 * restrict y, int k) { + assert(k % QK4_2 == 0); + + const int nb = k / QK4_2; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + + for (int l = 0; l < QK4_2; l++) { + const float v = x[i*QK4_2 + l]; + amax = MAX(amax, fabsf(v)); + } + + const float d = amax / ((1 << 3) - 1); + + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + + for (int l = 0; l < QK4_2; l += 2) { + const float v0 = x[i*QK4_2 + l + 0]*id; + const float v1 = x[i*QK4_2 + l + 1]*id; + + const uint8_t vi0 = (uint8_t)(v0 + 8.5f); + const uint8_t vi1 = (uint8_t)(v1 + 8.5f); + + assert(vi0 < 16); + assert(vi1 < 16); + + y[i].qs[l/2] = vi0 | (vi1 << 4); + } + } +} + +static inline int nearest_int(float fval) { + assert(fval <= 4194303.f); + float val = fval + 12582912.f; + int i; memcpy(&i, &val, sizeof(int)); + return (i & 0x007fffff) - 0x00400000; +} + +static float kquantize_q4_with_bounds(int n, int nmin, int nmax, const float * restrict X, int nCandidates, + const float * restrict candidates, int8_t * restrict L) { + assert (nmin >= INT8_MIN); + assert (nmax <= INT8_MAX); + float amax = 0; + for (int i=0; i sumlxM2*suml2P) { + if (sumlxP2 > best*suml2P) { + best = sumlxP2/suml2P; bestScale = iscale; + } + } else { + if (sumlxM2 > best*suml2M) { + best = sumlxM2/suml2M; bestScale = -iscale; + } + } + } + float sumlx = 0; int suml2 = 0; + for (int i=0; i max) max = v; + } + + const float d = (max - min) / ((1 << 4) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + y[i].m = GGML_FP32_TO_FP16(min); + + for (int l = 0; l < QK4_3; l += 2) { + const float v0 = (x[i*QK4_3 + l + 0] - min)*id; + const float v1 = (x[i*QK4_3 + l + 1] - min)*id; + + const uint8_t vi0 = (int) (v0 + 0.5f); + const uint8_t vi1 = (int) (v1 + 0.5f); + + assert(vi0 < 16); + assert(vi1 < 16); + + y[i].qs[l/2] = vi0 | (vi1 << 4); + } + } +} + +static void quantize_row_q4_3(const float * restrict x, void * restrict vy, int k) { + assert(k % QK4_3 == 0); + + block_q4_3 * restrict y = vy; + + quantize_row_q4_3_reference(x, y, k); +} + +// reference implementation for deterministic creation of model files +static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k) { + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + + for (int l = 0; l < QK8_0; l++) { + const float v = x[i*QK8_0 + l]; + amax = MAX(amax, fabsf(v)); + } + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + for (int l = 0; l < QK8_0; ++l) { + const float v = x[i*QK8_0 + l]*id; + y[i].qs[l] = roundf(v); + } + } +} + +static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) { + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + block_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + for (int i = 0; i < nb; i++) { + float32x4_t srcv [8]; + float32x4_t asrcv[8]; + float32x4_t amaxv[8]; + + for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l); + for (int l = 0; l < 8; l++) asrcv[l] = vabsq_f32(srcv[l]); + + for (int l = 0; l < 4; l++) amaxv[2*l] = vmaxq_f32(asrcv[2*l], asrcv[2*l+1]); + for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]); + for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]); + + const float amax = vmaxvq_f32(amaxv[0]); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + for (int l = 0; l < 8; l++) { + const float32x4_t v = vmulq_n_f32(srcv[l], id); + const int32x4_t vi = vcvtnq_s32_f32(v); + + y[i].qs[4*l + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3); + } + } +#elif defined(__AVX2__) || defined(__AVX__) + for (int i = 0; i < nb; i++) { + // Load elements into 4 AVX vectors + __m256 v0 = _mm256_loadu_ps( x ); + __m256 v1 = _mm256_loadu_ps( x + 8 ); + __m256 v2 = _mm256_loadu_ps( x + 16 ); + __m256 v3 = _mm256_loadu_ps( x + 24 ); + x += 32; + + // Compute max(abs(e)) for the block + const __m256 signBit = _mm256_set1_ps( -0.0f ); + __m256 maxAbs = _mm256_andnot_ps( signBit, v0 ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) ); + + __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) ); + max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); + max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); + const float maxScalar = _mm_cvtss_f32( max4 ); + + // Quantize these floats + const float d = maxScalar / 127.f; + y[i].d = d; + const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; + const __m256 mul = _mm256_set1_ps( id ); + + // Apply the multiplier + v0 = _mm256_mul_ps( v0, mul ); + v1 = _mm256_mul_ps( v1, mul ); + v2 = _mm256_mul_ps( v2, mul ); + v3 = _mm256_mul_ps( v3, mul ); + + // Round to nearest integer + v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); + v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); + v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); + v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); + + // Convert floats to integers + __m256i i0 = _mm256_cvtps_epi32( v0 ); + __m256i i1 = _mm256_cvtps_epi32( v1 ); + __m256i i2 = _mm256_cvtps_epi32( v2 ); + __m256i i3 = _mm256_cvtps_epi32( v3 ); + +#if defined(__AVX2__) + // Convert int32 to int16 + i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 + i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 + // Convert int16 to int8 + i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 + + // We got our precious signed bytes, but the order is now wrong + // These AVX2 pack instructions process 16-byte pieces independently + // The following instruction is fixing the order + const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); + i0 = _mm256_permutevar8x32_epi32( i0, perm ); + + _mm256_storeu_si256((__m256i *)y[i].qs, i0); +#else + // Since we don't have in AVX some necessary functions, + // we split the registers in half and call AVX2 analogs from SSE + __m128i ni0 = _mm256_castsi256_si128( i0 ); + __m128i ni1 = _mm256_extractf128_si256( i0, 1); + __m128i ni2 = _mm256_castsi256_si128( i1 ); + __m128i ni3 = _mm256_extractf128_si256( i1, 1); + __m128i ni4 = _mm256_castsi256_si128( i2 ); + __m128i ni5 = _mm256_extractf128_si256( i2, 1); + __m128i ni6 = _mm256_castsi256_si128( i3 ); + __m128i ni7 = _mm256_extractf128_si256( i3, 1); + + // Convert int32 to int16 + ni0 = _mm_packs_epi32( ni0, ni1 ); + ni2 = _mm_packs_epi32( ni2, ni3 ); + ni4 = _mm_packs_epi32( ni4, ni5 ); + ni6 = _mm_packs_epi32( ni6, ni7 ); + // Convert int16 to int8 + ni0 = _mm_packs_epi16( ni0, ni2 ); + ni4 = _mm_packs_epi16( ni4, ni6 ); + + _mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0); + _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4); +#endif + } +#else + // scalar + quantize_row_q8_0_reference(x, y, k); +#endif +} + static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, int k) { - assert(k % QK == 0); - const int nb = k / QK; + assert(k % QK4_0 == 0); + const int nb = k / QK4_0; const block_q4_0 * restrict x = vx; @@ -983,9 +1445,9 @@ static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, in const uint8_t * restrict pp = x[i].qs; - for (int l = 0; l < QK; l += 32) { + for (int l = 0; l < QK4_0; l += 32) { // Load 32x4-bit integers into 32x8-bit integers - __m256i vx8 = bytesFromNibbles(pp+l/2); + __m256i vx8 = bytes_from_nibbles_32(pp+l/2); // Subtract 8 from the integers vx8 = _mm256_sub_epi8(vx8, _mm256_set1_epi8(8)); @@ -1005,7 +1467,7 @@ static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, in // Scale and store for (int j = 0; j < 4; j++) { const __m256 result = _mm256_mul_ps(vf[j], d_v); - _mm256_storeu_ps(y + i * QK + l + j*8, result); + _mm256_storeu_ps(y + i * QK4_0 + l + j*8, result); } } } @@ -1015,7 +1477,7 @@ static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, in const uint8_t * restrict pp = x[i].qs; - for (int l = 0; l < QK; l += 16) { + for (int l = 0; l < QK4_0; l += 16) { // Load 16x4-bit integers into 8x8-bit integers const uint8x8_t v8 = vld1_u8(pp + l/2); @@ -1054,10 +1516,10 @@ static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, in const float32x4_t r3 = vmulq_f32(vf_3, vd); // Store - vst1q_f32(y + i*QK + l + 0, r0); - vst1q_f32(y + i*QK + l + 4, r1); - vst1q_f32(y + i*QK + l + 8, r2); - vst1q_f32(y + i*QK + l + 12, r3); + vst1q_f32(y + i*QK4_0 + l + 0, r0); + vst1q_f32(y + i*QK4_0 + l + 4, r1); + vst1q_f32(y + i*QK4_0 + l + 8, r2); + vst1q_f32(y + i*QK4_0 + l + 12, r3); } } #else @@ -1067,7 +1529,7 @@ static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, in const uint8_t * restrict pp = x[i].qs; - for (int l = 0; l < QK; l += 2) { + for (int l = 0; l < QK4_0; l += 2) { const uint8_t vi = pp[l/2]; const int8_t vi0 = vi & 0xf; @@ -1078,19 +1540,19 @@ static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, in //printf("d = %f, vi = %d, vi0 = %d, vi1 = %d, v0 = %f, v1 = %f\n", d, vi, vi0, vi1, v0, v1); - y[i*QK + l + 0] = v0; - y[i*QK + l + 1] = v1; + y[i*QK4_0 + l + 0] = v0; + y[i*QK4_0 + l + 1] = v1; - assert(!isnan(y[i*QK + l + 0])); - assert(!isnan(y[i*QK + l + 1])); + assert(!isnan(y[i*QK4_0 + l + 0])); + assert(!isnan(y[i*QK4_0 + l + 1])); } } #endif } static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, int k) { - assert(k % QK == 0); - const int nb = k / QK; + assert(k % QK4_1 == 0); + const int nb = k / QK4_1; const block_q4_1 * restrict x = vx; @@ -1101,9 +1563,9 @@ static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, in const uint8_t * restrict pp = x[i].qs; - for (int l = 0; l < QK; l += 32) { + for (int l = 0; l < QK4_1; l += 32) { // Load 32x4-bit integers into 32x8-bit integers - __m256i vx8 = bytesFromNibbles(pp+l/2); + __m256i vx8 = bytes_from_nibbles_32(pp+l/2); // Convert to 16-bit int const __m256i vx16_lo = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 0)); @@ -1120,7 +1582,7 @@ static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, in // Scale, add m and store for (int j = 0; j < 4; j++) { const __m256 result = _mm256_add_ps(_mm256_mul_ps(vf[j], d_v), d_m); - _mm256_storeu_ps(y + i * QK + l + j*8, result); + _mm256_storeu_ps(y + i * QK4_1 + l + j*8, result); } } } @@ -1131,7 +1593,7 @@ static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, in const uint8_t * restrict pp = x[i].qs; - for (int l = 0; l < QK; l += 16) { + for (int l = 0; l < QK4_1; l += 16) { // Load 16x4-bit integers into 8x8-bit integers const uint8x8_t v8 = vld1_u8(pp + l/2); @@ -1162,10 +1624,10 @@ static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, in const float32x4_t r3 = vmlaq_f32(vm, vf_3, vd); // Store - vst1q_f32(y + i*QK + l + 0, r0); - vst1q_f32(y + i*QK + l + 4, r1); - vst1q_f32(y + i*QK + l + 8, r2); - vst1q_f32(y + i*QK + l + 12, r3); + vst1q_f32(y + i*QK4_1 + l + 0, r0); + vst1q_f32(y + i*QK4_1 + l + 4, r1); + vst1q_f32(y + i*QK4_1 + l + 8, r2); + vst1q_f32(y + i*QK4_1 + l + 12, r3); } } #else @@ -1175,7 +1637,7 @@ static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, in const uint8_t * restrict pp = x[i].qs; - for (int l = 0; l < QK; l += 2) { + for (int l = 0; l < QK4_1; l += 2) { const uint8_t vi = pp[l/2]; const int8_t vi0 = vi & 0xf; @@ -1184,16 +1646,125 @@ static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, in const float v0 = vi0*d + m; const float v1 = vi1*d + m; - y[i*QK + l + 0] = v0; - y[i*QK + l + 1] = v1; + y[i*QK4_1 + l + 0] = v0; + y[i*QK4_1 + l + 1] = v1; - assert(!isnan(y[i*QK + l + 0])); - assert(!isnan(y[i*QK + l + 1])); + assert(!isnan(y[i*QK4_1 + l + 0])); + assert(!isnan(y[i*QK4_1 + l + 1])); } } #endif } +static void dequantize_row_q4_2(const void * restrict vx, float * restrict y, int k) { + assert(k % QK4_2 == 0); + const int nb = k / QK4_2; + + const block_q4_2 * restrict x = vx; + + for (int i = 0; i < nb; i++) { + const float d = GGML_FP16_TO_FP32(x[i].d); + + const uint8_t * restrict pp = x[i].qs; + + for (int l = 0; l < QK4_2; l += 2) { + const uint8_t vi = pp[l/2]; + + const int8_t vi0 = vi & 0xf; + const int8_t vi1 = vi >> 4; + + const float v0 = (vi0 - 8)*d; + const float v1 = (vi1 - 8)*d; + + y[i*QK4_2 + l + 0] = v0; + y[i*QK4_2 + l + 1] = v1; + + assert(!isnan(y[i*QK4_2 + l + 0])); + assert(!isnan(y[i*QK4_2 + l + 1])); + } + } +} + +static void dequantize_row_q4_3(const void * restrict vx, float * restrict y, int k) { + assert(k % QK4_3 == 0); + const int nb = k / QK4_3; + + const block_q4_3 * restrict x = vx; + + for (int i = 0; i < nb; i++) { + const float d = GGML_FP16_TO_FP32(x[i].d); + const float m = GGML_FP16_TO_FP32(x[i].m); + + const uint8_t * restrict pp = x[i].qs; + + for (int l = 0; l < QK4_3; l += 2) { + const uint8_t vi = pp[l/2]; + + const int8_t vi0 = vi & 0xf; + const int8_t vi1 = vi >> 4; + + const float v0 = vi0*d + m; + const float v1 = vi1*d + m; + + y[i*QK4_3 + l + 0] = v0; + y[i*QK4_3 + l + 1] = v1; + + assert(!isnan(y[i*QK4_3 + l + 0])); + assert(!isnan(y[i*QK4_3 + l + 1])); + } + } +} + +static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); +static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); +static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); +static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); + +static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = { + [GGML_TYPE_Q4_0] = { + .dequantize_row_q = dequantize_row_q4_0, + .quantize_row_q = quantize_row_q4_0, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_0_reference, + .quantize_row_q_dot = quantize_row_q8_0, + .vec_dot_q = ggml_vec_dot_q4_0_q8_0, + }, + [GGML_TYPE_Q4_1] = { + .dequantize_row_q = dequantize_row_q4_1, + .quantize_row_q = quantize_row_q4_1, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_1_reference, + .quantize_row_q_dot = quantize_row_q8_0, + .vec_dot_q = ggml_vec_dot_q4_1_q8_0, + }, + [GGML_TYPE_Q4_2] = { + .dequantize_row_q = dequantize_row_q4_2, + .quantize_row_q = quantize_row_q4_2, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_2_rmse, //quantize_row_q4_2_reference, + .quantize_row_q_dot = quantize_row_q8_0, + .vec_dot_q = ggml_vec_dot_q4_2_q8_0, + }, + [GGML_TYPE_Q4_3] = { + .dequantize_row_q = dequantize_row_q4_3, + .quantize_row_q = quantize_row_q4_3, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_3_reference, // TODO: RMSE optimization + .quantize_row_q_dot = quantize_row_q8_0, + .vec_dot_q = ggml_vec_dot_q4_3_q8_0, + }, + [GGML_TYPE_Q8_0] = { + .dequantize_row_q = NULL, // TODO + .quantize_row_q = quantize_row_q8_0, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q8_0_reference, + .quantize_row_q_dot = quantize_row_q8_0, + .vec_dot_q = NULL, // TODO + }, +}; + +// For internal test use +quantize_fns_t ggml_internal_get_quantize_fn(size_t i) { + GGML_ASSERT(i < GGML_TYPE_COUNT); + return quantize_fns[i]; +} + + // // simd mappings // @@ -1226,15 +1797,7 @@ static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, in #define GGML_F32x4_FMA(a, b, c) vfmaq_f32(a, b, c) #define GGML_F32x4_ADD vaddq_f32 #define GGML_F32x4_MUL vmulq_f32 -#if defined(__ARM_FEATURE_QRDMX) - #define GGML_F32x4_REDUCE_ONE(x) vaddvq_f32(x) -#else - #define GGML_F32x4_REDUCE_ONE(x) \ - (vgetq_lane_f32(x, 0) + \ - vgetq_lane_f32(x, 1) + \ - vgetq_lane_f32(x, 2) + \ - vgetq_lane_f32(x, 3)) -#endif +#define GGML_F32x4_REDUCE_ONE(x) vaddvq_f32(x) #define GGML_F32x4_REDUCE(res, x) \ { \ for (int i = 0; i < GGML_F32_ARR/2; ++i) { \ @@ -1758,37 +2321,6 @@ inline static void ggml_vec_dot_f32(const int n, float * restrict s, const float *s = sumf; } -#if __AVX512F__ && QK == 32 -static inline __m512 dot_q4_0_oneblock_avx512( - __m512 acc, - const block_q4_0 * restrict x, - const block_q4_0 * restrict y, - int i -) { - // Compute combined scale for the block - __m512 d = _mm512_set1_ps( x[i].d * y[i].d ); - - __m256i bx = bytesFromNibbles( x[i].qs ); - __m256i by = bytesFromNibbles( y[i].qs ); - - // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. - const __m256i off = _mm256_set1_epi8( 8 ); - bx = _mm256_sub_epi8( bx, off ); - by = _mm256_sub_epi8( by, off ); - - // Sign-extend 16 signed bytes into int16_t - __m512i x32 = _mm512_cvtepi8_epi16( bx ); - __m512i y32 = _mm512_cvtepi8_epi16( by ); - // Compute products of int16_t integers, add pairwise - __m512i i64 = _mm512_madd_epi16( x32, y32 ); - - // Convert int32_t to float - __m512 p = _mm512_cvtepi32_ps( i64 ); - // Apply the scale, and accumulate - return _mm512_fmadd_ps( d, p, acc ); -} -#endif - inline static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y) { ggml_float sumf = 0.0; @@ -1825,208 +2357,121 @@ inline static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t *s = sumf; } -static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK; +static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int nb = n / QK8_0; - assert(n % QK == 0); + assert(n % QK8_0 == 0); assert(nb % 2 == 0); const block_q4_0 * restrict x = vx; - const block_q4_0 * restrict y = vy; + const block_q8_0 * restrict y = vy; float sumf = 0.0; #if defined(__ARM_NEON) - float sum0 = 0.0f; - float sum1 = 0.0f; + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); for (int i = 0; i < nb; i += 2) { const block_q4_0 * restrict x0 = &x[i + 0]; - const block_q4_0 * restrict y0 = &y[i + 0]; const block_q4_0 * restrict x1 = &x[i + 1]; - const block_q4_0 * restrict y1 = &y[i + 1]; + const block_q8_0 * restrict y0 = &y[i + 0]; + const block_q8_0 * restrict y1 = &y[i + 1]; - const uint8x16_t m4b = vdupq_n_u8(0xf); - const int8x16_t s8b = vdupq_n_s8(0x8); + const uint8x16_t m4b = vdupq_n_u8(0xf); + const int8x16_t s8b = vdupq_n_s8(0x8); const uint8x16_t v0_0 = vld1q_u8(x0->qs); - const uint8x16_t v1_0 = vld1q_u8(y0->qs); const uint8x16_t v0_1 = vld1q_u8(x1->qs); - const uint8x16_t v1_1 = vld1q_u8(y1->qs); // 4-bit -> 8-bit - const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8(v0_0, m4b)); - const int8x16_t v1_0l = vreinterpretq_s8_u8(vandq_u8(v1_0, m4b)); - + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); - const int8x16_t v1_0h = vreinterpretq_s8_u8(vshrq_n_u8(v1_0, 4)); - - const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8(v0_1, m4b)); - const int8x16_t v1_1l = vreinterpretq_s8_u8(vandq_u8(v1_1, m4b)); - + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - const int8x16_t v1_1h = vreinterpretq_s8_u8(vshrq_n_u8(v1_1, 4)); // sub 8 const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b); - const int8x16_t v1_0ls = vsubq_s8(v1_0l, s8b); - const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b); - const int8x16_t v1_0hs = vsubq_s8(v1_0h, s8b); - const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b); - const int8x16_t v1_1ls = vsubq_s8(v1_1l, s8b); - const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b); - const int8x16_t v1_1hs = vsubq_s8(v1_1h, s8b); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + + // interleave + const int8x16_t v1_0ls = vuzp1q_s8(v1_0l, v1_0h); + const int8x16_t v1_0hs = vuzp2q_s8(v1_0l, v1_0h); + const int8x16_t v1_1ls = vuzp1q_s8(v1_1l, v1_1h); + const int8x16_t v1_1hs = vuzp2q_s8(v1_1l, v1_1h); #if defined(__ARM_FEATURE_DOTPROD) - // dot product into int16x8_t - int32x4_t p_0 = vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls); - int32x4_t p_1 = vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls); + // dot product into int32x4_t + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls), v0_0hs, v1_0hs); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls), v0_1hs, v1_1hs); - p_0 = vdotq_s32(p_0, v0_0hs, v1_0hs); - p_1 = vdotq_s32(p_1, v0_1hs, v1_1hs); - - // scalar -#if defined(__ARM_FEATURE_QRDMX) - sum0 += x0->d * y0->d * vaddvq_s32(p_0); - sum1 += x1->d * y1->d * vaddvq_s32(p_1); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); #else - sum0 += x0->d * y0->d * (vgetq_lane_s32(p_0, 0) + vgetq_lane_s32(p_0, 1) + vgetq_lane_s32(p_0, 2) + vgetq_lane_s32(p_0, 3)); - sum1 += x1->d * y1->d * (vgetq_lane_s32(p_1, 0) + vgetq_lane_s32(p_1, 1) + vgetq_lane_s32(p_1, 2) + vgetq_lane_s32(p_1, 3)); -#endif -#else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls)); + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls)); const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0hs)); const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0hs)); const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1ls)); const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1ls)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1hs)); const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1hs)); - const int16x8_t pl_0 = vaddq_s16(pl0l, pl0h); - const int16x8_t ph_0 = vaddq_s16(ph0l, ph0h); + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - const int16x8_t pl_1 = vaddq_s16(pl1l, pl1h); - const int16x8_t ph_1 = vaddq_s16(ph1l, ph1h); - - const int16x8_t p_0 = vaddq_s16(pl_0, ph_0); - const int16x8_t p_1 = vaddq_s16(pl_1, ph_1); - - // scalar -#if defined(__ARM_FEATURE_QRDMX) - sum0 += x0->d * y0->d * vaddvq_s16(p_0); - sum1 += x1->d * y1->d * vaddvq_s16(p_1); -#else - sum0 += x0->d * y0->d * (vgetq_lane_s16(p_0, 0) + vgetq_lane_s16(p_0, 1) + vgetq_lane_s16(p_0, 2) + vgetq_lane_s16(p_0, 3) + vgetq_lane_s16(p_0, 4) + vgetq_lane_s16(p_0, 5) + vgetq_lane_s16(p_0, 6) + vgetq_lane_s16(p_0, 7)); - sum1 += x1->d * y1->d * (vgetq_lane_s16(p_1, 0) + vgetq_lane_s16(p_1, 1) + vgetq_lane_s16(p_1, 2) + vgetq_lane_s16(p_1, 3) + vgetq_lane_s16(p_1, 4) + vgetq_lane_s16(p_1, 5) + vgetq_lane_s16(p_1, 6) + vgetq_lane_s16(p_1, 7)); -#endif + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d); #endif } - sumf = sum0 + sum1; -#elif defined(__AVX512F__) - // Initialize accumulator with zeros - __m512 acc0 = _mm512_setzero_ps(); - __m512 acc1 = _mm512_setzero_ps(); - - const int superblock_size = 8; - const int superblock_count = nb / superblock_size; - - for (int superblock_ix = 0; superblock_ix < superblock_count; superblock_ix += 1) { - int i = superblock_ix * superblock_size; - - acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i+0 ); - acc1 = dot_q4_0_oneblock_avx512( acc1, x, y, i+1 ); - acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i+2 ); - acc1 = dot_q4_0_oneblock_avx512( acc1, x, y, i+3 ); - acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i+4 ); - acc1 = dot_q4_0_oneblock_avx512( acc1, x, y, i+5 ); - acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i+6 ); - acc1 = dot_q4_0_oneblock_avx512( acc1, x, y, i+7 ); - } - - // Remainders - for (int i = superblock_count * superblock_size; i < nb; ++i) { - acc0 = dot_q4_0_oneblock_avx512( acc0, x, y, i ); - } - - // Horizontal sum of all lanes of the accumulator - sumf = _mm512_reduce_add_ps( acc0 ) + _mm512_reduce_add_ps( acc1 ); + sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); #elif defined(__AVX2__) // Initialize accumulator with zeros __m256 acc = _mm256_setzero_ps(); - /* Prepare the constants we will need during execution */ - const __m256i lowMask = _mm256_set1_epi8( 0xF ); - const __m256i offset_8 = _mm256_set1_epi16( 8 ); - -#define UNROLL_COUNT 8 - // make sure we only unroll multiples of the block count - assert(nb % UNROLL_COUNT == 0); - // Main loop - for (int i = 0; i < nb; i+=UNROLL_COUNT) { + for (int i = 0; i < nb; ++i) { + /* Compute combined scale for the block */ + const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); - // This loop will be unrolled by the compiler - for (int u=0;u we now have a vector of 8 int_32t */ - __m256i xy_q = _mm256_add_epi32( xy_high_q, xy_low_q ); + // Perform multiplication and create 16-bit values + const __m256i dot = _mm256_maddubs_epi16(ax, sy); - /* Convert to vectore of 8 int32_t to 8 floats */ - __m256 q = _mm256_cvtepi32_ps( xy_q ); + const __m256i ones = _mm256_set1_epi16(1); + __m256i xy_q = _mm256_madd_epi16(ones, dot); - /* Multiply q with scale and accumulate */ - acc = _mm256_fmadd_ps( scale, q, acc ); - } - - } + /* Convert to vectore of 8 int32_t to 8 floats */ + __m256 q = _mm256_cvtepi32_ps( xy_q ); + + /* Multiply q with scale and accumulate */ + acc = _mm256_fmadd_ps( d, q, acc ); + } // Return horizontal sum of the acc vector __m128 res = _mm256_extractf128_ps( acc, 1 ); @@ -2047,13 +2492,12 @@ static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * rest __m128i i32[2]; for (int j = 0; j < 2; ++j) { // Load 8 bytes, and unpack 4 bit fields into bytes, making 16 bytes - __m128i bx = bytesFromNibbles( x[i].qs + 8*j ); - __m128i by = bytesFromNibbles( y[i].qs + 8*j ); + __m128i bx = bytes_from_nibbles_16(x[i].qs + 8*j); + __m128i by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16*j)); // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. const __m128i off = _mm_set1_epi8( 8 ); bx = _mm_sub_epi8( bx, off ); - by = _mm_sub_epi8( by, off ); // Get absolute values of x vectors const __m128i ax = _mm_sign_epi8(bx, bx); @@ -2081,86 +2525,6 @@ static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * rest res = _mm_add_ss( res, _mm_movehdup_ps( res ) ); sumf = _mm_cvtss_f32( res ); -#elif defined(__wasm_simd128__) - // wasm simd - float sum0 = 0.0f; - float sum1 = 0.0f; - - for (int i = 0; i < nb; i += 2) { - const block_q4_0 * restrict x0 = &px[i + 0]; - const block_q4_0 * restrict y0 = &py[i + 0]; - const block_q4_0 * restrict x1 = &px[i + 1]; - const block_q4_0 * restrict y1 = &py[i + 1]; - - const v128_t m4b = wasm_u8x16_splat(0xf); - const v128_t s8b = wasm_i8x16_splat(0x8); - - const v128_t v0_0 = wasm_v128_load(x0.qs); - const v128_t v0_1 = wasm_v128_load(y0.qs); - const v128_t v1_0 = wasm_v128_load(x1.qs); - const v128_t v1_1 = wasm_v128_load(y1.qs); - - // 4-bit -> 8-bit - const v128_t v0_0l = wasm_v128_and(v0_0, m4b); - const v128_t v1_0l = wasm_v128_and(v1_0, m4b); - - const v128_t v0_0h = wasm_u8x16_shr(v0_0, 4); - const v128_t v1_0h = wasm_u8x16_shr(v1_0, 4); - - const v128_t v0_1l = wasm_v128_and(v0_1, m4b); - const v128_t v1_1l = wasm_v128_and(v1_1, m4b); - - const v128_t v0_1h = wasm_u8x16_shr(v0_1, 4); - const v128_t v1_1h = wasm_u8x16_shr(v1_1, 4); - - // sub 8 - const v128_t v0_0ls = wasm_i8x16_sub(v0_0l, s8b); - const v128_t v1_0ls = wasm_i8x16_sub(v1_0l, s8b); - - const v128_t v0_0hs = wasm_i8x16_sub(v0_0h, s8b); - const v128_t v1_0hs = wasm_i8x16_sub(v1_0h, s8b); - - const v128_t v0_1ls = wasm_i8x16_sub(v0_1l, s8b); - const v128_t v1_1ls = wasm_i8x16_sub(v1_1l, s8b); - - const v128_t v0_1hs = wasm_i8x16_sub(v0_1h, s8b); - const v128_t v1_1hs = wasm_i8x16_sub(v1_1h, s8b); - - // dot product into int16x8_t - const v128_t pl0l = wasm_i16x8_mul(wasm_i16x8_extend_low_i8x16(v0_0ls), wasm_i16x8_extend_low_i8x16(v1_0ls)); - const v128_t pl0h = wasm_i16x8_mul(wasm_i16x8_extend_high_i8x16(v0_0ls), wasm_i16x8_extend_high_i8x16(v1_0ls)); - - const v128_t ph0l = wasm_i16x8_mul(wasm_i16x8_extend_low_i8x16(v0_0hs), wasm_i16x8_extend_low_i8x16(v1_0hs)); - const v128_t ph0h = wasm_i16x8_mul(wasm_i16x8_extend_high_i8x16(v0_0hs), wasm_i16x8_extend_high_i8x16(v1_0hs)); - - const v128_t pl1l = wasm_i16x8_mul(wasm_i16x8_extend_low_i8x16(v0_1ls), wasm_i16x8_extend_low_i8x16(v1_1ls)); - const v128_t pl1h = wasm_i16x8_mul(wasm_i16x8_extend_high_i8x16(v0_1ls), wasm_i16x8_extend_high_i8x16(v1_1ls)); - - const v128_t ph1l = wasm_i16x8_mul(wasm_i16x8_extend_low_i8x16(v0_1hs), wasm_i16x8_extend_low_i8x16(v1_1hs)); - const v128_t ph1h = wasm_i16x8_mul(wasm_i16x8_extend_high_i8x16(v0_1hs), wasm_i16x8_extend_high_i8x16(v1_1hs)); - - const v128_t pl_0 = wasm_i16x8_add(pl0l, pl0h); - const v128_t ph_0 = wasm_i16x8_add(ph0l, ph0h); - - const v128_t pl_1 = wasm_i16x8_add(pl1l, pl1h); - const v128_t ph_1 = wasm_i16x8_add(ph1l, ph1h); - - const v128_t p_0 = wasm_i16x8_add(pl_0, ph_0); - const v128_t p_1 = wasm_i16x8_add(pl_1, ph_1); - - sum0 += x0->d * y0->d * ( - wasm_i16x8_extract_lane(p_0, 0) + wasm_i16x8_extract_lane(p_0, 1) + - wasm_i16x8_extract_lane(p_0, 2) + wasm_i16x8_extract_lane(p_0, 3) + - wasm_i16x8_extract_lane(p_0, 4) + wasm_i16x8_extract_lane(p_0, 5) + - wasm_i16x8_extract_lane(p_0, 6) + wasm_i16x8_extract_lane(p_0, 7)); - sum1 += x1->d * y1->d * ( - wasm_i16x8_extract_lane(p_1, 0) + wasm_i16x8_extract_lane(p_1, 1) + - wasm_i16x8_extract_lane(p_1, 2) + wasm_i16x8_extract_lane(p_1, 3) + - wasm_i16x8_extract_lane(p_1, 4) + wasm_i16x8_extract_lane(p_1, 5) + - wasm_i16x8_extract_lane(p_1, 6) + wasm_i16x8_extract_lane(p_1, 7)); - } - - sumf = sum0 + sum1; #else // scalar for (int i = 0; i < nb; i++) { @@ -2168,96 +2532,159 @@ static void ggml_vec_dot_q4_0(const int n, float * restrict s, const void * rest const float d1 = y[i].d; const uint8_t * restrict p0 = x[i].qs; - const uint8_t * restrict p1 = y[i].qs; + const int8_t * restrict p1 = y[i].qs; - for (int j = 0; j < QK/2; j++) { + int sumi = 0; + for (int j = 0; j < QK8_0/2; j++) { const uint8_t v0 = p0[j]; - const uint8_t v1 = p1[j]; - const float f0 = d0*((int8_t) (v0 & 0xf) - 8); - const float f1 = d0*((int8_t) (v0 >> 4) - 8); + const int i0 = (int8_t) (v0 & 0xf) - 8; + const int i1 = (int8_t) (v0 >> 4) - 8; - const float f2 = d1*((int8_t) (v1 & 0xf) - 8); - const float f3 = d1*((int8_t) (v1 >> 4) - 8); + const int i2 = p1[2*j + 0]; + const int i3 = p1[2*j + 1]; - sumf += f0*f2 + f1*f3; + sumi += i0*i2 + i1*i3; } + sumf += d0*d1*sumi; } #endif *s = sumf; } -static void ggml_vec_dot_q4_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK; +static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int nb = n / QK8_0; + + assert(n % QK8_0 == 0); + assert(nb % 2 == 0); const block_q4_1 * restrict x = vx; - const block_q4_1 * restrict y = vy; + const block_q8_0 * restrict y = vy; float sumf = 0.0; -#if defined(__AVX2__) + // TODO: add AVX / WASM SIMD / etc +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + for (int i = 0; i < nb; i += 2) { + const block_q4_1 * restrict x0 = &x[i + 0]; + const block_q4_1 * restrict x1 = &x[i + 1]; + const block_q8_0 * restrict y0 = &y[i + 0]; + const block_q8_0 * restrict y1 = &y[i + 1]; + + const uint8x16_t m4b = vdupq_n_u8(0xf); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); + + // 4-bit -> 8-bit + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + + // interleave + const int8x16_t v1_0ls = vuzp1q_s8(v1_0l, v1_0h); + const int8x16_t v1_0hs = vuzp2q_s8(v1_0l, v1_0h); + const int8x16_t v1_1ls = vuzp1q_s8(v1_1l, v1_1h); + const int8x16_t v1_1hs = vuzp2q_s8(v1_1l, v1_1h); + + const int16x8_t s0i = vaddq_s16( + vaddq_s16(vmovl_s8(vget_low_s8(v1_0ls)), vmovl_s8(vget_high_s8(v1_0ls))), + vaddq_s16(vmovl_s8(vget_low_s8(v1_0hs)), vmovl_s8(vget_high_s8(v1_0hs)))); + + const int16x8_t s1i = vaddq_s16( + vaddq_s16(vmovl_s8(vget_low_s8(v1_1ls)), vmovl_s8(vget_high_s8(v1_1ls))), + vaddq_s16(vmovl_s8(vget_low_s8(v1_1hs)), vmovl_s8(vget_high_s8(v1_1hs)))); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddl_s16(vget_low_s16(s0i), vget_high_s16(s0i))), x0->m*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddl_s16(vget_low_s16(s1i), vget_high_s16(s1i))), x1->m*y1->d); + +#if defined(__ARM_FEATURE_DOTPROD) + // dot product into int32x4_t + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0ls), v0_0h, v1_0hs); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1ls), v0_1h, v1_1hs); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0ls)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0ls)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0hs)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0hs)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1ls)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1ls)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1hs)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1hs)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d); +#endif + } + + sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#elif defined(__AVX2__) // Initialize accumulator with zeros __m256 acc = _mm256_setzero_ps(); - // Accumulator for constant offsets - float acc_offset = 0.0f; // Main loop for (int i = 0; i < nb; ++i) { const float * d0 = &x[i].d; const float * d1 = &y[i].d; - const float * m0 = &x[i].m; - const float * m1 = &y[i].m; const __m256 d0v = _mm256_broadcast_ss( d0 ); const __m256 d1v = _mm256_broadcast_ss( d1 ); const __m256 m0v = _mm256_broadcast_ss( m0 ); - const __m256 m1v = _mm256_broadcast_ss( m1 ); - // Compute combined scale for the block - const __m256 scale_01 = _mm256_mul_ps( d0v, d1v ); - - // Compute cross scales for the block - const __m256 scale_0 = _mm256_mul_ps( d0v, m1v ); - const __m256 scale_1 = _mm256_mul_ps( m0v, d1v ); - const __m256 cross_scales = _mm256_blend_ps( scale_0, scale_1, 0xAA /* 0b10101010 */ ); + // Compute combined scales + const __m256 d0d1 = _mm256_mul_ps( d0v, d1v ); + const __m256 d1m0 = _mm256_mul_ps( d1v, m0v ); // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes - __m256i bx = bytesFromNibbles( x[i].qs ); - __m256i by = bytesFromNibbles( y[i].qs ); + const __m256i bx = bytes_from_nibbles_32(x[i].qs); + const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs ); - // Now we have a vector with bytes in [ 0 .. 15 ] interval. + // Get absolute values of x vectors + const __m256i ax = _mm256_sign_epi8( bx, bx ); - // Sign-extend first 16 signed bytes into int16_t - __m256i x16 = _mm256_cvtepi8_epi16( _mm256_castsi256_si128( bx ) ); - __m256i y16 = _mm256_cvtepi8_epi16( _mm256_castsi256_si128( by ) ); - // Compute products of int16_t integers, add pairwise - __m256i i32 = _mm256_madd_epi16( x16, y16 ); + // Sign the values of the y vectors + const __m256i sy = _mm256_sign_epi8( by, bx ); - // Sign-extend last 16 signed bytes into int16_t vectors - __m256i x16_h = _mm256_cvtepi8_epi16( _mm256_extracti128_si256( bx, 1 ) ); - __m256i y16_h = _mm256_cvtepi8_epi16( _mm256_extracti128_si256( by, 1 ) ); - // Accumulate products of int16_t integers - i32 = _mm256_add_epi32( i32, _mm256_madd_epi16( x16_h, y16_h ) ); + // Perform multiplication and create 16-bit values + const __m256i dot = _mm256_maddubs_epi16( ax, sy ); + const __m256i ones = _mm256_set1_epi16( 1 ); + const __m256i xy_q = _mm256_madd_epi16( ones, dot ); - // compute sums of unsigned bytes in bx, by in blocks of 8. - // This results in a layout like X100 0000 X200 0000 X300 0000 X400 0000, - // which we then interleave as X100 Y100 X200 Y200 X300 Y300 X400 Y400. - // so if we then cast to 8 singles, we get 8 floats like [ x0_7, y0_7, x8_15, y8_15, x16_23, y16_23, x24_31, y24_31 ] - __m256i xsumi = _mm256_sad_epu8( bx, _mm256_setzero_si256() ); - __m256i ysumi = _mm256_sad_epu8( by, _mm256_setzero_si256() ); - __m256i sumsi = _mm256_or_si256( xsumi, _mm256_slli_si256( ysumi, 4 ) ); - __m256 sums = _mm256_cvtepi32_ps( sumsi ); + // Convert to vector of 8 int32_t to 8 floats + const __m256 xy = _mm256_cvtepi32_ps( xy_q ); - // Convert int32_t to float - __m256 p = _mm256_cvtepi32_ps( i32 ); - // Apply the scale, and accumulate - // acc += d0*d1*x*y + d0*m1*x + d1*m0*y - acc = _mm256_fmadd_ps( scale_01, p, acc ); - acc = _mm256_fmadd_ps( cross_scales, sums, acc ); - // acc_offset += m0*m1 (for each entry in the block) - acc_offset += (*m0)*(*m1); + // Accumulate d0*d1*x*y + acc = _mm256_fmadd_ps( d0d1, xy, acc ); + + // Compute sum of y values + const __m256i y16_l = _mm256_cvtepi8_epi16( _mm256_castsi256_si128( by ) ); + const __m256i y16_h = _mm256_cvtepi8_epi16( _mm256_extracti128_si256( by, 1 ) ); + const __m256i ysumi = _mm256_madd_epi16( _mm256_add_epi16(y16_l, y16_h), ones ); + const __m256 ysum = _mm256_cvtepi32_ps( ysumi ); + + // Accumulate d1*m0*y + acc = _mm256_fmadd_ps( d1m0, ysum, acc ); } // Return horizontal sum of the acc vector @@ -2266,67 +2693,26 @@ static void ggml_vec_dot_q4_1(const int n, float * restrict s, const void * rest res = _mm_add_ps( res, _mm_movehl_ps( res, res ) ); res = _mm_add_ss( res, _mm_movehdup_ps( res ) ); - sumf = _mm_cvtss_f32( res ) + acc_offset * QK; -#elif defined(__ARM_NEON) - float sum00 = 0.0f; - float sum01 = 0.0f; - float sum10 = 0.0f; - float sum11 = 0.0f; - - for (int i = 0; i < nb; ++i) { - const block_q4_1 * restrict x0 = &x[i + 0]; - const block_q4_1 * restrict y0 = &y[i + 0]; - - const uint8x16_t m4b = vdupq_n_u8(0xf); - - const uint8x16_t v0_0 = vld1q_u8(x0->qs); - const uint8x16_t v1_0 = vld1q_u8(y0->qs); - - // and with 0xf - const uint8x16_t v0_0l = vandq_u8(v0_0, m4b); - const uint8x16_t v1_0l = vandq_u8(v1_0, m4b); - - const uint8x16_t v0_0h = vshrq_n_u8(v0_0, 4); - const uint8x16_t v1_0h = vshrq_n_u8(v1_0, 4); - - // dot product into uint16x8_t - const uint16x8_t pl0l = vmull_u8(vget_low_u8 (v0_0l), vget_low_u8 (v1_0l)); - const uint16x8_t pl0h = vmull_u8(vget_high_u8(v0_0l), vget_high_u8(v1_0l)); - - const uint16x8_t ph0l = vmull_u8(vget_low_u8 (v0_0h), vget_low_u8 (v1_0h)); - const uint16x8_t ph0h = vmull_u8(vget_high_u8(v0_0h), vget_high_u8(v1_0h)); - - const uint16x8_t pl0 = vaddq_u16(pl0l, pl0h); - const uint16x8_t ph0 = vaddq_u16(ph0l, ph0h); - - sum00 += x0->m*y0->m; - sum01 += y0->m*x0->d*(vaddvq_u8(v0_0l) + vaddvq_u8(v0_0h)); - sum10 += x0->m*y0->d*(vaddvq_u8(v1_0l) + vaddvq_u8(v1_0h)); - sum11 += x0->d*y0->d*vaddvq_u16(vaddq_u16(pl0, ph0)); - } - - sumf = QK*sum00 + sum01 + sum10 + sum11; + sumf = _mm_cvtss_f32( res ); #else // scalar for (int i = 0; i < nb; i++) { const float d0 = x[i].d; + const float m0 = x[i].m; const float d1 = y[i].d; - const float m0 = x[i].m; - const float m1 = y[i].m; - const uint8_t * restrict p0 = x[i].qs; - const uint8_t * restrict p1 = y[i].qs; + const int8_t * restrict p1 = y[i].qs; - for (int j = 0; j < QK/2; j++) { + // TODO: this is very slow .. + for (int j = 0; j < QK8_0/2; j++) { const uint8_t v0 = p0[j]; - const uint8_t v1 = p1[j]; const float f0 = d0*(v0 & 0xf) + m0; const float f1 = d0*(v0 >> 4) + m0; - const float f2 = d1*(v1 & 0xf) + m1; - const float f3 = d1*(v1 >> 4) + m1; + const float f2 = d1*p1[2*j + 0]; + const float f3 = d1*p1[2*j + 1]; sumf += f0*f2 + f1*f3; } @@ -2336,6 +2722,330 @@ static void ggml_vec_dot_q4_1(const int n, float * restrict s, const void * rest *s = sumf; } +static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int nb = n / QK8_0; + + assert(n % QK8_0 == 0); + assert(nb % 2 == 0); + assert(QK8_0 == 2*QK4_2); + + const block_q4_2 * restrict x = vx; + const block_q8_0 * restrict y = vy; + + float sumf = 0.0; + +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + for (int i = 0; i < nb; i += 2) { + const block_q4_2 * restrict x0_0 = &x[2*(i + 0) + 0]; + const block_q4_2 * restrict x0_1 = &x[2*(i + 0) + 1]; + const block_q4_2 * restrict x1_0 = &x[2*(i + 1) + 0]; + const block_q4_2 * restrict x1_1 = &x[2*(i + 1) + 1]; + + const block_q8_0 * restrict y0 = &y[i + 0]; + const block_q8_0 * restrict y1 = &y[i + 1]; + + const uint8x16_t m4b = vdupq_n_u8(0xf); + const int8x16_t s8b = vdupq_n_s8(0x8); + + const uint8x16_t v0_0 = vcombine_u8(vld1_u8(x0_0->qs), vld1_u8(x0_1->qs)); + const uint8x16_t v0_1 = vcombine_u8(vld1_u8(x1_0->qs), vld1_u8(x1_1->qs)); + + // 4-bit -> 8-bit + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // sub 8 + const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b); + const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b); + const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b); + const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b); + + // interleave + const int8x16_t v0_0lz = vzip1q_s8(v0_0ls, v0_0hs); + const int8x16_t v0_0hz = vzip2q_s8(v0_0ls, v0_0hs); + const int8x16_t v0_1lz = vzip1q_s8(v0_1ls, v0_1hs); + const int8x16_t v0_1hz = vzip2q_s8(v0_1ls, v0_1hs); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + +#if defined(__ARM_FEATURE_DOTPROD) + sumv0 = vmlaq_n_f32(sumv0, vaddq_f32( + vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l)), GGML_FP16_TO_FP32(x0_0->d)), + vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0hz, v1_0h)), GGML_FP16_TO_FP32(x0_1->d))), y0->d); + + sumv1 = vmlaq_n_f32(sumv1, vaddq_f32( + vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l)), GGML_FP16_TO_FP32(x1_0->d)), + vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1hz, v1_1h)), GGML_FP16_TO_FP32(x1_1->d))), y1->d); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); + + sumv0 = vmlaq_n_f32(sumv0, vaddq_f32( + vmulq_n_f32(vcvtq_f32_s32(pl0), GGML_FP16_TO_FP32(x0_0->d)), + vmulq_n_f32(vcvtq_f32_s32(ph0), GGML_FP16_TO_FP32(x0_1->d))), y0->d); + + sumv1 = vmlaq_n_f32(sumv1, vaddq_f32( + vmulq_n_f32(vcvtq_f32_s32(pl1), GGML_FP16_TO_FP32(x1_0->d)), + vmulq_n_f32(vcvtq_f32_s32(ph1), GGML_FP16_TO_FP32(x1_1->d))), y1->d); +#endif + } + + sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + // Main loop + for (int i = 0; i < nb; i++) { + /* Compute combined scale for the block */ + const __m128 d0 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 0].d)); + const __m128 d1 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 1].d)); + const __m256 d = _mm256_mul_ps(_mm256_set_m128(d1, d0), _mm256_broadcast_ss(&y[i].d)); + + __m128i bx0 = bytes_from_nibbles_16(x[2*i + 0].qs); + __m128i bx1 = bytes_from_nibbles_16(x[2*i + 1].qs); + __m256i bx = _mm256_set_m128i(bx1, bx0); + + // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. + const __m256i off = _mm256_set1_epi8(8); + bx = _mm256_sub_epi8(bx, off); + + __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + // Get absolute values of x vectors + const __m256i ax = _mm256_sign_epi8(bx, bx); + // Sign the values of the y vectors + const __m256i sy = _mm256_sign_epi8(by, bx); + // Perform multiplication and create 16-bit values + const __m256i dot = _mm256_maddubs_epi16(ax, sy); + + const __m256i ones = _mm256_set1_epi16(1); + __m256i xy_q = _mm256_madd_epi16(ones, dot); + + /* Convert to vectore of 8 int32_t to 8 floats */ + __m256 q = _mm256_cvtepi32_ps(xy_q); + + /* Multiply q with scale and accumulate */ + acc = _mm256_fmadd_ps(d, q, acc); + } + + // Return horizontal sum of the acc vector + __m128 res = _mm256_extractf128_ps(acc, 1); + res = _mm_add_ps(res, _mm256_castps256_ps128(acc)); + res = _mm_add_ps(res, _mm_movehl_ps(res, res)); + res = _mm_add_ss(res, _mm_movehdup_ps(res)); + + sumf = _mm_cvtss_f32(res); +#else + // scalar + for (int i = 0; i < nb; i++) { + const uint8_t * restrict x0 = x[2*i + 0].qs; + const uint8_t * restrict x1 = x[2*i + 1].qs; + const int8_t * restrict y0 = y[i].qs; + + const float d0 = GGML_FP16_TO_FP32(x[2*i + 0].d); + const float d1 = GGML_FP16_TO_FP32(x[2*i + 1].d); + + int sumi_0 = 0; + int sumi_1 = 0; + + for (int j = 0; j < QK8_0/4; j++) { + const uint8_t v0 = x0[j]; + const uint8_t v1 = x1[j]; + + const int i0_0 = (int8_t) (v0 & 0xf) - 8; + const int i1_0 = (int8_t) (v0 >> 4) - 8; + + const int i0_1 = (int8_t) (v1 & 0xf) - 8; + const int i1_1 = (int8_t) (v1 >> 4) - 8; + + const int i2_0 = y0[2*j + 0]; + const int i3_0 = y0[2*j + 1]; + + const int i2_1 = y0[2*(j + QK8_0/4) + 0]; + const int i3_1 = y0[2*(j + QK8_0/4) + 1]; + + sumi_0 += i0_0*i2_0 + i1_0*i3_0; + sumi_1 += i0_1*i2_1 + i1_1*i3_1; + } + + sumf += (d0 * y[i].d) * sumi_0; + sumf += (d1 * y[i].d) * sumi_1; + } +#endif + + *s = sumf; +} + +static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int nb = n / QK8_0; + + assert(n % QK8_0 == 0); + assert(nb % 2 == 0); + assert(QK8_0 == 2*QK4_2); + + const block_q4_3 * restrict x = vx; + const block_q8_0 * restrict y = vy; + + float sumf = 0.0; + +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + for (int i = 0; i < nb; i += 2) { + const block_q4_3 * restrict x0_0 = &x[2*(i + 0) + 0]; + const block_q4_3 * restrict x0_1 = &x[2*(i + 0) + 1]; + const block_q4_3 * restrict x1_0 = &x[2*(i + 1) + 0]; + const block_q4_3 * restrict x1_1 = &x[2*(i + 1) + 1]; + + const block_q8_0 * restrict y0 = &y[i + 0]; + const block_q8_0 * restrict y1 = &y[i + 1]; + + const uint8x16_t m4b = vdupq_n_u8(0xf); + + const float x0_0d = GGML_FP16_TO_FP32(x0_0->d); + const float x0_1d = GGML_FP16_TO_FP32(x0_1->d); + const float x1_0d = GGML_FP16_TO_FP32(x1_0->d); + const float x1_1d = GGML_FP16_TO_FP32(x1_1->d); + + const float x0_0m = GGML_FP16_TO_FP32(x0_0->m); + const float x0_1m = GGML_FP16_TO_FP32(x0_1->m); + const float x1_0m = GGML_FP16_TO_FP32(x1_0->m); + const float x1_1m = GGML_FP16_TO_FP32(x1_1->m); + + const uint8x16_t v0_0 = vcombine_u8(vld1_u8(x0_0->qs), vld1_u8(x0_1->qs)); + const uint8x16_t v0_1 = vcombine_u8(vld1_u8(x1_0->qs), vld1_u8(x1_1->qs)); + + // 4-bit -> 8-bit + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // interleave + const int8x16_t v0_0lz = vzip1q_s8(v0_0l, v0_0h); + const int8x16_t v0_0hz = vzip2q_s8(v0_0l, v0_0h); + const int8x16_t v0_1lz = vzip1q_s8(v0_1l, v0_1h); + const int8x16_t v0_1hz = vzip2q_s8(v0_1l, v0_1h); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + + const int16x8_t sy0_0 = vaddq_s16(vmovl_s8(vget_low_s8(v1_0l)), vmovl_s8(vget_high_s8(v1_0l))); + const int16x8_t sy0_1 = vaddq_s16(vmovl_s8(vget_low_s8(v1_0h)), vmovl_s8(vget_high_s8(v1_0h))); + + const int16x8_t sy1_0 = vaddq_s16(vmovl_s8(vget_low_s8(v1_1l)), vmovl_s8(vget_high_s8(v1_1l))); + const int16x8_t sy1_1 = vaddq_s16(vmovl_s8(vget_low_s8(v1_1h)), vmovl_s8(vget_high_s8(v1_1h))); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy0_0), vget_high_s16(sy0_0))), x0_0m*y0->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy0_1), vget_high_s16(sy0_1))), x0_1m*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy1_0), vget_high_s16(sy1_0))), x1_0m*y1->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy1_1), vget_high_s16(sy1_1))), x1_1m*y1->d); + +#if defined(__ARM_FEATURE_DOTPROD) + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l)), x0_0d*y0->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0hz, v1_0h)), x0_1d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l)), x1_0d*y1->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1hz, v1_1h)), x1_1d*y1->d); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(pl0), x0_0d*y0->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(ph0), x0_1d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(pl1), x1_0d*y1->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(ph1), x1_1d*y1->d); +#endif + } + + sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#else + // scalar + for (int i = 0; i < nb; i++) { + const uint8_t * restrict x0 = x[2*i + 0].qs; + const uint8_t * restrict x1 = x[2*i + 1].qs; + const int8_t * restrict y0 = y[i].qs; + + const float d0 = GGML_FP16_TO_FP32(x[2*i + 0].d); + const float m0 = GGML_FP16_TO_FP32(x[2*i + 0].m); + const float d1 = GGML_FP16_TO_FP32(x[2*i + 1].d); + const float m1 = GGML_FP16_TO_FP32(x[2*i + 1].m); + + int sy_0 = 0; + int sy_1 = 0; + + int sxy_0 = 0; + int sxy_1 = 0; + + for (int j = 0; j < QK8_0/4; j++) { + const uint8_t v0 = x0[j]; + const uint8_t v1 = x1[j]; + + const int x0_0 = v0 & 0xf; + const int x1_0 = v0 >> 4; + + const int x0_1 = v1 & 0xf; + const int x1_1 = v1 >> 4; + + const int y0_0 = y0[2*j + 0]; + const int y1_0 = y0[2*j + 1]; + + const int y0_1 = y0[2*(j + QK8_0/4) + 0]; + const int y1_1 = y0[2*(j + QK8_0/4) + 1]; + + sy_0 += y0_0 + y1_0; + sy_1 += y0_1 + y1_1; + + sxy_0 += x0_0*y0_0 + x1_0*y1_0; + sxy_1 += x0_1*y0_1 + x1_1*y1_1; + } + + sumf += (d0*sxy_0 + m0*sy_0)*y[i].d; + sumf += (d1*sxy_1 + m1*sy_1)*y[i].d; + } +#endif + + *s = sumf; +} + + // compute GGML_VEC_DOT_UNROLL dot products at once // xs - x row stride in bytes inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * restrict s, void * restrict xv, ggml_fp16_t * restrict y) { @@ -2578,29 +3288,61 @@ inline static void ggml_vec_norm_inv_f32(const int n, float * s, const float * x // static const int GGML_BLCK_SIZE[GGML_TYPE_COUNT] = { - QK, - QK, - 1, - 1, - 1, - 1, - 1, + [GGML_TYPE_F32] = 1, + [GGML_TYPE_F16] = 1, + [GGML_TYPE_Q4_0] = QK4_0, + [GGML_TYPE_Q4_1] = QK4_1, + [GGML_TYPE_Q4_2] = QK4_2, + [GGML_TYPE_Q4_3] = QK4_3, + [GGML_TYPE_Q8_0] = QK8_0, + [GGML_TYPE_I8] = 1, + [GGML_TYPE_I16] = 1, + [GGML_TYPE_I32] = 1, }; - -static_assert(GGML_TYPE_COUNT == 7, "GGML_TYPE_COUNT != 5"); +static_assert(GGML_TYPE_COUNT == 10, "GGML_BLCK_SIZE is outdated"); static const size_t GGML_TYPE_SIZE[GGML_TYPE_COUNT] = { - sizeof(block_q4_0), - sizeof(block_q4_1), - sizeof(int8_t ), - sizeof(int16_t), - sizeof(int32_t), - sizeof(ggml_fp16_t), - sizeof(float ), + [GGML_TYPE_F32] = sizeof(float), + [GGML_TYPE_F16] = sizeof(ggml_fp16_t), + [GGML_TYPE_Q4_0] = sizeof(block_q4_0), + [GGML_TYPE_Q4_1] = sizeof(block_q4_1), + [GGML_TYPE_Q4_2] = sizeof(block_q4_2), + [GGML_TYPE_Q4_3] = sizeof(block_q4_3), + [GGML_TYPE_Q8_0] = sizeof(block_q8_0), + [GGML_TYPE_I8] = sizeof(int8_t), + [GGML_TYPE_I16] = sizeof(int16_t), + [GGML_TYPE_I32] = sizeof(int32_t), }; +static_assert(GGML_TYPE_COUNT == 10, "GGML_TYPE_SIZE is outdated"); -// don't forget to update the array above when adding new types -static_assert(GGML_TYPE_COUNT == 7, "GGML_TYPE_COUNT != 5"); + +static const char * GGML_TYPE_NAME[GGML_TYPE_COUNT] = { + [GGML_TYPE_F32] = "f32", + [GGML_TYPE_F16] = "f16", + [GGML_TYPE_Q4_0] = "q4_0", + [GGML_TYPE_Q4_1] = "q4_1", + [GGML_TYPE_Q4_2] = "q4_2", + [GGML_TYPE_Q4_3] = "q4_3", + [GGML_TYPE_Q8_0] = "q8_0", + [GGML_TYPE_I8] = "i8", + [GGML_TYPE_I16] = "i16", + [GGML_TYPE_I32] = "i32", +}; +static_assert(GGML_TYPE_COUNT == 10, "GGML_TYPE_NAME is outdated"); + +static bool GGML_IS_QUANTIZED[GGML_TYPE_COUNT] = { + [GGML_TYPE_F32] = false, + [GGML_TYPE_F16] = false, + [GGML_TYPE_Q4_0] = true, + [GGML_TYPE_Q4_1] = true, + [GGML_TYPE_Q4_2] = true, + [GGML_TYPE_Q4_3] = true, + [GGML_TYPE_Q8_0] = true, + [GGML_TYPE_I8] = false, + [GGML_TYPE_I16] = false, + [GGML_TYPE_I32] = false, +}; +static_assert(GGML_TYPE_COUNT == 10, "GGML_IS_QUANTIZED is outdated"); static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { "NONE", @@ -2629,6 +3371,7 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { "SCALE", "CPY", + "CONT", "RESHAPE", "VIEW", "PERMUTE", @@ -2642,9 +3385,12 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { "FLASH_ATTN", "FLASH_FF", + + "MAP_UNARY", + "MAP_BINARY", }; -static_assert(GGML_OP_COUNT == 35, "GGML_OP_COUNT != 35"); +static_assert(GGML_OP_COUNT == 38, "GGML_OP_COUNT != 38"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -2673,6 +3419,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "x*v", "x-\\>y", + "cont(x)", "reshape(x)", "view(x)", "permute(x)", @@ -2686,24 +3433,12 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "flash_attn(x)", "flash_ff(x)", + + "f(x)", + "f(x,y)", }; -static_assert(GGML_OP_COUNT == 35, "GGML_OP_COUNT != 35"); - -// -// ggml object -// - -struct ggml_object { - size_t offs; - size_t size; - - struct ggml_object * next; - - char padding[8]; -}; - -static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object); +static_assert(GGML_OP_COUNT == 38, "GGML_OP_COUNT != 38"); static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN"); static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN"); @@ -2716,7 +3451,6 @@ struct ggml_context { size_t mem_size; void * mem_buffer; bool mem_buffer_owned; - bool mem_buffer_mlocked; bool no_alloc; int n_objects; @@ -2834,6 +3568,11 @@ float ggml_type_sizef(enum ggml_type type) { return ((float)(GGML_TYPE_SIZE[type]))/GGML_BLCK_SIZE[type]; } +const char * ggml_type_name(enum ggml_type type) { + return GGML_TYPE_NAME[type]; +} + + size_t ggml_element_size(const struct ggml_tensor * tensor) { return GGML_TYPE_SIZE[tensor->type]; } @@ -2865,6 +3604,10 @@ static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct (t0->ne[3] == t1->ne[3]); } +bool ggml_is_quantized(enum ggml_type type) { + return GGML_IS_QUANTIZED[type]; +} + static inline bool ggml_is_transposed(const struct ggml_tensor * tensor) { return tensor->nb[0] > tensor->nb[1]; } @@ -2975,6 +3718,11 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f); } + // initialize cuBLAS + #if defined(GGML_USE_CUBLAS) + init_cublas(); + #endif + is_first_call = false; } @@ -2999,11 +3747,12 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { return NULL; } + const size_t mem_size = (params.mem_size + GGML_MEM_ALIGN - 1) & ~(GGML_MEM_ALIGN - 1); + *ctx = (struct ggml_context) { - /*.mem_size =*/ params.mem_size, - /*.mem_buffer =*/ params.mem_buffer ? params.mem_buffer : malloc(params.mem_size), + /*.mem_size =*/ mem_size, + /*.mem_buffer =*/ params.mem_buffer ? params.mem_buffer : GGML_ALIGNED_MALLOC(mem_size), /*.mem_buffer_owned =*/ params.mem_buffer ? false : true, - /*.mem_buffer_mlocked =*/ false, /*.no_alloc =*/ params.no_alloc, /*.n_objects =*/ 0, /*.objects_begin =*/ NULL, @@ -3012,7 +3761,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { /*.scratch_save =*/ { 0, 0, NULL, }, }; - GGML_ASSERT(ctx->mem_buffer != NULL); // check for allocation failure + GGML_ASSERT(ctx->mem_buffer != NULL); ggml_assert_aligned(ctx->mem_buffer); @@ -3036,16 +3785,8 @@ void ggml_free(struct ggml_context * ctx) { GGML_PRINT_DEBUG("%s: context %d with %d objects has been freed. memory used = %zu\n", __func__, i, ctx->n_objects, ctx->objects_end->offs + ctx->objects_end->size); -#if GGML_MLOCK_SUPPORT - if (ctx->mem_buffer_mlocked) { - if (munlock(ctx->mem_buffer, ctx->mem_size)) { - fprintf(stderr, "%s: failed to munlock buffer: %s\n", __func__, strerror(errno)); - } - } -#endif - if (ctx->mem_buffer_owned) { - free(ctx->mem_buffer); + GGML_ALIGNED_FREE(ctx->mem_buffer); } found = true; @@ -3072,48 +3813,6 @@ size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch) return result; } -#ifdef __APPLE__ -#define MLOCK_SUGGESTION \ - "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \ - "decreasing 'vm.global_no_user_wire_amount'. Also try increasing RLIMIT_MLOCK (ulimit -l).\n" -#else -#define MLOCK_SUGGESTION \ - "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n" -#endif - -bool ggml_mlock_supported(void) { - return GGML_MLOCK_SUPPORT; -} - -bool ggml_mlock( - struct ggml_context * ctx, - const void *opt_extra_addr, - size_t opt_extra_len, - char **err_p) { - // TODO: Use SetProcessWorkingSetSize() + VirtualLock() on WIN32 -#if GGML_MLOCK_SUPPORT - if (ctx->mem_buffer_mlocked) { - return true; - } - if (mlock(ctx->mem_buffer, ctx->mem_size) || - (opt_extra_len && - mlock(opt_extra_addr, opt_extra_len))) { - if ((*err_p = malloc(1024))) { - snprintf(*err_p, 1024, - "failed to mlock %zu-byte buffer: %s\n" MLOCK_SUGGESTION, - ctx->mem_size + opt_extra_len, - strerror(errno)); - } - return false; - } - ctx->mem_buffer_mlocked = true; - return true; -#else // GGML_MLOCK_SUPPORT - *err_p = strdup("can't mlock because it's not supported on this system"); - return false; -#endif // GGML_MLOCK_SUPPORT -} - //////////////////////////////////////////////////////////////////////////////// struct ggml_tensor * ggml_new_tensor_impl( @@ -3219,7 +3918,8 @@ struct ggml_tensor * ggml_new_tensor_impl( /*.pad =*/ { 0 }, }; - ggml_assert_aligned(result->data); + // TODO: this should not be needed as long as we don't rely on aligned SIMD loads + //ggml_assert_aligned(result->data); for (int i = 0; i < n_dims; i++) { result->ne[i] = ne[i]; @@ -3324,14 +4024,6 @@ struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value) { char * const data = tensor->data; switch (tensor->type) { - case GGML_TYPE_Q4_0: - { - GGML_ASSERT(false); - } break; - case GGML_TYPE_Q4_1: - { - GGML_ASSERT(false); - } break; case GGML_TYPE_I8: { assert(tensor->nb[0] == sizeof(int8_t)); @@ -3367,7 +4059,7 @@ struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value) { ggml_vec_set_f32(nc, (float *)(data + i*n1), value); } } break; - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -3384,14 +4076,6 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) { char * const data = tensor->data; switch (tensor->type) { - case GGML_TYPE_Q4_0: - { - GGML_ASSERT(false); - } break; - case GGML_TYPE_Q4_1: - { - GGML_ASSERT(false); - } break; case GGML_TYPE_I8: { assert(tensor->nb[0] == sizeof(int8_t)); @@ -3427,7 +4111,7 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) { ggml_vec_set_f32(nc, (float *)(data + i*n1), value); } } break; - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -3438,14 +4122,6 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) { int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) { switch (tensor->type) { - case GGML_TYPE_Q4_0: - { - GGML_ASSERT(false); - } break; - case GGML_TYPE_Q4_1: - { - GGML_ASSERT(false); - } break; case GGML_TYPE_I8: { GGML_ASSERT(tensor->nb[0] == sizeof(int8_t)); @@ -3471,7 +4147,7 @@ int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) { GGML_ASSERT(tensor->nb[0] == sizeof(float)); return ((float *)(tensor->data))[i]; } break; - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -3482,14 +4158,6 @@ int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) { void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) { switch (tensor->type) { - case GGML_TYPE_Q4_0: - { - GGML_ASSERT(false); - } break; - case GGML_TYPE_Q4_1: - { - GGML_ASSERT(false); - } break; case GGML_TYPE_I8: { GGML_ASSERT(tensor->nb[0] == sizeof(int8_t)); @@ -3515,7 +4183,7 @@ void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) { GGML_ASSERT(tensor->nb[0] == sizeof(float)); ((float *)(tensor->data))[i] = value; } break; - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -3524,14 +4192,6 @@ void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) { float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) { switch (tensor->type) { - case GGML_TYPE_Q4_0: - { - GGML_ASSERT(false); - } break; - case GGML_TYPE_Q4_1: - { - GGML_ASSERT(false); - } break; case GGML_TYPE_I8: { GGML_ASSERT(tensor->nb[0] == sizeof(int8_t)); @@ -3557,7 +4217,7 @@ float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) { GGML_ASSERT(tensor->nb[0] == sizeof(float)); return ((float *)(tensor->data))[i]; } break; - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -3568,14 +4228,6 @@ float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) { void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) { switch (tensor->type) { - case GGML_TYPE_Q4_0: - { - GGML_ASSERT(false); - } break; - case GGML_TYPE_Q4_1: - { - GGML_ASSERT(false); - } break; case GGML_TYPE_I8: { GGML_ASSERT(tensor->nb[0] == sizeof(int8_t)); @@ -3601,7 +4253,7 @@ void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) { GGML_ASSERT(tensor->nb[0] == sizeof(float)); ((float *)(tensor->data))[i] = value; } break; - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -3620,7 +4272,14 @@ float * ggml_get_data_f32(const struct ggml_tensor * tensor) { struct ggml_tensor * ggml_view_tensor( struct ggml_context * ctx, const struct ggml_tensor * src) { - return ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src->data); + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src->data); + + result->nb[0] = src->nb[0]; + result->nb[1] = src->nb[1]; + result->nb[2] = src->nb[2]; + result->nb[3] = src->nb[3]; + + return result; } //////////////////////////////////////////////////////////////////////////////// @@ -4380,6 +5039,41 @@ struct ggml_tensor * ggml_cpy_inplace( return ggml_cpy_impl(ctx, a, b, true); } +// ggml_cont + +struct ggml_tensor * ggml_cont_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + bool inplace) { + bool is_node = false; + + if (!inplace && a->grad) { + GGML_ASSERT(false); // TODO: implement backward + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_CONT; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = NULL; + + return result; +} + +struct ggml_tensor * ggml_cont( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_cont_impl(ctx, a, false); +} + +struct ggml_tensor * ggml_cont_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_cont_impl(ctx, a, true); +} + // ggml_reshape struct ggml_tensor * ggml_reshape( @@ -4510,6 +5204,37 @@ struct ggml_tensor * ggml_view_2d( return result; } +// ggml_view_3d + +struct ggml_tensor * ggml_view_3d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + size_t nb1, + size_t nb2, + size_t offset) { + if (a->grad) { + GGML_ASSERT(false); // gradient propagation is not supported + } + + const int64_t ne[GGML_MAX_DIMS] = { ne0, ne1, ne2, 1 }; + + struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, (char *) a->data + offset); + + result->nb[1] = nb1; + result->nb[2] = nb2; + result->nb[3] = result->nb[2]*ne2; + + result->op = GGML_OP_VIEW; + result->grad = NULL; + result->src0 = a; + result->src1 = NULL; // TODO: maybe store the offset here? + + return result; +} + // ggml_permute struct ggml_tensor * ggml_permute( @@ -4827,6 +5552,90 @@ struct ggml_tensor * ggml_flash_ff( return result; } +// ggml_map_unary + +struct ggml_tensor * ggml_map_unary_impl_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + const ggml_unary_op_f32_t fun, + bool inplace) { + bool is_node = false; + + if (!inplace && a->grad) { + is_node = true; + } + + struct ggml_tensor * addr_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(void *) / sizeof(int32_t)); + *((void (**)(void))addr_tensor->data) = (void (*)(void))fun; + struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_MAP_UNARY; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->opt[0] = addr_tensor; + + return result; +} + +struct ggml_tensor * ggml_map_unary_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + const ggml_unary_op_f32_t fun) { + return ggml_map_unary_impl_f32(ctx, a, fun, false); +} + +struct ggml_tensor * ggml_map_unary_inplace_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + const ggml_unary_op_f32_t fun) { + return ggml_map_unary_impl_f32(ctx, a, fun, true); +} + +// ggml_map_binary + +struct ggml_tensor * ggml_map_binary_impl_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + const ggml_binary_op_f32_t fun, + bool inplace) { + GGML_ASSERT(ggml_are_same_shape(a, b)); + + bool is_node = false; + + if (!inplace && (a->grad || b->grad)) { + is_node = true; + } + + struct ggml_tensor * addr_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(void *) / sizeof(int32_t)); + *((void (**)(void))addr_tensor->data) = (void (*)(void))fun; + struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_MAP_BINARY; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src0 = a; + result->src1 = b; + result->opt[0] = addr_tensor; + + return result; +} + +struct ggml_tensor * ggml_map_binary_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + const ggml_binary_op_f32_t fun) { + return ggml_map_binary_impl_f32(ctx, a, b, fun, false); +} + +struct ggml_tensor * ggml_map_binary_inplace_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + const ggml_binary_op_f32_t fun) { + return ggml_map_binary_impl_f32(ctx, a, b, fun, true); +} + //////////////////////////////////////////////////////////////////////////////// void ggml_set_param( @@ -4844,8 +5653,6 @@ static void ggml_compute_forward_dup_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, struct ggml_tensor * dst) { - GGML_ASSERT(params->ith == 0); - GGML_ASSERT(ggml_is_contiguous(dst)); GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { @@ -4857,90 +5664,287 @@ static void ggml_compute_forward_dup_f16( const int64_t ne02 = src0->ne[2]; const int64_t ne03 = src0->ne[3]; + const int64_t ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t ne3 = dst->ne[3]; + const size_t nb00 = src0->nb[0]; const size_t nb01 = src0->nb[1]; const size_t nb02 = src0->nb[2]; const size_t nb03 = src0->nb[3]; - if (ggml_is_contiguous(src0) && src0->type == dst->type) { - memcpy(dst->data, src0->data, ggml_nelements(dst) * GGML_TYPE_SIZE[src0->type]); + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) { + // parallelize by elements + const int ne = ggml_nelements(dst); + const int dr = (ne + nth - 1) / nth; + const int ie0 = dr * ith; + const int ie1 = MIN(ie0 + dr, ne); + + memcpy( + ((char *) dst->data + ie0*nb0), + ((char *) src0->data + ie0*nb00), + (ie1 - ie0) * GGML_TYPE_SIZE[src0->type]); + return; } - if (src0->nb[0] == sizeof(ggml_fp16_t)) { - if (dst->type == GGML_TYPE_F16) { - size_t id = 0; - const size_t rs = ne00*nb00; + // parallelize by rows + const int nr = ne01; + // number of 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); - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; - char * dst_ptr = (char *) dst->data + id*rs; - - memcpy(dst_ptr, src0_ptr, rs); - - id++; - } + if (src0->type == dst->type && + ne00 == ne0 && + nb00 == GGML_TYPE_SIZE[src0->type] && nb0 == GGML_TYPE_SIZE[dst->type]) { + // copy by rows + const size_t rs = ne00*nb00; + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ir0; i01 < ir1; i01++) { + memcpy( + ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), + rs); } } - } else if (dst->type == GGML_TYPE_F32) { - size_t id = 0; - float * dst_ptr = (float *) dst->data; + } + return; + } - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + // TODO: add more special-case implementations for tensor shapes/strides that can benefit from memcpy - dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr); - id++; + if (ggml_is_contiguous(dst)) { + if (nb00 == sizeof(ggml_fp16_t)) { + if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + const size_t rs = ne00 * nb00; + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, rs); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + for (int i00 = 0; i00 < ne00; i00++) { + dst_ptr[id] = GGML_FP16_TO_FP32(src0_ptr[i00]); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (ggml_is_quantized(dst->type)) { + quantize_row_q_t const quantize_row_q = quantize_fns[dst->type].quantize_row_q; + float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith; + + size_t id = 0; + size_t rs = nb0 * (ne00 / GGML_BLCK_SIZE[dst->type]); + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + + for (int i00 = 0; i00 < ne00; i00++) { + src0_f32[i00] = GGML_FP16_TO_FP32(src0_ptr[i00]); + } + + quantize_row_q(src0_f32, dst_ptr + id, ne00); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else { + GGML_ASSERT(false); // TODO: implement + } + } else { + //printf("%s: this is not optimal - fix me\n", __func__); + + if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = *src0_ptr; + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else { + GGML_ASSERT(false); // TODO: implement + } + } + return; + } + + // dst counters + int64_t i10 = 0; + int64_t i11 = 0; + int64_t i12 = 0; + int64_t i13 = 0; + + if (dst->type == GGML_TYPE_F16) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + memcpy(dst_ptr, src0_ptr, sizeof(ggml_fp16_t)); + + if (++i10 == ne00) { + i10 = 0; + if (++i11 == ne01) { + i11 = 0; + if (++i12 == ne02) { + i12 = 0; + if (++i13 == ne03) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else if (dst->type == GGML_TYPE_F32) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + *(float *) dst_ptr = GGML_FP16_TO_FP32(*(const ggml_fp16_t *) src0_ptr); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } } } } } - } else { - GGML_ASSERT(false); // TODO: implement } } else { - //printf("%s: this is not optimal - fix me\n", __func__); - - if (dst->type == GGML_TYPE_F32) { - size_t id = 0; - float * dst_ptr = (float *) dst->data; - - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - - dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr); - id++; - } - } - } - } - } else if (dst->type == GGML_TYPE_F16) { - size_t id = 0; - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; - - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - - dst_ptr[id] = *src0_ptr; - id++; - } - } - } - } - } else { - GGML_ASSERT(false); // TODO: implement - } + GGML_ASSERT(false); // TODO: implement } } @@ -4948,8 +5952,6 @@ static void ggml_compute_forward_dup_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, struct ggml_tensor * dst) { - GGML_ASSERT(params->ith == 0); - GGML_ASSERT(ggml_is_contiguous(dst)); GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { @@ -4961,90 +5963,283 @@ static void ggml_compute_forward_dup_f32( const int64_t ne02 = src0->ne[2]; const int64_t ne03 = src0->ne[3]; + const int64_t ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t ne3 = dst->ne[3]; + const size_t nb00 = src0->nb[0]; const size_t nb01 = src0->nb[1]; const size_t nb02 = src0->nb[2]; const size_t nb03 = src0->nb[3]; - if (ggml_is_contiguous(src0) && src0->type == dst->type) { - memcpy(dst->data, src0->data, ggml_nelements(dst) * GGML_TYPE_SIZE[src0->type]); + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + const size_t nb2 = dst->nb[2]; + const size_t nb3 = dst->nb[3]; + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) { + // parallelize by elements + const int ne = ggml_nelements(dst); + const int dr = (ne + nth - 1) / nth; + const int ie0 = dr * ith; + const int ie1 = MIN(ie0 + dr, ne); + + memcpy( + ((char *) dst->data + ie0*nb0), + ((char *) src0->data + ie0*nb00), + (ie1 - ie0) * GGML_TYPE_SIZE[src0->type]); + return; } - if (src0->nb[0] == sizeof(float)) { - if (dst->type == GGML_TYPE_F32) { - size_t id = 0; - const size_t rs = ne00*nb00; + // parallelize by rows + const int nr = ne01; + // number of 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); - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; - char * dst_ptr = (char *) dst->data + id*rs; - - memcpy(dst_ptr, src0_ptr, rs); - - id++; - } + if (src0->type == dst->type && + ne00 == ne0 && + nb00 == GGML_TYPE_SIZE[src0->type] && nb0 == GGML_TYPE_SIZE[dst->type]) { + // copy by rows + const size_t rs = ne00*nb00; + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ir0; i01 < ir1; i01++) { + memcpy( + ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), + rs); } } - } else if (dst->type == GGML_TYPE_F16) { - size_t id = 0; - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; + } + return; + } - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + if (ggml_is_contiguous(dst)) { + // TODO: simplify + if (nb00 == sizeof(float)) { + if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + const size_t rs = ne00 * nb00; + char * dst_ptr = (char *) dst->data; - dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr); - id++; + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, rs); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (ggml_is_quantized(dst->type)) { + quantize_row_q_t const quantize_row_q = quantize_fns[dst->type].quantize_row_q; + + size_t id = 0; + size_t rs = nb0 * (ne00 / GGML_BLCK_SIZE[dst->type]); + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const float * src0_ptr = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + quantize_row_q(src0_ptr, dst_ptr + id, ne00); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else { + GGML_ASSERT(false); // TODO: implement + } + } else { + //printf("%s: this is not optimal - fix me\n", __func__); + + if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = *src0_ptr; + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else { + GGML_ASSERT(false); // TODO: implement + } + } + + return; + } + + // dst counters + + int64_t i10 = 0; + int64_t i11 = 0; + int64_t i12 = 0; + int64_t i13 = 0; + + if (dst->type == GGML_TYPE_F32) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + memcpy(dst_ptr, src0_ptr, sizeof(float)); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else if (dst->type == GGML_TYPE_F16) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(*(const float *) src0_ptr); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } } } } } - } else { - GGML_ASSERT(false); // TODO: implement } } else { - //printf("%s: this is not optimal - fix me\n", __func__); - - if (dst->type == GGML_TYPE_F32) { - size_t id = 0; - float * dst_ptr = (float *) dst->data; - - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - - dst_ptr[id] = *src0_ptr; - id++; - } - } - } - } - } else if (dst->type == GGML_TYPE_F16) { - size_t id = 0; - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; - - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - - dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr); - id++; - } - } - } - } - } else { - GGML_ASSERT(false); // TODO: implement - } + GGML_ASSERT(false); // TODO: implement } } @@ -5061,12 +6256,7 @@ static void ggml_compute_forward_dup( { ggml_compute_forward_dup_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5105,14 +6295,18 @@ static void ggml_compute_forward_add_f32( GGML_ASSERT(nb00 == sizeof(float)); if (nb10 == sizeof(float)) { - const int j0 = (n/nth)*ith; - const int j1 = ith == nth - 1 ? n : (n/nth)*(ith + 1); - - for (int j = j0; j < j1; j++) { + for (int j = ith; j < n; j += nth) { +#ifdef GGML_USE_ACCELERATE + vDSP_vadd( + (float *) ((char *) src0->data + j*nb01), 1, + (float *) ((char *) src1->data + j*nb11), 1, + (float *) ((char *) dst->data + j*nb1), 1, nc); +#else ggml_vec_add_f32(nc, (float *) ((char *) dst->data + j*nb1), (float *) ((char *) src0->data + j*nb01), (float *) ((char *) src1->data + j*nb11)); +#endif } } else { // src1 is not contiguous @@ -5128,6 +6322,212 @@ static void ggml_compute_forward_add_f32( } } +static void ggml_compute_forward_add_f16_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int ith = params->ith; + const int nth = params->nth; + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + + const size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F16); + + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + if (nb10 == sizeof(float)) { + for (int j = ith; j < n; j += nth) { + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + j*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01); + for (int i = 0; i < nc; i++) { + float * src1_ptr = (float *) ((char *) src1->data + j*nb11 + i*nb10); + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + *src1_ptr); + } + } + } + else { + // src1 is not contiguous + GGML_ASSERT(false); + } +} + +static void ggml_compute_forward_add_f16_f16( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int ith = params->ith; + const int nth = params->nth; + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + + const size_t nb10 = src1->nb[0]; + const size_t nb11 = src1->nb[1]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F16); + + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + if (nb10 == sizeof(ggml_fp16_t)) { + for (int j = ith; j < n; j += nth) { + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + j*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01); + for (int i = 0; i < nc; i++) { + ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + j*nb11 + i*nb10); + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + GGML_FP16_TO_FP32(*src1_ptr)); + } + } + } + else { + // src1 is not contiguous + GGML_ASSERT(false); + } +} + +static void ggml_compute_forward_add_q_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst) { + GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + const int64_t ne03 = src0->ne[3]; + + //const int64_t ne10 = src1->ne[0]; + //const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + const int64_t ne13 = src1->ne[3]; + + //const int64_t ne0 = dst->ne[0]; + //const int64_t ne1 = dst->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t ne3 = dst->ne[3]; + + const int nb00 = src0->nb[0]; + const int nb01 = src0->nb[1]; + const int nb02 = src0->nb[2]; + const int nb03 = src0->nb[3]; + + const int nb10 = src1->nb[0]; + const int nb11 = src1->nb[1]; + const int nb12 = src1->nb[2]; + const int nb13 = src1->nb[3]; + + const int nb0 = dst->nb[0]; + const int nb1 = dst->nb[1]; + const int nb2 = dst->nb[2]; + const int nb3 = dst->nb[3]; + + const int ith = params->ith; + const int nth = params->nth; + + GGML_ASSERT(ne02 == ne12); + GGML_ASSERT(ne03 == ne13); + GGML_ASSERT(ne2 == ne12); + GGML_ASSERT(ne3 == ne13); + + const enum ggml_type type = src0->type; + dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q; + quantize_row_q_t const quantize_row_q = quantize_fns[type].quantize_row_q; + + // we don't support permuted src0 or src1 + GGML_ASSERT(nb00 == (int) GGML_TYPE_SIZE[type]); + GGML_ASSERT(nb10 == sizeof(float)); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + GGML_ASSERT(ggml_is_quantized(src0->type)); + GGML_ASSERT(dst->type == src0->type); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + // total rows in src0 + const int nr = ne01*ne02*ne03; + + // 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 * wdata = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith; + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 indices + const int i03 = ir/(ne02*ne01); + const int i02 = (ir - i03*ne02*ne01)/ne01; + const int i01 = (ir - i03*ne02*ne01 - i02*ne01); + + // src1 and dst are same shape as src0 => same indices + const int i13 = i03; + const int i12 = i02; + const int i11 = i01; + + const int i3 = i03; + const int i2 = i02; + const int i1 = i01; + + void * src0_row = (void *) ((char *) src0->data + (i01*nb01 + i02*nb02 + i03*nb03)); + float * src1_row = (float *)((char *) src1->data + (i11*nb11 + i12*nb12 + i13*nb13)); + void * dst_row = (void *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb0)); + + assert(ne00 % 32 == 0); + + // unquantize row from src0 to temp buffer + dequantize_row_q(src0_row, wdata, ne00); + // add src1 + ggml_vec_acc_f32(ne00, wdata, src1_row); + // quantize row to dst + quantize_row_q(wdata, dst_row, ne00); + } +} + static void ggml_compute_forward_add( const struct ggml_compute_params * params, const struct ggml_tensor * src0, @@ -5138,13 +6538,26 @@ static void ggml_compute_forward_add( { ggml_compute_forward_add_f32(params, src0, src1, dst); } break; + case GGML_TYPE_F16: + { + if (src1->type == GGML_TYPE_F16) { + ggml_compute_forward_add_f16_f16(params, src0, src1, dst); + } + else if (src1->type == GGML_TYPE_F32) { + ggml_compute_forward_add_f16_f32(params, src0, src1, dst); + } + else { + GGML_ASSERT(false); + } + } break; case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + case GGML_TYPE_Q4_2: + case GGML_TYPE_Q4_3: + { + ggml_compute_forward_add_q_f32(params, src0, src1, dst); + } break; + default: { GGML_ASSERT(false); } break; @@ -5190,13 +6603,7 @@ static void ggml_compute_forward_sub( { ggml_compute_forward_sub_f32(params, src0, src1, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5242,13 +6649,7 @@ static void ggml_compute_forward_mul( { ggml_compute_forward_mul_f32(params, src0, src1, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5294,13 +6695,7 @@ static void ggml_compute_forward_div( { ggml_compute_forward_div_f32(params, src0, src1, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5342,13 +6737,7 @@ static void ggml_compute_forward_sqr( { ggml_compute_forward_sqr_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5390,13 +6779,7 @@ static void ggml_compute_forward_sqrt( { ggml_compute_forward_sqrt_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5448,13 +6831,7 @@ static void ggml_compute_forward_sum( { ggml_compute_forward_sum_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5525,13 +6902,7 @@ static void ggml_compute_forward_mean( { ggml_compute_forward_mean_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5589,13 +6960,7 @@ static void ggml_compute_forward_repeat( { ggml_compute_forward_repeat_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5637,13 +7002,7 @@ static void ggml_compute_forward_abs( { ggml_compute_forward_abs_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5685,13 +7044,7 @@ static void ggml_compute_forward_sgn( { ggml_compute_forward_sgn_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5733,13 +7086,7 @@ static void ggml_compute_forward_neg( { ggml_compute_forward_neg_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5781,13 +7128,7 @@ static void ggml_compute_forward_step( { ggml_compute_forward_step_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5829,13 +7170,7 @@ static void ggml_compute_forward_relu( { ggml_compute_forward_relu_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5894,13 +7229,7 @@ static void ggml_compute_forward_gelu( { ggml_compute_forward_gelu_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -5961,13 +7290,7 @@ static void ggml_compute_forward_silu( { ggml_compute_forward_silu_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -6047,13 +7370,7 @@ static void ggml_compute_forward_norm( { ggml_compute_forward_norm_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -6127,13 +7444,7 @@ static void ggml_compute_forward_rms_norm( { ggml_compute_forward_rms_norm_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -6143,7 +7454,7 @@ static void ggml_compute_forward_rms_norm( // ggml_compute_forward_mul_mat -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) // helper function to determine if it is better to use BLAS or not // for large matrices, BLAS is faster static bool ggml_compute_forward_mul_mat_use_blas( @@ -6183,7 +7494,7 @@ static void ggml_compute_forward_mul_mat_f32( const int64_t ne02 = src0->ne[2]; const int64_t ne03 = src0->ne[3]; -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) const int64_t ne10 = src1->ne[0]; #endif const int64_t ne11 = src1->ne[1]; @@ -6240,7 +7551,7 @@ static void ggml_compute_forward_mul_mat_f32( // nb01 >= nb00 - src0 is not transposed // compute by src0 rows -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) { if (params->ith != 0) { return; @@ -6254,6 +7565,21 @@ static void ggml_compute_forward_mul_mat_f32( return; } +#if defined(GGML_USE_CUBLAS) + float *d_X = NULL; + float *d_Y = NULL; + float *d_D = NULL; + const float alpha = 1.0f; + const float beta = 0.0f; + const int x_ne = ne01 * ne10; + const int y_ne = ne11 * ne10; + const int d_ne = ne11 * ne01; + + CUDA_CHECK(cudaMalloc((void **)(&d_X), sizeof(float) * x_ne)); + CUDA_CHECK(cudaMalloc((void **)(&d_Y), sizeof(float) * y_ne)); + CUDA_CHECK(cudaMalloc((void **)(&d_D), sizeof(float) * d_ne)); +#endif + for (int64_t i03 = 0; i03 < ne03; i03++) { for (int64_t i02 = 0; i02 < ne02; i02++) { const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03); @@ -6261,15 +7587,37 @@ static void ggml_compute_forward_mul_mat_f32( float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); +#if defined(GGML_USE_CUBLAS) + // copy data to device + CUDA_CHECK(cudaMemcpyAsync(d_X, x, sizeof(float) * x_ne, cudaMemcpyHostToDevice, cudaStream)); + CUDA_CHECK(cudaMemcpyAsync(d_Y, y, sizeof(float) * y_ne, cudaMemcpyHostToDevice, cudaStream)); + + // compute + CUBLAS_CHECK( + cublasSgemm(cublasH, CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + &alpha, d_X, ne00, + d_Y, ne10, + &beta, d_D, ne01)); + + // copy data to host + CUDA_CHECK(cudaMemcpyAsync(d, d_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream)); +#else // zT = y * xT cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans, ne11, ne01, ne10, 1.0f, y, ne10, - x, ne10, + x, ne00, 0.0f, d, ne01); +#endif } } - +#if defined(GGML_USE_CUBLAS) + CUDA_CHECK(cudaStreamSynchronize(cudaStream)); + CUDA_CHECK(cudaFree(d_X)); + CUDA_CHECK(cudaFree(d_Y)); + CUDA_CHECK(cudaFree(d_D)); +#endif //printf("CBLAS F32 = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3); return; @@ -6399,7 +7747,7 @@ static void ggml_compute_forward_mul_mat_f16_f32( // nb01 >= nb00 - src0 is not transposed // compute by src0 rows -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) { GGML_ASSERT(nb10 == sizeof(float)); @@ -6415,10 +7763,37 @@ static void ggml_compute_forward_mul_mat_f16_f32( return; } - float * const wdata = params->wdata; +#if defined(GGML_USE_CUBLAS) + ggml_fp16_t * const wdata = params->wdata; + float *d_X = NULL; + float *d_Y = NULL; + float *d_D = NULL; + const float alpha = 1.0f; + const float beta = 0.0f; + const int x_ne = ne01 * ne10; + const int y_ne = ne11 * ne10; + const int d_ne = ne11 * ne01; + + CUDA_CHECK(cudaMalloc((void **)(&d_X), sizeof(ggml_fp16_t) * x_ne)); + CUDA_CHECK(cudaMalloc((void **)(&d_Y), sizeof(float) * y_ne)); + CUDA_CHECK(cudaMalloc((void **)(&d_D), sizeof(float) * d_ne)); +#else + float * const wdata = params->wdata; +#endif for (int64_t i03 = 0; i03 < ne03; i03++) { for (int64_t i02 = 0; i02 < ne02; i02++) { +#if defined(GGML_USE_CUBLAS) + // with cuBlAS, instead of converting src0 to fp32, we convert src1 to fp16 + { + size_t id = 0; + for (int64_t i01 = 0; i01 < ne11; ++i01) { + for (int64_t i00 = 0; i00 < ne10; ++i00) { + wdata[id++] = GGML_FP32_TO_FP16(*(float *) ((char *) src1->data + i03*nb13 + i02*nb12 + i01*nb11 + i00*nb10)); + } + } + } +#else { size_t id = 0; for (int64_t i01 = 0; i01 < ne01; ++i01) { @@ -6427,7 +7802,31 @@ static void ggml_compute_forward_mul_mat_f16_f32( } } } +#endif +#if defined(GGML_USE_CUBLAS) + const ggml_fp16_t * x = (ggml_fp16_t *) ((char *) src0->data + i02*nb02 + i03*nb03); + const ggml_fp16_t * y = (ggml_fp16_t *) wdata; + + float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); + + // copy data to device + CUDA_CHECK(cudaMemcpyAsync(d_X, x, sizeof(ggml_fp16_t) * x_ne, cudaMemcpyHostToDevice, cudaStream)); + CUDA_CHECK(cudaMemcpyAsync(d_Y, y, sizeof(ggml_fp16_t) * y_ne, cudaMemcpyHostToDevice, cudaStream)); + + // compute + CUBLAS_CHECK( + cublasGemmEx(cublasH, CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + &alpha, d_X, CUDA_R_16F, ne00, + d_Y, CUDA_R_16F, ne10, + &beta, d_D, CUDA_R_32F, ne01, + CUBLAS_COMPUTE_32F, + CUBLAS_GEMM_DEFAULT)); + + // copy data to host + CUDA_CHECK(cudaMemcpyAsync(d, d_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream)); +#else const float * x = wdata; const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13); @@ -6437,11 +7836,18 @@ static void ggml_compute_forward_mul_mat_f16_f32( cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans, ne11, ne01, ne10, 1.0f, y, ne10, - x, ne10, + x, ne00, 0.0f, d, ne01); +#endif } } +#if defined(GGML_USE_CUBLAS) + CUDA_CHECK(cudaStreamSynchronize(cudaStream)); + CUDA_CHECK(cudaFree(d_X)); + CUDA_CHECK(cudaFree(d_Y)); + CUDA_CHECK(cudaFree(d_D)); +#endif /*printf("CBLAS F16 = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);*/ return; @@ -6525,29 +7931,6 @@ static void ggml_compute_forward_mul_mat_f16_f32( //} } -typedef void (*dequantize_row_q_t)(const void * restrict x, float * restrict y, int k); -typedef void (*quantize_row_q_t)(const float * restrict x, void * restrict y, int k); -typedef void (*vec_dot_q_t)(const int n, float * restrict s, const void * restrict x, const void * restrict y); - -typedef struct { - dequantize_row_q_t dequantize_row_q; - quantize_row_q_t quantize_row_q; - vec_dot_q_t vec_dot_q; -} quantize_fns_t; - -static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = { - [GGML_TYPE_Q4_0] = { - .dequantize_row_q = dequantize_row_q4_0, - .quantize_row_q = quantize_row_q4_0, - .vec_dot_q = ggml_vec_dot_q4_0, - }, - [GGML_TYPE_Q4_1] = { - .dequantize_row_q = dequantize_row_q4_1, - .quantize_row_q = quantize_row_q4_1, - .vec_dot_q = ggml_vec_dot_q4_1, - }, -}; - static void ggml_compute_forward_mul_mat_q_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, @@ -6595,8 +7978,8 @@ static void ggml_compute_forward_mul_mat_q_f32( GGML_ASSERT(ne3 == ne13); const enum ggml_type type = src0->type; - quantize_row_q_t const quantize_row_q = quantize_fns[type].quantize_row_q; - vec_dot_q_t const vec_dot_q = quantize_fns[type].vec_dot_q; + quantize_row_q_t const quantize_row_q_dot = quantize_fns[type].quantize_row_q_dot; + vec_dot_q_t const vec_dot_q = quantize_fns[type].vec_dot_q; // we don't support permuted src0 or src1 GGML_ASSERT(nb00 == (int) GGML_TYPE_SIZE[type]); @@ -6616,7 +7999,7 @@ static void ggml_compute_forward_mul_mat_q_f32( // nb01 >= nb00 - src0 is not transposed // compute by src0 rows -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) { if (params->ith != 0) { return; @@ -6630,11 +8013,55 @@ static void ggml_compute_forward_mul_mat_q_f32( return; } +#if defined(GGML_USE_CUBLAS) + float *d_X = NULL; + float *d_Y = NULL; + float *d_D = NULL; + float *d_Q = NULL; + const float alpha = 1.0f; + const float beta = 0.0f; + const int x_ne = ne01 * ne10; + const int y_ne = ne11 * ne10; + const int d_ne = ne11 * ne01; + + CUDA_CHECK(cudaMalloc((void **)(&d_X), sizeof(float) * x_ne)); + CUDA_CHECK(cudaMalloc((void **)(&d_Y), sizeof(float) * y_ne)); + CUDA_CHECK(cudaMalloc((void **)(&d_D), sizeof(float) * d_ne)); + CUDA_CHECK(cudaMalloc((void **)(&d_Q), GGML_TYPE_SIZE[type] * x_ne / GGML_BLCK_SIZE[type])); + + void (*dequantize_row_q_cuda)(const void * x, float * y, int k, cudaStream_t stream) = NULL; + if (type == GGML_TYPE_Q4_0) { + dequantize_row_q_cuda = dequantize_row_q4_0_cuda; + } + else if (type == GGML_TYPE_Q4_1) { + dequantize_row_q_cuda = dequantize_row_q4_1_cuda; + } + else if (type == GGML_TYPE_Q4_2) { + dequantize_row_q_cuda = dequantize_row_q4_2_cuda; + } + else { + GGML_ASSERT(false); + } +#else float * const wdata = params->wdata; dequantize_row_q_t const dequantize_row_q = quantize_fns[type].dequantize_row_q; +#endif for (int64_t i03 = 0; i03 < ne03; i03++) { for (int64_t i02 = 0; i02 < ne02; i02++) { + const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13); + + float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); + +#if defined(GGML_USE_CUBLAS) + // copy and dequantize on device + CUDA_CHECK( + cudaMemcpyAsync(d_Q, (char *) src0->data + i03*nb03 + i02*nb02, + GGML_TYPE_SIZE[type] * x_ne / GGML_BLCK_SIZE[type], cudaMemcpyHostToDevice, cudaStream)); + + dequantize_row_q_cuda(d_Q, d_X, ne01 * ne00, cudaStream); + CUDA_CHECK(cudaGetLastError()); +#else { size_t id = 0; for (int64_t i01 = 0; i01 < ne01; ++i01) { @@ -6642,21 +8069,42 @@ static void ggml_compute_forward_mul_mat_q_f32( id += ne00; } } - const float * x = wdata; - const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13); +#endif - float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); +#if defined(GGML_USE_CUBLAS) + // copy data to device + CUDA_CHECK(cudaMemcpyAsync(d_Y, y, sizeof(float) * y_ne, cudaMemcpyHostToDevice, cudaStream)); + + // compute + CUBLAS_CHECK( + cublasSgemm(cublasH, CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + &alpha, d_X, ne00, + d_Y, ne10, + &beta, d_D, ne01)); + + // copy data to host + CUDA_CHECK(cudaMemcpyAsync(d, d_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream)); +#else // zT = y * xT cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans, ne11, ne01, ne10, 1.0f, y, ne10, - x, ne10, + x, ne00, 0.0f, d, ne01); +#endif } } +#if defined(GGML_USE_CUBLAS) + CUDA_CHECK(cudaStreamSynchronize(cudaStream)); + CUDA_CHECK(cudaFree(d_X)); + CUDA_CHECK(cudaFree(d_Y)); + CUDA_CHECK(cudaFree(d_D)); + CUDA_CHECK(cudaFree(d_Q)); +#endif //printf("CBLAS = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3); return; @@ -6665,12 +8113,12 @@ static void ggml_compute_forward_mul_mat_q_f32( if (params->type == GGML_TASK_INIT) { char * wdata = params->wdata; - const size_t row_size = ne10*GGML_TYPE_SIZE[type]/GGML_BLCK_SIZE[type]; + const size_t row_size = ne10*GGML_TYPE_SIZE[GGML_TYPE_Q8_0]/GGML_BLCK_SIZE[GGML_TYPE_Q8_0]; for (int64_t i13 = 0; i13 < ne13; ++i13) { for (int64_t i12 = 0; i12 < ne12; ++i12) { for (int64_t i11 = 0; i11 < ne11; ++i11) { - quantize_row_q((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) wdata, ne10); + quantize_row_q_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) wdata, ne10); wdata += row_size; } } @@ -6696,7 +8144,7 @@ static void ggml_compute_forward_mul_mat_q_f32( const int ir1 = MIN(ir0 + dr, nr); void * wdata = params->wdata; - const size_t row_size = ne00*GGML_TYPE_SIZE[type]/GGML_BLCK_SIZE[type]; + const size_t row_size = ne00*GGML_TYPE_SIZE[GGML_TYPE_Q8_0]/GGML_BLCK_SIZE[GGML_TYPE_Q8_0]; for (int ir = ir0; ir < ir1; ++ir) { // src0 indices @@ -6744,6 +8192,9 @@ static void ggml_compute_forward_mul_mat( switch (src0->type) { case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: + case GGML_TYPE_Q4_2: + case GGML_TYPE_Q4_3: + case GGML_TYPE_Q8_0: { ggml_compute_forward_mul_mat_q_f32(params, src0, src1, dst); } break; @@ -6755,42 +8206,11 @@ static void ggml_compute_forward_mul_mat( { ggml_compute_forward_mul_mat_f32(params, src0, src1, dst); } break; - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; } - -#if 0 - if (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_Q4_1) { - static int first = 8; - printf("src0: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", src0->ne[0], src0->ne[1], src0->ne[2]); - printf("src1: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", src1->ne[0], src1->ne[1], src1->ne[2]); - printf("dst: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", dst->ne[0], dst->ne[1], dst->ne[2]); - if (first) { - --first; - } else { - for (int k = 0; k < dst->ne[1]; ++k) { - for (int j = 0; j < dst->ne[0]/16; ++j) { - for (int i = 0; i < 16; ++i) { - printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]); - } - printf("\n"); - } - printf("\n"); - } - printf("\n"); - exit(0); - } - } else { - printf("aaaa src0: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", src0->ne[0], src0->ne[1], src0->ne[2]); - printf("aaaa src1: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", src1->ne[0], src1->ne[1], src1->ne[2]); - printf("aaaa dst: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", dst->ne[0], dst->ne[1], dst->ne[2]); - } -#endif } // ggml_compute_forward_scale @@ -6840,13 +8260,7 @@ static void ggml_compute_forward_scale( { ggml_compute_forward_scale_f32(params, src0, src1, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -6862,6 +8276,15 @@ static void ggml_compute_forward_cpy( ggml_compute_forward_dup(params, src0, dst); } +// ggml_compute_forward_cont + +static void ggml_compute_forward_cont( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + ggml_compute_forward_dup(params, src0, dst); +} + // ggml_compute_forward_reshape static void ggml_compute_forward_reshape( @@ -6998,6 +8421,9 @@ static void ggml_compute_forward_get_rows( switch (src0->type) { case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: + case GGML_TYPE_Q4_2: + case GGML_TYPE_Q4_3: + case GGML_TYPE_Q8_0: { ggml_compute_forward_get_rows_q(params, src0, src1, dst); } break; @@ -7009,10 +8435,7 @@ static void ggml_compute_forward_get_rows( { ggml_compute_forward_get_rows_f32(params, src0, src1, dst); } break; - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -7085,13 +8508,7 @@ static void ggml_compute_forward_diag_mask_inf( { ggml_compute_forward_diag_mask_inf_f32(params, src0, src1, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -7179,13 +8596,7 @@ static void ggml_compute_forward_soft_max( { ggml_compute_forward_soft_max_f32(params, src0, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_F16: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -7199,7 +8610,6 @@ static void ggml_compute_forward_rope_f32( const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { - assert(params->ith == 0); assert(src1->type == GGML_TYPE_I32); assert(ggml_nelements(src1) == 3); @@ -7226,25 +8636,59 @@ static void ggml_compute_forward_rope_f32( assert(nb0 == sizeof(float)); - // TODO: optimize + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + // 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); + + // row index used to determine which thread to use + int ir = 0; + + const float theta_scale = powf(10000.0, -2.0f/n_dims); + + const bool is_neox = mode & 2; + for (int64_t i3 = 0; i3 < ne3; i3++) { - for (int64_t i2 = (mode == 0 ? 0 : n_past); i2 < ne2; i2++) { - const int p = (mode == 0 ? n_past + i2 : i2); + for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) { + const int p = ((mode & 1) == 0 ? n_past + i2 : i2); for (int64_t i1 = 0; i1 < ne1; i1++) { + if (ir++ < ir0) continue; + if (ir > ir1) break; + + float theta = (float)p; + for (int i0 = 0; i0 < n_dims; i0 += 2) { - const float theta = powf(10000.0, ((float)-i0)/n_dims); + const float cos_theta = cosf(theta); + const float sin_theta = sinf(theta); - const float cos_theta = cosf(p*theta); - const float sin_theta = sinf(p*theta); + theta *= theta_scale; - const float * const src = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + if (!is_neox) { + const float * const src = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - const float x0 = src[0]; - const float x1 = src[1]; + const float x0 = src[0]; + const float x1 = src[1]; - dst_data[0] = x0*cos_theta - x1*sin_theta; - dst_data[1] = x0*sin_theta + x1*cos_theta; + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[1] = x0*sin_theta + x1*cos_theta; + } else { + const float * const src = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims/2]; + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta; + } } } } @@ -7256,7 +8700,6 @@ static void ggml_compute_forward_rope_f16( const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { - assert(params->ith == 0); assert(src1->type == GGML_TYPE_I32); assert(ggml_nelements(src1) == 3); @@ -7283,24 +8726,59 @@ static void ggml_compute_forward_rope_f16( assert(nb0 == sizeof(ggml_fp16_t)); + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + // 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); + + // row index used to determine which thread to use + int ir = 0; + + const float theta_scale = powf(10000.0, -2.0f/n_dims); + + const bool is_neox = mode & 2; + for (int64_t i3 = 0; i3 < ne3; i3++) { - for (int64_t i2 = (mode == 0 ? 0 : n_past); i2 < ne2; i2++) { - const int p = (mode == 0 ? n_past + i2 : i2); + for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) { + const int p = ((mode & 1) == 0 ? n_past + i2 : i2); for (int64_t i1 = 0; i1 < ne1; i1++) { + if (ir++ < ir0) continue; + if (ir > ir1) break; + + float theta = (float)p; + for (int i0 = 0; i0 < n_dims; i0 += 2) { - const float theta = powf(10000.0, ((float)-i0)/n_dims); + const float cos_theta = cosf(theta); + const float sin_theta = sinf(theta); - const float cos_theta = cosf(p*theta); - const float sin_theta = sinf(p*theta); + theta *= theta_scale; - const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + if (!is_neox) { + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - const float x0 = ggml_fp16_to_fp32(src[0]); - const float x1 = ggml_fp16_to_fp32(src[1]); + const float x0 = GGML_FP16_TO_FP32(src[0]); + const float x1 = GGML_FP16_TO_FP32(src[1]); - dst_data[0] = ggml_fp32_to_fp16(x0*cos_theta - x1*sin_theta); - dst_data[1] = ggml_fp32_to_fp16(x0*sin_theta + x1*cos_theta); + dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + } else { + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0); + + const float x0 = GGML_FP16_TO_FP32(src[0]); + const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]); + + dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + } } } } @@ -7321,12 +8799,7 @@ static void ggml_compute_forward_rope( { ggml_compute_forward_rope_f32(params, src0, src1, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -7589,12 +9062,7 @@ static void ggml_compute_forward_conv_1d_1s( { ggml_compute_forward_conv_1d_1s_f32(params, src0, src1, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -7857,12 +9325,7 @@ static void ggml_compute_forward_conv_1d_2s( { ggml_compute_forward_conv_1d_2s_f32(params, src0, src1, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -8342,12 +9805,7 @@ static void ggml_compute_forward_flash_attn( { ggml_compute_forward_flash_attn_f32(params, q, k, v, masked, dst); } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_COUNT: + default: { GGML_ASSERT(false); } break; @@ -8553,12 +10011,100 @@ static void ggml_compute_forward_flash_ff( { GGML_ASSERT(false); // TODO } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_COUNT: + default: + { + GGML_ASSERT(false); + } break; + } +} + +// ggml_compute_forward_map_unary + +static void ggml_compute_forward_map_unary_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst, + const ggml_unary_op_f32_t fun) { + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + assert( dst->nb[0] == sizeof(float)); + assert(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + fun(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + + +static void ggml_compute_forward_map_unary( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst, + const ggml_unary_op_f32_t fun) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_map_unary_f32(params, src0, dst, fun); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + +// ggml_compute_forward_map_binary + +static void ggml_compute_forward_map_binary_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst, + const ggml_binary_op_f32_t fun) { + assert(params->ith == 0); + assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + assert( dst->nb[0] == sizeof(float)); + assert(src0->nb[0] == sizeof(float)); + assert(src1->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + fun(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1])), + (float *) ((char *) src1->data + i*(src1->nb[1]))); + } +} + + +static void ggml_compute_forward_map_binary( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + struct ggml_tensor * dst, + const ggml_binary_op_f32_t fun) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_map_binary_f32(params, src0, src1, dst, fun); + } break; + default: { GGML_ASSERT(false); } break; @@ -8659,6 +10205,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_cpy(params, tensor->src0, tensor); } break; + case GGML_OP_CONT: + { + ggml_compute_forward_cont(params, tensor->src0, tensor); + } break; case GGML_OP_RESHAPE: { ggml_compute_forward_reshape(params, tensor->src0, tensor); @@ -8710,6 +10260,18 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_flash_ff(params, tensor->src0, tensor->src1, tensor->opt[0], tensor->opt[1], tensor->opt[2], tensor); } break; + case GGML_OP_MAP_UNARY: + { + const ggml_unary_op_f32_t fun = *((ggml_unary_op_f32_t *)tensor->opt[0]->data); + ggml_compute_forward_map_unary(params, tensor->src0, tensor, fun); + } + break; + case GGML_OP_MAP_BINARY: + { + const ggml_binary_op_f32_t fun = *((ggml_binary_op_f32_t *)tensor->opt[0]->data); + ggml_compute_forward_map_binary(params, tensor->src0, tensor->src1, tensor, fun); + } + break; case GGML_OP_NONE: { // nop @@ -8903,8 +10465,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor src1->grad = ggml_add_impl(ctx, src1->grad, - // TODO: fix transpose, the node will break the graph connections - ggml_mul_mat(ctx, ggml_transpose(ctx, src0), tensor->grad), + ggml_mul_mat(ctx, + ggml_cont(ctx, ggml_transpose(ctx, src0)), + tensor->grad), inplace); } } break; @@ -8916,6 +10479,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { GGML_ASSERT(false); // TODO: not implemented } break; + case GGML_OP_CONT: + { + GGML_ASSERT(false); // TODO: not implemented + } break; case GGML_OP_RESHAPE: { GGML_ASSERT(false); // TODO: not implemented @@ -8964,6 +10531,11 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { GGML_ASSERT(false); // not supported } break; + case GGML_OP_MAP_UNARY: + case GGML_OP_MAP_BINARY: + { + GGML_ASSERT(false); // not supported + } break; case GGML_OP_NONE: { // nop @@ -9054,7 +10626,7 @@ struct ggml_cgraph ggml_build_forward(struct ggml_tensor * tensor) { struct ggml_cgraph result = { /*.n_nodes =*/ 0, /*.n_leafs =*/ 0, - /*.n_threads =*/ 0, + /*.n_threads =*/ GGML_DEFAULT_N_THREADS, /*.work_size =*/ 0, /*.work =*/ NULL, /*.nodes =*/ { NULL }, @@ -9282,13 +10854,29 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) struct ggml_tensor * node = cgraph->nodes[i]; switch (node->op) { + case GGML_OP_CPY: case GGML_OP_DUP: { - node->n_tasks = 1; + node->n_tasks = n_threads; + + size_t cur = 0; + if (ggml_is_quantized(node->type)) { + cur = GGML_TYPE_SIZE[GGML_TYPE_F32] * node->ne[0] * n_threads; + } + + work_size = MAX(work_size, cur); } break; case GGML_OP_ADD: { node->n_tasks = n_threads; + + size_t cur = 0; + + if (ggml_is_quantized(node->src0->type)) { + cur = GGML_TYPE_SIZE[GGML_TYPE_F32] * node->src0->ne[0] * n_threads; + } + + work_size = MAX(work_size, cur); } break; case GGML_OP_SUB: case GGML_OP_MUL: @@ -9333,7 +10921,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) size_t cur = 0; if (node->src0->type == GGML_TYPE_F16 && node->src1->type == GGML_TYPE_F32) { -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) { node->n_tasks = 1; // TODO: this actually is doing nothing // the threads are still spinning @@ -9349,15 +10937,15 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) #endif } else if (node->src0->type == GGML_TYPE_F32 && node->src1->type == GGML_TYPE_F32) { cur = 0; - } else if (quantize_fns[node->src0->type].vec_dot_q && node->src1->type == GGML_TYPE_F32) { -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) + } else if (ggml_is_quantized(node->src0->type) && node->src1->type == GGML_TYPE_F32) { +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) { node->n_tasks = 1; cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]); } else #endif { - cur = GGML_TYPE_SIZE[node->src0->type]*ggml_nelements(node->src1)/GGML_BLCK_SIZE[node->src0->type]; + cur = GGML_TYPE_SIZE[GGML_TYPE_Q8_0]*ggml_nelements(node->src1)/GGML_BLCK_SIZE[GGML_TYPE_Q8_0]; } } else { GGML_ASSERT(false); @@ -9369,7 +10957,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) { node->n_tasks = n_threads; } break; - case GGML_OP_CPY: + case GGML_OP_CONT: case GGML_OP_RESHAPE: case GGML_OP_VIEW: case GGML_OP_PERMUTE: @@ -9385,7 +10973,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) } break; case GGML_OP_ROPE: { - node->n_tasks = 1; + node->n_tasks = n_threads; } break; case GGML_OP_CONV_1D_1S: case GGML_OP_CONV_1D_2S: @@ -9455,6 +11043,11 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph) work_size = MAX(work_size, cur); } break; + case GGML_OP_MAP_UNARY: + case GGML_OP_MAP_BINARY: + { + node->n_tasks = 1; + } break; case GGML_OP_NONE: { node->n_tasks = 1; @@ -9673,8 +11266,8 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) { GGML_PRINT("=== GRAPH ===\n"); - GGML_PRINT_DEBUG("n_threads = %d\n", cgraph->n_threads); - GGML_PRINT_DEBUG("total work size = %zu bytes\n",cgraph->work_size); + GGML_PRINT_DEBUG("n_threads = %d\n", cgraph->n_threads); + GGML_PRINT_DEBUG("total work size = %zu bytes\n", cgraph->work_size); GGML_PRINT("n_nodes = %d\n", cgraph->n_nodes); for (int i = 0; i < cgraph->n_nodes; i++) { @@ -10526,16 +12119,16 @@ enum ggml_opt_result ggml_opt( //////////////////////////////////////////////////////////////////////////////// size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist) { - assert(k % QK == 0); - const int nb = k / QK; + assert(k % QK4_0 == 0); + const int nb = k / QK4_0; for (int j = 0; j < n; j += k) { - block_q4_0 * restrict y = (block_q4_0 *)dst + j/QK; + block_q4_0 * restrict y = (block_q4_0 *)dst + j/QK4_0; quantize_row_q4_0_reference(src + j, y, k); for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK; l += 2) { + for (int l = 0; l < QK4_0; l += 2) { const uint8_t vi0 = y[i].qs[l/2] & 0xF; const uint8_t vi1 = y[i].qs[l/2] >> 4; @@ -10545,20 +12138,20 @@ size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * } } - return (n/QK*sizeof(block_q4_0)); + return (n/QK4_0*sizeof(block_q4_0)); } size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist) { - assert(k % QK == 0); - const int nb = k / QK; + assert(k % QK4_1 == 0); + const int nb = k / QK4_1; for (int j = 0; j < n; j += k) { - block_q4_1 * restrict y = (block_q4_1 *)dst + j/QK; + block_q4_1 * restrict y = (block_q4_1 *)dst + j/QK4_1; quantize_row_q4_1_reference(src + j, y, k); for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK; l += 2) { + for (int l = 0; l < QK4_1; l += 2) { const uint8_t vi0 = y[i].qs[l/2] & 0xF; const uint8_t vi1 = y[i].qs[l/2] >> 4; @@ -10568,7 +12161,87 @@ size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * } } - return (n/QK*sizeof(block_q4_1)); + return (n/QK4_1*sizeof(block_q4_1)); +} + +size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % QK4_2 == 0); + const int nb = k / QK4_2; + + for (int j = 0; j < n; j += k) { + block_q4_2 * restrict y = (block_q4_2 *)dst + j/QK4_2; + + //quantize_row_q4_2_reference(src + j, y, k); + quantize_row_q4_2_rmse(src + j, y, k); + + for (int i = 0; i < nb; i++) { + for (int l = 0; l < QK4_2; l += 2) { + const uint8_t vi0 = y[i].qs[l/2] & 0xF; + const uint8_t vi1 = y[i].qs[l/2] >> 4; + + hist[vi0]++; + hist[vi1]++; + } + } + } + + return (n/QK4_2*sizeof(block_q4_2)); +} + +size_t ggml_quantize_q4_3(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % QK4_3 == 0); + const int nb = k / QK4_3; + + for (int j = 0; j < n; j += k) { + block_q4_3 * restrict y = (block_q4_3 *)dst + j/QK4_3; + + quantize_row_q4_3_reference(src + j, y, k); + + for (int i = 0; i < nb; i++) { + for (int l = 0; l < QK4_3; l += 2) { + const uint8_t vi0 = y[i].qs[l/2] & 0xF; + const uint8_t vi1 = y[i].qs[l/2] >> 4; + + hist[vi0]++; + hist[vi1]++; + } + } + } + + return (n/QK4_3*sizeof(block_q4_3)); +} + +size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist) { + size_t result = 0; + switch (type) { + case GGML_TYPE_Q4_0: + { + GGML_ASSERT(start % QK4_0 == 0); + block_q4_0 * block = (block_q4_0*)dst + start / QK4_0; + result = ggml_quantize_q4_0(src + start, block, n, n, hist); + } break; + case GGML_TYPE_Q4_1: + { + GGML_ASSERT(start % QK4_1 == 0); + block_q4_1 * block = (block_q4_1*)dst + start / QK4_1; + result = ggml_quantize_q4_1(src + start, block, n, n, hist); + } break; + case GGML_TYPE_Q4_2: + { + GGML_ASSERT(start % QK4_2 == 0); + block_q4_2 * block = (block_q4_2*)dst + start / QK4_2; + result = ggml_quantize_q4_2(src + start, block, n, n, hist); + } break; + case GGML_TYPE_Q4_3: + { + GGML_ASSERT(start % QK4_3 == 0); + block_q4_3 * block = (block_q4_3*)dst + start / QK4_3; + result = ggml_quantize_q4_3(src + start, block, n, n, hist); + } break; + default: + assert(false); + } + return result; } //////////////////////////////////////////////////////////////////////////////// @@ -10597,6 +12270,22 @@ int ggml_cpu_has_avx512(void) { #endif } +int ggml_cpu_has_avx512_vbmi(void) { +#if defined(__AVX512VBMI__) + return 1; +#else + return 0; +#endif +} + +int ggml_cpu_has_avx512_vnni(void) { +#if defined(__AVX512VNNI__) + return 1; +#else + return 0; +#endif +} + int ggml_cpu_has_fma(void) { #if defined(__FMA__) return 1; @@ -10646,7 +12335,15 @@ int ggml_cpu_has_wasm_simd(void) { } int ggml_cpu_has_blas(void) { -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) + return 1; +#else + return 0; +#endif +} + +int ggml_cpu_has_cublas(void) { +#if defined(GGML_USE_CUBLAS) return 1; #else return 0; diff --git a/ggml.h b/ggml.h index ad962b109..460d4ffe0 100644 --- a/ggml.h +++ b/ggml.h @@ -177,11 +177,12 @@ extern "C" { #include #include -#define GGML_MAX_DIMS 4 -#define GGML_MAX_NODES 4096 -#define GGML_MAX_PARAMS 16 -#define GGML_MAX_CONTEXTS 64 -#define GGML_MAX_OPT 4 +#define GGML_MAX_DIMS 4 +#define GGML_MAX_NODES 4096 +#define GGML_MAX_PARAMS 16 +#define GGML_MAX_CONTEXTS 64 +#define GGML_MAX_OPT 4 +#define GGML_DEFAULT_N_THREADS 4 #ifdef __ARM_NEON // we use the built-in 16-bit float type @@ -198,13 +199,17 @@ struct ggml_object; struct ggml_context; enum ggml_type { - GGML_TYPE_Q4_0, - GGML_TYPE_Q4_1, + // explicitly numbered values are used in llama.cpp files + GGML_TYPE_F32 = 0, + GGML_TYPE_F16 = 1, + GGML_TYPE_Q4_0 = 2, + GGML_TYPE_Q4_1 = 3, + GGML_TYPE_Q4_2 = 4, + GGML_TYPE_Q4_3 = 5, + GGML_TYPE_Q8_0 = 6, GGML_TYPE_I8, GGML_TYPE_I16, GGML_TYPE_I32, - GGML_TYPE_F16, - GGML_TYPE_F32, GGML_TYPE_COUNT, }; @@ -236,6 +241,7 @@ enum ggml_op { GGML_OP_SCALE, GGML_OP_CPY, + GGML_OP_CONT, GGML_OP_RESHAPE, GGML_OP_VIEW, GGML_OP_PERMUTE, @@ -250,9 +256,25 @@ enum ggml_op { GGML_OP_FLASH_ATTN, GGML_OP_FLASH_FF, + GGML_OP_MAP_UNARY, + GGML_OP_MAP_BINARY, + GGML_OP_COUNT, }; + +// ggml object +struct ggml_object { + size_t offs; + size_t size; + + struct ggml_object * next; + + char padding[8]; +}; + +static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object); + // n-dimensional tensor struct ggml_tensor { enum ggml_type type; @@ -335,8 +357,12 @@ int ggml_blck_size (enum ggml_type type); size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block float ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float +const char * ggml_type_name(enum ggml_type type); + size_t ggml_element_size(const struct ggml_tensor * tensor); +bool ggml_is_quantized(enum ggml_type type); + struct ggml_context * ggml_init(struct ggml_init_params params); void ggml_free(struct ggml_context * ctx); @@ -344,13 +370,6 @@ size_t ggml_used_mem(const struct ggml_context * ctx); size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch); -bool ggml_mlock_supported(void); -bool ggml_mlock( - struct ggml_context * ctx, - const void *opt_extra_addr, - size_t opt_extra_len, - char **err_p); - struct ggml_tensor * ggml_new_tensor( struct ggml_context * ctx, enum ggml_type type, @@ -415,6 +434,12 @@ struct ggml_tensor * ggml_add( struct ggml_tensor * a, struct ggml_tensor * b); + +struct ggml_tensor * ggml_add_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b); + struct ggml_tensor * ggml_sub( struct ggml_context * ctx, struct ggml_tensor * a, @@ -519,6 +544,11 @@ struct ggml_tensor * ggml_cpy( struct ggml_tensor * a, struct ggml_tensor * b); +// make contiguous +struct ggml_tensor * ggml_cont( + struct ggml_context * ctx, + struct ggml_tensor * a); + // return view(a), b specifies the new shape // TODO: when we start computing gradient, make a copy instead of view struct ggml_tensor * ggml_reshape( @@ -558,6 +588,16 @@ struct ggml_tensor * ggml_view_2d( size_t nb1, // row stride in bytes size_t offset); +struct ggml_tensor * ggml_view_3d( + struct ggml_context * ctx, + struct ggml_tensor * a, + int64_t ne0, + int64_t ne1, + int64_t ne2, + size_t nb1, // row stride in bytes + size_t nb2, // slice stride in bytes + size_t offset); + struct ggml_tensor * ggml_permute( struct ggml_context * ctx, struct ggml_tensor * a, @@ -590,7 +630,8 @@ struct ggml_tensor * ggml_soft_max( // rotary position embedding // in-place, returns view(a) -// if mode == 1, skip n_past elements +// if mode & 1 == 1, skip n_past elements +// if mode & 2 == 1, GPT-NeoX style // TODO: avoid creating a new tensor every time struct ggml_tensor * ggml_rope( struct ggml_context * ctx, @@ -628,6 +669,21 @@ struct ggml_tensor * ggml_flash_ff( struct ggml_tensor * c0, struct ggml_tensor * c1); +// Mapping operations +typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *); +typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *); + +struct ggml_tensor * ggml_map_unary_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + const ggml_unary_op_f32_t fun); + +struct ggml_tensor * ggml_map_binary_f32( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + const ggml_binary_op_f32_t fun); + // // automatic differentiation // @@ -755,6 +811,10 @@ enum ggml_opt_result ggml_opt( size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist); size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist); +size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t * hist); +size_t ggml_quantize_q4_3(const float * src, void * dst, int n, int k, int64_t * hist); + +size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist); // // system info @@ -763,6 +823,8 @@ size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * int ggml_cpu_has_avx(void); int ggml_cpu_has_avx2(void); int ggml_cpu_has_avx512(void); +int ggml_cpu_has_avx512_vbmi(void); +int ggml_cpu_has_avx512_vnni(void); int ggml_cpu_has_fma(void); int ggml_cpu_has_neon(void); int ggml_cpu_has_arm_fma(void); @@ -770,9 +832,35 @@ int ggml_cpu_has_f16c(void); int ggml_cpu_has_fp16_va(void); int ggml_cpu_has_wasm_simd(void); int ggml_cpu_has_blas(void); +int ggml_cpu_has_cublas(void); int ggml_cpu_has_sse3(void); int ggml_cpu_has_vsx(void); + +// +// Internal types and functions exposed for tests and benchmarks +// + +#ifdef __cplusplus +// restrict not standard in C++ +#define GGML_RESTRICT +#else +#define GGML_RESTRICT restrict +#endif +typedef void (*dequantize_row_q_t)(const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); +typedef void (*quantize_row_q_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); +typedef void (*vec_dot_q_t)(const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y); + +typedef struct { + dequantize_row_q_t dequantize_row_q; + quantize_row_q_t quantize_row_q; + quantize_row_q_t quantize_row_q_reference; + quantize_row_q_t quantize_row_q_dot; + vec_dot_q_t vec_dot_q; +} quantize_fns_t; + +quantize_fns_t ggml_internal_get_quantize_fn(size_t i); + #ifdef __cplusplus } #endif diff --git a/llama.cpp b/llama.cpp index 854bb8993..4a646eb91 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1,49 +1,36 @@ +// Defines fileno on msys: +#ifndef _GNU_SOURCE +#define _GNU_SOURCE +#include +#include +#endif + +#include "llama_util.h" #include "llama.h" #include "ggml.h" +#include +#include #include #include #include #include #include #include -#include #include #include - -#if defined(_WIN32) && !defined(_POSIX_MAPPED_FILES) -#define WIN32_LEAN_AND_MEAN -#include -#else -#include -#include -#include -#include -#endif - -#define Min(X, Y) ((Y) > (X) ? (X) : (Y)) -#define Max(X, Y) ((Y) < (X) ? (X) : (Y)) +#include +#include +#include +#include +#include +#include +#include #define LLAMA_USE_SCRATCH #define LLAMA_MAX_SCRATCH_BUFFERS 16 -#define LLAMA_ASSERT(x) \ - do { \ - if (!(x)) { \ - fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \ - abort(); \ - } \ - } while (0) - - -// determine number of model parts based on the dimension -static const std::unordered_map LLAMA_N_PARTS = { - { 4096, 1 }, - { 5120, 2 }, - { 6656, 4 }, - { 8192, 8 }, -}; // available llama models enum e_model { @@ -60,47 +47,67 @@ static const size_t MB = 1024*1024; // TODO: dynamically determine these sizes // needs modifications in ggml -static const std::map MEM_REQ_SCRATCH0 = { - { MODEL_7B, 512ull*MB }, - { MODEL_13B, 512ull*MB }, - { MODEL_30B, 512ull*MB }, - { MODEL_65B, 512ull*MB }, -}; +static const std::map & MEM_REQ_SCRATCH0() +{ + static std::map _MEM_REQ_SCRATCH0 = { + { MODEL_7B, 512ull * MB }, + { MODEL_13B, 512ull * MB }, + { MODEL_30B, 512ull * MB }, + { MODEL_65B, 512ull * MB }, + }; + return _MEM_REQ_SCRATCH0; +} -static const std::map MEM_REQ_SCRATCH1 = { - { MODEL_7B, 512ull*MB }, - { MODEL_13B, 512ull*MB }, - { MODEL_30B, 512ull*MB }, - { MODEL_65B, 512ull*MB }, +static const std::map & MEM_REQ_SCRATCH1() +{ + static std::map _MEM_REQ_SCRATCH1 = { + { MODEL_7B, 512ull * MB }, + { MODEL_13B, 512ull * MB }, + { MODEL_30B, 512ull * MB }, + { MODEL_65B, 512ull * MB }, + }; + return _MEM_REQ_SCRATCH1; }; // 2*n_embd*n_ctx*n_layer*sizeof(float16) -static const std::map MEM_REQ_KV_SELF = { - { MODEL_7B, 1026ull*MB }, - { MODEL_13B, 1608ull*MB }, - { MODEL_30B, 3124ull*MB }, - { MODEL_65B, 5120ull*MB }, +static const std::map & MEM_REQ_KV_SELF() +{ + static std::map _MEM_REQ_KV_SELF = { + { MODEL_7B, 1026ull * MB }, + { MODEL_13B, 1608ull * MB }, + { MODEL_30B, 3124ull * MB }, + { MODEL_65B, 5120ull * MB }, + }; + return _MEM_REQ_KV_SELF; }; // this is mostly needed for temporary mul_mat buffers to dequantize the data // not actually needed if BLAS is disabled -static const std::map MEM_REQ_EVAL = { - { MODEL_7B, 768ull*MB }, - { MODEL_13B, 1024ull*MB }, - { MODEL_30B, 1280ull*MB }, - { MODEL_65B, 1536ull*MB }, +static const std::map & MEM_REQ_EVAL() +{ + static std::map _MEM_REQ_EVAL = { + { MODEL_7B, 768ull * MB }, + { MODEL_13B, 1024ull * MB }, + { MODEL_30B, 1280ull * MB }, + { MODEL_65B, 1536ull * MB }, + }; + return _MEM_REQ_EVAL; }; // default hparams (LLaMA 7B) struct llama_hparams { - int32_t n_vocab = 32000; - int32_t n_ctx = 512; // this is provided as user input? - int32_t n_embd = 4096; - int32_t n_mult = 256; - int32_t n_head = 32; - int32_t n_layer = 32; - int32_t n_rot = 64; - int32_t f16 = 1; + uint32_t n_vocab = 32000; + uint32_t n_ctx = 512; // this is provided as user input? + uint32_t n_embd = 4096; + uint32_t n_mult = 256; + uint32_t n_head = 32; + uint32_t n_layer = 32; + uint32_t n_rot = 64; + enum llama_ftype ftype = LLAMA_FTYPE_MOSTLY_F16; + + bool operator!=(const llama_hparams & other) const { + return memcmp(this, &other, sizeof(llama_hparams)); + } }; struct llama_layer { @@ -126,11 +133,17 @@ struct llama_kv_cache { struct ggml_tensor * k; struct ggml_tensor * v; - struct ggml_context * ctx; + struct ggml_context * ctx = NULL; - std::vector buf; + llama_buffer buf; int n; // number of tokens currently in the cache + + ~llama_kv_cache() { + if (ctx) { + ggml_free(ctx); + } + } }; struct llama_model { @@ -146,22 +159,30 @@ struct llama_model { std::vector layers; // context - struct ggml_context * ctx; + struct ggml_context * ctx = NULL; // key + value cache for the self attention // TODO: move to llama_state struct llama_kv_cache kv_self; // the model memory buffer - std::vector buf; + llama_buffer buf; // model memory mapped file - void * mm_addr = NULL; - uint64_t mm_length = 0; + std::unique_ptr mapping; - // tensors - int n_loaded; - std::unordered_map tensors; + // objects representing data potentially being locked in memory + llama_mlock mlock_buf; + llama_mlock mlock_mmap; + + // for quantize-stats only + std::vector> tensors_by_name; + + ~llama_model() { + if (ctx) { + ggml_free(ctx); + } + } }; struct llama_vocab { @@ -206,8 +227,8 @@ struct llama_context { // memory buffers used to evaluate the model // TODO: move in llama_state - std::vector buf_compute; - std::vector buf_scratch[LLAMA_MAX_SCRATCH_BUFFERS]; + llama_buffer buf_compute; + llama_buffer buf_scratch[LLAMA_MAX_SCRATCH_BUFFERS]; int buf_last = 0; size_t buf_max_size[LLAMA_MAX_SCRATCH_BUFFERS] = { 0 }; @@ -220,11 +241,11 @@ struct llama_context { last_size = ggml_set_scratch(ctx, { 0, 0, nullptr, }); } else { auto & buf = buf_scratch[i]; - last_size = ggml_set_scratch(ctx, { 0, buf.size(), buf.data(), }); + last_size = ggml_set_scratch(ctx, { 0, buf.size, buf.addr, }); } if (buf_last >= 0) { - buf_max_size[buf_last] = Max(buf_max_size[buf_last], last_size); + buf_max_size[buf_last] = std::max(buf_max_size[buf_last], last_size); } buf_last = i; @@ -244,6 +265,504 @@ struct llama_context { } }; +template +static T checked_mul(T a, T b) { + T ret = a * b; + if (a != 0 && ret / a != b) { + throw format("overflow multiplying %llu * %llu", + (unsigned long long) a, (unsigned long long) b); + } + return ret; +} + +static size_t checked_div(size_t a, size_t b) { + if (b == 0 || a % b != 0) { + throw format("error dividing %zu / %zu", a, b); + } + return a / b; +} + +static std::string llama_format_tensor_shape(const std::vector & ne) { + char buf[256]; + snprintf(buf, sizeof(buf), "%5u", ne.at(0)); + for (size_t i = 1; i < ne.size(); i++) { + snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), " x %5u", ne.at(i)); + } + return buf; +} + +static size_t llama_calc_tensor_size(const std::vector & ne, enum ggml_type type) { + size_t size = ggml_type_size(type); + for (uint32_t dim : ne) { + size = checked_mul(size, dim); + } + return size / ggml_blck_size(type); +} + +struct llama_load_tensor_shard { + std::vector ne; + size_t size; + enum ggml_type type; + size_t file_idx; + size_t file_off; + + void calc_size() { + size = llama_calc_tensor_size(ne, type); + } +}; + +enum llama_split_type { + SPLIT_NONE, + SPLIT_BY_COLUMNS, + SPLIT_BY_ROWS +}; + +struct llama_load_tensor { + std::vector shards; + + std::string name; + enum ggml_type type = GGML_TYPE_F32; + llama_split_type split_type = SPLIT_NONE; + std::vector ne; + size_t size; + struct ggml_tensor * ggml_tensor = NULL; + uint8_t * data; + + llama_load_tensor(const std::string & name) : name(name) {} + + void calc_all() { + calc_type(); + calc_split_type(); + calc_ne(); + calc_size(); + } + + void calc_type() { + const auto & first_shard = shards.at(0); + for (const auto & shard : shards) { + if (shard.type != first_shard.type) { + throw format("inconsistent tensor shard type in '%s'", name.c_str()); + } + } + type = first_shard.type; + } + + void calc_split_type() { + if (shards.at(0).ne.size() == 1 || // 1D tensors are just duplicated in every file + shards.size() == 1) { // only one file? + split_type = SPLIT_NONE; + } else if (name.find("tok_embeddings.") == 0 || + name.find(".attention.wo.weight") != std::string::npos || + name.find(".feed_forward.w2.weight") != std::string::npos) { + split_type = SPLIT_BY_COLUMNS; + } else { + split_type = SPLIT_BY_ROWS; + } + } + + void calc_ne() { + const auto & first_shard = shards.at(0); + for (const auto & shard : shards) { + if (shard.ne != first_shard.ne) { + throw format("inconsistent tensor shard shape in '%s': first was %s, other was %s", + name.c_str(), llama_format_tensor_shape(first_shard.ne).c_str(), llama_format_tensor_shape(shard.ne).c_str()); + } + } + ne = first_shard.ne; + LLAMA_ASSERT(shards.size() <= UINT32_MAX); + uint32_t n_shards = (uint32_t) shards.size(); + switch (split_type) { + case SPLIT_NONE: + ne = first_shard.ne; + break; + case SPLIT_BY_COLUMNS: + ne = {checked_mul(first_shard.ne[0], n_shards), + first_shard.ne[1]}; + break; + case SPLIT_BY_ROWS: + ne = {first_shard.ne[0], + checked_mul(first_shard.ne[1], n_shards)}; + break; + } + } + + void calc_size() { + size = llama_calc_tensor_size(ne, type); + } +}; + +struct llama_load_tensors_map { + // tensors is kept in a separate vector to preserve file order + std::vector tensors; + std::unordered_map name_to_idx; +}; + +enum llama_file_version { + LLAMA_FILE_VERSION_GGML, + LLAMA_FILE_VERSION_GGMF_V1, // added version field and scores in vocab + LLAMA_FILE_VERSION_GGJT_V1, // added padding +}; + +struct llama_file_loader { + llama_file file; + llama_file_version file_version; + llama_hparams hparams; + llama_vocab vocab; + + llama_file_loader(const char * fname, size_t file_idx, llama_load_tensors_map & tensors_map) + : file(fname, "rb") { + fprintf(stderr, "llama.cpp: loading model from %s\n", fname); + read_magic(); + read_hparams(); + read_vocab(); + read_tensor_metadata(file_idx, tensors_map); + } + void read_magic() { + uint32_t magic = file.read_u32(); + uint32_t version = 0; + + if (magic != 'ggml') { + version = file.read_u32(); + } + + if (magic == 'ggml' && version == 0) { + file_version = LLAMA_FILE_VERSION_GGML; + } else if (magic == 'ggmf' && version == 1) { + file_version = LLAMA_FILE_VERSION_GGMF_V1; + } else if (magic == 'ggjt' && version == 1) { + file_version = LLAMA_FILE_VERSION_GGJT_V1; + } else { + throw format("unknown (magic, version) combination: %08x, %08x; is this really a GGML file?", + magic, version); + } + } + void read_hparams() { + hparams.n_vocab = file.read_u32(); + hparams.n_embd = file.read_u32(); + hparams.n_mult = file.read_u32(); + hparams.n_head = file.read_u32(); + hparams.n_layer = file.read_u32(); + hparams.n_rot = file.read_u32(); + hparams.ftype = (enum llama_ftype) file.read_u32(); + } + void read_vocab() { + vocab.id_to_token.resize(hparams.n_vocab); + + for (uint32_t i = 0; i < hparams.n_vocab; i++) { + uint32_t len = file.read_u32(); + std::string word = file.read_string(len); + + float score = 0.0f; + if (file_version >= LLAMA_FILE_VERSION_GGMF_V1) { + file.read_raw(&score, sizeof(score)); + } + + vocab.token_to_id[word] = i; + + auto & tok_score = vocab.id_to_token[i]; + tok_score.tok = std::move(word); + tok_score.score = score; + } + } + void read_tensor_metadata(size_t file_idx, llama_load_tensors_map & tensors_map) { + while (file.tell() < file.size) { + llama_load_tensor_shard shard; + uint32_t n_dims = file.read_u32(); + uint32_t name_len = file.read_u32(); + shard.type = (enum ggml_type) file.read_u32(); + shard.ne.resize(n_dims); + file.read_raw(shard.ne.data(), sizeof(shard.ne[0]) * n_dims); + std::string name = file.read_string(name_len); + if (n_dims < 1 || n_dims > 2) { + throw format("llama.cpp: tensor '%s' should not be %u-dimensional", name.c_str(), n_dims); + } + switch (shard.type) { + case GGML_TYPE_F32: + case GGML_TYPE_F16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q4_2: + case GGML_TYPE_Q4_3: + break; + default: { + throw format("unrecognized tensor type %u\n", shard.type); + } + } + + if (file_version >= LLAMA_FILE_VERSION_GGJT_V1) { + // skip to the next multiple of 32 bytes + file.seek(-file.tell() & 31, SEEK_CUR); + } + shard.file_idx = file_idx; + shard.file_off = file.tell(); + + shard.calc_size(); + file.seek(shard.size, SEEK_CUR); + + auto it = tensors_map.name_to_idx.find(name); + size_t idx; + if (it != tensors_map.name_to_idx.end()) { + idx = it->second; + } else { + tensors_map.tensors.emplace_back(name); + idx = tensors_map.tensors.size() - 1; + tensors_map.name_to_idx.emplace(name, idx); + } + tensors_map.tensors.at(idx).shards.push_back(shard); + } + } +}; + +struct llama_file_saver { + llama_file file; + llama_file_loader * any_file_loader; + llama_file_saver(const char * fname, llama_file_loader * any_file_loader, enum llama_ftype new_ftype) + : file(fname, "wb"), any_file_loader(any_file_loader) { + fprintf(stderr, "llama.cpp: saving model to %s\n", fname); + write_magic(); + write_hparams(new_ftype); + write_vocab(); + } + void write_magic() { + file.write_u32('ggjt'); // magic + file.write_u32(1); // version + } + void write_hparams(enum llama_ftype new_ftype) { + const llama_hparams & hparams = any_file_loader->hparams; + file.write_u32(hparams.n_vocab); + file.write_u32(hparams.n_embd); + file.write_u32(hparams.n_mult); + file.write_u32(hparams.n_head); + file.write_u32(hparams.n_layer); + file.write_u32(hparams.n_rot); + file.write_u32(new_ftype); + } + void write_vocab() { + if (any_file_loader->file_version == LLAMA_FILE_VERSION_GGML) { + fprintf(stderr, "llama.cpp: WARNING: input is an old file that doesn't have scores; will add dummy scores\n"); + } + uint32_t n_vocab = any_file_loader->hparams.n_vocab; + for (uint32_t i = 0; i < n_vocab; i++) { + const auto & token_score = any_file_loader->vocab.id_to_token.at(i); + file.write_u32((uint32_t) token_score.tok.size()); + file.write_raw(token_score.tok.data(), token_score.tok.size()); + file.write_raw(&token_score.score, sizeof(token_score.score)); + } + } + void write_tensor(llama_load_tensor & tensor, enum ggml_type new_type, const void * new_data, size_t new_size) { + switch (new_type) { + case GGML_TYPE_F32: + case GGML_TYPE_F16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q4_2: + case GGML_TYPE_Q4_3: + break; + default: LLAMA_ASSERT(false); + } + file.write_u32((uint32_t) tensor.ne.size()); + file.write_u32((uint32_t) tensor.name.size()); + file.write_u32(new_type); + file.write_raw(tensor.ne.data(), sizeof(tensor.ne[0]) * tensor.ne.size()); + file.write_raw(tensor.name.data(), tensor.name.size()); + file.seek(-file.tell() & 31, SEEK_CUR); + LLAMA_ASSERT(new_size == llama_calc_tensor_size(tensor.ne, new_type)); + file.write_raw(new_data, new_size); + } +}; + +struct llama_model_loader { + std::vector> file_loaders; + llama_load_tensors_map tensors_map; + bool use_mmap; + size_t num_ggml_tensors_created = 0; + struct ggml_context * ggml_ctx = NULL; + std::unique_ptr mapping; + + llama_model_loader(const std::string & fname_base, bool use_mmap, bool vocab_only) { + auto first_file = new llama_file_loader(fname_base.c_str(), 0, tensors_map); + file_loaders.emplace_back(first_file); + uint32_t n_parts = vocab_only ? 1 : guess_n_parts(); + for (uint32_t i = 1; i < n_parts; i++) { + std::string fname = fname_base + "." + std::to_string(i); + auto ith_file = new llama_file_loader(fname.c_str(), i, tensors_map); + file_loaders.emplace_back(ith_file); + if (ith_file->hparams != first_file->hparams) { + throw format("llama.cpp: hparams inconsistent between files"); + } + } + if (!llama_mmap::SUPPORTED) { + use_mmap = false; + } + if (use_mmap && alignment_prevents_mmap()) { + fprintf(stderr, "llama.cpp: can't use mmap because tensors are not aligned; convert to new format to avoid this\n"); + use_mmap = false; + } + this->use_mmap = use_mmap; + for (llama_load_tensor & lt : tensors_map.tensors) { + lt.calc_all(); + } + } + + bool alignment_prevents_mmap() { + for (const llama_load_tensor & lt : tensors_map.tensors) { + for (const llama_load_tensor_shard & shard : lt.shards) { + if (shard.file_off & 3) { + return true; + } + } + } + return false; + } + + uint32_t guess_n_parts() const { + auto it = tensors_map.name_to_idx.find("tok_embeddings.weight"); + if (it == tensors_map.name_to_idx.end()) { + throw std::string("missing tok_embeddings.weight"); + } + const llama_load_tensor & lt = tensors_map.tensors.at(it->second); + return file_loaders.at(0)->hparams.n_embd / lt.shards.at(0).ne.at(0); + } + + void calc_sizes(size_t * ctx_size_p, size_t * mmapped_size_p) const { + *ctx_size_p = *mmapped_size_p = 0; + for (const llama_load_tensor & lt : tensors_map.tensors) { + *ctx_size_p += sizeof(struct ggml_tensor) + GGML_OBJECT_SIZE; + *(use_mmap ? mmapped_size_p : ctx_size_p) += lt.size; + } + } + + struct ggml_tensor * get_tensor(const std::string & name, std::vector ne) { + auto it = tensors_map.name_to_idx.find(name); + if (it == tensors_map.name_to_idx.end()) { + throw format("llama.cpp: tensor '%s' is missing from model", name.c_str()); + } + llama_load_tensor & lt = tensors_map.tensors.at(it->second); + if (lt.ne != ne) { + throw format("llama.cpp: tensor '%s' has wrong shape; expected %s, got %s", + name.c_str(), llama_format_tensor_shape(ne).c_str(), llama_format_tensor_shape(lt.ne).c_str()); + } + + return get_tensor_for(lt); + } + + struct ggml_tensor * get_tensor_for(llama_load_tensor & lt) { + struct ggml_tensor * tensor; + if (lt.ne.size() == 2) { + tensor = ggml_new_tensor_2d(ggml_ctx, lt.type, lt.ne.at(0), lt.ne.at(1)); + } else { + LLAMA_ASSERT(lt.ne.size() == 1); + tensor = ggml_new_tensor_1d(ggml_ctx, lt.type, lt.ne.at(0)); + } + LLAMA_ASSERT(lt.ggml_tensor == NULL); // if this fails, we called get_tensor twice on the same tensor + lt.ggml_tensor = tensor; + num_ggml_tensors_created++; + return tensor; + } + + void done_getting_tensors() { + if (num_ggml_tensors_created != tensors_map.tensors.size()) { + throw std::string("llama.cpp: file contained more tensors than expected"); + } + } + + void load_all_data(llama_progress_callback progress_callback, void * progress_callback_user_data, llama_mlock * lmlock) { + size_t data_size = 0; + for (const llama_load_tensor & lt : tensors_map.tensors) { + data_size += lt.size; + } + + if (use_mmap) { + mapping.reset(new llama_mmap(&file_loaders.at(0)->file)); + if (!lmlock) { + // Don't call the callback since the actual loading will be lazy + // and we can't measure it. + progress_callback = NULL; + } + if (lmlock) { + lmlock->init(mapping->addr); + } + } + + size_t done_size = 0; + for (llama_load_tensor & lt : tensors_map.tensors) { + if (progress_callback) { + progress_callback((float) done_size / data_size, progress_callback_user_data); + } + LLAMA_ASSERT(lt.ggml_tensor); // unused tensors should have been caught by load_data already + lt.data = (uint8_t *) lt.ggml_tensor->data; + load_data_for(lt); + lt.ggml_tensor->data = lt.data; + done_size += lt.size; + if (use_mmap && lmlock) { + lmlock->grow_to(done_size); + } + } + if (progress_callback) { + progress_callback(1.0f, progress_callback_user_data); + } + } + + void load_data_for(llama_load_tensor & lt) { + if (use_mmap) { + LLAMA_ASSERT(lt.shards.size() == 1); + lt.data = (uint8_t *) mapping->addr + lt.shards.at(0).file_off; + } else if (lt.split_type == SPLIT_NONE) { + llama_file & file = file_loaders.at(lt.shards.at(0).file_idx)->file; + file.seek(lt.shards.at(0).file_off, SEEK_SET); + file.read_raw(lt.data, lt.size); + } else if (lt.split_type == SPLIT_BY_ROWS) { + size_t offset = 0; + for (llama_load_tensor_shard & shard : lt.shards) { + llama_file & file = file_loaders.at(shard.file_idx)->file; + file.seek(shard.file_off, SEEK_SET); + file.read_raw(lt.data + offset, shard.size); + offset += shard.size; + } + LLAMA_ASSERT(offset == lt.size); + } else if (lt.split_type == SPLIT_BY_COLUMNS) { + // Let's load the data into temporary buffers to ensure the OS performs large loads. + std::vector tmp_bufs; + tmp_bufs.resize(lt.shards.size()); + for (size_t i = 0; i < lt.shards.size(); i++) { + llama_load_tensor_shard & shard = lt.shards.at(i); + llama_file & file = file_loaders.at(shard.file_idx)->file; + file.seek(shard.file_off, SEEK_SET); + tmp_bufs.at(i).resize(shard.size); + file.read_raw(tmp_bufs.at(i).addr, shard.size); + } + // Then reshape. + size_t num_rows = lt.ne.at(1); + size_t per_shard_row_size = lt.shards.at(0).size / num_rows; + size_t out_offset = 0; + for (size_t row = 0; row < num_rows; row++) { + for (llama_buffer & tmp_buf : tmp_bufs) { + memcpy(lt.data + out_offset, + tmp_buf.addr + row * per_shard_row_size, + per_shard_row_size); + out_offset += per_shard_row_size; + } + } + LLAMA_ASSERT(out_offset == lt.size); + } + if (0) { + print_checksum(lt); + } + } + + static void print_checksum(llama_load_tensor & lt) { + uint32_t sum = 0; + for (size_t i = 0; i < lt.size; i++) { + uint8_t byte = lt.data[i]; + sum = byte + (sum << 6) + (sum << 16) - sum; // sdbm hash + } + fprintf(stderr, "%s checksum: %#08x (%s, size %zu)\n", lt.name.c_str(), sum, + llama_format_tensor_shape(lt.ne).c_str(), lt.size); + } + +}; + + // // kv cache // @@ -262,8 +781,8 @@ static bool kv_cache_init( cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB); struct ggml_init_params params; - params.mem_size = cache.buf.size(); - params.mem_buffer = cache.buf.data(); + params.mem_size = cache.buf.size; + params.mem_buffer = cache.buf.addr; params.no_alloc = false; cache.ctx = ggml_init(params); @@ -279,13 +798,6 @@ static bool kv_cache_init( return true; } -static void kv_cache_free(struct llama_kv_cache & cache) { - if (cache.ctx) { - ggml_free(cache.ctx); - cache.ctx = nullptr; - } -} - struct llama_context_params llama_context_default_params() { struct llama_context_params result = { /*.n_ctx =*/ 512, @@ -294,6 +806,7 @@ struct llama_context_params llama_context_default_params() { /*.f16_kv =*/ false, /*.logits_all =*/ false, /*.vocab_only =*/ false, + /*.use_mmap =*/ true, /*.use_mlock =*/ false, /*.embedding =*/ false, /*.progress_callback =*/ nullptr, @@ -303,243 +816,108 @@ struct llama_context_params llama_context_default_params() { return result; } +bool llama_mmap_supported() { + return llama_mmap::SUPPORTED; +} + +bool llama_mlock_supported() { + return llama_mlock::SUPPORTED; +} + // // model loading // -static void *mmap_file(const char *fname, uint64_t *mm_length) { -#if defined(_WIN32) && !defined(_POSIX_MAPPED_FILES) - HANDLE hFile = CreateFileA(fname, - GENERIC_READ, - FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE, - NULL, - OPEN_EXISTING, - FILE_ATTRIBUTE_NORMAL | FILE_ATTRIBUTE_NOT_CONTENT_INDEXED, - NULL); - if (hFile == INVALID_HANDLE_VALUE) return 0; - LARGE_INTEGER fileSize; - fileSize.QuadPart = -1; - GetFileSizeEx(hFile, &fileSize); - int64_t length = fileSize.QuadPart; - HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL); - CloseHandle(hFile); - if (!hMapping) return 0; - void *addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0); - CloseHandle(hMapping); - if (!addr) return 0; -#else - int fd = open(fname, O_RDONLY); - if (fd == -1) return 0; - int64_t length = lseek(fd, 0, SEEK_END); - void *addr = mmap(NULL, length, PROT_READ, MAP_SHARED, fd, 0); - close(fd); - if (addr == MAP_FAILED) return 0; -#endif - *mm_length = length; - return addr; +static const char *llama_file_version_name(llama_file_version version) { + switch (version) { + case LLAMA_FILE_VERSION_GGML: return "'ggml' (old version with low tokenizer quality and no mmap support)"; + case LLAMA_FILE_VERSION_GGMF_V1: return "ggmf v1 (old version with no mmap support)"; + case LLAMA_FILE_VERSION_GGJT_V1: return "ggjt v1 (latest)"; + default: LLAMA_ASSERT(false); + } } -static void munmap_file(void * addr, size_t length) { -#if defined(_WIN32) && !defined(_POSIX_MAPPED_FILES) - UnmapViewOfFile(addr); -#else - munmap(addr, length); -#endif +static const char *llama_ftype_name(enum llama_ftype ftype) { + switch (ftype) { + case LLAMA_FTYPE_ALL_F32: return "all F32"; + case LLAMA_FTYPE_MOSTLY_F16: return "mostly F16"; + case LLAMA_FTYPE_MOSTLY_Q4_0: return "mostly Q4_0"; + case LLAMA_FTYPE_MOSTLY_Q4_1: return "mostly Q4_1"; + case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16: + return "mostly Q4_1, some F16"; + case LLAMA_FTYPE_MOSTLY_Q4_2: return "mostly Q4_2"; + case LLAMA_FTYPE_MOSTLY_Q4_3: return "mostly Q4_3"; + default: return "unknown, may not work"; + } } -static bool report_bad_magic(const char *path, uint32_t got, uint32_t want) { - fprintf(stderr, - "%s: invalid model file (bad magic [got %#x want %#x])\n" - "\tyou most likely need to regenerate your ggml files\n" - "\tthe benefit is you'll get 10-100x faster load times\n" - "\tsee https://github.com/ggerganov/llama.cpp/issues/91\n" - "\tuse convert-pth-to-ggml.py to regenerate from original pth\n" - "\tuse migrate-ggml-2023-03-30-pr613.py if you deleted originals\n", - path, got, want); - return false; +static const char *llama_model_type_name(e_model type) { + switch (type) { + case MODEL_7B: return "7B"; + case MODEL_13B: return "13B"; + case MODEL_30B: return "30B"; + case MODEL_65B: return "65B"; + default: LLAMA_ASSERT(false); + } } -static bool llama_model_load( +static void llama_model_load_internal( const std::string & fname, llama_context & lctx, int n_ctx, - int n_parts, ggml_type memory_type, + bool use_mmap, + bool use_mlock, bool vocab_only, llama_progress_callback progress_callback, - void *progress_callback_user_data) { - fprintf(stderr, "%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str()); + void * progress_callback_user_data) { lctx.t_start_us = ggml_time_us(); + std::unique_ptr ml(new llama_model_loader(fname, use_mmap, vocab_only)); + + lctx.vocab = std::move(ml->file_loaders.at(0)->vocab); auto & model = lctx.model; - auto & vocab = lctx.vocab; + model.hparams = ml->file_loaders.at(0)->hparams; + llama_file_version file_version = ml->file_loaders.at(0)->file_version; + auto & hparams = model.hparams; + uint32_t n_ff = ((2*(4*hparams.n_embd)/3 + hparams.n_mult - 1)/hparams.n_mult)*hparams.n_mult; - auto fin = std::ifstream(fname, std::ios::binary); - if (!fin) { - fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str()); - return false; - } - - std::vector f_buf(1024*1024); - fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size()); - - fin.seekg(0, fin.end); - const size_t file_size = fin.tellg(); - fin.seekg(0); - - // verify magic { - uint32_t magic; - fin.read((char *) &magic, sizeof(magic)); - if (magic == LLAMA_FILE_MAGIC_UNVERSIONED) { - fprintf(stderr, "%s: invalid model file '%s' (too old, regenerate your model files or convert them with convert-unversioned-ggml-to-ggml.py!)\n", - __func__, fname.c_str()); - return false; + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_13B; break; + case 60: model.type = e_model::MODEL_30B; break; + case 80: model.type = e_model::MODEL_65B; break; } - if (magic != LLAMA_FILE_MAGIC) { - return report_bad_magic(fname.c_str(), magic, LLAMA_FILE_MAGIC); - } - - uint32_t format_version; - fin.read((char *) &format_version, sizeof(format_version)); - - if (format_version != LLAMA_FILE_VERSION) { - fprintf(stderr, "%s: invalid model file '%s' (unsupported format version %" PRIu32 ", expected %d)\n", - __func__, fname.c_str(), format_version, LLAMA_FILE_VERSION); - return false; - } - } - - int n_ff = 0; - - // load hparams - { - auto & hparams = model.hparams; - - fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab)); - //fin.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx)); - fin.read((char *) &hparams.n_embd, sizeof(hparams.n_embd)); - fin.read((char *) &hparams.n_mult, sizeof(hparams.n_mult)); - fin.read((char *) &hparams.n_head, sizeof(hparams.n_head)); - fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer)); - fin.read((char *) &hparams.n_rot, sizeof(hparams.n_rot)); - fin.read((char *) &hparams.f16, sizeof(hparams.f16)); hparams.n_ctx = n_ctx; - - n_ff = ((2*(4*hparams.n_embd)/3 + hparams.n_mult - 1)/hparams.n_mult)*hparams.n_mult; - - if (n_parts < 1) { - n_parts = LLAMA_N_PARTS.at(hparams.n_embd); - } - - // temp warning to tell the user to use "--n_parts" - if (hparams.f16 == 4 && n_parts != 1) { - fprintf(stderr, "%s: GPTQ model detected - are you sure n_parts should be %d? we normally expect it to be 1\n", __func__, n_parts); - fprintf(stderr, "%s: use '--n_parts 1' if necessary\n", __func__); - } - - if (hparams.n_layer == 32) { - model.type = e_model::MODEL_7B; - } - - if (hparams.n_layer == 40) { - model.type = e_model::MODEL_13B; - } - - if (hparams.n_layer == 60) { - model.type = e_model::MODEL_30B; - } - - if (hparams.n_layer == 80) { - model.type = e_model::MODEL_65B; - } - - fprintf(stderr, "%s: n_vocab = %d\n", __func__, hparams.n_vocab); - fprintf(stderr, "%s: n_ctx = %d\n", __func__, hparams.n_ctx); - fprintf(stderr, "%s: n_embd = %d\n", __func__, hparams.n_embd); - fprintf(stderr, "%s: n_mult = %d\n", __func__, hparams.n_mult); - fprintf(stderr, "%s: n_head = %d\n", __func__, hparams.n_head); - fprintf(stderr, "%s: n_layer = %d\n", __func__, hparams.n_layer); - fprintf(stderr, "%s: n_rot = %d\n", __func__, hparams.n_rot); - fprintf(stderr, "%s: f16 = %d\n", __func__, hparams.f16); - fprintf(stderr, "%s: n_ff = %d\n", __func__, n_ff); - fprintf(stderr, "%s: n_parts = %d\n", __func__, n_parts); - fprintf(stderr, "%s: type = %d\n", __func__, model.type); } - // load vocab { - std::string word; - vocab.id_to_token.resize(model.hparams.n_vocab); - std::vector tmp(64); - - for (int i = 0; i < model.hparams.n_vocab; i++) { - uint32_t len; - fin.read((char *) &len, sizeof(len)); - - word.resize(len); - if (len > 0) { - tmp.resize(len); - fin.read(tmp.data(), len); - word.assign(tmp.data(), len); - } else { - word.clear(); - } - - float score; - fin.read((char *) &score, sizeof(score)); - - vocab.token_to_id[word] = i; - - auto &tok_score = vocab.id_to_token[i]; - tok_score.tok = word; - tok_score.score = score; - } + fprintf(stderr, "%s: format = %s\n", __func__, llama_file_version_name(file_version)); + fprintf(stderr, "%s: n_vocab = %u\n", __func__, hparams.n_vocab); + fprintf(stderr, "%s: n_ctx = %u\n", __func__, hparams.n_ctx); + fprintf(stderr, "%s: n_embd = %u\n", __func__, hparams.n_embd); + fprintf(stderr, "%s: n_mult = %u\n", __func__, hparams.n_mult); + fprintf(stderr, "%s: n_head = %u\n", __func__, hparams.n_head); + fprintf(stderr, "%s: n_layer = %u\n", __func__, hparams.n_layer); + fprintf(stderr, "%s: n_rot = %u\n", __func__, hparams.n_rot); + fprintf(stderr, "%s: ftype = %u (%s)\n", __func__, hparams.ftype, llama_ftype_name(hparams.ftype)); + fprintf(stderr, "%s: n_ff = %u\n", __func__, n_ff); + fprintf(stderr, "%s: n_parts = %zu\n", __func__, ml->file_loaders.size()); + fprintf(stderr, "%s: model size = %s\n", __func__, llama_model_type_name(model.type)); } if (vocab_only) { - return true; + return; } - // for the big tensors, we have the option to store the data in 16-bit floats or quantized - // in order to save memory and also to speed up the computation - // wtype is for per-layer weights, while vtype is for other weights - ggml_type wtype, vtype; - switch (model.hparams.f16) { - case 0: wtype = vtype = GGML_TYPE_F32; break; - case 1: wtype = vtype = GGML_TYPE_F16; break; - case 2: wtype = vtype = GGML_TYPE_Q4_0; break; - case 3: wtype = vtype = GGML_TYPE_Q4_1; break; - case 4: wtype = GGML_TYPE_Q4_1; vtype = GGML_TYPE_F16; break; - default: - { - fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n", - __func__, fname.c_str(), model.hparams.f16); - return false; - } - } - - // map model into memory - char *mm_addr = NULL; - model.mm_addr = mmap_file(fname.c_str(), &model.mm_length); - if (model.mm_addr == NULL) { - fprintf(stderr, "%s: failed to mmap '%s'\n", __func__, fname.c_str()); - return false; - } - mm_addr = (char *)model.mm_addr; - fprintf(stderr, "%s: ggml map size = %6.2f MB\n", __func__, model.mm_length/(1024.0*1024.0)); - auto & ctx = model.ctx; - size_t ctx_size = 0; - { - const auto &hparams = model.hparams; - const int n_layer = hparams.n_layer; - ctx_size += (5 + 10*n_layer)*256; // object overhead - fprintf(stderr, "%s: ggml ctx size = %6.2f KB\n", __func__, ctx_size/1024.0); - } + size_t ctx_size, mmapped_size; + ml->calc_sizes(&ctx_size, &mmapped_size); + fprintf(stderr, "%s: ggml ctx size = %6.2f KB\n", __func__, ctx_size/1024.0); // print memory requirements { @@ -548,14 +926,14 @@ static bool llama_model_load( // this is the total memory required to run the inference const size_t mem_required = ctx_size + - model.mm_length + - MEM_REQ_SCRATCH0.at(model.type) + - MEM_REQ_SCRATCH1.at(model.type) + - MEM_REQ_EVAL.at (model.type); + mmapped_size + + MEM_REQ_SCRATCH0().at(model.type) + + MEM_REQ_SCRATCH1().at(model.type) + + MEM_REQ_EVAL().at(model.type); // this is the memory required by one llama_state const size_t mem_required_state = - scale*MEM_REQ_KV_SELF.at(model.type); + scale*MEM_REQ_KV_SELF().at(model.type); fprintf(stderr, "%s: mem required = %7.2f MB (+ %7.2f MB per state)\n", __func__, mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0); @@ -564,17 +942,20 @@ static bool llama_model_load( // create the ggml context { lctx.model.buf.resize(ctx_size); + if (use_mlock) { + lctx.model.mlock_buf.init(lctx.model.buf.addr); + lctx.model.mlock_buf.grow_to(lctx.model.buf.size); + } struct ggml_init_params params = { - /*.mem_size =*/ lctx.model.buf.size(), - /*.mem_buffer =*/ lctx.model.buf.data(), - /*.no_alloc =*/ true, + /*.mem_size =*/ lctx.model.buf.size, + /*.mem_buffer =*/ lctx.model.buf.addr, + /*.no_alloc =*/ ml->use_mmap, }; model.ctx = ggml_init(params); if (!model.ctx) { - fprintf(stderr, "%s: ggml_init() failed\n", __func__); - return false; + throw format("ggml_init() failed"); } } @@ -582,161 +963,71 @@ static bool llama_model_load( { const auto & hparams = model.hparams; - const int n_embd = hparams.n_embd; - const int n_layer = hparams.n_layer; - const int n_vocab = hparams.n_vocab; + const uint32_t n_embd = hparams.n_embd; + const uint32_t n_layer = hparams.n_layer; + const uint32_t n_vocab = hparams.n_vocab; + + ml->ggml_ctx = ctx; + + model.tok_embeddings = ml->get_tensor("tok_embeddings.weight", {n_embd, n_vocab}); + model.norm = ml->get_tensor("norm.weight", {n_embd}); + model.output = ml->get_tensor("output.weight", {n_embd, n_vocab}); model.layers.resize(n_layer); - - model.tok_embeddings = ggml_new_tensor_2d(ctx, vtype, n_embd, n_vocab); - - model.norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); - model.output = ggml_new_tensor_2d(ctx, vtype, n_embd, n_vocab); - - // map by name - model.tensors["tok_embeddings.weight"] = model.tok_embeddings; - - model.tensors["norm.weight"] = model.norm; - model.tensors["output.weight"] = model.output; - - for (int i = 0; i < n_layer; ++i) { + for (uint32_t i = 0; i < n_layer; ++i) { auto & layer = model.layers[i]; - layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + std::string layers_i = "layers." + std::to_string(i); - layer.wq = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); - layer.wk = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); - layer.wv = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); - layer.wo = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + layer.attention_norm = ml->get_tensor(layers_i + ".attention_norm.weight", {n_embd}); - layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + layer.wq = ml->get_tensor(layers_i + ".attention.wq.weight", {n_embd, n_embd}); + layer.wk = ml->get_tensor(layers_i + ".attention.wk.weight", {n_embd, n_embd}); + layer.wv = ml->get_tensor(layers_i + ".attention.wv.weight", {n_embd, n_embd}); + layer.wo = ml->get_tensor(layers_i + ".attention.wo.weight", {n_embd, n_embd}); - layer.w1 = ggml_new_tensor_2d(ctx, wtype, n_embd, n_ff); - layer.w2 = ggml_new_tensor_2d(ctx, wtype, n_ff, n_embd); - layer.w3 = ggml_new_tensor_2d(ctx, wtype, n_embd, n_ff); + layer.ffn_norm = ml->get_tensor(layers_i + ".ffn_norm.weight", {n_embd}); - // map by name - model.tensors["layers." + std::to_string(i) + ".attention_norm.weight"] = layer.attention_norm; - - model.tensors["layers." + std::to_string(i) + ".attention.wq.weight"] = layer.wq; - model.tensors["layers." + std::to_string(i) + ".attention.wk.weight"] = layer.wk; - model.tensors["layers." + std::to_string(i) + ".attention.wv.weight"] = layer.wv; - model.tensors["layers." + std::to_string(i) + ".attention.wo.weight"] = layer.wo; - - model.tensors["layers." + std::to_string(i) + ".ffn_norm.weight"] = layer.ffn_norm; - - model.tensors["layers." + std::to_string(i) + ".feed_forward.w1.weight"] = layer.w1; - model.tensors["layers." + std::to_string(i) + ".feed_forward.w2.weight"] = layer.w2; - model.tensors["layers." + std::to_string(i) + ".feed_forward.w3.weight"] = layer.w3; + layer.w1 = ml->get_tensor(layers_i + ".feed_forward.w1.weight", {n_embd, n_ff}); + layer.w2 = ml->get_tensor(layers_i + ".feed_forward.w2.weight", { n_ff, n_embd}); + layer.w3 = ml->get_tensor(layers_i + ".feed_forward.w3.weight", {n_embd, n_ff}); } } - std::vector tmp; + ml->done_getting_tensors(); - if (progress_callback) { - progress_callback(0.0, progress_callback_user_data); + // populate `tensors_by_name` + for (llama_load_tensor & lt : ml->tensors_map.tensors) { + model.tensors_by_name.emplace_back(lt.name, lt.ggml_tensor); } - fprintf(stderr, "%s: loading tensors from '%s'\n", __func__, fname.c_str()); + ml->load_all_data(progress_callback, progress_callback_user_data, use_mlock ? &lctx.model.mlock_mmap : NULL); - // load weights - { - size_t total_size = 0; - model.n_loaded = 0; - - while (true) { - int32_t n_dims; - int32_t length; - int32_t ftype; - - fin.read(reinterpret_cast(&n_dims), sizeof(n_dims)); - fin.read(reinterpret_cast(&length), sizeof(length)); - fin.read(reinterpret_cast(&ftype), sizeof(ftype)); - - if (fin.eof()) { - break; - } - - int32_t nelements = 1; - int32_t ne[2] = { 1, 1 }; - for (int i = 0; i < n_dims; ++i) { - fin.read(reinterpret_cast(&ne[i]), sizeof(ne[i])); - nelements *= ne[i]; - } - - std::string name(length, 0); - fin.read(&name[0], length); - - if (model.tensors.find(name.data()) == model.tensors.end()) { - fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data()); - return false; - } - - auto tensor = model.tensors[name.data()]; - - if (ggml_nelements(tensor) != nelements) { - fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data()); - return false; - } - if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) { - fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%" PRId64 ", %" PRId64 "], expected [%d, %d]\n", - __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]); - return false; - } - if (0) { - static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", }; - fprintf(stderr, "%24s - [%5d, %5d], type = %6s\n", name.data(), ne[0], ne[1], ftype_str[ftype]); - } - - switch (ftype) { - case 0: // f32 - case 1: // f16 - break; - case 2: // q4_0 - case 3: // q4_1 - assert(ne[0] % 64 == 0); - break; - default: - fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype); - return false; - }; - - // load the tensor data into memory without copying or reading it - size_t offset = fin.tellg(); - size_t tensor_data_size = ggml_nbytes(tensor); - offset = (offset + 31) & -32; - tensor->data = mm_addr + offset; - fin.seekg(offset + tensor_data_size); - total_size += tensor_data_size; - model.n_loaded++; - - // progress - if (progress_callback) { - double current_progress = size_t(fin.tellg()) / double(file_size); - progress_callback(current_progress, progress_callback_user_data); - } - } - - fin.close(); - - fprintf(stderr, "%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, model.n_loaded); - if (model.n_loaded == 0) { - fprintf(stderr, "%s: WARN no tensors loaded from model file - assuming empty model for testing\n", __func__); - } else if (model.n_loaded != (int) model.tensors.size()) { - fprintf(stderr, "%s: ERROR not all tensors loaded from model file - expected %zu, got %d\n", __func__, model.tensors.size(), model.n_loaded); - return false; - } - } + model.mapping = std::move(ml->mapping); // loading time will be recalculate after the first eval, so // we take page faults deferred by mmap() into consideration lctx.t_load_us = ggml_time_us() - lctx.t_start_us; +} - if (progress_callback) { - progress_callback(1.0, progress_callback_user_data); +static bool llama_model_load( + const std::string & fname, + llama_context & lctx, + int n_ctx, + ggml_type memory_type, + bool use_mmap, + bool use_mlock, + bool vocab_only, + llama_progress_callback progress_callback, + void *progress_callback_user_data) { + try { + llama_model_load_internal(fname, lctx, n_ctx, memory_type, use_mmap, use_mlock, + vocab_only, progress_callback, progress_callback_user_data); + return true; + } catch (const std::string & err) { + fprintf(stderr, "error loading model: %s\n", err.c_str()); + return false; } - - return true; } // evaluate the transformer @@ -774,8 +1065,8 @@ static bool llama_eval_internal( auto & buf_compute = lctx.buf_compute; struct ggml_init_params params = { - /*.mem_size =*/ buf_compute.size(), - /*.mem_buffer =*/ buf_compute.data(), + /*.mem_size =*/ buf_compute.size, + /*.mem_buffer =*/ buf_compute.addr, /*.no_alloc =*/ false, }; @@ -784,7 +1075,7 @@ static bool llama_eval_internal( // for big prompts, if BLAS is enabled, it is better to use only one thread // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance ggml_cgraph gf = {}; - gf.n_threads = N >= 32 && ggml_cpu_has_blas() ? 1 : n_threads; + gf.n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_cublas() ? 1 : n_threads; struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); memcpy(embd->data, tokens, N*ggml_element_size(embd)); @@ -810,37 +1101,35 @@ static bool llama_eval_internal( // self-attention { - struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); - struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); - struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0); + struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0); // store key and value to memory - if (N >= 1) { - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past)); - struct ggml_tensor * v = ggml_view_1d(ctx0, kv_self.v, N*n_embd, (ggml_element_size(kv_self.v)*n_embd)*(il*n_ctx + n_past)); + { + // compute the transposed [N, n_embd] V matrix + struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), n_embd, N)); + struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past)); + struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd, + ( n_ctx)*ggml_element_size(kv_self.v), + (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v)); + + // important: storing RoPE-ed version of K in the KV cache! ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k)); ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v)); } - // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3) struct ggml_tensor * Q = ggml_permute(ctx0, - ggml_rope(ctx0, - ggml_cpy(ctx0, - Qcur, - ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)), - n_past, n_rot, 0), + Qcur, 0, 2, 1, 3); - // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3) struct ggml_tensor * K = ggml_permute(ctx0, - ggml_rope(ctx0, - ggml_reshape_3d(ctx0, - ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd), - n_embd/n_head, n_head, n_past + N), - n_past, n_rot, 1), + ggml_reshape_3d(ctx0, + ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd), + n_embd/n_head, n_head, n_past + N), 0, 2, 1, 3); // K * Q @@ -858,18 +1147,23 @@ static bool llama_eval_internal( // KQ = soft_max(KQ_masked) struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous() - struct ggml_tensor * V_trans = - ggml_cpy(ctx0, - ggml_permute(ctx0, - ggml_reshape_3d(ctx0, - ggml_view_1d(ctx0, kv_self.v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.v)*n_embd), - n_embd/n_head, n_head, n_past + N), - 1, 2, 0, 3), - ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd/n_head, n_head)); + // split cached V into n_head heads + struct ggml_tensor * V = + ggml_view_3d(ctx0, kv_self.v, + n_past + N, n_embd/n_head, n_head, + n_ctx*ggml_element_size(kv_self.v), + n_ctx*ggml_element_size(kv_self.v)*n_embd/n_head, + il*n_ctx*ggml_element_size(kv_self.v)*n_embd); - // KQV = transpose(V) * KQ_soft_max - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max); +#if 1 + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); +#else + // make V contiguous in memory to speed up the matmul, however we waste time on the copy + // on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation + // is there a better way? + struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd/n_head, n_head)); + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max); +#endif // KQV_merged = KQV.permute(0, 2, 1, 3) struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); @@ -955,9 +1249,13 @@ static bool llama_eval_internal( ggml_build_forward_expand(&gf, inpL); ggml_graph_compute (ctx0, &gf); + // print timing information per ggml operation (for debugging purposes) + // requires GGML_PERF to be defined + //ggml_graph_print(&gf); + + // plot the computation graph in dot format (for debugging purposes) //if (n_past%100 == 0) { - // ggml_graph_print (&gf); - // ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot"); + // ggml_graph_dump_dot(&gf, NULL, "llama.dot"); //} //embd_w.resize(n_vocab*N); @@ -1054,7 +1352,7 @@ struct llama_tokenizer { size_t offs = 0; while (offs < text.size()) { llama_sp_symbol sym; - size_t char_len = Min(text.size() - offs, utf8_len(text[offs])); + size_t char_len = std::min(text.size() - offs, utf8_len(text[offs])); sym.text = text.c_str() + offs; sym.n = char_len; offs += char_len; @@ -1229,17 +1527,13 @@ static llama_vocab::id llama_sample_top_p_top_k( } } - sample_top_k(logits_id, top_k); - - float maxl = -std::numeric_limits::infinity(); - for (const auto & kv : logits_id) { - maxl = Max(maxl, kv.first); - } + sample_top_k(logits_id, top_k > 0 ? std::min(top_k, n_logits) : n_logits); // compute probs for the top k tokens std::vector probs; probs.reserve(logits_id.size()); + float maxl = logits_id[0].first; double sum = 0.0; for (const auto & kv : logits_id) { const float p = expf(kv.first - maxl); @@ -1262,16 +1556,11 @@ static llama_vocab::id llama_sample_top_p_top_k( break; } } - - cumsum = 1.0/cumsum; - for (int i = 0; i < (int) probs.size(); i++) { - probs[i] *= cumsum; - } } //printf("\n"); //for (int i = 0; i < (int) 10; i++) { - // printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]); + // printf("%d: '%s' %f\n", i, lctx.vocab.id_to_token.at(logits_id[i].second).tok.c_str(), probs[i]); //} //printf("\n\n"); //exit(0); @@ -1286,298 +1575,152 @@ static llama_vocab::id llama_sample_top_p_top_k( // quantization // -// TODO: reuse code from the llama_model_load() somehow -static bool llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, int itype) { - ggml_type type = GGML_TYPE_Q4_1; - - switch (itype) { - case 2: type = GGML_TYPE_Q4_0; break; - case 3: type = GGML_TYPE_Q4_1; break; - default: fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype); return 1; +static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, enum llama_ftype ftype, int nthread) { + ggml_type quantized_type; + switch (ftype) { + case LLAMA_FTYPE_MOSTLY_Q4_0: quantized_type = GGML_TYPE_Q4_0; break; + case LLAMA_FTYPE_MOSTLY_Q4_1: quantized_type = GGML_TYPE_Q4_1; break; + case LLAMA_FTYPE_MOSTLY_Q4_2: quantized_type = GGML_TYPE_Q4_2; break; + case LLAMA_FTYPE_MOSTLY_Q4_3: quantized_type = GGML_TYPE_Q4_3; break; + default: throw format("invalid output file type %d\n", ftype); }; - if (type != GGML_TYPE_Q4_0 && type != GGML_TYPE_Q4_1) { - fprintf(stderr, "%s: invalid quantization type %d\n", __func__, type); - return false; + if (nthread <= 0) { + nthread = std::thread::hardware_concurrency(); } - llama_vocab vocab; + std::unique_ptr model_loader(new llama_model_loader(fname_inp.c_str(), /*use_mmap*/ false, + /*vocab_only*/ false)); + llama_file_saver file_saver(fname_out.c_str(), model_loader->file_loaders.at(0).get(), ftype); - printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str()); + size_t total_size_org = 0; + size_t total_size_new = 0; + std::vector hist_all(1 << 4, 0); - auto finp = std::ifstream(fname_inp, std::ios::binary); - if (!finp) { - fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str()); - return false; - } + std::vector workers; + std::mutex mutex; - auto fout = std::ofstream(fname_out, std::ios::binary); - if (!fout) { - fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str()); - return false; - } + size_t idx = 0; + for (llama_load_tensor & tensor : model_loader->tensors_map.tensors) { + llama_buffer read_data; + read_data.resize(tensor.size); + tensor.data = read_data.addr; + model_loader->load_data_for(tensor); - // verify magic - { - uint32_t magic; - finp.read((char *) &magic, sizeof(magic)); - if (magic == LLAMA_FILE_MAGIC_UNVERSIONED) { - fprintf(stderr, "%s: invalid model file '%s' (too old, regenerate your model files!)\n", - __func__, fname_inp.c_str()); - return false; - } - if (magic != LLAMA_FILE_MAGIC) { - return report_bad_magic(fname_inp.c_str(), magic, LLAMA_FILE_MAGIC); - } + printf("[%4zu/%4zu] %36s - %16s, type = %6s, ", + ++idx, model_loader->tensors_map.tensors.size(), + tensor.name.c_str(), llama_format_tensor_shape(tensor.ne).c_str(), + ggml_type_name(tensor.type)); - fout.write((char *) &magic, sizeof(magic)); + // This used to be a regex, but has an extreme cost to compile times. + bool quantize = tensor.name.rfind("weight") == tensor.name.size() - 6; // ends with 'weight'? - uint32_t format_version; - finp.read((char *) &format_version, sizeof(format_version)); + // quantize only 2D tensors + quantize &= (tensor.ne.size() == 2); - if (format_version != LLAMA_FILE_VERSION) { - fprintf(stderr, "%s: invalid model file '%s' (unsupported format version %" PRIu32 ", expected %d)\n", - __func__, fname_inp.c_str(), format_version, LLAMA_FILE_VERSION); - return false; - } + // GG: uncomment this to keep the output layer in FP16 + //if (tensor.name.rfind("output")) { + // quantize = false; + //} - fout.write((char *) &format_version, sizeof(format_version)); - } + enum ggml_type new_type; + void * new_data; + size_t new_size; + llama_buffer work; - llama_hparams hparams; - - // load hparams - { - finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab)); - //finp.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx)); - finp.read((char *) &hparams.n_embd, sizeof(hparams.n_embd)); - finp.read((char *) &hparams.n_mult, sizeof(hparams.n_mult)); - finp.read((char *) &hparams.n_head, sizeof(hparams.n_head)); - finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer)); - finp.read((char *) &hparams.n_rot, sizeof(hparams.n_rot)); - finp.read((char *) &hparams.f16, sizeof(hparams.f16)); - - printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab); - printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx); - printf("%s: n_embd = %d\n", __func__, hparams.n_embd); - printf("%s: n_mult = %d\n", __func__, hparams.n_mult); - printf("%s: n_head = %d\n", __func__, hparams.n_head); - printf("%s: n_layer = %d\n", __func__, hparams.n_layer); - printf("%s: f16 = %d\n", __func__, hparams.f16); - - fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab)); - //fout.write((char *) &hparams.n_ctx, sizeof(hparams.n_ctx)); - fout.write((char *) &hparams.n_embd, sizeof(hparams.n_embd)); - fout.write((char *) &hparams.n_mult, sizeof(hparams.n_mult)); - fout.write((char *) &hparams.n_head, sizeof(hparams.n_head)); - fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer)); - fout.write((char *) &hparams.n_rot, sizeof(hparams.n_rot)); - fout.write((char *) &itype, sizeof(hparams.f16)); - } - - // load vocab - { - const int32_t n_vocab = hparams.n_vocab; - - if (n_vocab != hparams.n_vocab) { - fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n", - __func__, fname_inp.c_str(), n_vocab, hparams.n_vocab); - return false; - } - - std::vector word(32); - vocab.id_to_token.resize(n_vocab); - for (int i = 0; i < n_vocab; i++) { - uint32_t len; - finp.read ((char *) &len, sizeof(len)); - fout.write((char *) &len, sizeof(len)); - - word.resize(len); - finp.read ((char *) &word[0], len); - fout.write((char *) &word[0], len); - - float score; - finp.read ((char *) &score, sizeof(score)); - fout.write((char *) &score, sizeof(score)); - - vocab.token_to_id[word.data()] = i; - - auto &tok_score = vocab.id_to_token[i]; - tok_score.tok = word.data(); - tok_score.score = score; - } - } - - // load weights - { - size_t total_size_org = 0; - size_t total_size_new = 0; - - std::vector work; - - std::vector data_u8; - std::vector data_f16; - std::vector data_f32; - - std::vector hist_all(1 << 4, 0); - - while (true) { - int32_t n_dims; - int32_t length; - int32_t ftype; - - finp.read(reinterpret_cast(&n_dims), sizeof(n_dims)); - finp.read(reinterpret_cast(&length), sizeof(length)); - finp.read(reinterpret_cast(&ftype), sizeof(ftype)); - - if (finp.eof()) { - break; - } - - int32_t nelements = 1; - int32_t ne[2] = { 1, 1 }; - for (int i = 0; i < n_dims; ++i) { - finp.read (reinterpret_cast(&ne[i]), sizeof(ne[i])); - nelements *= ne[i]; - } - - std::string name(length, 0); - finp.read (&name[0], length); - - { - // ensure tensor data is aligned - uint64_t offset = finp.tellg(); - offset = (offset + 31) & -32; - finp.seekg(offset); - } - - { - static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", }; - printf("%48s - [%5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ftype_str[ftype]); - } - - // regexes of tensor names to be quantized - const std::vector k_names = { - ".*weight", - }; - - bool quantize = false; - for (const auto & s : k_names) { - if (std::regex_match(name, std::regex(s))) { - quantize = true; - break; + if (!quantize) { + new_type = tensor.type; + new_data = tensor.data; + new_size = tensor.size; + printf("size = %8.3f MB\n", tensor.size/1024.0/1024.0); + } else { + new_type = quantized_type; + float * f32_data; + size_t nelements = tensor.ne.at(0) * tensor.ne.at(1); + llama_buffer f32_conv_buf; + if (tensor.type == GGML_TYPE_F32) { + f32_data = (float *) tensor.data; + } else if (tensor.type == GGML_TYPE_F16) { + f32_conv_buf.resize(nelements * sizeof(float)); + f32_data = (float *) f32_conv_buf.addr; + auto f16_data = (const ggml_fp16_t *) tensor.data; + for (size_t i = 0; i < nelements; i++) { + f32_data[i] = ggml_fp16_to_fp32(f16_data[i]); } - } - - // quantize only 2D tensors - quantize &= (n_dims == 2); - - if (quantize) { - if (ftype != 0 && ftype != 1) { - fprintf(stderr, "%s: unsupported ftype %d for integer quantization\n", __func__, ftype); - return false; - } - - if (ftype == 1) { - data_f16.resize(nelements); - finp.read(reinterpret_cast(data_f16.data()), nelements * sizeof(ggml_fp16_t)); - data_f32.resize(nelements); - for (int i = 0; i < nelements; ++i) { - data_f32[i] = ggml_fp16_to_fp32(data_f16[i]); - } - } else { - data_f32.resize(nelements); - finp.read(reinterpret_cast(data_f32.data()), nelements * sizeof(float)); - } - - ftype = itype; } else { - const int bpe = (ftype == 0) ? sizeof(float) : sizeof(uint16_t); - - data_u8.resize(nelements*bpe); - finp.read(reinterpret_cast(data_u8.data()), nelements * bpe); + throw format("type %s unsupported for integer quantization", ggml_type_name(tensor.type)); } - fout.write(reinterpret_cast(&n_dims), sizeof(n_dims)); - fout.write(reinterpret_cast(&length), sizeof(length)); - fout.write(reinterpret_cast(&ftype), sizeof(ftype)); - for (int i = 0; i < n_dims; ++i) { - fout.write(reinterpret_cast(&ne[i]), sizeof(ne[i])); - } - fout.write(&name[0], length); + printf("quantizing .. "); + fflush(stdout); - { - // ensure tensor data is aligned - uint64_t offset = fout.tellp(); - offset = (offset + 31) & -32; - fout.seekp(offset); - } + work.resize(nelements * 4); // upper bound on size + new_data = work.addr; + std::vector hist_cur(1 << 4, 0); - if (quantize) { - printf("quantizing .. "); - work.resize(nelements); // for quantization - - size_t cur_size = 0; - std::vector hist_cur(1 << 4, 0); - - switch (type) { - case GGML_TYPE_Q4_0: - { - cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q4_1: - { - cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data()); - } break; - default: - { - fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, type); - return false; + int chunk_size = 32 * 512; + const int nchunk = (nelements + chunk_size - 1)/chunk_size; + const int nthread_use = nthread > 1 ? std::max(1, std::min(nthread, nchunk)) : 1; + if (nthread_use < 2) { + new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nelements, hist_cur.data()); + } else { + size_t counter = 0; + new_size = 0; + auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, nelements, chunk_size] () { + std::vector local_hist; + size_t local_size = 0; + while (true) { + std::unique_lock lock(mutex); + size_t first = counter; counter += chunk_size; + if (first >= nelements) { + if (!local_hist.empty()) { + for (int j=0; j(work.data()), cur_size); - total_size_new += cur_size; - - printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0); - for (int i = 0; i < (int) hist_cur.size(); ++i) { - hist_all[i] += hist_cur[i]; - } - - for (int i = 0; i < (int) hist_cur.size(); ++i) { - printf("%5.3f ", hist_cur[i] / float(nelements)); - } - printf("\n"); - } else { - printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0); - fout.write(reinterpret_cast(data_u8.data()), data_u8.size()); - total_size_new += data_u8.size(); + lock.unlock(); + size_t last = std::min(nelements, first + chunk_size); + if (local_hist.empty()) local_hist.resize(hist_cur.size(), 0); + local_size += ggml_quantize_chunk(new_type, f32_data, new_data, first, last - first, local_hist.data()); + } + }; + if (int(workers.size()) < nthread_use - 1) workers.resize(nthread_use - 1); + for (int it = 0; it < nthread_use - 1; ++it) workers[it] = std::thread(compute); + compute(); + for (int it = 0; it < nthread_use - 1; ++it) workers[it].join(); } - total_size_org += nelements * sizeof(float); - } - - printf("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0); - printf("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0); - - { - int64_t sum_all = 0; - for (int i = 0; i < (int) hist_all.size(); ++i) { - sum_all += hist_all[i]; + printf("size = %8.2f MB -> %8.2f MB | hist: ", tensor.size/1024.0/1024.0, new_size/1024.0/1024.0); + for (size_t i = 0; i < hist_cur.size(); i++) { + hist_all[i] += hist_cur[i]; } - printf("%s: hist: ", __func__); - for (int i = 0; i < (int) hist_all.size(); ++i) { - printf("%5.3f ", hist_all[i] / float(sum_all)); + for (size_t i = 0; i < hist_cur.size(); i++) { + printf("%5.3f ", hist_cur[i] / float(nelements)); } printf("\n"); } + total_size_org += tensor.size; + total_size_new += new_size; + file_saver.write_tensor(tensor, new_type, new_data, new_size); } - finp.close(); - fout.close(); + printf("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0); + printf("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0); - return true; + { + int64_t sum_all = 0; + for (size_t i = 0; i < hist_all.size(); i++) { + sum_all += hist_all[i]; + } + + printf("%s: hist: ", __func__); + for (size_t i = 0; i < hist_all.size(); i++) { + printf("%5.3f ", hist_all[i] / float(sum_all)); + } + printf("\n"); + } } // @@ -1595,32 +1738,36 @@ struct llama_context * llama_init_from_file( params.seed = time(NULL); } + unsigned cur_percentage = 0; + if (params.progress_callback == NULL) { + params.progress_callback_user_data = &cur_percentage; + params.progress_callback = [](float progress, void * ctx) { + unsigned * cur_percentage_p = (unsigned *) ctx; + unsigned percentage = (unsigned) (100 * progress); + while (percentage > *cur_percentage_p) { + ++*cur_percentage_p; + fprintf(stderr, "."); + fflush(stderr); + if (percentage >= 100) { + fprintf(stderr, "\n"); + } + } + }; + } + ctx->rng = std::mt19937(params.seed); ctx->logits_all = params.logits_all; ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32; - if (!llama_model_load(path_model, *ctx, params.n_ctx, params.n_parts, memory_type, - params.vocab_only, params.progress_callback, - params.progress_callback_user_data)) { + if (!llama_model_load(path_model, *ctx, params.n_ctx, memory_type, + params.use_mmap, params.use_mlock, params.vocab_only, + params.progress_callback, params.progress_callback_user_data)) { fprintf(stderr, "%s: failed to load model\n", __func__); llama_free(ctx); return nullptr; } - if (params.use_mlock) { - char *err; - if (!ggml_mlock(ctx->model.ctx, - ctx->model.mm_addr, - ctx->model.mm_length, - &err)) { - fprintf(stderr, "%s\n", err); - free(err); - llama_free(ctx); - return nullptr; - } - } - // reserve memory for context buffers if (!params.vocab_only) { if (!kv_cache_init(ctx->model.hparams, ctx->model.kv_self, memory_type, ctx->model.hparams.n_ctx)) { @@ -1647,50 +1794,290 @@ struct llama_context * llama_init_from_file( ctx->embedding.resize(hparams.n_embd); } - ctx->buf_compute.resize(MEM_REQ_EVAL.at(ctx->model.type)); + ctx->buf_compute.resize(MEM_REQ_EVAL().at(ctx->model.type)); - ctx->buf_scratch[0].resize(MEM_REQ_SCRATCH0.at(ctx->model.type)); - ctx->buf_scratch[1].resize(MEM_REQ_SCRATCH1.at(ctx->model.type)); + ctx->buf_scratch[0].resize(MEM_REQ_SCRATCH0().at(ctx->model.type)); + ctx->buf_scratch[1].resize(MEM_REQ_SCRATCH1().at(ctx->model.type)); } return ctx; } void llama_free(struct llama_context * ctx) { - kv_cache_free(ctx->model.kv_self); - - if (ctx->model.ctx) { - ggml_free(ctx->model.ctx); - } - - if (ctx->model.mm_addr) { - munmap_file(ctx->model.mm_addr, ctx->model.mm_length); - } - delete ctx; } int llama_model_quantize( const char * fname_inp, const char * fname_out, - int itype) { - if (!llama_model_quantize_internal(fname_inp, fname_out, itype)) { - fprintf(stderr, "%s: failed to quantize\n", __func__); + enum llama_ftype ftype, + int nthread) { + try { + llama_model_quantize_internal(fname_inp, fname_out, ftype, nthread); + return 0; + } catch (const std::string & err) { + fprintf(stderr, "%s: failed to quantize: %s\n", __func__, err.c_str()); + return 1; + } +} + +int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char * path_lora, const char * path_base_model, int n_threads) { + fprintf(stderr, "%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora); + + auto & model = ctx->model; + + const int64_t t_start_lora_us = ggml_time_us(); + + auto fin = std::ifstream(path_lora, std::ios::binary); + if (!fin) { + fprintf(stderr, "%s: failed to open '%s'\n", __func__, path_lora); return 1; } + // verify magic and version + { + uint32_t magic; + fin.read((char *) &magic, sizeof(magic)); + if (magic != 'ggla') { + fprintf(stderr, "%s: bad file magic\n", __func__); + return 1; + } + uint32_t format_version; + fin.read((char *) &format_version, sizeof(format_version)); + + if (format_version != 1) { + fprintf(stderr, "%s: unsupported file version\n", __func__ ); + return 1; + } + } + + int32_t lora_r; + int32_t lora_alpha; + fin.read((char *) &lora_r, sizeof(lora_r)); + fin.read((char *) &lora_alpha, sizeof(lora_alpha)); + float scaling = (float)lora_alpha / (float)lora_r; + + fprintf(stderr, "%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling); + + + // create a temporary ggml context to store the lora tensors + // todo: calculate size from biggest possible tensor + std::vector lora_buf(1024ull * 1024ull * 1024ull); + struct ggml_init_params params; + params.mem_size = lora_buf.size(); + params.mem_buffer = lora_buf.data(); + params.no_alloc = false; + + ggml_context * lora_ctx = ggml_init(params); + std::unordered_map lora_tensors; + + // create a name -> tensor map of the model to accelerate lookups + std::unordered_map model_tensors; + for (auto & kv: model.tensors_by_name) { + model_tensors.insert(kv); + } + + + // load base model + std::unique_ptr model_loader; + ggml_context * base_ctx = NULL; + llama_buffer base_buf; + if (path_base_model) { + fprintf(stderr, "%s: loading base model from '%s'\n", __func__, path_base_model); + model_loader.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*vocab_only*/ false)); + + size_t ctx_size, mmapped_size; + model_loader->calc_sizes(&ctx_size, &mmapped_size); + base_buf.resize(ctx_size); + + ggml_init_params base_params; + base_params.mem_size = base_buf.size; + base_params.mem_buffer = base_buf.addr; + base_params.no_alloc = model_loader->use_mmap; + + base_ctx = ggml_init(base_params); + + model_loader->ggml_ctx = base_ctx; + + // maybe this should in llama_model_loader + if (model_loader->use_mmap) { + model_loader->mapping.reset(new llama_mmap(&model_loader->file_loaders.at(0)->file, /* prefetch */ false)); + } + } + + // read tensors and apply + bool warned = false; + int n_tensors = 0; + while (true) { + int32_t n_dims; + int32_t length; + int32_t ftype; + + fin.read(reinterpret_cast(&n_dims), sizeof(n_dims)); + fin.read(reinterpret_cast(&length), sizeof(length)); + fin.read(reinterpret_cast(&ftype), sizeof(ftype)); + if (fin.eof()) { + break; + } + + int32_t ne[2] = { 1, 1 }; + for (int i = 0; i < n_dims; ++i) { + fin.read(reinterpret_cast(&ne[i]), sizeof(ne[i])); + } + + std::string name(length, 0); + fin.read(&name[0], length); + + // check for lora suffix and get the type of tensor + const std::string lora_suffix = ".lora"; + size_t pos = name.rfind(lora_suffix); + if (pos == std::string::npos) { + fprintf(stderr, "%s: error: '%s' is not a lora tensor\n", __func__, name.c_str()); + return 1; + } + + std::string lora_type = name.substr(pos + lora_suffix.length()); + std::string base_name = name; + base_name.erase(pos); + // fprintf(stderr, "%s: %s => %s (lora type %s) ", __func__, name.c_str(),base_name.c_str(), lora_type.c_str()); + + if (model_tensors.find(base_name.data()) == model_tensors.end()) { + fprintf(stderr, "%s: unknown tensor '%s' in lora adapter\n", __func__, name.data()); + return 1; + } + + // create ggml tensor + ggml_type wtype; + switch (ftype) { + case 0: wtype = GGML_TYPE_F32; break; + case 1: wtype = GGML_TYPE_F16; break; + default: + { + fprintf(stderr, "%s: invalid tensor data type '%d'\n", + __func__, ftype); + return false; + } + } + ggml_tensor* lora_tensor; + if (n_dims == 2) { + lora_tensor = ggml_new_tensor_2d(lora_ctx, wtype, ne[0], ne[1]); + } + else { + fprintf(stderr, "%s: unsupported tensor dimension %d\n", __func__, n_dims); + return 1; + } + + // load tensor data + size_t offset = fin.tellg(); + size_t tensor_data_size = ggml_nbytes(lora_tensor); + offset = (offset + 31) & -32; + fin.seekg(offset); + fin.read((char*)lora_tensor->data, tensor_data_size); + + lora_tensors[name] = lora_tensor; + + // check if we have both A and B tensors and apply + if (lora_tensors.find(base_name + ".loraA") != lora_tensors.end() && + lora_tensors.find(base_name + ".loraB") != lora_tensors.end()) { + + ggml_tensor * dest_t = model_tensors[base_name]; + ggml_tensor * base_t; + if (model_loader) { + // load from base model + if (model_loader->tensors_map.name_to_idx.find(base_name) == model_loader->tensors_map.name_to_idx.end()) { + fprintf(stderr, "%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str()); + return 1; + } + size_t idx = model_loader->tensors_map.name_to_idx[base_name]; + llama_load_tensor & lt = model_loader->tensors_map.tensors[idx]; + base_t = model_loader->get_tensor(base_name, { (uint32_t)dest_t->ne[0], (uint32_t)dest_t->ne[1] }); + lt.data = (uint8_t *) lt.ggml_tensor->data; + model_loader->load_data_for(lt); + lt.ggml_tensor->data = lt.data; + } + else { + base_t = dest_t; + } + + if (ggml_is_quantized(base_t->type)) { + if (!warned) { + fprintf(stderr, "%s: warning: using a lora adapter with a quantized model may result in poor quality, " + "use a f16 or f32 base model with --lora-base\n", __func__); + warned = true; + } + } + + ggml_tensor * loraA = lora_tensors[base_name + ".loraA"]; + ggml_tensor * loraB = lora_tensors[base_name + ".loraB"]; + + if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) { + fprintf(stderr, "%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");" + " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]); + return 1; + } + + // w = w + BA*s + ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB); + + if (scaling != 1.0f) { + ggml_tensor * scale_tensor = ggml_new_f32(lora_ctx, scaling); + BA = ggml_scale(lora_ctx, BA, scale_tensor); + } + + ggml_tensor * r; + if (base_t == dest_t) { + r = ggml_add_inplace(lora_ctx, dest_t, BA); + } + else { + r = ggml_add(lora_ctx, base_t, BA); + r = ggml_cpy(lora_ctx, r, dest_t); + } + + struct ggml_cgraph gf = ggml_build_forward(r); + gf.n_threads = n_threads; + ggml_graph_compute(lora_ctx, &gf); + + // we won't need these tensors again, reset the context to save memory + ggml_free(lora_ctx); + lora_ctx = ggml_init(params); + lora_tensors.clear(); + + n_tensors++; + if (n_tensors % 4 == 0) + fprintf(stderr, "."); + } + } + + // TODO: this should be in a destructor, it will leak on failure + ggml_free(lora_ctx); + if (base_ctx) { + ggml_free(base_ctx); + } + + const int64_t t_lora_us = ggml_time_us() - t_start_lora_us; + fprintf(stderr, " done (%.2f ms)\n", t_lora_us / 1000.0); + return 0; } +int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lora, const char * path_base_model, int n_threads) { + try { + return llama_apply_lora_from_file_internal(ctx, path_lora, path_base_model, n_threads); + } catch (const std::string & err) { + fprintf(stderr, "%s: failed to apply lora adapter: %s\n", __func__, err.c_str()); + return 1; + } +} + // Returns the KV cache that will contain the context for the // ongoing prediction with the model. const uint8_t * llama_get_kv_cache(struct llama_context * ctx) { - return ctx->model.kv_self.buf.data(); + return ctx->model.kv_self.buf.addr; } // Returns the size of the KV cache size_t llama_get_kv_cache_size(struct llama_context * ctx) { - return ctx->model.kv_self.buf.size(); + return ctx->model.kv_self.buf.size; } int llama_get_kv_cache_token_count(struct llama_context * ctx) { @@ -1704,8 +2091,8 @@ void llama_set_kv_cache( size_t n_size, int n_token_count) { // Make sure we have the same kv cache setup - LLAMA_ASSERT(ctx->model.kv_self.buf.size() == n_size); - memcpy(ctx->model.kv_self.buf.data(), kv_cache, n_size); + LLAMA_ASSERT(ctx->model.kv_self.buf.size == n_size); + memcpy(ctx->model.kv_self.buf.addr, kv_cache, n_size); ctx->model.kv_self.n = n_token_count; } @@ -1816,9 +2203,9 @@ llama_token llama_sample_top_p_top_k( void llama_print_timings(struct llama_context * ctx) { const int64_t t_end_us = ggml_time_us(); - const int32_t n_sample = Max(1, ctx->n_sample); - const int32_t n_eval = Max(1, ctx->n_eval); - const int32_t n_p_eval = Max(1, ctx->n_p_eval); + const int32_t n_sample = std::max(1, ctx->n_sample); + const int32_t n_eval = std::max(1, ctx->n_eval); + const int32_t n_p_eval = std::max(1, ctx->n_p_eval); fprintf(stderr, "\n"); fprintf(stderr, "%s: load time = %8.2f ms\n", __func__, ctx->t_load_us / 1000.0); @@ -1839,18 +2226,25 @@ const char * llama_print_system_info(void) { static std::string s; s = ""; - s += "AVX = " + std::to_string(ggml_cpu_has_avx()) + " | "; - s += "AVX2 = " + std::to_string(ggml_cpu_has_avx2()) + " | "; - s += "AVX512 = " + std::to_string(ggml_cpu_has_avx512()) + " | "; - s += "FMA = " + std::to_string(ggml_cpu_has_fma()) + " | "; - s += "NEON = " + std::to_string(ggml_cpu_has_neon()) + " | "; - s += "ARM_FMA = " + std::to_string(ggml_cpu_has_arm_fma()) + " | "; - s += "F16C = " + std::to_string(ggml_cpu_has_f16c()) + " | "; - s += "FP16_VA = " + std::to_string(ggml_cpu_has_fp16_va()) + " | "; - s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | "; - s += "BLAS = " + std::to_string(ggml_cpu_has_blas()) + " | "; - s += "SSE3 = " + std::to_string(ggml_cpu_has_sse3()) + " | "; - s += "VSX = " + std::to_string(ggml_cpu_has_vsx()) + " | "; + s += "AVX = " + std::to_string(ggml_cpu_has_avx()) + " | "; + s += "AVX2 = " + std::to_string(ggml_cpu_has_avx2()) + " | "; + s += "AVX512 = " + std::to_string(ggml_cpu_has_avx512()) + " | "; + s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | "; + s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | "; + s += "FMA = " + std::to_string(ggml_cpu_has_fma()) + " | "; + s += "NEON = " + std::to_string(ggml_cpu_has_neon()) + " | "; + s += "ARM_FMA = " + std::to_string(ggml_cpu_has_arm_fma()) + " | "; + s += "F16C = " + std::to_string(ggml_cpu_has_f16c()) + " | "; + s += "FP16_VA = " + std::to_string(ggml_cpu_has_fp16_va()) + " | "; + s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | "; + s += "BLAS = " + std::to_string(ggml_cpu_has_blas()) + " | "; + s += "SSE3 = " + std::to_string(ggml_cpu_has_sse3()) + " | "; + s += "VSX = " + std::to_string(ggml_cpu_has_vsx()) + " | "; return s.c_str(); } + +// For internal test use +std::vector>& llama_internal_get_tensor_map(struct llama_context * ctx) { + return ctx->model.tensors_by_name; +} diff --git a/llama.h b/llama.h index 04e2bf71c..e95ff73b8 100644 --- a/llama.h +++ b/llama.h @@ -55,6 +55,7 @@ extern "C" { bool f16_kv; // use fp16 for KV cache bool logits_all; // the llama_eval() call computes all logits, not just the last one bool vocab_only; // only load the vocabulary, no weights + bool use_mmap; // use mmap if possible bool use_mlock; // force system to keep model in RAM bool embedding; // embedding mode only @@ -64,8 +65,22 @@ extern "C" { void * progress_callback_user_data; }; + // model file types + enum llama_ftype { + LLAMA_FTYPE_ALL_F32 = 0, + LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16 + LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q4_3 = 6, // except 1d tensors + }; + LLAMA_API struct llama_context_params llama_context_default_params(); + LLAMA_API bool llama_mmap_supported(); + LLAMA_API bool llama_mlock_supported(); + // Various functions for loading a ggml llama model. // Allocate (almost) all memory needed for the model. // Return NULL on failure @@ -78,10 +93,24 @@ extern "C" { // TODO: not great API - very likely to change // Returns 0 on success + // nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given LLAMA_API int llama_model_quantize( const char * fname_inp, const char * fname_out, - int itype); + enum llama_ftype ftype, + int nthread); + + // Apply a LoRA adapter to a loaded model + // path_base_model is the path to a higher quality model to use as a base for + // the layers modified by the adapter. Can be NULL to use the current loaded model. + // The model needs to be reloaded before applying a new adapter, otherwise the adapter + // will be applied on top of the previous one + // Returns 0 on success + LLAMA_API int llama_apply_lora_from_file( + struct llama_context * ctx, + const char * path_lora, + const char * path_base_model, + int n_threads); // Returns the KV cache that will contain the context for the // ongoing prediction with the model. @@ -166,4 +195,15 @@ extern "C" { } #endif +// Internal API to be implemented by llama.cpp and used by tests/benchmarks only +#ifdef LLAMA_API_INTERNAL + +#include +#include +struct ggml_tensor; + +std::vector>& llama_internal_get_tensor_map(struct llama_context * ctx); + #endif + +#endif // LLAMA_H diff --git a/llama_util.h b/llama_util.h new file mode 100755 index 000000000..eba14656a --- /dev/null +++ b/llama_util.h @@ -0,0 +1,395 @@ +// Internal header to be included only by llama.cpp. +// Contains wrappers around OS interfaces. + +#ifndef LLAMA_UTIL_H +#define LLAMA_UTIL_H + +#include +#include +#include +#include +#include +#include +#include + +#include +#include + +#ifdef __has_include + #if __has_include() + #include + #if defined(_POSIX_MAPPED_FILES) + #include + #endif + #endif +#endif + +#if defined(_WIN32) + #define WIN32_LEAN_AND_MEAN + #ifndef NOMINMAX + #define NOMINMAX + #endif + #include + #include + #include // for _fseeki64 +#endif + +#define LLAMA_ASSERT(x) \ + do { \ + if (!(x)) { \ + fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \ + abort(); \ + } \ + } while (0) + +#ifdef __GNUC__ +#ifdef __MINGW32__ +__attribute__((format(gnu_printf, 1, 2))) +#else +__attribute__((format(printf, 1, 2))) +#endif +#endif +static std::string format(const char * fmt, ...) { + va_list ap, ap2; + va_start(ap, fmt); + va_copy(ap2, ap); + int size = vsnprintf(NULL, 0, fmt, ap); + LLAMA_ASSERT(size >= 0 && size < INT_MAX); + std::vector buf(size + 1); + int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2); + LLAMA_ASSERT(size2 == size); + va_end(ap2); + va_end(ap); + return std::string(buf.data(), size); +} + +struct llama_file { + // use FILE * so we don't have to re-open the file to mmap + FILE * fp; + size_t size; + + llama_file(const char * fname, const char * mode) { + fp = std::fopen(fname, mode); + if (fp == NULL) { + throw format("failed to open %s: %s", fname, std::strerror(errno)); + } + seek(0, SEEK_END); + size = tell(); + seek(0, SEEK_SET); + } + + size_t tell() const { +#ifdef _WIN32 + __int64 ret = _ftelli64(fp); +#else + long ret = std::ftell(fp); +#endif + LLAMA_ASSERT(ret != -1); // this really shouldn't fail + return (size_t) ret; + } + + void seek(size_t offset, int whence) { +#ifdef _WIN32 + int ret = _fseeki64(fp, (__int64) offset, whence); +#else + int ret = std::fseek(fp, (long) offset, whence); +#endif + LLAMA_ASSERT(ret == 0); // same + } + + void read_raw(void * ptr, size_t size) { + if (size == 0) { + return; + } + errno = 0; + std::size_t ret = std::fread(ptr, size, 1, fp); + if (ferror(fp)) { + throw format("read error: %s", strerror(errno)); + } + if (ret != 1) { + throw std::string("unexpectedly reached end of file"); + } + } + + std::uint32_t read_u32() { + std::uint32_t ret; + read_raw(&ret, sizeof(ret)); + return ret; + } + + std::string read_string(std::uint32_t len) { + std::vector chars(len); + read_raw(chars.data(), len); + return std::string(chars.data(), len); + } + + void write_raw(const void * ptr, size_t size) { + if (size == 0) { + return; + } + errno = 0; + size_t ret = std::fwrite(ptr, size, 1, fp); + if (ret != 1) { + throw format("write error: %s", strerror(errno)); + } + } + + void write_u32(std::uint32_t val) { + write_raw(&val, sizeof(val)); + } + + ~llama_file() { + if (fp) { + std::fclose(fp); + } + } +}; + +#if defined(_WIN32) +static std::string llama_format_win_err(DWORD err) { + LPSTR buf; + size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, + NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL); + if (!size) { + return "FormatMessageA failed"; + } + std::string ret(buf, size); + LocalFree(buf); + return ret; +} +#endif + +struct llama_mmap { + void * addr; + size_t size; + + llama_mmap(const llama_mmap &) = delete; + +#ifdef _POSIX_MAPPED_FILES + static constexpr bool SUPPORTED = true; + + llama_mmap(struct llama_file * file, bool prefetch = true) { + size = file->size; + int fd = fileno(file->fp); + int flags = MAP_SHARED; +#ifdef __linux__ + flags |= MAP_POPULATE; +#endif + addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0); + if (addr == MAP_FAILED) { + throw format("mmap failed: %s", strerror(errno)); + } + + if (prefetch) { + // Advise the kernel to preload the mapped memory + if (madvise(addr, file->size, MADV_WILLNEED)) { + fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n", + strerror(errno)); + } + } + } + + ~llama_mmap() { + munmap(addr, size); + } +#elif defined(_WIN32) + static constexpr bool SUPPORTED = true; + + llama_mmap(struct llama_file * file, bool prefetch = true) { + size = file->size; + + HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp)); + + HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL); + DWORD error = GetLastError(); + + if (hMapping == NULL) { + throw format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()); + } + + addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0); + error = GetLastError(); + CloseHandle(hMapping); + + if (addr == NULL) { + throw format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()); + } + + #if _WIN32_WINNT >= _WIN32_WINNT_WIN8 + if (prefetch) { + // Advise the kernel to preload the mapped memory + WIN32_MEMORY_RANGE_ENTRY range; + range.VirtualAddress = addr; + range.NumberOfBytes = (SIZE_T)size; + if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) { + fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + } + } + #else + #pragma message("warning: You are building for pre-Windows 8; prefetch not supported") + #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8 + } + + ~llama_mmap() { + if (!UnmapViewOfFile(addr)) { + fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + } + } +#else + static constexpr bool SUPPORTED = false; + + llama_mmap(struct llama_file *) { + throw std::string("mmap not supported"); + } +#endif +}; + +// Represents some region of memory being locked using mlock or VirtualLock; +// will automatically unlock on destruction. +struct llama_mlock { + void * addr = NULL; + size_t size = 0; + bool failed_already = false; + + llama_mlock() {} + llama_mlock(const llama_mlock &) = delete; + + ~llama_mlock() { + if (size) { + raw_unlock(addr, size); + } + } + + void init(void * addr) { + LLAMA_ASSERT(this->addr == NULL && this->size == 0); + this->addr = addr; + } + + void grow_to(size_t target_size) { + LLAMA_ASSERT(addr); + if (failed_already) { + return; + } + size_t granularity = lock_granularity(); + target_size = (target_size + granularity - 1) & ~(granularity - 1); + if (target_size > size) { + if (raw_lock((uint8_t *) addr + size, target_size - size)) { + size = target_size; + } else { + failed_already = true; + } + } + } + +#ifdef _POSIX_MEMLOCK_RANGE + static constexpr bool SUPPORTED = true; + + size_t lock_granularity() { + return (size_t) sysconf(_SC_PAGESIZE); + } + + #ifdef __APPLE__ + #define MLOCK_SUGGESTION \ + "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \ + "decreasing 'vm.global_no_user_wire_amount'. Also try increasing RLIMIT_MLOCK (ulimit -l).\n" + #else + #define MLOCK_SUGGESTION \ + "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n" + #endif + + bool raw_lock(const void * addr, size_t size) { + if (!mlock(addr, size)) { + return true; + } else { + fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n" MLOCK_SUGGESTION, + size, this->size, std::strerror(errno)); + return false; + } + } + + #undef MLOCK_SUGGESTION + + void raw_unlock(void * addr, size_t size) { + if (munlock(addr, size)) { + fprintf(stderr, "warning: failed to munlock buffer: %s\n", std::strerror(errno)); + } + } +#elif defined(_WIN32) + static constexpr bool SUPPORTED = true; + + size_t lock_granularity() { + SYSTEM_INFO si; + GetSystemInfo(&si); + return (size_t) si.dwPageSize; + } + + bool raw_lock(void * addr, size_t size) { + for (int tries = 1; ; tries++) { + if (VirtualLock(addr, size)) { + return true; + } + if (tries == 2) { + fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n", + size, this->size, llama_format_win_err(GetLastError()).c_str()); + return false; + } + + // It failed but this was only the first try; increase the working + // set size and try again. + SIZE_T min_ws_size, max_ws_size; + if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) { + fprintf(stderr, "warning: GetProcessWorkingSetSize failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + return false; + } + // Per MSDN: "The maximum number of pages that a process can lock + // is equal to the number of pages in its minimum working set minus + // a small overhead." + // Hopefully a megabyte is enough overhead: + size_t increment = size + 1048576; + // The minimum must be <= the maximum, so we need to increase both: + min_ws_size += increment; + max_ws_size += increment; + if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) { + fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + return false; + } + } + } + + void raw_unlock(void * addr, size_t size) { + if (!VirtualUnlock(addr, size)) { + fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n", + llama_format_win_err(GetLastError()).c_str()); + } + } +#else + static constexpr bool SUPPORTED = false; + + void raw_lock(const void * addr, size_t size) { + fprintf(stderr, "warning: mlock not supported on this system\n"); + } + + void raw_unlock(const void * addr, size_t size) {} +#endif +}; + +// Replacement for std::vector that doesn't require zero-initialization. +struct llama_buffer { + uint8_t * addr = NULL; + size_t size = 0; + + void resize(size_t size) { + delete[] addr; + addr = new uint8_t[size]; + this->size = size; + } + + ~llama_buffer() { + delete[] addr; + } +}; +#endif diff --git a/media/llama-leader.jpeg b/media/llama-leader.jpeg new file mode 100644 index 000000000..0b4e6e1cf Binary files /dev/null and b/media/llama-leader.jpeg differ diff --git a/media/llama0-banner.png b/media/llama0-banner.png new file mode 100644 index 000000000..cee3a87f1 Binary files /dev/null and b/media/llama0-banner.png differ diff --git a/media/llama0-logo.png b/media/llama0-logo.png new file mode 100644 index 000000000..e55b38bd9 Binary files /dev/null and b/media/llama0-logo.png differ diff --git a/media/llama1-banner.png b/media/llama1-banner.png new file mode 100644 index 000000000..1e469584e Binary files /dev/null and b/media/llama1-banner.png differ diff --git a/media/llama1-logo.png b/media/llama1-logo.png new file mode 100644 index 000000000..365c5b865 Binary files /dev/null and b/media/llama1-logo.png differ diff --git a/migrate-ggml-2023-03-30-pr613.py b/migrate-ggml-2023-03-30-pr613.py deleted file mode 100644 index b6ef2476e..000000000 --- a/migrate-ggml-2023-03-30-pr613.py +++ /dev/null @@ -1,311 +0,0 @@ -# Migrate ggml file(s) with ggmf magic to ggml file with ggjt magic -# -# We caused a breaking change to the file format on 2023-03-30 in: -# https://github.com/ggerganov/llama.cpp/pull/613 -# -# (1) If you still have the Meta LLaMA .pth files, then close this -# file now; you can just run `convert-pth-to-ggml.py` again to -# migrate to the new format. The tool is easier to use too. It -# isn't necessary anymore to manage split output files because -# the new format always combines things into a single file. -# -# (2) If you deleted the Meta LLaMA .pth files due to save on disk -# space, then this tool is intended to help you. Please check -# out the instructions below. -# -# USAGE -# -# python migrate-ggml-2023-03-30-pr613.py INPUT OUTPUT -# -# PREREQUISITES -# -# pip install numpy -# cd llama.cpp -# make -j4 -# -# EXAMPLE (7B MODEL) -# -# # you can replace all the 'f16' with 'q4_0' if you're using quantized weights -# python migrate-ggml-2023-03-30-pr613.py models/7B/ggml-model-f16.bin models/7B/ggml-model-f16-ggjt.bin -# -# # check that it works -# ./main -m models/7B/ggml-model-f16-ggjt.bin -p 'Question: Do you love me?' -# -# # you can delete the old files -# rm -f models/7B/ggml-model-f16.bin -# mv models/7B/ggml-model-f16-ggjt.bin models/7B/ggml-model-f16.bin -# -# EXAMPLE (13B MODEL) -# -# # you can replace all the 'f16' with 'q4_0' if you're using quantized weights -# python migrate-ggml-2023-03-30-pr613.py models/13B/ggml-model-f16.bin models/13B/ggml-model-f16-ggjt.bin -# -# # check that it works -# ./main -m models/13B/ggml-model-f16-ggjt.bin -p 'Question: Do you love me?' -# -# # you can delete the old files -# rm -f models/13B/ggml-model-f16.bin* -# mv models/13B/ggml-model-f16-ggjt.bin models/13B/ggml-model-f16.bin -# - -import argparse -import os -import sys -import json -import struct -import numpy as np - -QK = 32 - -GGML_TYPE_Q4_0 = 0 -GGML_TYPE_Q4_1 = 1 -GGML_TYPE_I8 = 2 -GGML_TYPE_I16 = 3 -GGML_TYPE_I32 = 4 -GGML_TYPE_F16 = 5 -GGML_TYPE_F32 = 6 - -WTYPE_NAMES = { - 0: "F32", - 1: "F16", - 2: "Q4_0", - 3: "Q4_1", -} - -WTYPES = { - 0: GGML_TYPE_F32, - 1: GGML_TYPE_F16, - 2: GGML_TYPE_Q4_0, - 3: GGML_TYPE_Q4_1, -} - -GGML_BLCK_SIZE = { - GGML_TYPE_Q4_0: QK, - GGML_TYPE_Q4_1: QK, - GGML_TYPE_I8: 1, - GGML_TYPE_I16: 1, - GGML_TYPE_I32: 1, - GGML_TYPE_F16: 1, - GGML_TYPE_F32: 1, -} - -GGML_TYPE_SIZE = { - GGML_TYPE_Q4_0: 4 + QK//2, - GGML_TYPE_Q4_1: 4*2 + QK//2, - GGML_TYPE_I8: 1, - GGML_TYPE_I16: 2, - GGML_TYPE_I32: 4, - GGML_TYPE_F16: 2, - GGML_TYPE_F32: 4, -} - -HPARAMS = [ - 'magic', # int32 - 'version', # int32 - 'n_vocab', # int32 - 'n_embd', # int32 - 'n_mult', # int32 - 'n_head', # int32 - 'n_layer', # int32 - 'n_rot', # int32 - 'f16', # int32 -] - -def read_hparams(fin): - struct_fmt = "i" * len(HPARAMS) - struct_size = struct.calcsize(struct_fmt) - buf = fin.read(struct_size) - ints = struct.unpack(struct_fmt, buf) - hparams = dict(zip(HPARAMS, ints)) - return hparams - -def write_hparams(fout, hparams): - struct_fmt = "i" * len(HPARAMS) - struct_size = struct.calcsize(struct_fmt) - ints = [hparams[h] for h in HPARAMS] - fout.write(struct.pack(struct_fmt, *ints)) - -def read_tokens(fin, hparams): - tokens = [] - for i in range(hparams['n_vocab']): - len_b = fin.read(4) - (length,) = struct.unpack("i", len_b) - word = fin.read(length) - score_b = fin.read(4) - (score,) = struct.unpack("f", score_b) - tokens.append((word, score)) - return tokens - -def write_tokens(fout, tokens): - for word, score in tokens: - fout.write(struct.pack("i", len(word))) - fout.write(word) - fout.write(struct.pack("f", score)) - -def ggml_nelements(shape): - r = 1 - for i in shape: - r *= i - return r - -def ggml_nbytes(shape, ftype): - x = ggml_nelements(shape) - t = WTYPES[ftype] - x *= GGML_TYPE_SIZE[t] - x //= GGML_BLCK_SIZE[t] - return x - -def copy_tensors(fin, fout, part_id, n_parts): - while True: - - b = fin.read(4) - if not b: break - (n_dims,) = struct.unpack("i", b) - b = fin.read(4) - (length,) = struct.unpack("i", b) - b = fin.read(4) - (ftype,) = struct.unpack("i", b) - - assert n_dims in (1, 2) - - partshape = list(range(n_dims)) - for i in range(n_dims): - b = fin.read(4) - partshape[i] = struct.unpack("i", b)[0] - partshape = list(reversed(partshape)) - - name = fin.read(length) - data = fin.read(ggml_nbytes(partshape, ftype)) - - blck_size = GGML_BLCK_SIZE[WTYPES[ftype]] - type_size = GGML_TYPE_SIZE[WTYPES[ftype]] - - print(f"Processing tensor {name} with shape: {partshape} and type: {WTYPE_NAMES[ftype]}") - - # determine dimension along which multipart tensor is sharded - # - # split_dim 0 regex: - # - output.* - # - layers.*.attention.wq.weight - # - layers.*.attention.wk.weight - # - layers.*.attention.wv.weight - # - layers.*.feed_forward.w1.weight - # - layers.*.feed_forward.w3.weight - # - # split_dim 1 regex: - # - tok_embeddings.* - # - layers.*.attention.wo.weight - # - layers.*.feed_forward.w2.weight - # - if n_dims > 1: - split_dim = 1 - if b"tok_embeddings" in name: - split_dim = 1 - elif b"layers" in name: - if b"attention.wo.weight" in name: - split_dim = 1 - elif b"feed_forward.w2.weight" in name: - split_dim = 1 - else: - split_dim = 0 - elif b"output" in name: - split_dim = 0 - - # output tensor header - fullshape = list(partshape) - if n_dims > 1: - fullshape[split_dim] *= n_parts - fout.write(struct.pack("iii", n_dims, len(name), ftype)) - for dim in reversed(fullshape): - fout.write(struct.pack("i", dim)) - fout.write(name) - - # ensure tensor data is aligned - tensor_data_offset = fout.tell() - while tensor_data_offset % QK != 0: - fout.write(struct.pack("B", 0)) - tensor_data_offset += 1 - - # output unified mappable tensor data - if n_dims == 1 or n_parts == 1: - # copy tensor which we thankfully received in one piece - if part_id == 0: - fout.write(data) - elif split_dim == 0: - # reassemble multifile tensor containing some of the rows - rows_per_chunk = partshape[0] - current_row = part_id * rows_per_chunk - bytes_per_row = fullshape[1] // blck_size * type_size - offset = current_row * bytes_per_row - fout.seek(tensor_data_offset + offset) - fout.write(data) - elif split_dim == 1: - # reassemble multifile tensor containing some of the cols - cols_per_chunk = partshape[1] - current_col = part_id * cols_per_chunk - bpr = partshape[1] // blck_size * type_size - bytes_per_row = fullshape[1] // blck_size * type_size - offset_current_col = current_col // blck_size * type_size - for row in range(partshape[0]): - offset_row = row * bytes_per_row - offset = offset_row + offset_current_col - fout.seek(tensor_data_offset + offset) - fout.write(data[row * bpr:row * bpr + bpr]) - - # advance file position to next tensor - fout.seek(tensor_data_offset + ggml_nbytes(fullshape, ftype)) - -def parse_args(): - parser = argparse.ArgumentParser(description='Migrate from GGML to new GGJT file format') - parser.add_argument('fin_path', help='your old ggml file (leave out the .1 .2 etc.)') - parser.add_argument('fout_path', help='your new ggjt file name') - return parser.parse_args() - -def main(): - args = parse_args() - assert args.fin_path - assert args.fout_path - assert args.fin_path != args.fout_path - - with open(args.fin_path, "rb") as fin: - hparams = read_hparams(fin) - tokens = read_tokens(fin, hparams) - - if hparams['magic'] == 0x67676a74: # ggjt - print(f"{args.fin_path}: input ggml has already been converted to 'ggjt' magic\n") - sys.exit(1) - - if hparams['magic'] != 0x67676d66: # ggmf - print(f"{args.fin_path}: input ggml file doesn't have expected 'ggmf' magic: {hparams['magic']:#x}\n") - sys.exit(1) - - hparams['magic'] = 0x67676a74 # ggjt - - # count number of multipart files by convention - n_parts = 1 - while True: - if os.path.exists(f"{args.fin_path}.{n_parts}"): - n_parts += 1 - else: - break - - # we output a single file for ggml - with open(args.fout_path, "wb") as fout: - write_hparams(fout, hparams) - write_tokens(fout, tokens) - offset_of_tensors = fout.tell() - # the tensors we load could be split across multiple files - for part_id in range(n_parts): - fout.seek(offset_of_tensors) - print(f"Processing part {part_id+1} of {n_parts}\n") - fin_path = args.fin_path - if part_id > 0: - fin_path += f".{part_id}" - with open(fin_path, "rb") as fin: - read_tokens(fin, read_hparams(fin)) - copy_tensors(fin, fout, part_id, n_parts) - - print(f"Done. Output file: {args.fout_path}\n") - -if __name__ == "__main__": - main() diff --git a/pocs/CMakeLists.txt b/pocs/CMakeLists.txt new file mode 100644 index 000000000..03e1d2c04 --- /dev/null +++ b/pocs/CMakeLists.txt @@ -0,0 +1,12 @@ +# dependencies + +find_package(Threads REQUIRED) + +# third-party + +include_directories(${CMAKE_CURRENT_SOURCE_DIR}) + +if (EMSCRIPTEN) +else() + add_subdirectory(vdot) +endif() diff --git a/pocs/vdot/CMakeLists.txt b/pocs/vdot/CMakeLists.txt new file mode 100644 index 000000000..cbc852236 --- /dev/null +++ b/pocs/vdot/CMakeLists.txt @@ -0,0 +1,4 @@ +set(TARGET vdot) +add_executable(${TARGET} vdot.cpp) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/pocs/vdot/vdot.cpp b/pocs/vdot/vdot.cpp new file mode 100644 index 000000000..26bf50c9a --- /dev/null +++ b/pocs/vdot/vdot.cpp @@ -0,0 +1,305 @@ +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include + +constexpr int kVecSize = 1 << 18; + +float drawFromGaussianPdf(std::mt19937& rndm) { + constexpr double kScale = 1./(1. + std::mt19937::max()); + constexpr double kTwoPiTimesScale = 6.28318530717958647692*kScale; + static float lastX; + static bool haveX = false; + if (haveX) { haveX = false; return lastX; } + auto r = sqrt(-2*log(1 - kScale*rndm())); + auto phi = kTwoPiTimesScale * rndm(); + lastX = r*sin(phi); + haveX = true; + return r*cos(phi); +} +void fillRandomGaussianFloats(std::vector& values, std::mt19937& rndm, float mean = 0) { + for (auto& v : values) v = mean + drawFromGaussianPdf(rndm); +} + +// Copy-pasted from ggml.c +#define QK4_0 32 +typedef struct { + float d; // delta + uint8_t qs[QK4_0 / 2]; // nibbles / quants +} block_q4_0; +static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding"); + +#define QK4_1 32 +typedef struct { + float d; // delta + float m; // min + uint8_t qs[QK4_1 / 2]; // nibbles / quants +} block_q4_1; +static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding"); + +// Copy-pasted from ggml.c +#define QK8_0 32 +typedef struct { + float d; // delta + int8_t qs[QK8_0]; // quants +} block_q8_0; +static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding"); + +// "Scalar" dot product between the quantized vector x and float vector y +inline double dot(int n, const block_q4_0* x, const float* y) { + const static float kValues[16] = {-8.f, -7.f, -6.f, -5.f, -4.f, -3.f, -2.f, -1.f, 0.f, 1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f}; + constexpr uint32_t kMask1 = 0x0f0f0f0f; + uint32_t u1, u2; + auto q1 = (const uint8_t*)&u1; + auto q2 = (const uint8_t*)&u2; + double sum = 0; + for (int i=0; id; + auto u = (const uint32_t*)x->qs; + float s = 0; + for (int k=0; k<4; ++k) { + u1 = u[k] & kMask1; + u2 = (u[k] >> 4) & kMask1; + s += y[0]*kValues[q1[0]] + y[1]*kValues[q2[0]] + + y[2]*kValues[q1[1]] + y[3]*kValues[q2[1]] + + y[4]*kValues[q1[2]] + y[5]*kValues[q2[2]] + + y[6]*kValues[q1[3]] + y[7]*kValues[q2[3]]; + y += 8; + } + sum += s*d; + ++x; + } + return sum; +} +// Alternative version of the above. Faster on my Mac (~45 us vs ~55 us per dot product), +// but about the same on X86_64 (Ryzen 7950X CPU). +inline double dot3(int n, const block_q4_0* x, const float* y) { + const static std::pair kValues[256] = { + {-8.f, -8.f}, {-7.f, -8.f}, {-6.f, -8.f}, {-5.f, -8.f}, {-4.f, -8.f}, {-3.f, -8.f}, {-2.f, -8.f}, {-1.f, -8.f}, + { 0.f, -8.f}, { 1.f, -8.f}, { 2.f, -8.f}, { 3.f, -8.f}, { 4.f, -8.f}, { 5.f, -8.f}, { 6.f, -8.f}, { 7.f, -8.f}, + {-8.f, -7.f}, {-7.f, -7.f}, {-6.f, -7.f}, {-5.f, -7.f}, {-4.f, -7.f}, {-3.f, -7.f}, {-2.f, -7.f}, {-1.f, -7.f}, + { 0.f, -7.f}, { 1.f, -7.f}, { 2.f, -7.f}, { 3.f, -7.f}, { 4.f, -7.f}, { 5.f, -7.f}, { 6.f, -7.f}, { 7.f, -7.f}, + {-8.f, -6.f}, {-7.f, -6.f}, {-6.f, -6.f}, {-5.f, -6.f}, {-4.f, -6.f}, {-3.f, -6.f}, {-2.f, -6.f}, {-1.f, -6.f}, + { 0.f, -6.f}, { 1.f, -6.f}, { 2.f, -6.f}, { 3.f, -6.f}, { 4.f, -6.f}, { 5.f, -6.f}, { 6.f, -6.f}, { 7.f, -6.f}, + {-8.f, -5.f}, {-7.f, -5.f}, {-6.f, -5.f}, {-5.f, -5.f}, {-4.f, -5.f}, {-3.f, -5.f}, {-2.f, -5.f}, {-1.f, -5.f}, + { 0.f, -5.f}, { 1.f, -5.f}, { 2.f, -5.f}, { 3.f, -5.f}, { 4.f, -5.f}, { 5.f, -5.f}, { 6.f, -5.f}, { 7.f, -5.f}, + {-8.f, -4.f}, {-7.f, -4.f}, {-6.f, -4.f}, {-5.f, -4.f}, {-4.f, -4.f}, {-3.f, -4.f}, {-2.f, -4.f}, {-1.f, -4.f}, + { 0.f, -4.f}, { 1.f, -4.f}, { 2.f, -4.f}, { 3.f, -4.f}, { 4.f, -4.f}, { 5.f, -4.f}, { 6.f, -4.f}, { 7.f, -4.f}, + {-8.f, -3.f}, {-7.f, -3.f}, {-6.f, -3.f}, {-5.f, -3.f}, {-4.f, -3.f}, {-3.f, -3.f}, {-2.f, -3.f}, {-1.f, -3.f}, + { 0.f, -3.f}, { 1.f, -3.f}, { 2.f, -3.f}, { 3.f, -3.f}, { 4.f, -3.f}, { 5.f, -3.f}, { 6.f, -3.f}, { 7.f, -3.f}, + {-8.f, -2.f}, {-7.f, -2.f}, {-6.f, -2.f}, {-5.f, -2.f}, {-4.f, -2.f}, {-3.f, -2.f}, {-2.f, -2.f}, {-1.f, -2.f}, + { 0.f, -2.f}, { 1.f, -2.f}, { 2.f, -2.f}, { 3.f, -2.f}, { 4.f, -2.f}, { 5.f, -2.f}, { 6.f, -2.f}, { 7.f, -2.f}, + {-8.f, -1.f}, {-7.f, -1.f}, {-6.f, -1.f}, {-5.f, -1.f}, {-4.f, -1.f}, {-3.f, -1.f}, {-2.f, -1.f}, {-1.f, -1.f}, + { 0.f, -1.f}, { 1.f, -1.f}, { 2.f, -1.f}, { 3.f, -1.f}, { 4.f, -1.f}, { 5.f, -1.f}, { 6.f, -1.f}, { 7.f, -1.f}, + {-8.f, 0.f}, {-7.f, 0.f}, {-6.f, 0.f}, {-5.f, 0.f}, {-4.f, 0.f}, {-3.f, 0.f}, {-2.f, 0.f}, {-1.f, 0.f}, + { 0.f, 0.f}, { 1.f, 0.f}, { 2.f, 0.f}, { 3.f, 0.f}, { 4.f, 0.f}, { 5.f, 0.f}, { 6.f, 0.f}, { 7.f, 0.f}, + {-8.f, 1.f}, {-7.f, 1.f}, {-6.f, 1.f}, {-5.f, 1.f}, {-4.f, 1.f}, {-3.f, 1.f}, {-2.f, 1.f}, {-1.f, 1.f}, + { 0.f, 1.f}, { 1.f, 1.f}, { 2.f, 1.f}, { 3.f, 1.f}, { 4.f, 1.f}, { 5.f, 1.f}, { 6.f, 1.f}, { 7.f, 1.f}, + {-8.f, 2.f}, {-7.f, 2.f}, {-6.f, 2.f}, {-5.f, 2.f}, {-4.f, 2.f}, {-3.f, 2.f}, {-2.f, 2.f}, {-1.f, 2.f}, + { 0.f, 2.f}, { 1.f, 2.f}, { 2.f, 2.f}, { 3.f, 2.f}, { 4.f, 2.f}, { 5.f, 2.f}, { 6.f, 2.f}, { 7.f, 2.f}, + {-8.f, 3.f}, {-7.f, 3.f}, {-6.f, 3.f}, {-5.f, 3.f}, {-4.f, 3.f}, {-3.f, 3.f}, {-2.f, 3.f}, {-1.f, 3.f}, + { 0.f, 3.f}, { 1.f, 3.f}, { 2.f, 3.f}, { 3.f, 3.f}, { 4.f, 3.f}, { 5.f, 3.f}, { 6.f, 3.f}, { 7.f, 3.f}, + {-8.f, 4.f}, {-7.f, 4.f}, {-6.f, 4.f}, {-5.f, 4.f}, {-4.f, 4.f}, {-3.f, 4.f}, {-2.f, 4.f}, {-1.f, 4.f}, + { 0.f, 4.f}, { 1.f, 4.f}, { 2.f, 4.f}, { 3.f, 4.f}, { 4.f, 4.f}, { 5.f, 4.f}, { 6.f, 4.f}, { 7.f, 4.f}, + {-8.f, 5.f}, {-7.f, 5.f}, {-6.f, 5.f}, {-5.f, 5.f}, {-4.f, 5.f}, {-3.f, 5.f}, {-2.f, 5.f}, {-1.f, 5.f}, + { 0.f, 5.f}, { 1.f, 5.f}, { 2.f, 5.f}, { 3.f, 5.f}, { 4.f, 5.f}, { 5.f, 5.f}, { 6.f, 5.f}, { 7.f, 5.f}, + {-8.f, 6.f}, {-7.f, 6.f}, {-6.f, 6.f}, {-5.f, 6.f}, {-4.f, 6.f}, {-3.f, 6.f}, {-2.f, 6.f}, {-1.f, 6.f}, + { 0.f, 6.f}, { 1.f, 6.f}, { 2.f, 6.f}, { 3.f, 6.f}, { 4.f, 6.f}, { 5.f, 6.f}, { 6.f, 6.f}, { 7.f, 6.f}, + {-8.f, 7.f}, {-7.f, 7.f}, {-6.f, 7.f}, {-5.f, 7.f}, {-4.f, 7.f}, {-3.f, 7.f}, {-2.f, 7.f}, {-1.f, 7.f}, + { 0.f, 7.f}, { 1.f, 7.f}, { 2.f, 7.f}, { 3.f, 7.f}, { 4.f, 7.f}, { 5.f, 7.f}, { 6.f, 7.f}, { 7.f, 7.f} + }; + double sum = 0; + for (int i=0; id; + auto q = x->qs; + float s = 0; + for (int k=0; k<4; ++k) { + s += y[0]*kValues[q[0]].first + y[1]*kValues[q[0]].second + + y[2]*kValues[q[1]].first + y[3]*kValues[q[1]].second + + y[4]*kValues[q[2]].first + y[5]*kValues[q[2]].second + + y[6]*kValues[q[3]].first + y[7]*kValues[q[3]].second; + y += 8; q += 4; + } + sum += s*d; + ++x; + } + return sum; +} + +inline double dot41(int n, const block_q4_1* x, const float* y) { + const static float kValues[16] = {0.f, 1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f, 8.f, 9.f, 10.f, 11.f, 12.f, 13.f, 14.f, 15.f}; + constexpr uint32_t kMask1 = 0x0f0f0f0f; + uint32_t u1, u2; + auto q1 = (const uint8_t*)&u1; + auto q2 = (const uint8_t*)&u2; + double sum = 0; + for (int i=0; iqs; + float s = 0, s1 = 0; + for (int k=0; k<4; ++k) { + u1 = u[k] & kMask1; + u2 = (u[k] >> 4) & kMask1; + s += y[0]*kValues[q1[0]] + y[1]*kValues[q2[0]] + + y[2]*kValues[q1[1]] + y[3]*kValues[q2[1]] + + y[4]*kValues[q1[2]] + y[5]*kValues[q2[2]] + + y[6]*kValues[q1[3]] + y[7]*kValues[q2[3]]; + s1 += y[0] + y[1] + y[2] + y[3] + y[4] + y[5] + y[6] + y[7]; + y += 8; + } + sum += s*x->d + s1*x->m; + ++x; + } + return sum; +} + +// Copy-pasted from ggml.c +static void quantize_row_q8_0_reference(const float *x, block_q8_0 *y, int k) { + assert(k % QK8_0 == 0); + const int nb = k / QK8_0; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + + for (int l = 0; l < QK8_0; l++) { + const float v = x[i*QK8_0 + l]; + amax = std::max(amax, fabsf(v)); + } + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + for (int l = 0; l < QK8_0; ++l) { + const float v = x[i*QK8_0 + l]*id; + y[i].qs[l] = roundf(v); + } + } +} + +// Copy-pasted from ggml.c +static void dot_q4_q8(const int n, float* s, const void* vx, const void* vy) { + const int nb = n / QK8_0; + const block_q4_0* x = (const block_q4_0*)vx; + const block_q8_0* y = (const block_q8_0*)vy; + float sumf = 0; + for (int i = 0; i < nb; i++) { + const float d0 = x[i].d; + const float d1 = y[i].d; + + const uint8_t * p0 = x[i].qs; + const int8_t * p1 = y[i].qs; + + int sumi = 0; + for (int j = 0; j < QK8_0/2; j++) { + const uint8_t v0 = p0[j]; + + const int i0 = (int8_t) (v0 & 0xf) - 8; + const int i1 = (int8_t) (v0 >> 4) - 8; + + const int i2 = p1[2*j + 0]; + const int i3 = p1[2*j + 1]; + + sumi += i0*i2 + i1*i3; + } + sumf += d0*d1*sumi; + } + *s = sumf; +} + +int main(int argc, char** argv) { + + int nloop = argc > 1 ? atoi(argv[1]) : 10; + bool scalar = argc > 2 ? atoi(argv[2]) : false; + bool useQ4_1 = argc > 3 ? atoi(argv[3]) : false; + + if (scalar && useQ4_1) { + printf("It is not possible to use Q4_1 quantization and scalar implementations\n"); + return 1; + } + + std::mt19937 rndm(1234); + + std::vector x1(kVecSize), y1(kVecSize); + int n4 = useQ4_1 ? kVecSize / QK4_1 : kVecSize / QK4_0; n4 = 64*((n4 + 63)/64); + int n8 = kVecSize / QK8_0; n8 = 64*((n8 + 63)/64); + + auto funcs = useQ4_1 ? ggml_internal_get_quantize_fn(GGML_TYPE_Q4_1) : ggml_internal_get_quantize_fn(GGML_TYPE_Q4_0); + + std::vector q40; + std::vector q41; + if (useQ4_1) q41.resize(n4); + else q40.resize(n4); + std::vector q8(n8); + std::vector H(16, 0); + double sumt = 0, sumt2 = 0, maxt = 0; + double sumqt = 0, sumqt2 = 0, maxqt = 0; + double sum = 0, sumq = 0, exactSum = 0; + for (int iloop=0; iloop(t2-t1).count(); + sumt += t; sumt2 += t*t; maxt = std::max(maxt, t); + + // And now measure the time needed to quantize y and perform the dot product with the quantized y + t1 = std::chrono::high_resolution_clock::now(); + float result; + if (scalar) { + quantize_row_q8_0_reference(y1.data(), q8.data(), kVecSize); + dot_q4_q8(kVecSize, &result, q40.data(), q8.data()); + } + else { + funcs.quantize_row_q_dot(y1.data(), q8.data(), kVecSize); + if (useQ4_1) funcs.vec_dot_q(kVecSize, &result, q41.data(), q8.data()); + else funcs.vec_dot_q(kVecSize, &result, q40.data(), q8.data()); + } + sumq += result; + t2 = std::chrono::high_resolution_clock::now(); + t = 1e-3*std::chrono::duration_cast(t2-t1).count(); + sumqt += t; sumqt2 += t*t; maxqt = std::max(maxqt, t); + + } + + // Report the time (and the average of the dot products so the compiler does not come up with the idea + // of optimizing away the function calls after figuring that the result is not used). + sum /= nloop; sumq /= nloop; + exactSum /= nloop; + printf("Exact result: = %g\n",exactSum); + printf(" = %g, %g\n",sum,sumq); + sumt /= nloop; sumt2 /= nloop; sumt2 -= sumt*sumt; + if (sumt2 > 0) sumt2 = sqrt(sumt2); + printf("time = %g +/- %g us. maxt = %g us\n",sumt,sumt2,maxt); + sumqt /= nloop; sumqt2 /= nloop; sumqt2 -= sumqt*sumqt; + if (sumqt2 > 0) sumqt2 = sqrt(sumqt2); + printf("timeq = %g +/- %g us. maxt = %g us\n",sumqt,sumqt2,maxqt); + return 0; +} diff --git a/prompts/chat-with-bob.txt b/prompts/chat-with-bob.txt index 009da39ae..ad494d831 100644 --- a/prompts/chat-with-bob.txt +++ b/prompts/chat-with-bob.txt @@ -4,4 +4,4 @@ User: Hello, Bob. Bob: Hello. How may I help you today? User: Please tell me the largest city in Europe. Bob: Sure. The largest city in Europe is Moscow, the capital of Russia. -User: +User: \ No newline at end of file diff --git a/prompts/reason-act.txt b/prompts/reason-act.txt index 872016631..a4f4f4ee6 100644 --- a/prompts/reason-act.txt +++ b/prompts/reason-act.txt @@ -15,4 +15,4 @@ Answer: The calculate tool says it is 9.3333333333 Question: What is capital of france? Thought: Do I need to use an action? No, I know the answer Answer: Paris is the capital of France -Question: +Question: \ No newline at end of file diff --git a/requirements.txt b/requirements.txt new file mode 100644 index 000000000..6c32cbd04 --- /dev/null +++ b/requirements.txt @@ -0,0 +1,2 @@ +numpy==1.24 +sentencepiece==0.1.98 diff --git a/tests/test-tokenizer-0.cpp b/tests/test-tokenizer-0.cpp index 55b086dae..b08984571 100644 --- a/tests/test-tokenizer-0.cpp +++ b/tests/test-tokenizer-0.cpp @@ -5,13 +5,17 @@ #include #include -static const std::map> k_tests = { - { "Hello World", { 1, 10994, 2787, }, }, - { " Hello World", { 1, 15043, 2787, }, }, - { " Hello World!", { 1, 15043, 2787, 29991, }, }, - { " this is 🦙.cpp", { 1, 445, 338, 29871, 243, 162, 169, 156, 29889, 8223, }, }, - { "w048 7tuijk dsdfhu", { 1, 29893, 29900, 29946, 29947, 29871, 29955, 9161, 13535, 18031, 2176, 6905, }, }, - { "нещо на Български", { 1, 821, 4851, 665, 1386, 29713, 1305, }, }, +static const std::map> & k_tests() +{ + static std::map> _k_tests = { + { "Hello World", { 1, 10994, 2787, }, }, + { " Hello World", { 1, 15043, 2787, }, }, + { " Hello World!", { 1, 15043, 2787, 29991, }, }, + { " this is 🦙.cpp", { 1, 445, 338, 29871, 243, 162, 169, 156, 29889, 8223, }, }, + { "w048 7tuijk dsdfhu", { 1, 29893, 29900, 29946, 29947, 29871, 29955, 9161, 13535, 18031, 2176, 6905, }, }, + { "нещо на Български", { 1, 821, 4851, 665, 1386, 29713, 1305, }, }, + }; + return _k_tests; }; int main(int argc, char **argv) { @@ -47,7 +51,7 @@ int main(int argc, char **argv) { return 2; } - for (const auto & test_kv : k_tests) { + for (const auto & test_kv : k_tests()) { std::vector res(test_kv.first.size()); const int n = llama_tokenize(ctx, test_kv.first.c_str(), res.data(), res.size(), true); res.resize(n);