diff --git a/.devops/full-cuda.Dockerfile b/.devops/full-cuda.Dockerfile index 2a7da586a7086..61f671465ca71 100644 --- a/.devops/full-cuda.Dockerfile +++ b/.devops/full-cuda.Dockerfile @@ -6,7 +6,7 @@ ARG CUDA_VERSION=11.7.1 # Target the CUDA build image ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION} -FROM ${BASE_CUDA_DEV_CONTAINER} as build +FROM ${BASE_CUDA_DEV_CONTAINER} AS build # Unless otherwise specified, we make a fat build. ARG CUDA_DOCKER_ARCH=all diff --git a/.devops/full-rocm.Dockerfile b/.devops/full-rocm.Dockerfile index 5cbd2e7a131aa..680d1cb92205d 100644 --- a/.devops/full-rocm.Dockerfile +++ b/.devops/full-rocm.Dockerfile @@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6 # Target the CUDA build image ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete -FROM ${BASE_ROCM_DEV_CONTAINER} as build +FROM ${BASE_ROCM_DEV_CONTAINER} AS build # Unless otherwise specified, we make a fat build. # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878 diff --git a/.devops/full.Dockerfile b/.devops/full.Dockerfile index 6f19afa9c37f1..2a06f82b738ae 100644 --- a/.devops/full.Dockerfile +++ b/.devops/full.Dockerfile @@ -1,6 +1,6 @@ ARG UBUNTU_VERSION=22.04 -FROM ubuntu:$UBUNTU_VERSION as build +FROM ubuntu:$UBUNTU_VERSION AS build RUN apt-get update && \ apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1 diff --git a/.devops/llama-cli-cuda.Dockerfile b/.devops/llama-cli-cuda.Dockerfile index bff946cbc2405..8eda63a89504d 100644 --- a/.devops/llama-cli-cuda.Dockerfile +++ b/.devops/llama-cli-cuda.Dockerfile @@ -6,7 +6,7 @@ ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VER # Target the CUDA runtime image ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION} -FROM ${BASE_CUDA_DEV_CONTAINER} as build +FROM ${BASE_CUDA_DEV_CONTAINER} AS build # Unless otherwise specified, we make a fat build. ARG CUDA_DOCKER_ARCH=all @@ -25,7 +25,7 @@ ENV GGML_CUDA=1 RUN make -j$(nproc) llama-cli -FROM ${BASE_CUDA_RUN_CONTAINER} as runtime +FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime RUN apt-get update && \ apt-get install -y libgomp1 diff --git a/.devops/llama-cli-intel.Dockerfile b/.devops/llama-cli-intel.Dockerfile index bd816f9f5a1e6..79dba06a77d6e 100644 --- a/.devops/llama-cli-intel.Dockerfile +++ b/.devops/llama-cli-intel.Dockerfile @@ -1,6 +1,6 @@ ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04 -FROM intel/oneapi-basekit:$ONEAPI_VERSION as build +FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build ARG GGML_SYCL_F16=OFF RUN apt-get update && \ @@ -14,10 +14,12 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \ echo "GGML_SYCL_F16 is set" && \ export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \ fi && \ - cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \ + echo "Building with static libs" && \ + cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \ + ${OPT_SYCL_F16} -DBUILD_SHARED_LIBS=OFF && \ cmake --build build --config Release --target llama-cli -FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime +FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime COPY --from=build /app/build/bin/llama-cli /llama-cli diff --git a/.devops/llama-cli-rocm.Dockerfile b/.devops/llama-cli-rocm.Dockerfile index caa507b081885..c3d1ab06702ec 100644 --- a/.devops/llama-cli-rocm.Dockerfile +++ b/.devops/llama-cli-rocm.Dockerfile @@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6 # Target the CUDA build image ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete -FROM ${BASE_ROCM_DEV_CONTAINER} as build +FROM ${BASE_ROCM_DEV_CONTAINER} AS build # Unless otherwise specified, we make a fat build. # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878 diff --git a/.devops/llama-cli-vulkan.Dockerfile b/.devops/llama-cli-vulkan.Dockerfile index 6155d588147c9..9b0dad8bf7a13 100644 --- a/.devops/llama-cli-vulkan.Dockerfile +++ b/.devops/llama-cli-vulkan.Dockerfile @@ -1,6 +1,6 @@ ARG UBUNTU_VERSION=jammy -FROM ubuntu:$UBUNTU_VERSION as build +FROM ubuntu:$UBUNTU_VERSION AS build # Install build tools RUN apt update && apt install -y git build-essential cmake wget libgomp1 diff --git a/.devops/llama-cli.Dockerfile b/.devops/llama-cli.Dockerfile index 38382bfc9add2..7f741aa46ecf0 100644 --- a/.devops/llama-cli.Dockerfile +++ b/.devops/llama-cli.Dockerfile @@ -1,6 +1,6 @@ ARG UBUNTU_VERSION=22.04 -FROM ubuntu:$UBUNTU_VERSION as build +FROM ubuntu:$UBUNTU_VERSION AS build RUN apt-get update && \ apt-get install -y build-essential git @@ -11,7 +11,7 @@ COPY . . RUN make -j$(nproc) llama-cli -FROM ubuntu:$UBUNTU_VERSION as runtime +FROM ubuntu:$UBUNTU_VERSION AS runtime RUN apt-get update && \ apt-get install -y libgomp1 diff --git a/.devops/llama-server-cuda.Dockerfile b/.devops/llama-server-cuda.Dockerfile index d7eaa0925a4a5..67328cf1c1788 100644 --- a/.devops/llama-server-cuda.Dockerfile +++ b/.devops/llama-server-cuda.Dockerfile @@ -6,7 +6,7 @@ ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VER # Target the CUDA runtime image ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION} -FROM ${BASE_CUDA_DEV_CONTAINER} as build +FROM ${BASE_CUDA_DEV_CONTAINER} AS build # Unless otherwise specified, we make a fat build. ARG CUDA_DOCKER_ARCH=all @@ -27,7 +27,7 @@ ENV LLAMA_CURL=1 RUN make -j$(nproc) llama-server -FROM ${BASE_CUDA_RUN_CONTAINER} as runtime +FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime RUN apt-get update && \ apt-get install -y libcurl4-openssl-dev libgomp1 curl diff --git a/.devops/llama-server-intel.Dockerfile b/.devops/llama-server-intel.Dockerfile index 8f8fef8c09730..f525658dddfe5 100644 --- a/.devops/llama-server-intel.Dockerfile +++ b/.devops/llama-server-intel.Dockerfile @@ -1,6 +1,6 @@ ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04 -FROM intel/oneapi-basekit:$ONEAPI_VERSION as build +FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build ARG GGML_SYCL_F16=OFF RUN apt-get update && \ @@ -14,10 +14,11 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \ echo "GGML_SYCL_F16 is set" && \ export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \ fi && \ + echo "Building with dynamic libs" && \ cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \ cmake --build build --config Release --target llama-server -FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime +FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime RUN apt-get update && \ apt-get install -y libcurl4-openssl-dev curl diff --git a/.devops/llama-server-rocm.Dockerfile b/.devops/llama-server-rocm.Dockerfile index af96c332595b1..763b4cd3f1c2e 100644 --- a/.devops/llama-server-rocm.Dockerfile +++ b/.devops/llama-server-rocm.Dockerfile @@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6 # Target the CUDA build image ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete -FROM ${BASE_ROCM_DEV_CONTAINER} as build +FROM ${BASE_ROCM_DEV_CONTAINER} AS build # Unless otherwise specified, we make a fat build. # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878 diff --git a/.devops/llama-server-vulkan.Dockerfile b/.devops/llama-server-vulkan.Dockerfile index 49062f84bf314..13a61ffd8454b 100644 --- a/.devops/llama-server-vulkan.Dockerfile +++ b/.devops/llama-server-vulkan.Dockerfile @@ -1,6 +1,6 @@ ARG UBUNTU_VERSION=jammy -FROM ubuntu:$UBUNTU_VERSION as build +FROM ubuntu:$UBUNTU_VERSION AS build # Install build tools RUN apt update && apt install -y git build-essential cmake wget diff --git a/.devops/llama-server.Dockerfile b/.devops/llama-server.Dockerfile index a53a5c999c8cd..b631d580639b2 100644 --- a/.devops/llama-server.Dockerfile +++ b/.devops/llama-server.Dockerfile @@ -1,6 +1,6 @@ ARG UBUNTU_VERSION=22.04 -FROM ubuntu:$UBUNTU_VERSION as build +FROM ubuntu:$UBUNTU_VERSION AS build RUN apt-get update && \ apt-get install -y build-essential git libcurl4-openssl-dev curl @@ -13,7 +13,7 @@ ENV LLAMA_CURL=1 RUN make -j$(nproc) llama-server -FROM ubuntu:$UBUNTU_VERSION as runtime +FROM ubuntu:$UBUNTU_VERSION AS runtime RUN apt-get update && \ apt-get install -y libcurl4-openssl-dev libgomp1 diff --git a/.devops/nix/apps.nix b/.devops/nix/apps.nix index 897fce4d324c1..0ecf19fc56d55 100644 --- a/.devops/nix/apps.nix +++ b/.devops/nix/apps.nix @@ -10,7 +10,6 @@ "llama-embedding" "llama-server" "llama-quantize" - "llama-train-text-from-scratch" ]; mkApp = name: { type = "app"; diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index 00be596ce3fb7..911c42ecb16cf 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -18,6 +18,7 @@ vulkan-headers, vulkan-loader, curl, + shaderc, useBlas ? builtins.all (x: !x) [ useCuda useMetalKit @@ -146,6 +147,7 @@ let vulkanBuildInputs = [ vulkan-headers vulkan-loader + shaderc ]; in diff --git a/.devops/tools.sh b/.devops/tools.sh index 335382f69ee1f..24dcfd35079cb 100755 --- a/.devops/tools.sh +++ b/.devops/tools.sh @@ -8,13 +8,11 @@ arg1="$1" shift if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then - python3 ./convert-hf-to-gguf.py "$@" + python3 ./convert_hf_to_gguf.py "$@" elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then ./llama-quantize "$@" elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then ./llama-cli "$@" -elif [[ "$arg1" == '--finetune' || "$arg1" == '-f' ]]; then - ./llama-finetune "$@" elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then echo "Converting PTH to GGML..." for i in `ls $1/$2/ggml-model-f16.bin*`; do @@ -36,8 +34,6 @@ else echo " ex: --outtype f16 \"/models/7B/\" " echo " --quantize (-q): Optimize with quantization process ggml" echo " ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2" - echo " --finetune (-f): Run finetune command to create a lora finetune of the model" - echo " See documentation for finetune for command-line parameters" echo " --all-in-one (-a): Execute --convert & --quantize" echo " ex: \"/models/\" 7B" echo " --server (-s): Run a model on the server" diff --git a/.github/labeler.yml b/.github/labeler.yml index 9c0397d16e1a0..89436740d1ffb 100644 --- a/.github/labeler.yml +++ b/.github/labeler.yml @@ -16,7 +16,9 @@ SYCL: - any-glob-to-any-file: - ggml/include/ggml-sycl.h - ggml/src/ggml-sycl.cpp - - README-sycl.md + - ggml/src/ggml-sycl/** + - docs/backend/SYCL.md + - examples/sycl/** Nvidia GPU: - changed-files: - any-glob-to-any-file: diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 1e344db6b09e9..a1e183d118268 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -355,8 +355,10 @@ jobs: - name: Dependencies id: depends run: | - sudo apt-get update - sudo apt-get install build-essential libvulkan-dev + wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add - + sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list + sudo apt-get update -y + sudo apt-get install -y build-essential vulkan-sdk - name: Build id: cmake_build @@ -858,7 +860,7 @@ jobs: mkdir build cd build cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON - cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS} + cmake --build . --config Release -j $((${env:NUMBER_OF_PROCESSORS} - 1)) - name: Determine tag name id: tag diff --git a/.gitignore b/.gitignore index 4866f612277f7..7c7dee0c607aa 100644 --- a/.gitignore +++ b/.gitignore @@ -61,6 +61,11 @@ llama-batched-swift out/ tmp/ +# Deprecated + +/main +/server + # CI !.github/workflows/*.yml diff --git a/CMakeLists.txt b/CMakeLists.txt index 4f6cd68724aa6..7937091227ad4 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -50,9 +50,6 @@ endif() # option list # -# general -option(LLAMA_CCACHE "llama: use ccache if available" ON) - # debug option(LLAMA_ALL_WARNINGS "llama: enable all compiler warnings" ON) option(LLAMA_ALL_WARNINGS_3RD_PARTY "llama: enable all compiler warnings in 3rd party libs" OFF) @@ -77,7 +74,6 @@ option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF) include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake) # override ggml options -set(GGML_CCACHE ${LLAMA_CCACHE}) set(GGML_SANITIZE_THREAD ${LLAMA_SANITIZE_THREAD}) set(GGML_SANITIZE_ADDRESS ${LLAMA_SANITIZE_ADDRESS}) set(GGML_SANITIZE_UNDEFINED ${LLAMA_SANITIZE_UNDEFINED}) @@ -110,12 +106,16 @@ llama_option_depr(WARNING LLAMA_NATIVE GGML_NATIVE) llama_option_depr(WARNING LLAMA_RPC GGML_RPC) llama_option_depr(WARNING LLAMA_SYCL GGML_SYCL) llama_option_depr(WARNING LLAMA_SYCL_F16 GGML_SYCL_F16) +llama_option_depr(WARNING LLAMA_CANN GGML_CANN) # # build the library # -add_subdirectory(ggml) +if (NOT TARGET ggml) + add_subdirectory(ggml) + # ... otherwise assume ggml is added by a parent CMakeLists.txt +endif() add_subdirectory(src) # @@ -133,7 +133,16 @@ set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location o set(LLAMA_LIB_INSTALL_DIR ${CMAKE_INSTALL_LIBDIR} CACHE PATH "Location of library files") set(LLAMA_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR} CACHE PATH "Location of binary files") -get_directory_property(LLAMA_TRANSIENT_DEFINES COMPILE_DEFINITIONS) + +# At the moment some compile definitions are placed within the ggml/src +# directory but not exported on the `ggml` target. This could be improved by +# determining _precisely_ which defines are necessary for the llama-config +# package. +# +get_directory_property(GGML_DIR_DEFINES DIRECTORY ggml/src COMPILE_DEFINITIONS) +get_target_property(GGML_TARGET_DEFINES ggml COMPILE_DEFINITIONS) +set(GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES} ${GGML_DIR_DEFINES}) +get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES) set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h) install(TARGETS llama LIBRARY PUBLIC_HEADER) diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index 9ad5160495a8e..b688f78ecf195 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -1,13 +1,17 @@ -# Pull requests +# Pull requests (for contributors) -- Always squash-merge the PR before merging -- Use the following format for your final commit: ` : (#)`. For example: `utils : fix typo in utils.py (#1234)` - Test your changes: - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library - Execute [the full CI locally on your machine](ci/README.md) before publishing -- If the pull request contains only documentation changes (e.g., updating READMEs, adding new wiki pages), please add `[no ci]` to the commit title. This will skip unnecessary CI checks and help reduce build times - Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs. - - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your conveience + - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience +- If your PR becomes stale, don't hesitate to ping the maintainers in the comments + +# Pull requests (for collaborators) + +- Squash-merge PRs +- Use the following format for the squashed commit title: ` : (#)`. For example: `utils : fix typo in utils.py (#1234)` +- Optionally, pick a `` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules # Coding guidelines diff --git a/Makefile b/Makefile index bb6e2f968cf0f..8d2ccddc469f9 100644 --- a/Makefile +++ b/Makefile @@ -11,7 +11,6 @@ BUILD_TARGETS = \ llama-embedding \ llama-eval-callback \ llama-export-lora \ - llama-finetune \ llama-gbnf-validator \ llama-gguf \ llama-gguf-hash \ @@ -37,7 +36,6 @@ BUILD_TARGETS = \ llama-simple \ llama-speculative \ llama-tokenize \ - llama-train-text-from-scratch \ llama-vdot \ llama-cvector-generator \ tests/test-c.o @@ -64,9 +62,13 @@ TEST_TARGETS = \ tests/test-tokenizer-1-spm # Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned -LEGACY_TARGETS = main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \ +LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \ simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \ - retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm + retrieval speculative infill tokenize benchmark-matmult parallel export-lora lookahead lookup passkey gritlm + +# Legacy build targets that were renamed in #7809, but we want to build binaries that for them that output a deprecation warning if people try to use them. +# We don't want to clutter things too much, so we only build replacements for the most commonly used binaries. +LEGACY_TARGETS_BUILD = main quantize perplexity embedding server # Deprecation aliases ifdef LLAMA_CUBLAS @@ -193,7 +195,11 @@ ifdef GGML_RPC BUILD_TARGETS += rpc-server endif -default: $(BUILD_TARGETS) +ifdef GGML_VULKAN + BUILD_TARGETS += vulkan-shaders-gen +endif + +default: $(BUILD_TARGETS) $(LEGACY_TARGETS_BUILD) test: $(TEST_TARGETS) @failures=0; \ @@ -228,7 +234,7 @@ test: $(TEST_TARGETS) fi @echo 'All tests passed.' -all: $(BUILD_TARGETS) $(TEST_TARGETS) +all: $(BUILD_TARGETS) $(TEST_TARGETS) $(LEGACY_TARGETS_BUILD) ifdef RISCV_CROSS_COMPILE CC := riscv64-unknown-linux-gnu-gcc @@ -245,17 +251,22 @@ MK_CFLAGS = -std=c11 -fPIC MK_CXXFLAGS = -std=c++11 -fPIC MK_NVCCFLAGS = -std=c++11 -ifndef LLAMA_NO_CCACHE +ifdef LLAMA_NO_CCACHE +GGML_NO_CCACHE := 1 +DEPRECATE_WARNING := 1 +endif + +ifndef GGML_NO_CCACHE CCACHE := $(shell which ccache) ifdef CCACHE export CCACHE_SLOPPINESS = time_macros -$(info I ccache found, compilation results will be cached. Disable with LLAMA_NO_CCACHE.) +$(info I ccache found, compilation results will be cached. Disable with GGML_NO_CCACHE.) CC := $(CCACHE) $(CC) CXX := $(CCACHE) $(CXX) else $(info I ccache not found. Consider installing it for faster compilation.) endif # CCACHE -endif # LLAMA_NO_CCACHE +endif # GGML_NO_CCACHE # clock_gettime came in POSIX.1b (1993) # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional @@ -538,14 +549,20 @@ ifdef GGML_OPENBLAS64 endif # GGML_OPENBLAS64 ifdef GGML_BLIS - MK_CPPFLAGS += -DGGML_USE_BLAS -I/usr/local/include/blis -I/usr/include/blis + MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_BLIS -I/usr/local/include/blis -I/usr/include/blis MK_LDFLAGS += -lblis -L/usr/local/lib OBJ_GGML += ggml/src/ggml-blas.o endif # GGML_BLIS +ifdef GGML_NVPL + MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_NVPL -DNVPL_ILP64 -I/usr/local/include/nvpl_blas -I/usr/include/nvpl_blas + MK_LDFLAGS += -L/usr/local/lib -lnvpl_blas_core -lnvpl_blas_ilp64_gomp + OBJ_GGML += ggml/src/ggml-blas.o +endif # GGML_NVPL + ifndef GGML_NO_LLAMAFILE MK_CPPFLAGS += -DGGML_USE_LLAMAFILE - OBJ_GGML += ggml/src/sgemm.o + OBJ_GGML += ggml/src/llamafile/sgemm.o endif ifdef GGML_RPC @@ -695,8 +712,8 @@ endif # GGML_CUDA ifdef GGML_VULKAN MK_CPPFLAGS += -DGGML_USE_VULKAN - MK_LDFLAGS += -lvulkan - OBJ_GGML += ggml/src/ggml-vulkan.o + MK_LDFLAGS += $(shell pkg-config --libs vulkan) + OBJ_GGML += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o ifdef GGML_VULKAN_CHECK_RESULTS MK_CPPFLAGS += -DGGML_VULKAN_CHECK_RESULTS @@ -718,10 +735,28 @@ ifdef GGML_VULKAN_RUN_TESTS MK_CPPFLAGS += -DGGML_VULKAN_RUN_TESTS endif -ggml/src/ggml-vulkan.o: \ - ggml/src/ggml-vulkan.cpp \ - ggml/include/ggml-vulkan.h - $(CXX) $(CXXFLAGS) -c $< -o $@ +GLSLC_CMD = glslc +_ggml_vk_genshaders_cmd = $(shell pwd)/vulkan-shaders-gen +_ggml_vk_header = ggml/src/ggml-vulkan-shaders.hpp +_ggml_vk_source = ggml/src/ggml-vulkan-shaders.cpp +_ggml_vk_input_dir = ggml/src/vulkan-shaders +_ggml_vk_shader_deps = $(echo $(_ggml_vk_input_dir)/*.comp) + +ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source) + $(CXX) $(CXXFLAGS) $(shell pkg-config --cflags vulkan) -c $< -o $@ + +$(_ggml_vk_header): $(_ggml_vk_source) + +$(_ggml_vk_source): $(_ggml_vk_shader_deps) vulkan-shaders-gen + $(_ggml_vk_genshaders_cmd) \ + --glslc $(GLSLC_CMD) \ + --input-dir $(_ggml_vk_input_dir) \ + --target-hpp $(_ggml_vk_header) \ + --target-cpp $(_ggml_vk_source) + +vulkan-shaders-gen: ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp + $(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp + endif # GGML_VULKAN ifdef GGML_HIPBLAS @@ -758,6 +793,14 @@ ifdef GGML_CUDA_FORCE_DMMV HIPFLAGS += -DGGML_CUDA_FORCE_DMMV endif # GGML_CUDA_FORCE_DMMV +ifdef GGML_CUDA_FORCE_MMQ + HIPFLAGS += -DGGML_CUDA_FORCE_MMQ +endif # GGML_CUDA_FORCE_MMQ + +ifdef GGML_CUDA_FORCE_CUBLAS + HIPFLAGS += -DGGML_CUDA_FORCE_CUBLAS +endif # GGML_CUDA_FORCE_CUBLAS + ifdef GGML_CUDA_NO_PEER_COPY HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY endif # GGML_CUDA_NO_PEER_COPY @@ -826,10 +869,14 @@ OBJ_GGML += \ ggml/src/ggml.o \ ggml/src/ggml-alloc.o \ ggml/src/ggml-backend.o \ - ggml/src/ggml-quants.o + ggml/src/ggml-quants.o \ + ggml/src/ggml-aarch64.o OBJ_LLAMA = \ src/llama.o \ + src/llama-vocab.o \ + src/llama-grammar.o \ + src/llama-sampling.o \ src/unicode.o \ src/unicode-data.o @@ -926,6 +973,7 @@ $(info - LLAMA_NO_LLAMAFILE) $(info - LLAMA_NO_ACCELERATE) $(info - LLAMA_NO_OPENMP) $(info - LLAMA_NO_METAL) +$(info - LLAMA_NO_CCACHE) $(info ) endif @@ -959,15 +1007,22 @@ ggml/src/ggml-quants.o: \ ggml/src/ggml-common.h $(CC) $(CFLAGS) -c $< -o $@ +ggml/src/ggml-aarch64.o: \ + ggml/src/ggml-aarch64.c \ + ggml/include/ggml.h \ + ggml/src/ggml-aarch64.h \ + ggml/src/ggml-common.h + $(CC) $(CFLAGS) -c $< -o $@ + ggml/src/ggml-blas.o: \ ggml/src/ggml-blas.cpp \ ggml/include/ggml-blas.h $(CXX) $(CXXFLAGS) -c $< -o $@ ifndef GGML_NO_LLAMAFILE -ggml/src/sgemm.o: \ - ggml/src/sgemm.cpp \ - ggml/src/sgemm.h \ +ggml/src/llamafile/sgemm.o: \ + ggml/src/llamafile/sgemm.cpp \ + ggml/src/llamafile/sgemm.h \ ggml/include/ggml.h $(CXX) $(CXXFLAGS) -c $< -o $@ endif # GGML_NO_LLAMAFILE @@ -1001,6 +1056,10 @@ src/unicode-data.o: \ src/llama.o: \ src/llama.cpp \ + src/llama-impl.h \ + src/llama-vocab.h \ + src/llama-grammar.h \ + src/llama-sampling.h \ src/unicode.h \ include/llama.h \ ggml/include/ggml-cuda.h \ @@ -1010,6 +1069,29 @@ src/llama.o: \ ggml/include/ggml-backend.h $(CXX) $(CXXFLAGS) -c $< -o $@ +src/llama-vocab.o: \ + src/llama-vocab.cpp \ + src/llama-vocab.h \ + src/llama-impl.h \ + include/llama.h + $(CXX) $(CXXFLAGS) -c $< -o $@ + +src/llama-grammar.o: \ + src/llama-grammar.cpp \ + src/llama-grammar.h \ + src/llama-impl.h \ + src/llama-vocab.h \ + src/llama-sampling.h \ + include/llama.h + $(CXX) $(CXXFLAGS) -c $< -o $@ + +src/llama-sampling.o: \ + src/llama-sampling.cpp \ + src/llama-sampling.h \ + src/llama-impl.h \ + include/llama.h + $(CXX) $(CXXFLAGS) -c $< -o $@ + $(LIB_LLAMA): \ $(OBJ_LLAMA) \ $(LIB_GGML) @@ -1092,7 +1174,8 @@ clean: rm -vrf ggml/src/ggml-cuda/template-instances/*.o rm -rvf $(BUILD_TARGETS) rm -rvf $(TEST_TARGETS) - rm -rvf $(LEGACY_TARGETS) + rm -f vulkan-shaders-gen ggml/src/ggml-vulkan-shaders.hpp ggml/src/ggml-vulkan-shaders.cpp + rm -rvf $(LEGACY_TARGETS_CLEAN) find examples pocs -type f -name "*.o" -delete # @@ -1211,11 +1294,6 @@ llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp \ $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp \ - $(OBJ_ALL) - $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) - llama-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp \ $(OBJ_GGML) $(OBJ_LLAMA) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) @@ -1231,13 +1309,8 @@ llama-baby-llama: examples/baby-llama/baby-llama.cpp \ $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-finetune: examples/finetune/finetune.cpp \ - $(OBJ_ALL) - $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) - llama-export-lora: examples/export-lora/export-lora.cpp \ - $(OBJ_GGML) common/log.h + $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -1384,7 +1457,7 @@ run-benchmark-matmult: llama-benchmark-matmult .PHONY: run-benchmark-matmult swift tests/test-llama-grammar: tests/test-llama-grammar.cpp \ - $(OBJ_GGML) $(OBJ_COMMON) src/unicode.o src/unicode-data.o + $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -1488,3 +1561,51 @@ llama-q8dot: pocs/vdot/q8dot.cpp ggml/src/ggml.o \ $(OBJ_GGML) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + +# +# Deprecated binaries that we want to keep around long enough for people to migrate to the new filenames, then these can be removed. +# +# Mark legacy binary targets as .PHONY so that they are always checked. +.PHONY: main quantize perplexity embedding server + +# NOTE: We currently will always build the deprecation-warning `main` and `server` binaries to help users migrate. +# Eventually we will want to remove these target from building all the time. +main: examples/deprecation-warning/deprecation-warning.cpp + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + @echo "NOTICE: The 'main' binary is deprecated. Please use 'llama-cli' instead." + +server: examples/deprecation-warning/deprecation-warning.cpp + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + @echo "NOTICE: The 'server' binary is deprecated. Please use 'llama-server' instead." + +quantize: examples/deprecation-warning/deprecation-warning.cpp +ifneq (,$(wildcard quantize)) + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + @echo "#########" + @echo "WARNING: The 'quantize' binary is deprecated. Please use 'llama-quantize' instead." + @echo " Remove the 'quantize' binary to remove this warning." + @echo "#########" +endif + +perplexity: examples/deprecation-warning/deprecation-warning.cpp +ifneq (,$(wildcard perplexity)) + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + @echo "#########" + @echo "WARNING: The 'perplexity' binary is deprecated. Please use 'llama-perplexity' instead." + @echo " Remove the 'perplexity' binary to remove this warning." + @echo "#########" +endif + +embedding: examples/deprecation-warning/deprecation-warning.cpp +ifneq (,$(wildcard embedding)) + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + @echo "#########" + @echo "WARNING: The 'embedding' binary is deprecated. Please use 'llama-embedding' instead." + @echo " Remove the 'embedding' binary to remove this warning." + @echo "#########" +endif diff --git a/Package.swift b/Package.swift index 77fed86df3105..1d90b47bfa3da 100644 --- a/Package.swift +++ b/Package.swift @@ -4,12 +4,16 @@ import PackageDescription var sources = [ "src/llama.cpp", + "src/llama-vocab.cpp", + "src/llama-grammar.cpp", + "src/llama-sampling.cpp", "src/unicode.cpp", "src/unicode-data.cpp", "ggml/src/ggml.c", "ggml/src/ggml-alloc.c", "ggml/src/ggml-backend.c", "ggml/src/ggml-quants.c", + "ggml/src/ggml-aarch64.c", ] var resources: [Resource] = [] diff --git a/README.md b/README.md index 5498d5fc6755a..d0ae2efb9123e 100644 --- a/README.md +++ b/README.md @@ -3,7 +3,7 @@ ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png) [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) -[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml) +[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml) [![Conan Center](https://shields.io/conan/v/llama-cpp)](https://conan.io/center/llama-cpp) [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml) @@ -98,7 +98,7 @@ Typically finetunes of the base models below are supported as well. - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia) - [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b) -(instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md)) +(instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md)) **Multimodal models:** @@ -138,6 +138,7 @@ Typically finetunes of the base models below are supported as well. Unless otherwise noted these projects are open-source with permissive licensing: +- [MindWorkAI/AI-Studio](https://github.com/MindWorkAI/AI-Studio) (FSL-1.1-MIT) - [iohub/collama](https://github.com/iohub/coLLaMA) - [janhq/jan](https://github.com/janhq/jan) (AGPL) - [nat/openplayground](https://github.com/nat/openplayground) @@ -181,6 +182,9 @@ Unless otherwise noted these projects are open-source with permissive licensing: - [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp +**Games:** +- [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you. + ## Demo
@@ -453,7 +457,7 @@ To learn more how to measure perplexity using llama.cpp, [read this documentatio - [How to build](./docs/build.md) - [Running on Docker](./docs/docker.md) - [Build on Android](./docs/android.md) -- [Performance troubleshooting](./docs/token_generation_performance_tips.md) +- [Performance troubleshooting](./docs/development/token_generation_performance_tips.md) - [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks) **Seminal papers and background on the models** diff --git a/cmake/llama-config.cmake.in b/cmake/llama-config.cmake.in index 2e7da2f8ebd69..f072b76a39d2e 100644 --- a/cmake/llama-config.cmake.in +++ b/cmake/llama-config.cmake.in @@ -8,6 +8,13 @@ set(GGML_CUDA @GGML_CUDA@) set(GGML_METAL @GGML_METAL@) set(GGML_HIPBLAS @GGML_HIPBLAS@) set(GGML_ACCELERATE @GGML_ACCELERATE@) +set(GGML_VULKAN @GGML_VULKAN@) +set(GGML_VULKAN_CHECK_RESULTS @GGML_VULKAN_CHECK_RESULTS@) +set(GGML_VULKAN_DEBUG @GGML_VULKAN_DEBUG@) +set(GGML_VULKAN_MEMORY_DEBUG @GGML_VULKAN_MEMORY_DEBUG@) +set(GGML_VULKAN_VALIDATE @GGML_VULKAN_VALIDATE@) +set(GGML_SYCL @GGML_SYCL@) +set(GGML_OPENMP @GGML_OPENMP@) @PACKAGE_INIT@ @@ -37,18 +44,36 @@ if (GGML_METAL) find_library(METALKIT_FRAMEWORK MetalKit REQUIRED) endif() +if (GGML_VULKAN) + find_package(Vulkan REQUIRED) +endif() + if (GGML_HIPBLAS) find_package(hip REQUIRED) find_package(hipblas REQUIRED) find_package(rocblas REQUIRED) endif() +if (GGML_SYCL) + find_package(IntelSYCL REQUIRED) + find_package(MKL REQUIRED) +endif() + +if (GGML_OPENMP) + find_package(OpenMP REQUIRED) +endif() + + +find_library(ggml_LIBRARY ggml + REQUIRED + HINTS ${LLAMA_LIB_DIR}) + find_library(llama_LIBRARY llama REQUIRED HINTS ${LLAMA_LIB_DIR}) -set(_llama_link_deps "Threads::Threads" "@LLAMA_EXTRA_LIBS@") -set(_llama_transient_defines "@LLAMA_TRANSIENT_DEFINES@") +set(_llama_link_deps "${ggml_LIBRARY}" "@GGML_LINK_LIBRARIES@") +set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@") add_library(llama UNKNOWN IMPORTED) diff --git a/common/common.cpp b/common/common.cpp index fc0f3b350ec6e..ec44a05521c9d 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1,3 +1,7 @@ +#if defined(_MSC_VER) +#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING +#endif + #include "common.h" // Change JSON_ASSERT from assert() to GGML_ASSERT: #define JSON_ASSERT GGML_ASSERT @@ -681,7 +685,6 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa if (arg == "--lora") { CHECK_ARG params.lora_adapter.emplace_back(argv[i], 1.0f); - params.use_mmap = false; return true; } if (arg == "--lora-scaled") { @@ -689,12 +692,6 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa const char* lora_adapter = argv[i]; CHECK_ARG params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i])); - params.use_mmap = false; - return true; - } - if (arg == "--lora-base") { - CHECK_ARG - params.lora_base = argv[i]; return true; } if (arg == "--control-vector") { @@ -793,6 +790,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa params.cont_batching = true; return true; } + if (arg == "-nocb" || arg == "--no-cont-batching") { + params.cont_batching = false; + return true; + } if (arg == "-fa" || arg == "--flash-attn") { params.flash_attn = true; return true; @@ -1268,6 +1269,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa CHECK_ARG params.out_file = argv[i]; params.cvector_outfile = argv[i]; + params.lora_outfile = argv[i]; return true; } if (arg == "-ofreq" || arg == "--output-frequency") { @@ -1534,6 +1536,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "*", "-np, --parallel N", "number of parallel sequences to decode (default: %d)", params.n_parallel }); options.push_back({ "*", "-ns, --sequences N", "number of sequences to decode (default: %d)", params.n_sequences }); options.push_back({ "*", "-cb, --cont-batching", "enable continuous batching (a.k.a dynamic batching) (default: %s)", params.cont_batching ? "enabled" : "disabled" }); + options.push_back({ "*", "-nocb, --no-cont-batching", "disable continuous batching" }); options.push_back({ "multi-modality" }); options.push_back({ "*", " --mmproj FILE", "path to a multimodal projector file for LLaVA. see examples/llava/README.md" }); @@ -1576,9 +1579,8 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "*", " --override-kv KEY=TYPE:VALUE", "advanced option to override model metadata by key. may be specified multiple times.\n" "types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false" }); - options.push_back({ "*", " --lora FNAME", "apply LoRA adapter (implies --no-mmap)" }); - options.push_back({ "*", " --lora-scaled FNAME S", "apply LoRA adapter with user defined scaling S (implies --no-mmap)" }); - options.push_back({ "*", " --lora-base FNAME", "optional model to use as a base for the layers modified by the LoRA adapter" }); + options.push_back({ "*", " --lora FNAME", "apply LoRA adapter (can be repeated to use multiple adapters)" }); + options.push_back({ "*", " --lora-scaled FNAME S", "apply LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)" }); options.push_back({ "*", " --control-vector FNAME", "add a control vector\n" "note: this argument can be repeated to add multiple control vectors" }); options.push_back({ "*", " --control-vector-scaled FNAME SCALE", @@ -1669,6 +1671,13 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "cvector", " --pca-iter N", "number of iterations used for PCA (default: %d)", params.n_pca_iterations }); options.push_back({ "cvector", " --method {pca,mean}", "dimensionality reduction method to be used (default: pca)" }); + options.push_back({ "export-lora" }); + options.push_back({ "export-lora", "-m, --model", "model path from which to load base model (default '%s')", params.model.c_str() }); + options.push_back({ "export-lora", " --lora FNAME", "path to LoRA adapter (can be repeated to use multiple adapters)" }); + options.push_back({ "export-lora", " --lora-scaled FNAME S", "path to LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)" }); + options.push_back({ "*", "-t, --threads N", "number of threads to use during computation (default: %d)", params.n_threads }); + options.push_back({ "export-lora", "-o, --output FNAME", "output file (default: '%s')", params.lora_outfile.c_str() }); + printf("usage: %s [options]\n", argv[0]); for (const auto & o : options) { @@ -2080,19 +2089,14 @@ std::tuple llama_init_from_gpt_par for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) { const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]); float lora_scale = std::get<1>(params.lora_adapter[i]); - int err = llama_model_apply_lora_from_file(model, - lora_adapter.c_str(), - lora_scale, - ((i > 0) || params.lora_base.empty()) - ? NULL - : params.lora_base.c_str(), - params.n_threads); - if (err != 0) { + auto adapter = llama_lora_adapter_init(model, lora_adapter.c_str()); + if (adapter == nullptr) { fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__); llama_free(lctx); llama_free_model(model); return std::make_tuple(nullptr, nullptr); } + llama_lora_adapter_set(lctx, adapter, lora_scale); } if (params.ignore_eos) { @@ -2719,7 +2723,7 @@ std::string llama_chat_format_single(const struct llama_model * model, const llama_chat_msg & new_msg, bool add_ass) { std::ostringstream ss; - auto fmt_past_msg = llama_chat_apply_template(model, tmpl, past_msg, false); + auto fmt_past_msg = past_msg.empty() ? "" : llama_chat_apply_template(model, tmpl, past_msg, false); std::vector chat_new(past_msg); // if the past_msg ends with a newline, we must preserve it in the formatted version if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') { @@ -3164,7 +3168,6 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l } fprintf(stream, " - %s: %f\n", std::get<0>(la).c_str(), std::get<1>(la)); } - fprintf(stream, "lora_base: %s\n", params.lora_base.c_str()); fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu); fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep); fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat); diff --git a/common/common.h b/common/common.h index 184a53dc09064..8240ff99b8e2a 100644 --- a/common/common.h +++ b/common/common.h @@ -128,7 +128,6 @@ struct gpt_params { // TODO: avoid tuple, use struct std::vector> lora_adapter; // lora adapter path with user defined scale - std::string lora_base = ""; // base model path for the lora adapter std::vector control_vectors; // control vector with user defined scale @@ -255,6 +254,8 @@ struct gpt_params { std::string cvector_negative_file = "examples/cvector-generator/negative.txt"; bool spm_infill = false; // suffix/prefix/middle pattern for infill + + std::string lora_outfile = "ggml-lora-merged-f16.gguf"; }; void gpt_params_handle_hf_token(gpt_params & params); diff --git a/common/ngram-cache.h b/common/ngram-cache.h index e4fa4cbd12f11..ab4c9b3766546 100644 --- a/common/ngram-cache.h +++ b/common/ngram-cache.h @@ -37,11 +37,18 @@ struct llama_ngram { } }; +struct llama_token_hash_function { + size_t operator()(const llama_token token) const { + // see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/ + return token * 11400714819323198485llu; + } +}; + struct llama_ngram_hash_function { size_t operator()(const llama_ngram & ngram) const { - size_t hash = 0; - for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) { - hash ^= std::hash{}(ngram.tokens[i]); + size_t hash = llama_token_hash_function{}(ngram.tokens[0]); + for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) { + hash ^= llama_token_hash_function{}(ngram.tokens[i]); } return hash; } diff --git a/common/sampling.cpp b/common/sampling.cpp index e9db2715f2649..079e405168dff 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -330,7 +330,7 @@ static llama_token llama_sampling_sample_impl( llama_token_data_array single_token_data_array = { &single_token_data, 1, false }; // Apply grammar constraints to the single token - llama_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar); + llama_grammar_sample(ctx_sampling->grammar, ctx_main, &single_token_data_array); // Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY bool is_valid = single_token_data_array.data[0].logit != -INFINITY; @@ -378,7 +378,7 @@ static llama_token_data_array llama_sampling_prepare_impl( if (ctx_sampling->grammar != NULL && !apply_grammar) { GGML_ASSERT(original_logits != NULL); // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this. - *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))}; + *original_logits = {logits, logits + n_vocab}; } // apply params.logit_bias map @@ -391,10 +391,10 @@ static llama_token_data_array llama_sampling_prepare_impl( llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale); } - cur.clear(); + cur.resize(n_vocab); for (llama_token token_id = 0; token_id < n_vocab; token_id++) { - cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f}); + cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f}; } llama_token_data_array cur_p = { cur.data(), cur.size(), false }; @@ -421,7 +421,7 @@ static llama_token_data_array llama_sampling_prepare_impl( // apply grammar checks before sampling logic if (apply_grammar && ctx_sampling->grammar != NULL) { - llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar); + llama_grammar_sample(ctx_sampling->grammar, ctx_main, &cur_p); } return cur_p; @@ -455,6 +455,6 @@ void llama_sampling_accept( ctx_sampling->prev.push_back(id); if (ctx_sampling->grammar != NULL && apply_grammar) { - llama_grammar_accept_token(ctx_main, ctx_sampling->grammar, id); + llama_grammar_accept_token(ctx_sampling->grammar, ctx_main, id); } } diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 40843e08ead3a..710779e814bdd 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -48,34 +48,39 @@ class Model: dir_model: Path ftype: gguf.LlamaFileType + fname_out: Path is_big_endian: bool endianess: gguf.GGUFEndian use_temp_file: bool lazy: bool - model_name: str | None part_names: list[str] is_safetensors: bool hparams: dict[str, Any] block_count: int tensor_map: gguf.TensorNameMap tensor_names: set[str] | None - fname_out: Path gguf_writer: gguf.GGUFWriter + model_name: str | None + metadata_override: Path | None + dir_model_card: Path # subclasses should define this! model_arch: gguf.MODEL_ARCH - def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool, - model_name: str | None, split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False): + def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False, + use_temp_file: bool = False, eager: bool = False, + metadata_override: Path | None = None, model_name: str | None = None, + split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False): if type(self) is Model: raise TypeError(f"{type(self).__name__!r} should not be directly instantiated") + self.dir_model = dir_model self.ftype = ftype + self.fname_out = fname_out self.is_big_endian = is_big_endian self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE self.use_temp_file = use_temp_file self.lazy = not eager - self.model_name = model_name self.part_names = Model.get_model_part_names(self.dir_model, "model", ".safetensors") self.is_safetensors = len(self.part_names) > 0 if not self.is_safetensors: @@ -84,6 +89,11 @@ def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"]) self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) self.tensor_names = None + self.metadata_override = metadata_override + self.model_name = model_name + self.dir_model_card = dir_model # overridden in convert_lora_to_gguf.py + + # Apply heuristics to figure out typical tensor encoding based on first layer tensor encoding type if self.ftype == gguf.LlamaFileType.GUESSED: # NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie. _, first_tensor = next(self.get_tensors()) @@ -93,10 +103,8 @@ def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, else: logger.info(f"choosing --outtype bf16 from first tensor type ({first_tensor.dtype})") self.ftype = gguf.LlamaFileType.MOSTLY_BF16 - ftype_up: str = self.ftype.name.partition("_")[2].upper() - ftype_lw: str = ftype_up.lower() - # allow templating the file name with the output ftype, useful with the "auto" ftype - self.fname_out = fname_out.parent / fname_out.name.format(ftype_lw, outtype=ftype_lw, ftype=ftype_lw, OUTTYPE=ftype_up, FTYPE=ftype_up) + + # Configure GGUF Writer self.gguf_writer = gguf.GGUFWriter(path=None, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file, split_max_tensors=split_max_tensors, split_max_size=split_max_size, dry_run=dry_run, small_first_shard=small_first_shard) @@ -148,9 +156,16 @@ def get_tensors(self) -> Iterator[tuple[str, Tensor]]: tensor_names_from_parts.update(model_part.keys()) for name in model_part.keys(): - data = model_part.get_tensor(name) if self.is_safetensors else model_part[name] - if self.lazy: - data = LazyTorchTensor.from_eager(data) + if self.is_safetensors: + if self.lazy: + data = model_part.get_slice(name) + data = LazyTorchTensor.from_safetensors_slice(data) + else: + data = model_part.get_tensor(name) + else: + data = model_part[name] + if self.lazy: + data = LazyTorchTensor.from_eager(data) yield name, data # only verify tensor name presence; it doesn't matter if they are not in the right files @@ -186,7 +201,6 @@ def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", " return new_name def set_gguf_parameters(self): - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_block_count(self.block_count) if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None: @@ -225,6 +239,10 @@ def set_gguf_parameters(self): self.gguf_writer.add_expert_used_count(n_experts_used) logger.info(f"gguf: experts used count = {n_experts_used}") + if (head_dim := self.hparams.get("head_dim")) is not None: + self.gguf_writer.add_key_length(head_dim) + self.gguf_writer.add_value_length(head_dim) + self.gguf_writer.add_file_type(self.ftype) logger.info(f"gguf: file type = {self.ftype}") @@ -243,7 +261,7 @@ def extra_f16_tensors(self, name: str, new_name: str, bid: int | None, n_dims: i return False - def write_tensors(self): + def prepare_tensors(self): max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,") for name, data_torch in self.get_tensors(): @@ -326,9 +344,62 @@ def write_tensors(self): self.gguf_writer.add_tensor(new_name, data, raw_dtype=data_qtype) + def set_type(self): + self.gguf_writer.add_type(gguf.GGUFType.MODEL) + + def prepare_metadata(self, vocab_only: bool): + + total_params, shared_params, expert_params, expert_count = self.gguf_writer.get_total_parameter_count() + + self.metadata = gguf.Metadata.load(self.metadata_override, self.dir_model_card, self.model_name, total_params) + + # Fallback to model directory name if metadata name is still missing + if self.metadata.name is None: + self.metadata.name = self.dir_model.name + + # Generate parameter weight class (useful for leader boards) if not yet determined + if self.metadata.size_label is None and total_params > 0: + self.metadata.size_label = gguf.size_label(total_params, shared_params, expert_params, expert_count) + + # Extract the encoding scheme from the file type name. e.g. 'gguf.LlamaFileType.MOSTLY_Q8_0' --> 'Q8_0' + output_type: str = self.ftype.name.partition("_")[2] + + # Filename Output + if self.fname_out.is_dir(): + # Generate default filename based on model specification and available metadata + if not vocab_only: + fname_default: str = gguf.naming_convention(self.metadata.name, self.metadata.basename, self.metadata.finetune, self.metadata.version, self.metadata.size_label, output_type, model_type="LoRA" if total_params < 0 else None) + else: + fname_default: str = gguf.naming_convention(self.metadata.name, self.metadata.basename, self.metadata.finetune, self.metadata.version, size_label=None, output_type=None, model_type="vocab") + + # Use the default filename + self.fname_out = self.fname_out / f"{fname_default}.gguf" + else: + # Output path is a custom defined templated filename + # Note: `not is_dir()` is used because `.is_file()` will not detect + # file template strings as it doesn't actually exist as a file + + # Process templated file name with the output ftype, useful with the "auto" ftype + self.fname_out = self.fname_out.parent / gguf.fill_templated_filename(self.fname_out.name, output_type) + + self.set_type() + + logger.info("Set meta model") + self.metadata.set_gguf_meta_model(self.gguf_writer) + + logger.info("Set model parameters") + self.set_gguf_parameters() + + logger.info("Set model tokenizer") + self.set_vocab() + + logger.info("Set model quantization version") + self.gguf_writer.add_quantization_version(gguf.GGML_QUANT_VERSION) + def write(self): - self.write_tensors() - self.gguf_writer.write_header_to_file(self.fname_out) + self.prepare_tensors() + self.prepare_metadata(vocab_only=False) + self.gguf_writer.write_header_to_file(path=self.fname_out) self.gguf_writer.write_kv_data_to_file() self.gguf_writer.write_tensors_to_file(progress=True) self.gguf_writer.close() @@ -336,7 +407,9 @@ def write(self): def write_vocab(self): if len(self.gguf_writer.tensors) != 1: raise ValueError('Splitting the vocabulary is not supported') - self.gguf_writer.write_header_to_file(self.fname_out) + + self.prepare_metadata(vocab_only=True) + self.gguf_writer.write_header_to_file(path=self.fname_out) self.gguf_writer.write_kv_data_to_file() self.gguf_writer.close() @@ -373,6 +446,29 @@ def from_model_architecture(cls, arch: str) -> type[Model]: except KeyError: raise NotImplementedError(f'Architecture {arch!r} not supported!') from None + def does_token_look_special(self, token: str | bytes) -> bool: + if isinstance(token, (bytes, bytearray)): + token_text = token.decode(encoding="utf-8") + elif isinstance(token, memoryview): + token_text = token.tobytes().decode(encoding="utf-8") + else: + token_text = token + + # Some models mark some added tokens which ought to be control tokens as not special. + # (e.g. command-r, command-r-plus, deepseek-coder, gemma{,-2}) + seems_special = token_text in ( + "", # deepseek-coder + "", "<2mass>", "[@BOS@]", # gemma{,-2} + ) + + seems_special = seems_special or (token_text.startswith("<|") and token_text.endswith("|>")) + seems_special = seems_special or (token_text.startswith("<|") and token_text.endswith("|>")) # deepseek-coder + + # TODO: should these be marked as UNUSED instead? (maybe not) + seems_special = seems_special or (token_text.startswith("")) # gemma{,-2} + + return seems_special + # used for GPT-2 BPE and WordPiece vocabs def get_vocab_base(self) -> tuple[list[str], list[int], str]: tokens: list[str] = [] @@ -391,16 +487,18 @@ def get_vocab_base(self) -> tuple[list[str], list[int], str]: for i in range(vocab_size): if i not in reverse_vocab: tokens.append(f"[PAD{i}]") - toktypes.append(gguf.TokenType.USER_DEFINED) - elif reverse_vocab[i] in added_vocab: - tokens.append(reverse_vocab[i]) - if tokenizer.added_tokens_decoder[i].special: - toktypes.append(gguf.TokenType.CONTROL) - else: - toktypes.append(gguf.TokenType.USER_DEFINED) + toktypes.append(gguf.TokenType.UNUSED) else: - tokens.append(reverse_vocab[i]) - toktypes.append(gguf.TokenType.NORMAL) + token: str = reverse_vocab[i] + if token in added_vocab: + if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token): + toktypes.append(gguf.TokenType.CONTROL) + else: + token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ") # pre-normalize user-defined spaces + toktypes.append(gguf.TokenType.USER_DEFINED) + else: + toktypes.append(gguf.TokenType.NORMAL) + tokens.append(token) return tokens, toktypes, tokpre @@ -487,6 +585,8 @@ def get_vocab_base_pre(self, tokenizer) -> str: if chkhsh == "7967bfa498ade6b757b064f31e964dddbb80f8f9a4d68d4ba7998fcf281c531a": # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-code res = "jina-v2-code" + if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b": + # ref: https://huggingface.co/THUDM/glm-4-9b-chat res = "chatglm-bpe" if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee": # ref: https://huggingface.co/LumiOpen/Viking-7B @@ -494,10 +594,19 @@ def get_vocab_base_pre(self, tokenizer) -> str: if chkhsh == "b53802fb28e26d645c3a310b34bfe07da813026ec7c7716883404d5e0f8b1901": # ref: https://huggingface.co/core42/jais-13b res = "jais" + if chkhsh == "7b3e7548e4308f52a76e8229e4e6cc831195d0d1df43aed21ac6c93da05fec5f": + # ref: https://huggingface.co/WisdomShell/CodeShell-7B + res = "codeshell" + if chkhsh == "63b97e4253352e6f357cc59ea5b583e3a680eaeaf2632188c2b952de2588485e": + # ref: https://huggingface.co/mistralai/Mistral-Nemo-Base-2407 + res = "tekken" + if chkhsh == "855059429035d75a914d1eda9f10a876752e281a054a7a3d421ef0533e5b6249": + # ref: https://huggingface.co/HuggingFaceTB/SmolLM-135M + res = "smollm" if chkhsh == "c7699093ba4255a91e702aa38a596aa81669f3525dae06c2953267dde580f448": # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-zh res = "jina-v2-zh" - + if res is None: logger.warning("\n") logger.warning("**************************************************************************************") @@ -560,7 +669,7 @@ def _set_vocab_qwen(self): for i in range(vocab_size): if i not in reverse_vocab: tokens.append(f"[PAD{i}]") - toktypes.append(gguf.TokenType.USER_DEFINED) + toktypes.append(gguf.TokenType.UNUSED) elif reverse_vocab[i] in added_vocab: tokens.append(reverse_vocab[i]) toktypes.append(gguf.TokenType.CONTROL) @@ -610,7 +719,7 @@ def _create_vocab_sentencepiece(self): tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] scores: list[float] = [-10000.0] * vocab_size - toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size + toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size for token_id in range(tokenizer.vocab_size()): piece = tokenizer.IdToPiece(token_id) @@ -637,7 +746,7 @@ def _create_vocab_sentencepiece(self): added_tokens_json = json.load(f) for key in added_tokens_json: token_id = added_tokens_json[key] - if (token_id >= vocab_size): + if token_id >= vocab_size: logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') continue @@ -645,6 +754,26 @@ def _create_vocab_sentencepiece(self): scores[token_id] = -1000.0 toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED + tokenizer_config_file = self.dir_model / 'tokenizer_config.json' + if tokenizer_config_file.is_file(): + with open(tokenizer_config_file, "r", encoding="utf-8") as f: + tokenizer_config_json = json.load(f) + added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {}) + for token_id, token_data in added_tokens_decoder.items(): + token_id = int(token_id) + token: str = token_data["content"] + if toktypes[token_id] != SentencePieceTokenTypes.UNUSED: + if tokens[token_id] != token.encode("utf-8"): + logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token!r}') + if token_data.get("special") or self.does_token_look_special(token): + toktypes[token_id] = SentencePieceTokenTypes.CONTROL + else: + token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ") # pre-normalize user-defined spaces + toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED + + scores[token_id] = -1000.0 + tokens[token_id] = token.encode("utf-8") + if vocab_size > len(tokens): pad_count = vocab_size - len(tokens) logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]") @@ -730,7 +859,6 @@ class GPTNeoXModel(Model): def set_gguf_parameters(self): block_count = self.hparams["num_hidden_layers"] - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) self.gguf_writer.add_block_count(block_count) @@ -786,7 +914,6 @@ class BloomModel(Model): model_arch = gguf.MODEL_ARCH.BLOOM def set_gguf_parameters(self): - self.gguf_writer.add_name("Bloom") n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed")) n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads")) self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed)) @@ -863,7 +990,6 @@ def set_vocab(self): def set_gguf_parameters(self): block_count = self.hparams["n_layers"] - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_context_length(self.hparams["max_seq_len"]) self.gguf_writer.add_embedding_length(self.hparams["d_model"]) self.gguf_writer.add_block_count(block_count) @@ -902,7 +1028,6 @@ def set_gguf_parameters(self): block_count = self.hparams["num_hidden_layers"] head_count = self.hparams["num_attention_heads"] head_count_kv = self.hparams.get("num_key_value_heads", head_count) - hf_repo = self.hparams.get("_name_or_path", "") ctx_length = 0 if "max_sequence_length" in self.hparams: @@ -915,8 +1040,6 @@ def set_gguf_parameters(self): raise ValueError("gguf: can not find ctx length parameter.") self.gguf_writer.add_file_type(self.ftype) - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) - self.gguf_writer.add_source_hf_repo(hf_repo) self.gguf_writer.add_tensor_data_layout("Meta AI original pth") self.gguf_writer.add_context_length(ctx_length) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) @@ -940,7 +1063,6 @@ def set_gguf_parameters(self): block_count = self.hparams["num_hidden_layers"] head_count = self.hparams["num_attention_heads"] head_count_kv = self.hparams.get("num_key_value_heads", head_count) - hf_repo = self.hparams.get("_name_or_path", "") ctx_length = 0 if "max_sequence_length" in self.hparams: @@ -952,8 +1074,6 @@ def set_gguf_parameters(self): else: raise ValueError("gguf: can not find ctx length parameter.") - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) - self.gguf_writer.add_source_hf_repo(hf_repo) self.gguf_writer.add_tensor_data_layout("Meta AI original pth") self.gguf_writer.add_context_length(ctx_length) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) @@ -1067,7 +1187,6 @@ def set_gguf_parameters(self): block_count = self.hparams["num_hidden_layers"] head_count = self.hparams["num_attention_heads"] head_count_kv = self.hparams.get("num_key_value_heads", head_count) - hf_repo = self.hparams.get("_name_or_path", "") ctx_length = 0 if "max_sequence_length" in self.hparams: @@ -1079,8 +1198,6 @@ def set_gguf_parameters(self): else: raise ValueError("gguf: can not find ctx length parameter.") - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) - self.gguf_writer.add_source_hf_repo(hf_repo) self.gguf_writer.add_tensor_data_layout("Meta AI original pth") self.gguf_writer.add_context_length(ctx_length) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) @@ -1139,7 +1256,6 @@ def set_gguf_parameters(self): if n_head_kv is None: n_head_kv = self.hparams.get("n_head_kv", 1) # old name - self.gguf_writer.add_name("Falcon") self.gguf_writer.add_context_length(2048) # not in config.json self.gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) @@ -1184,7 +1300,6 @@ class StarCoderModel(Model): def set_gguf_parameters(self): block_count = self.hparams["n_layer"] - self.gguf_writer.add_name("StarCoder") self.gguf_writer.add_context_length(self.hparams["n_positions"]) self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"]) @@ -1204,11 +1319,11 @@ def set_vocab(self): # TODO: how to determine special FIM tokens automatically? special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False, - special_token_types = ['prefix', 'suffix', 'middle', 'fsep', 'eot']) + special_token_types = ['prefix', 'suffix', 'middle', 'eot']) special_vocab._set_special_token("prefix", 1) special_vocab._set_special_token("suffix", 3) special_vocab._set_special_token("middle", 2) - special_vocab._set_special_token("fsep", 4) # is this correct? + special_vocab.chat_template = None # do not add it twice special_vocab.add_to_gguf(self.gguf_writer) def set_gguf_parameters(self): @@ -1220,7 +1335,6 @@ def set_gguf_parameters(self): block_count = self.hparams["n_layer"] - self.gguf_writer.add_name("Refact") # refact uses Alibi. So this is from config.json which might be used by training. self.gguf_writer.add_context_length(self.hparams["n_positions"]) self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) @@ -1268,14 +1382,13 @@ def set_vocab(self): if (self.dir_model / "tokenizer.json").is_file(): self._set_vocab_gpt2() else: - # StableLM 2 1.6B uses a vocab in a similar format to Qwen's vocab + # StableLM 2 1.6B used to have a vocab in a similar format to Qwen's vocab self._set_vocab_qwen() def set_gguf_parameters(self): hparams = self.hparams block_count = hparams["num_hidden_layers"] - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_context_length(hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(hparams["hidden_size"]) self.gguf_writer.add_block_count(block_count) @@ -1337,8 +1450,8 @@ def _stack_qk_norm(self, bid: int, n_head: int, norms: dict[str, Tensor], layer_ return [(new_name, data_torch)] - def write_tensors(self): - super().write_tensors() + def prepare_tensors(self): + super().prepare_tensors() if self._q_norms is not None or self._k_norms is not None: # flatten two `list[dict[str, Tensor]]` into a single `list[str]` @@ -1357,7 +1470,7 @@ class LlamaModel(Model): def set_vocab(self): try: - self. _set_vocab_sentencepiece() + self._set_vocab_sentencepiece() except FileNotFoundError: try: self._set_vocab_llama_hf() @@ -1381,7 +1494,12 @@ def set_gguf_parameters(self): super().set_gguf_parameters() hparams = self.hparams self.gguf_writer.add_vocab_size(hparams["vocab_size"]) - self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"]) + + if "head_dim" in hparams: + rope_dim = hparams["head_dim"] + else: + rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"] + self.gguf_writer.add_rope_dimension_count(rope_dim) if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: if self.hparams["rope_scaling"].get("type") == "linear": @@ -1454,8 +1572,8 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter return [(self.map_tensor_name(name), data_torch)] - def write_tensors(self): - super().write_tensors() + def prepare_tensors(self): + super().prepare_tensors() if self._experts is not None: # flatten `list[dict[str, Tensor]]` into `list[str]` @@ -1518,7 +1636,6 @@ def __init__(self, *args, **kwargs): def set_gguf_parameters(self): super().set_gguf_parameters() - self.gguf_writer.add_name("Grok") _experts: list[dict[str, Tensor]] | None = None @@ -1567,7 +1684,6 @@ class DbrxModel(Model): def set_gguf_parameters(self): ffn_config = self.hparams["ffn_config"] attn_config = self.hparams["attn_config"] - self.gguf_writer.add_name(self.hparams["model_type"]) self.gguf_writer.add_block_count(self.hparams["n_layers"]) self.gguf_writer.add_context_length(self.hparams["max_seq_len"]) @@ -1580,7 +1696,6 @@ def set_gguf_parameters(self): self.gguf_writer.add_rope_freq_base(attn_config["rope_theta"]) self.gguf_writer.add_clamp_kqv(attn_config["clip_qkv"]) - self.gguf_writer.add_file_type(self.ftype) self.gguf_writer.add_expert_count(ffn_config["moe_num_experts"]) self.gguf_writer.add_expert_used_count(ffn_config["moe_top_k"]) @@ -1637,7 +1752,6 @@ class MiniCPMModel(Model): def set_gguf_parameters(self): block_count = self.hparams["num_hidden_layers"] - self.gguf_writer.add_name("MiniCPM") self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) self.gguf_writer.add_block_count(block_count) @@ -1707,7 +1821,6 @@ def set_vocab(self): self._set_vocab_qwen() def set_gguf_parameters(self): - self.gguf_writer.add_name("Qwen") self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"]) self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"]) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) @@ -1783,8 +1896,8 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter return [(self.map_tensor_name(name), data_torch)] - def write_tensors(self): - super().write_tensors() + def prepare_tensors(self): + super().prepare_tensors() if self._experts is not None: # flatten `list[dict[str, Tensor]]` into `list[str]` @@ -1798,7 +1911,6 @@ class GPT2Model(Model): model_arch = gguf.MODEL_ARCH.GPT2 def set_gguf_parameters(self): - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_block_count(self.hparams["n_layer"]) self.gguf_writer.add_context_length(self.hparams["n_ctx"]) self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) @@ -1841,7 +1953,6 @@ def set_gguf_parameters(self): n_embd = self.find_hparam(["hidden_size", "n_embd"]) n_head = self.find_hparam(["num_attention_heads", "n_head"]) - self.gguf_writer.add_name("Phi2") self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"])) self.gguf_writer.add_embedding_length(n_embd) @@ -1874,7 +1985,7 @@ def set_vocab(self): tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] scores: list[float] = [-10000.0] * vocab_size - toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size + toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size for token_id in range(tokenizer.vocab_size()): @@ -1903,7 +2014,7 @@ def set_vocab(self): for key in added_tokens_json: token_id = added_tokens_json[key] - if (token_id >= vocab_size): + if token_id >= vocab_size: logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') continue @@ -1919,8 +2030,9 @@ def set_vocab(self): for token_id, foken_data in added_tokens_decoder.items(): token_id = int(token_id) token = foken_data["content"].encode("utf-8") - if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN: - assert tokens[token_id] == token + if toktypes[token_id] != SentencePieceTokenTypes.UNUSED: + if tokens[token_id] != token: + logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token.decode("utf-8")!r}') tokens[token_id] = token scores[token_id] = -1000.0 toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED @@ -1935,8 +2047,9 @@ def set_vocab(self): for foken_data in added_tokens: token_id = int(foken_data["id"]) token = foken_data["content"].encode("utf-8") - if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN: - assert tokens[token_id] == token + if toktypes[token_id] != SentencePieceTokenTypes.UNUSED: + if tokens[token_id] != token: + logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token.decode("utf-8")!r}') tokens[token_id] = token scores[token_id] = -1000.0 toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED @@ -1963,7 +2076,6 @@ def set_gguf_parameters(self): orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"]) rope_dims = n_embd // n_head - self.gguf_writer.add_name("Phi3") self.gguf_writer.add_context_length(max_pos_embds) self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds) self.gguf_writer.add_embedding_length(n_embd) @@ -1975,10 +2087,11 @@ def set_gguf_parameters(self): self.gguf_writer.add_rope_dimension_count(rope_dims) self.gguf_writer.add_rope_freq_base(self.find_hparam(["rope_theta"])) self.gguf_writer.add_file_type(self.ftype) + self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"])) # write rope scaling for long context (128k) model rope_scaling = self.find_hparam(['rope_scaling'], True) - if (rope_scaling is None): + if rope_scaling is None: return scale = max_pos_embds / orig_max_pos_embds @@ -2020,7 +2133,6 @@ def set_gguf_parameters(self): hparams = self.hparams block_count = hparams["num_hidden_layers"] - self.gguf_writer.add_name("PLaMo") self.gguf_writer.add_context_length(4096) # not in config.json self.gguf_writer.add_embedding_length(hparams["hidden_size"]) self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) @@ -2065,7 +2177,6 @@ class CodeShellModel(Model): def set_gguf_parameters(self): block_count = self.hparams["n_layer"] - self.gguf_writer.add_name("CodeShell") self.gguf_writer.add_context_length(self.hparams["n_positions"]) self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"]) @@ -2145,6 +2256,9 @@ def set_vocab(self): toktype = SentencePieceTokenTypes.UNUSED elif tokenizer.IsByte(token_id): toktype = SentencePieceTokenTypes.BYTE + # take care of ununsed raw token + if piece.startswith('[UNUSED'): + toktype = SentencePieceTokenTypes.UNUSED tokens.append(text) scores.append(score) @@ -2160,6 +2274,49 @@ def set_vocab(self): scores.append(-1000.0) toktypes.append(SentencePieceTokenTypes.USER_DEFINED) + chat_eos_token = '<|im_end|>' + chat_eos_token_id = None + + tokenizer_config_file = self.dir_model / 'tokenizer_config.json' + if tokenizer_config_file.is_file(): + with open(tokenizer_config_file, "r", encoding="utf-8") as f: + tokenizer_config_json = json.load(f) + added_tokens_decoder = tokenizer_config_json.get("added_tokens_decoder", {}) + for token_id, foken_data in added_tokens_decoder.items(): + token_id = int(token_id) + token = foken_data["content"] + if token == chat_eos_token: + chat_eos_token_id = token_id + token = token.encode("utf-8") + if toktypes[token_id] != SentencePieceTokenTypes.UNUSED: + if tokens[token_id] != token: + logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token.decode("utf-8")!r}') + tokens[token_id] = token + scores[token_id] = -1000.0 + toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED + if foken_data.get("special"): + toktypes[token_id] = SentencePieceTokenTypes.CONTROL + + tokenizer_file = self.dir_model / 'tokenizer.json' + if tokenizer_file.is_file(): + with open(tokenizer_file, "r", encoding="utf-8") as f: + tokenizer_json = json.load(f) + added_tokens = tokenizer_json.get("added_tokens", []) + for foken_data in added_tokens: + token_id = int(foken_data["id"]) + token = foken_data["content"] + if token == chat_eos_token: + chat_eos_token_id = token_id + token = token.encode("utf-8") + if toktypes[token_id] != SentencePieceTokenTypes.UNUSED: + if tokens[token_id] != token: + logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token.decode("utf-8")!r}') + tokens[token_id] = token + scores[token_id] = -1000.0 + toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED + if foken_data.get("special"): + toktypes[token_id] = SentencePieceTokenTypes.CONTROL + self.gguf_writer.add_tokenizer_model("llama") self.gguf_writer.add_tokenizer_pre("default") self.gguf_writer.add_token_list(tokens) @@ -2169,37 +2326,17 @@ def set_vocab(self): special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) old_eos = special_vocab.special_token_ids["eos"] - if "chat" in os.path.basename(self.dir_model.absolute()): + if chat_eos_token_id is not None: # For the chat model, we replace the eos with '<|im_end|>'. # TODO: this is a hack, should be fixed # https://github.com/ggerganov/llama.cpp/pull/6745#issuecomment-2067687048 - special_vocab.special_token_ids["eos"] = self._try_get_sft_eos(tokenizer) - logger.warning(f"Replace eos:{old_eos} with a special token:{special_vocab.special_token_ids['eos']} \ -in chat mode so that the conversation can end normally.") + special_vocab.special_token_ids["eos"] = chat_eos_token_id + logger.warning(f"Replace eos:{old_eos} with a special token:{chat_eos_token_id}" + " in chat mode so that the conversation can end normally.") special_vocab.add_to_gguf(self.gguf_writer) - def _try_get_sft_eos(self, tokenizer): - unused_145_list = tokenizer.Encode('[UNUSED_TOKEN_145]') - im_end_list = tokenizer.Encode('<|im_end|>') - eos_token = None - assert (len(unused_145_list) == 1) ^ (len(im_end_list) == 1) - if len(unused_145_list) == 1: - eos_token = unused_145_list[0] - if len(im_end_list) == 1: - eos_token = im_end_list[0] - assert eos_token - return eos_token - - def _hf_permute_qk(self, weights, n_head: int, n_head_kv: int): - if n_head_kv is not None and n_head != n_head_kv: - n_head = n_head_kv - return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) - .swapaxes(1, 2) - .reshape(weights.shape)) - def set_gguf_parameters(self): - self.gguf_writer.add_name("InternLM2") self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"]) self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"]) self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) @@ -2209,30 +2346,30 @@ def set_gguf_parameters(self): self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"]) self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"]) self.gguf_writer.add_file_type(self.ftype) + if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: + if self.hparams["rope_scaling"].get("type") == "linear": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: num_heads = self.hparams["num_attention_heads"] num_kv_heads = self.hparams["num_key_value_heads"] - hidden_size = self.hparams["hidden_size"] + n_embd = self.hparams["hidden_size"] q_per_kv = num_heads // num_kv_heads - head_dim = hidden_size // num_heads + head_dim = n_embd // num_heads num_groups = num_heads // q_per_kv - qkv_pattern = r"model\.layers\.(\d+)\.attention\.wqkv" - - if re.match(qkv_pattern, name): - bid = re.findall(qkv_pattern, name)[0] + if bid is not None and f"model.layers.{bid}.attention.wqkv" in name: qkv = data_torch - # qkv = rearrange(qkv.T, " o (g n i) ->o g n i", g=num_groups, n=q_per_kv + 2, i=head_dim) - qkv = qkv.T.reshape((-1, num_groups, q_per_kv + 2, head_dim)) - q, k, v = qkv[..., : q_per_kv, :], qkv[..., q_per_kv: q_per_kv + 1, :], qkv[..., q_per_kv + 1: q_per_kv + 2, :] + + qkv = qkv.reshape((num_groups, q_per_kv + 2, head_dim, n_embd)) + q, k, v = qkv[:, : q_per_kv], qkv[:, -2], qkv[:, -1] + # The model weights of q and k equire additional reshape. - # q = self._hf_permute_qk(rearrange(q, " o g n i -> o (g n i)").T, num_heads, num_heads) - q = self._hf_permute_qk(q.reshape((q.shape[0], -1)).T, num_heads, num_heads) - # k = self._hf_permute_qk(rearrange(k, " o g n i -> o (g n i)").T, num_heads, num_kv_heads) - k = self._hf_permute_qk(k.reshape((k.shape[0], -1)).T, num_heads, num_kv_heads) - # v = rearrange(v, " o g n i -> o (g n i)").T - v = v.reshape((v.shape[0], -1)).T + q = LlamaModel.permute(q.reshape((-1, q.shape[-1])), num_heads, num_heads) + k = LlamaModel.permute(k.reshape((-1, k.shape[-1])), num_heads, num_kv_heads) + v = v.reshape((-1, v.shape[-1])) + return [ (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), q), (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), k), @@ -2359,6 +2496,7 @@ def set_vocab(self): special_vocab._set_special_token("middle", 68) special_vocab._set_special_token("fsep", 70) special_vocab._set_special_token("eot", 107) + special_vocab.chat_template = None # do not add it twice special_vocab.add_to_gguf(self.gguf_writer) self.gguf_writer.add_add_space_prefix(False) @@ -2367,7 +2505,6 @@ def set_gguf_parameters(self): hparams = self.hparams block_count = hparams["num_hidden_layers"] - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_context_length(hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(hparams["hidden_size"]) self.gguf_writer.add_block_count(block_count) @@ -2400,19 +2537,7 @@ class Gemma2Model(Model): model_arch = gguf.MODEL_ARCH.GEMMA2 def set_vocab(self): - tokens, scores, toktypes = self._create_vocab_sentencepiece() - # hack: This is required so that we can properly use start/end-of-turn for chat template - for i in range(108): - # including , , - toktypes[i] = SentencePieceTokenTypes.CONTROL - self.gguf_writer.add_tokenizer_model("llama") - self.gguf_writer.add_tokenizer_pre("default") - self.gguf_writer.add_token_list(tokens) - self.gguf_writer.add_token_scores(scores) - self.gguf_writer.add_token_types(toktypes) - - special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) - special_vocab.add_to_gguf(self.gguf_writer) + self._set_vocab_sentencepiece() self.gguf_writer.add_add_space_prefix(False) @@ -2420,7 +2545,6 @@ def set_gguf_parameters(self): hparams = self.hparams block_count = hparams["num_hidden_layers"] - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_context_length(hparams["max_position_embeddings"]) self.gguf_writer.add_embedding_length(hparams["hidden_size"]) self.gguf_writer.add_block_count(block_count) @@ -2439,11 +2563,6 @@ def set_gguf_parameters(self): ) self.gguf_writer.add_sliding_window(self.hparams["sliding_window"]) - # sanity check - attn_scalar = self.hparams["query_pre_attn_scalar"] - if attn_scalar != hparams["hidden_size"] / hparams["num_attention_heads"]: - raise ValueError("query_pre_attn_scalar must be equal to n_embd / n_head") - def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: del bid # unused @@ -2500,7 +2619,6 @@ def set_gguf_parameters(self): # Fail early for models which don't have a block expansion factor of 2 assert d_inner == 2 * d_model - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default self.gguf_writer.add_embedding_length(d_model) self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading @@ -2620,7 +2738,7 @@ def get_tensors(self): yield name, data - def set_vocab(self, *args, **kwargs): + def set_vocab(self): tokenizer_class = 'BertTokenizer' with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f: tokenizer_class = json.load(f)['tokenizer_class'] @@ -2679,7 +2797,6 @@ def set_gguf_parameters(self): assert self.block_count == len(self._num_query_heads) assert self.block_count == len(self._ffn_dims) - self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name) self.gguf_writer.add_block_count(self.block_count) self.gguf_writer.add_context_length(self.hparams["max_context_length"]) self.gguf_writer.add_embedding_length(n_embd) @@ -2736,7 +2853,7 @@ def set_vocab(self): tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] scores: list[float] = [-10000.0] * vocab_size - toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size + toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size for token_id in range(tokenizer.vocab_size()): @@ -2769,7 +2886,7 @@ def set_vocab(self): added_tokens_decoder = tokenizer_config_json["added_tokens_decoder"] for token_id, token_json in added_tokens_decoder.items(): token_id = int(token_id) - if (token_id >= vocab_size): + if token_id >= vocab_size: logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') continue @@ -2853,8 +2970,8 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter return [(self.map_tensor_name(name), data_torch)] - def write_tensors(self): - super().write_tensors() + def prepare_tensors(self): + super().prepare_tensors() if self._experts is not None: # flatten `list[dict[str, Tensor]]` into `list[str]` @@ -2932,8 +3049,8 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter return [(self.map_tensor_name(name), data_torch)] - def write_tensors(self): - super().write_tensors() + def prepare_tensors(self): + super().prepare_tensors() if self._experts is not None: # flatten `list[dict[str, Tensor]]` into `list[str]` @@ -2991,7 +3108,7 @@ def set_vocab(self): tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] scores: list[float] = [-10000.0] * vocab_size - toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size + toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size for token_id in range(tokenizer.vocab_size()): piece = tokenizer.IdToPiece(token_id) @@ -3018,7 +3135,7 @@ def set_vocab(self): added_tokens_json = json.load(f) for key in added_tokens_json: token_id = added_tokens_json[key] - if (token_id >= vocab_size): + if token_id >= vocab_size: logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') continue @@ -3051,7 +3168,6 @@ def set_vocab(self): self.gguf_writer.add_add_eos_token(True) def set_gguf_parameters(self): - self.gguf_writer.add_name("T5") if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None: logger.warning("Couldn't find context length in config.json, assuming default value of 512") n_ctx = 512 @@ -3125,7 +3241,6 @@ def set_vocab(self): self._set_vocab_gpt2() def set_gguf_parameters(self): - self.gguf_writer.add_name(self.dir_model.name) self.gguf_writer.add_block_count(self.hparams["n_layer"]) self.gguf_writer.add_context_length(self.hparams["n_positions"]) self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) @@ -3171,8 +3286,8 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter return tensors - def write_tensors(self): - super().write_tensors() + def prepare_tensors(self): + super().prepare_tensors() self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias) @@ -3209,15 +3324,14 @@ def set_vocab_chatglm3(self): if len(piece) != 0 and token_id < tokenizer.tokenizer.sp_model.vocab_size(): score = tokenizer.tokenizer.sp_model.get_score(token_id) - if len(piece) == 0: - text = f"[PAD{token_id}]".encode("utf-8") - if token_id >= tokenizer.tokenizer.sp_model.vocab_size(): if piece in special_tokens: - # show special tokens in prompt - toktype = SentencePieceTokenTypes.USER_DEFINED + toktype = SentencePieceTokenTypes.CONTROL + elif len(piece) == 0: + text = f"[PAD{token_id}]".encode("utf-8") + toktype = SentencePieceTokenTypes.UNUSED else: - toktype = SentencePieceTokenTypes.UNKNOWN + toktype = SentencePieceTokenTypes.USER_DEFINED tokens.append(text) scores.append(score) toktypes.append(toktype) @@ -3306,7 +3420,7 @@ def set_vocab(self): for i in range(vocab_size): if i not in reverse_vocab: tokens.append(f"[PAD{i}]") - toktypes.append(gguf.TokenType.USER_DEFINED) + toktypes.append(gguf.TokenType.UNUSED) elif reverse_vocab[i] in added_vocab: tokens.append(reverse_vocab[i]) if tokenizer.added_tokens_decoder[i].special: @@ -3332,7 +3446,6 @@ def set_vocab(self): special_vocab.add_to_gguf(self.gguf_writer) def set_gguf_parameters(self): - self.gguf_writer.add_name(self.hparams["_name_or_path"].split("/")[1]) # THUDM/glm4-9b-chat or THUDM/chatglm3-6b n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed")) n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads")) n_head_kv = self.hparams.get("multi_query_group_num", n_head) @@ -3376,19 +3489,46 @@ class LazyTorchTensor(gguf.LazyBase): torch.float32: np.float32, } + # used for safetensors slices + # ref: https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/src/lib.rs#L1046 + # TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734 + _dtype_str_map: dict[str, torch.dtype] = { + "F64": torch.float64, + "F32": torch.float32, + "BF16": torch.bfloat16, + "F16": torch.float16, + # "U64": torch.uint64, + "I64": torch.int64, + # "U32": torch.uint32, + "I32": torch.int32, + # "U16": torch.uint16, + "I16": torch.int16, + "U8": torch.uint8, + "I8": torch.int8, + "BOOL": torch.bool, + "F8_E4M3": torch.float8_e4m3fn, + "F8_E5M2": torch.float8_e5m2, + } + def numpy(self) -> gguf.LazyNumpyTensor: dtype = self._dtype_map[self.dtype] return gguf.LazyNumpyTensor( meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(dtype, self.shape), - lazy=self._lazy, args=(self,), - func=(lambda s: s[0].numpy()) + func=(lambda s: s.numpy()) ) @classmethod - def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: torch.Size) -> Tensor: + def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: tuple[int, ...]) -> Tensor: return torch.empty(size=shape, dtype=dtype, device="meta") + @classmethod + def from_safetensors_slice(cls, st_slice: Any) -> Tensor: + dtype = cls._dtype_str_map[st_slice.get_dtype()] + shape: tuple[int, ...] = tuple(st_slice.get_shape()) + lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:]) + return cast(torch.Tensor, lazy) + @classmethod def __torch_function__(cls, func, types, args=(), kwargs=None): del types # unused @@ -3399,7 +3539,7 @@ def __torch_function__(cls, func, types, args=(), kwargs=None): if func is torch.Tensor.numpy: return args[0].numpy() - return LazyTorchTensor._wrap_fn(func)(*args, **kwargs) + return cls._wrap_fn(func)(*args, **kwargs) def parse_args() -> argparse.Namespace: @@ -3457,6 +3597,10 @@ def parse_args() -> argparse.Namespace: "--no-tensor-first-split", action="store_true", help="do not add tensors to the first split (disabled by default)" ) + parser.add_argument( + "--metadata", type=Path, + help="Specify the path for an authorship metadata override file" + ) return parser.parse_args() @@ -3482,7 +3626,10 @@ def split_str_to_n_bytes(split_str: str) -> int: def main() -> None: args = parse_args() - logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + if args.verbose: + logging.basicConfig(level=logging.DEBUG) + else: + logging.basicConfig(level=logging.INFO) dir_model = args.model @@ -3506,33 +3653,30 @@ def main() -> None: if args.outfile is not None: fname_out = args.outfile else: - # output in the same directory as the model by default - fname_out = dir_model / 'ggml-model-{ftype}.gguf' + fname_out = dir_model logger.info(f"Loading model: {dir_model.name}") hparams = Model.load_hparams(dir_model) with torch.inference_mode(): + output_type = ftype_map[args.outtype] + model_architecture = hparams["architectures"][0] + try: - model_class = Model.from_model_architecture(hparams["architectures"][0]) + model_class = Model.from_model_architecture(model_architecture) except NotImplementedError: - logger.error(f"Model {hparams['architectures'][0]} is not supported") + logger.error(f"Model {model_architecture} is not supported") sys.exit(1) - model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file, - args.no_lazy, args.model_name, split_max_tensors=args.split_max_tensors, + model_instance = model_class(dir_model=dir_model, ftype=output_type, fname_out=fname_out, + is_big_endian=args.bigendian, use_temp_file=args.use_temp_file, + eager=args.no_lazy, + metadata_override=args.metadata, model_name=args.model_name, + split_max_tensors=args.split_max_tensors, split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run, small_first_shard=args.no_tensor_first_split) - logger.info("Set model parameters") - model_instance.set_gguf_parameters() - - logger.info("Set model tokenizer") - model_instance.set_vocab() - - model_instance.gguf_writer.add_quantization_version(gguf.GGML_QUANT_VERSION) - if args.vocab_only: logger.info("Exporting model vocab...") model_instance.write_vocab() diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index 3bec0b1d41492..11702b5043580 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -50,7 +50,7 @@ class TOKENIZER_TYPE(IntEnum): # TODO: this string has to exercise as much pre-tokenizer functionality as possible # will be updated with time - contributions welcome -chktxt = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天~ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL' +CHK_TXT = '\n \n\n \n\n\n \t \t\t \t\n \n \n \n \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天~ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL' if len(sys.argv) == 2: token = sys.argv[1] @@ -91,6 +91,9 @@ class TOKENIZER_TYPE(IntEnum): {"name": "gemma-2", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", }, {"name": "jais", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", }, {"name": "t5", "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", }, + {"name": "codeshell", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", }, + {"name": "tekken", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", }, + {"name": "smollm", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", }, {"name": "jina-v2-zh", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-zh", }, ] @@ -100,8 +103,8 @@ def download_file_with_auth(url, token, save_path): response = sess.get(url, headers=headers) response.raise_for_status() os.makedirs(os.path.dirname(save_path), exist_ok=True) - with open(save_path, 'wb') as f: - f.write(response.content) + with open(save_path, 'wb') as downloaded_file: + downloaded_file.write(response.content) logger.info(f"File {save_path} downloaded successfully") @@ -160,7 +163,7 @@ def download_model(model): logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}") continue # Skip to the next model if the tokenizer can't be loaded - chktok = tokenizer.encode(chktxt) + chktok = tokenizer.encode(CHK_TXT) chkhsh = sha256(str(chktok).encode()).hexdigest() logger.info(f"model: {name}") @@ -192,7 +195,7 @@ def get_vocab_base_pre(self, tokenizer) -> str: # we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can # use in llama.cpp to implement the same pre-tokenizer - chktxt = {repr(chktxt)} + chktxt = {repr(CHK_TXT)} chktok = tokenizer.encode(chktxt) chkhsh = sha256(str(chktok).encode()).hexdigest() @@ -288,7 +291,7 @@ def get_vocab_base_pre(self, tokenizer) -> str: "333333333", "Cửa Việt", # llama-bpe fails on this " discards", - chktxt, + CHK_TXT, ] # write the tests to ./models/ggml-vocab-{name}.gguf.inp diff --git a/convert_llama_ggml_to_gguf.py b/convert_llama_ggml_to_gguf.py index 95ea831a50a9c..7b00b4398178b 100755 --- a/convert_llama_ggml_to_gguf.py +++ b/convert_llama_ggml_to_gguf.py @@ -132,6 +132,10 @@ def load(self, data, offset): class GGMLModel: + + file_format: GGMLFormat + format_version: int + def __init__(self): self.hyperparameters = None self.vocab = None @@ -290,7 +294,7 @@ def add_vocab(self, gguf_writer): if self.vocab_override is not None: vo = self.vocab_override logger.info('* Adding vocab item(s)') - for (idx, (vbytes, score, ttype)) in enumerate(vo.all_tokens()): + for (_, (vbytes, score, ttype)) in enumerate(vo.all_tokens()): tokens.append(vbytes) scores.append(score) toktypes.append(ttype) diff --git a/convert_lora_to_gguf.py b/convert_lora_to_gguf.py new file mode 100755 index 0000000000000..a88d0d4a978a9 --- /dev/null +++ b/convert_lora_to_gguf.py @@ -0,0 +1,393 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- + +from __future__ import annotations + +from dataclasses import dataclass +import logging +import argparse +import os +import sys +import json +from math import prod +from pathlib import Path +from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast + +import torch + +if TYPE_CHECKING: + from torch import Tensor + +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) +import gguf + +# reuse model definitions from convert_hf_to_gguf.py +from convert_hf_to_gguf import LazyTorchTensor, Model + +logger = logging.getLogger("lora-to-gguf") + + +@dataclass +class PartialLoraTensor: + A: Tensor | None = None + B: Tensor | None = None + + +# magic to support tensor shape modifications and splitting +class LoraTorchTensor: + _lora_A: Tensor # (n_rank, row_size) + _lora_B: Tensor # (col_size, n_rank) + _rank: int + + def __init__(self, A: Tensor, B: Tensor): + assert len(A.shape) == len(B.shape) + assert A.shape[-2] == B.shape[-1] + if A.dtype != B.dtype: + A = A.to(torch.float32) + B = B.to(torch.float32) + self._lora_A = A + self._lora_B = B + self._rank = B.shape[-1] + + def get_lora_A_B(self) -> tuple[Tensor, Tensor]: + return (self._lora_A, self._lora_B) + + def __getitem__( + self, + indices: ( + SupportsIndex + | slice + | tuple[SupportsIndex | slice | Tensor, ...] # TODO: add ellipsis in the type signature + ), + ) -> LoraTorchTensor: + shape = self.shape + if isinstance(indices, SupportsIndex): + if len(shape) > 2: + return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices]) + else: + raise NotImplementedError # can't return a vector + elif isinstance(indices, slice): + if len(shape) > 2: + return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices]) + else: + return LoraTorchTensor(self._lora_A, self._lora_B[indices]) + elif isinstance(indices, tuple): + assert len(indices) > 0 + if indices[-1] is Ellipsis: + return self[indices[:-1]] + # expand ellipsis + indices = tuple( + u + for v in ( + ( + (slice(None, None) for _ in range(len(indices) - 1)) + if i is Ellipsis + else (i,) + ) + for i in indices + ) + for u in v + ) + + if len(indices) < len(shape): + indices = (*indices, *(slice(None, None) for _ in range(len(indices), len(shape)))) + + # TODO: make sure this is correct + indices_A = ( + *( + ( + j.__index__() % self._lora_A.shape[i] + if isinstance(j, SupportsIndex) + else slice(None, None) + ) + for i, j in enumerate(indices[:-2]) + ), + slice(None, None), + indices[-1], + ) + indices_B = indices[:-1] + return LoraTorchTensor(self._lora_A[indices_A], self._lora_B[indices_B]) + else: + raise NotImplementedError # unknown indice type + + @property + def dtype(self) -> torch.dtype: + assert self._lora_A.dtype == self._lora_B.dtype + return self._lora_A.dtype + + @property + def shape(self) -> tuple[int, ...]: + assert len(self._lora_A.shape) == len(self._lora_B.shape) + return (*self._lora_B.shape[:-1], self._lora_A.shape[-1]) + + def size(self, dim=None): + assert dim is None + return self.shape + + def reshape(self, *shape: int | tuple[int, ...]) -> LoraTorchTensor: + if isinstance(shape[0], tuple): + new_shape: tuple[int, ...] = shape[0] + else: + new_shape = cast(tuple[int, ...], shape) + orig_shape = self.shape + if len(new_shape) < 2: + raise NotImplementedError # can't become a vector + + # expand -1 in the shape + if any(dim == -1 for dim in new_shape): + n_elems = prod(orig_shape) + n_new_elems = prod(dim if dim != -1 else 1 for dim in new_shape) + assert n_elems % n_new_elems == 0 + new_shape = (*(dim if dim != -1 else n_elems // n_new_elems for dim in new_shape),) + + if new_shape[-1] != orig_shape[-1]: + raise NotImplementedError # can't reshape the row size trivially + + shape_A = (*(1 for _ in new_shape[:-2]), self._rank, orig_shape[-1]) + shape_B = (*new_shape[:-1], self._rank) + return LoraTorchTensor( + self._lora_A.reshape(shape_A), + self._lora_B.reshape(shape_B), + ) + + def reshape_as(self, other: Tensor) -> LoraTorchTensor: + return self.reshape(*other.shape) + + def view(self, *size: int) -> LoraTorchTensor: + return self.reshape(*size) + + def permute(self, *dims: int) -> LoraTorchTensor: + shape = self.shape + dims = tuple(dim - len(shape) if dim >= 0 else dim for dim in dims) + if dims[-1] == -1: + # TODO: support higher dimensional A shapes bigger than 1 + assert all(dim == 1 for dim in self._lora_A.shape[:-2]) + return LoraTorchTensor(self._lora_A, self._lora_B.permute(*dims)) + if len(shape) == 2 and dims[-1] == -2 and dims[-2] == -1: + return LoraTorchTensor(self._lora_B.permute(*dims), self._lora_A.permute(*dims)) + else: + # TODO: compose the above two + raise NotImplementedError + + def transpose(self, dim0: int, dim1: int) -> LoraTorchTensor: + shape = self.shape + dims = [i for i in range(len(shape))] + dims[dim0], dims[dim1] = dims[dim1], dims[dim0] + return self.permute(*dims) + + def swapaxes(self, axis0: int, axis1: int) -> LoraTorchTensor: + return self.transpose(axis0, axis1) + + def to(self, *args, **kwargs): + return LoraTorchTensor(self._lora_A.to(*args, **kwargs), self._lora_B.to(*args, **kwargs)) + + @classmethod + def __torch_function__(cls, func: Callable, types, args=(), kwargs=None): + del types # unused + + if kwargs is None: + kwargs = {} + + if func is torch.permute: + return type(args[0]).permute(*args, **kwargs) + elif func is torch.reshape: + return type(args[0]).reshape(*args, **kwargs) + elif func is torch.stack: + assert isinstance(args[0], Sequence) + dim = kwargs.get("dim", 0) + assert dim == 0 + return LoraTorchTensor( + torch.stack([a._lora_A for a in args[0]], dim), + torch.stack([b._lora_B for b in args[0]], dim), + ) + elif func is torch.cat: + assert isinstance(args[0], Sequence) + dim = kwargs.get("dim", 0) + assert dim == 0 + if len(args[0][0].shape) > 2: + return LoraTorchTensor( + torch.cat([a._lora_A for a in args[0]], dim), + torch.cat([b._lora_B for b in args[0]], dim), + ) + elif all(torch.equal(args[0][0]._lora_A, t._lora_A) for t in args[0][1:]): + return LoraTorchTensor( + args[0][0]._lora_A, + torch.cat([b._lora_B for b in args[0]], dim), + ) + else: + raise NotImplementedError + else: + raise NotImplementedError + + +def get_base_tensor_name(lora_tensor_name: str) -> str: + base_name = lora_tensor_name.replace("base_model.model.", "") + base_name = base_name.replace(".lora_A.weight", ".weight") + base_name = base_name.replace(".lora_B.weight", ".weight") + return base_name + + +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser( + description="Convert a huggingface PEFT LoRA adapter to a GGML compatible file") + parser.add_argument( + "--outfile", type=Path, + help="path to write to; default: based on input. {ftype} will be replaced by the outtype.", + ) + parser.add_argument( + "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16", + help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type", + ) + parser.add_argument( + "--bigendian", action="store_true", + help="model is executed on big endian machine", + ) + parser.add_argument( + "--no-lazy", action="store_true", + help="use more RAM by computing all outputs before writing (use in case lazy evaluation is broken)", + ) + parser.add_argument( + "--verbose", action="store_true", + help="increase output verbosity", + ) + parser.add_argument( + "--dry-run", action="store_true", + help="only print out what will be done, without writing any new files", + ) + parser.add_argument( + "--base", type=Path, required=True, + help="directory containing base model file", + ) + parser.add_argument( + "lora_path", type=Path, + help="directory containing LoRA adapter file", + ) + + return parser.parse_args() + + +if __name__ == '__main__': + args = parse_args() + logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + + ftype_map: dict[str, gguf.LlamaFileType] = { + "f32": gguf.LlamaFileType.ALL_F32, + "f16": gguf.LlamaFileType.MOSTLY_F16, + "bf16": gguf.LlamaFileType.MOSTLY_BF16, + "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0, + "auto": gguf.LlamaFileType.GUESSED, + } + + ftype = ftype_map[args.outtype] + + dir_base_model: Path = args.base + dir_lora: Path = args.lora_path + lora_config = dir_lora / "adapter_config.json" + input_model = dir_lora / "adapter_model.safetensors" + + if args.outfile is not None: + fname_out = args.outfile + else: + # output in the same directory as the model by default + fname_out = dir_lora + + if os.path.exists(input_model): + # lazy import load_file only if lora is in safetensors format. + from safetensors.torch import load_file + + lora_model = load_file(input_model, device="cpu") + else: + input_model = os.path.join(dir_lora, "adapter_model.bin") + lora_model = torch.load(input_model, map_location="cpu", weights_only=True) + + # load base model + logger.info(f"Loading base model: {dir_base_model.name}") + hparams = Model.load_hparams(dir_base_model) + with torch.inference_mode(): + try: + model_class = Model.from_model_architecture(hparams["architectures"][0]) + except NotImplementedError: + logger.error(f"Model {hparams['architectures'][0]} is not supported") + sys.exit(1) + + class LoraModel(model_class): + model_arch = model_class.model_arch + + lora_alpha: float + + def __init__(self, *args, dir_lora_model: Path, lora_alpha: float, **kwargs): + + super().__init__(*args, **kwargs) + + self.dir_model_card = dir_lora_model + self.lora_alpha = float(lora_alpha) + + def set_type(self): + self.gguf_writer.add_type(gguf.GGUFType.ADAPTER) + self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora") + + def set_gguf_parameters(self): + self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha) + super().set_gguf_parameters() + + def get_tensors(self) -> Iterator[tuple[str, Tensor]]: + tensor_map: dict[str, PartialLoraTensor] = {} + + for name, tensor in lora_model.items(): + if self.lazy: + tensor = LazyTorchTensor.from_eager(tensor) + base_name = get_base_tensor_name(name) + is_lora_a = ".lora_A.weight" in name + is_lora_b = ".lora_B.weight" in name + if not is_lora_a and not is_lora_b: + if ".base_layer.weight" in name: + continue + logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor") + sys.exit(1) + + if base_name in tensor_map: + if is_lora_a: + tensor_map[base_name].A = tensor + else: + tensor_map[base_name].B = tensor + else: + if is_lora_a: + tensor_map[base_name] = PartialLoraTensor(A=tensor) + else: + tensor_map[base_name] = PartialLoraTensor(B=tensor) + + for name, tensor in tensor_map.items(): + assert tensor.A is not None + assert tensor.B is not None + yield (name, cast(torch.Tensor, LoraTorchTensor(tensor.A, tensor.B))) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + dest = super().modify_tensors(data_torch, name, bid) + for dest_name, dest_data in dest: + assert isinstance(dest_data, LoraTorchTensor) + lora_a, lora_b = dest_data.get_lora_A_B() + + yield (dest_name + ".lora_a", lora_a) + yield (dest_name + ".lora_b", lora_b) + + with open(lora_config, "r") as f: + lparams: dict[str, Any] = json.load(f) + + alpha: float = lparams["lora_alpha"] + + model_instance = LoraModel( + dir_base_model, + ftype, + fname_out, + is_big_endian=args.bigendian, + use_temp_file=False, + eager=args.no_lazy, + dry_run=args.dry_run, + dir_lora_model=dir_lora, + lora_alpha=alpha, + ) + + logger.info("Exporting model...") + model_instance.write() + logger.info(f"Model successfully exported to {model_instance.fname_out}") diff --git a/docs/backend/SYCL.md b/docs/backend/SYCL.md index 885983e92277e..d36ac0a158dd4 100644 --- a/docs/backend/SYCL.md +++ b/docs/backend/SYCL.md @@ -293,31 +293,26 @@ Similar to the native `sycl-ls`, available SYCL devices can be queried as follow ```sh ./build/bin/llama-ls-sycl-device ``` -A example of such log in a system with 1 *intel CPU* and 1 *intel GPU* can look like the following: +This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following: ``` -found 6 SYCL devices: +found 2 SYCL devices: + | | | |Compute |Max compute|Max work|Max sub| | |ID| Device Type| Name|capability|units |group |group |Global mem size| |--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------| | 0|[level_zero:gpu:0]| Intel(R) Arc(TM) A770 Graphics| 1.3| 512| 1024| 32| 16225243136| | 1|[level_zero:gpu:1]| Intel(R) UHD Graphics 770| 1.3| 32| 512| 32| 53651849216| -| 2| [opencl:gpu:0]| Intel(R) Arc(TM) A770 Graphics| 3.0| 512| 1024| 32| 16225243136| -| 3| [opencl:gpu:1]| Intel(R) UHD Graphics 770| 3.0| 32| 512| 32| 53651849216| -| 4| [opencl:cpu:0]| 13th Gen Intel(R) Core(TM) i7-13700K| 3.0| 24| 8192| 64| 67064815616| -| 5| [opencl:acc:0]| Intel(R) FPGA Emulation Device| 1.2| 24|67108864| 64| 67064815616| ``` -| Attribute | Note | -|------------------------|-------------------------------------------------------------| -| compute capability 1.3 | Level-zero driver/runtime, recommended | -| compute capability 3.0 | OpenCL driver/runtime, slower than level-zero in most cases | 4. Launch inference There are two device selection modes: - Single device: Use one device target specified by the user. -- Multiple devices: Automatically select the devices with the same largest Max compute-units. +- Multiple devices: Automatically choose the devices with the same backend. + +In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR. | Device selection | Parameter | |------------------|----------------------------------------| @@ -474,33 +469,26 @@ Similar to the native `sycl-ls`, available SYCL devices can be queried as follow build\bin\ls-sycl-device.exe ``` -The output of this command in a system with 1 *intel CPU* and 1 *intel GPU* would look like the following: +This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following: ``` -found 6 SYCL devices: +found 2 SYCL devices: | | | |Compute |Max compute|Max work|Max sub| | |ID| Device Type| Name|capability|units |group |group |Global mem size| |--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------| | 0|[level_zero:gpu:0]| Intel(R) Arc(TM) A770 Graphics| 1.3| 512| 1024| 32| 16225243136| | 1|[level_zero:gpu:1]| Intel(R) UHD Graphics 770| 1.3| 32| 512| 32| 53651849216| -| 2| [opencl:gpu:0]| Intel(R) Arc(TM) A770 Graphics| 3.0| 512| 1024| 32| 16225243136| -| 3| [opencl:gpu:1]| Intel(R) UHD Graphics 770| 3.0| 32| 512| 32| 53651849216| -| 4| [opencl:cpu:0]| 13th Gen Intel(R) Core(TM) i7-13700K| 3.0| 24| 8192| 64| 67064815616| -| 5| [opencl:acc:0]| Intel(R) FPGA Emulation Device| 1.2| 24|67108864| 64| 67064815616| ``` -| Attribute | Note | -|------------------------|-----------------------------------------------------------| -| compute capability 1.3 | Level-zero running time, recommended | -| compute capability 3.0 | OpenCL running time, slower than level-zero in most cases | - 4. Launch inference There are two device selection modes: -- Single device: Use one device assigned by user. -- Multiple devices: Automatically choose the devices with the same biggest Max compute units. +- Single device: Use one device assigned by user. Default device id is 0. +- Multiple devices: Automatically choose the devices with the same backend. + +In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR. | Device selection | Parameter | |------------------|----------------------------------------| diff --git a/docs/build.md b/docs/build.md index bf41bfdf9c2f8..d9d12c46707bd 100644 --- a/docs/build.md +++ b/docs/build.md @@ -16,7 +16,7 @@ In order to build llama.cpp you have four different options. make ``` - - On Windows: + - On Windows (x86/x64 only, arm64 requires cmake): 1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases). 2. Extract `w64devkit` on your pc. @@ -28,6 +28,7 @@ In order to build llama.cpp you have four different options. ``` - Notes: + - For `Q4_0_4_4` quantization type build, add the `GGML_NO_LLAMAFILE=1` flag. For example, use `make GGML_NO_LLAMAFILE=1`. - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel. - For faster repeated compilation, install [ccache](https://ccache.dev/). - For debug builds, run `make LLAMA_DEBUG=1` @@ -41,6 +42,7 @@ In order to build llama.cpp you have four different options. **Notes**: + - For `Q4_0_4_4` quantization type build, add the `-DGGML_LLAMAFILE=OFF` cmake option. For example, use `cmake -B build -DGGML_LLAMAFILE=OFF`. - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel. - For faster repeated compilation, install [ccache](https://ccache.dev/). - For debug builds, there are two cases: @@ -58,6 +60,17 @@ In order to build llama.cpp you have four different options. cmake -B build -G "Xcode" cmake --build build --config Debug ``` + - Building for Windows (x86, x64 and arm64) with MSVC or clang as compilers: + - Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/de/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...): + - Tab Workload: Desktop-development with C++ + - Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang) + - Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test + - For Windows on ARM (arm64, WoA) build with: + ```bash + cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF + cmake --build build-arm64-windows-llvm-release + ``` + Note: Building for arm64 could also be done just with MSVC (with the build-arm64-windows-MSVC preset, or the standard CMake build instructions). But MSVC does not support inline ARM assembly-code, used e.g. for the accelerated Q4_0_4_8 CPU kernels. - Using `gmake` (FreeBSD): @@ -240,6 +253,45 @@ The following compilation options are also available to tweak performance (yes, ### Vulkan +**Windows** + +#### w64devkit + +Download and extract [w64devkit](https://github.com/skeeto/w64devkit/releases). + +Download and install the [Vulkan SDK](https://vulkan.lunarg.com/sdk/home#windows). When selecting components, only the Vulkan SDK Core is required. + +Launch `w64devkit.exe` and run the following commands to copy Vulkan dependencies: +```sh +SDK_VERSION=1.3.283.0 +cp /VulkanSDK/$SDK_VERSION/Bin/glslc.exe $W64DEVKIT_HOME/bin/ +cp /VulkanSDK/$SDK_VERSION/Lib/vulkan-1.lib $W64DEVKIT_HOME/x86_64-w64-mingw32/lib/ +cp -r /VulkanSDK/$SDK_VERSION/Include/* $W64DEVKIT_HOME/x86_64-w64-mingw32/include/ +cat > $W64DEVKIT_HOME/x86_64-w64-mingw32/lib/pkgconfig/vulkan.pc < Params: return params -@dataclass -class Metadata: - name: Optional[str] = None - author: Optional[str] = None - version: Optional[str] = None - url: Optional[str] = None - description: Optional[str] = None - license: Optional[str] = None - source_url: Optional[str] = None - source_hf_repo: Optional[str] = None - - @staticmethod - def load(metadata_path: Path) -> Metadata: - if metadata_path is None or not metadata_path.exists(): - return Metadata() - - with open(metadata_path, 'r') as file: - data = json.load(file) - - # Create a new Metadata instance - metadata = Metadata() - - # Assigning values to Metadata attributes if they exist in the JSON file - # This is based on LLM_KV_NAMES mapping in llama.cpp - metadata.name = data.get("general.name") - metadata.author = data.get("general.author") - metadata.version = data.get("general.version") - metadata.url = data.get("general.url") - metadata.description = data.get("general.description") - metadata.license = data.get("general.license") - metadata.source_url = data.get("general.source.url") - metadata.source_hf_repo = data.get("general.source.huggingface.repository") - - return metadata - - # # data loading # TODO: reuse (probably move to gguf.py?) @@ -806,7 +770,7 @@ class OutputFile: def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE): self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess) - def add_meta_model(self, params: Params, metadata: Metadata | None) -> None: + def add_meta_model(self, params: Params, metadata: gguf.Metadata | None) -> None: # Metadata About The Model And Its Provenence name = "LLaMA" if metadata is not None and metadata.name is not None: @@ -824,16 +788,73 @@ def add_meta_model(self, params: Params, metadata: Metadata | None) -> None: self.gguf.add_author(metadata.author) if metadata.version is not None: self.gguf.add_version(metadata.version) - if metadata.url is not None: - self.gguf.add_url(metadata.url) + if metadata.organization is not None: + self.gguf.add_organization(metadata.organization) + + if metadata.finetune is not None: + self.gguf.add_finetune(metadata.finetune) + if metadata.basename is not None: + self.gguf.add_basename(metadata.basename) + if metadata.description is not None: self.gguf.add_description(metadata.description) + if metadata.quantized_by is not None: + self.gguf.add_quantized_by(metadata.quantized_by) + + if metadata.size_label is not None: + self.gguf.add_size_label(metadata.size_label) + if metadata.license is not None: - self.gguf.add_licence(metadata.license) + self.gguf.add_license(metadata.license) + if metadata.license_name is not None: + self.gguf.add_license_name(metadata.license_name) + if metadata.license_link is not None: + self.gguf.add_license_link(metadata.license_link) + + if metadata.url is not None: + self.gguf.add_url(metadata.url) + if metadata.doi is not None: + self.gguf.add_doi(metadata.doi) + if metadata.uuid is not None: + self.gguf.add_uuid(metadata.uuid) + if metadata.repo_url is not None: + self.gguf.add_repo_url(metadata.repo_url) + if metadata.source_url is not None: self.gguf.add_source_url(metadata.source_url) - if metadata.source_hf_repo is not None: - self.gguf.add_source_hf_repo(metadata.source_hf_repo) + if metadata.source_doi is not None: + self.gguf.add_source_doi(metadata.source_doi) + if metadata.source_uuid is not None: + self.gguf.add_source_uuid(metadata.source_uuid) + if metadata.source_repo_url is not None: + self.gguf.add_source_repo_url(metadata.source_repo_url) + + if metadata.base_models is not None: + self.gguf.add_base_model_count(len(metadata.base_models)) + for key, base_model_entry in enumerate(metadata.base_models): + if "name" in base_model_entry: + self.gguf.add_base_model_name(key, base_model_entry["name"]) + if "author" in base_model_entry: + self.gguf.add_base_model_author(key, base_model_entry["author"]) + if "version" in base_model_entry: + self.gguf.add_base_model_version(key, base_model_entry["version"]) + if "organization" in base_model_entry: + self.gguf.add_base_model_organization(key, base_model_entry["organization"]) + if "url" in base_model_entry: + self.gguf.add_base_model_url(key, base_model_entry["url"]) + if "doi" in base_model_entry: + self.gguf.add_base_model_doi(key, base_model_entry["doi"]) + if "uuid" in base_model_entry: + self.gguf.add_base_model_uuid(key, base_model_entry["uuid"]) + if "repo_url" in base_model_entry: + self.gguf.add_base_model_repo_url(key, base_model_entry["repo_url"]) + + if metadata.tags is not None: + self.gguf.add_tags(metadata.tags) + if metadata.languages is not None: + self.gguf.add_languages(metadata.languages) + if metadata.datasets is not None: + self.gguf.add_datasets(metadata.datasets) def add_meta_arch(self, params: Params) -> None: # Metadata About The Neural Architecture Itself @@ -944,7 +965,7 @@ def close(self) -> None: @staticmethod def write_vocab_only( fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, - endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: Metadata | None = None, + endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: gguf.Metadata | None = None, ) -> None: check_vocab_size(params, vocab, pad_vocab=pad_vocab) @@ -978,7 +999,7 @@ def write_all( fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, - metadata: Metadata | None = None, + metadata: gguf.Metadata | None = None, ) -> None: check_vocab_size(params, vocab, pad_vocab=pad_vocab) @@ -1021,35 +1042,32 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT raise ValueError(f"Unexpected combination of types: {name_to_type}") -def model_parameter_count(model: LazyModel) -> int: - total_model_parameters = 0 - for i, (name, lazy_tensor) in enumerate(model.items()): - sum_weights_in_tensor = 1 +def per_model_weight_count_estimation(tensors: Iterable[tuple[str, LazyTensor]]) -> tuple[int, int, int]: + total_params = 0 + shared_params = 0 + expert_params = 0 + + for name, lazy_tensor in tensors: + # We don't need these + if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")): + continue + + # Got A Tensor + sum_weights_in_tensor: int = 1 + + # Tensor Volume for dim in lazy_tensor.shape: sum_weights_in_tensor *= dim - total_model_parameters += sum_weights_in_tensor - return total_model_parameters - - -def model_parameter_count_rounded_notation(model_params_count: int) -> str: - if model_params_count > 1e12 : - # Trillions Of Parameters - scaled_model_params = model_params_count * 1e-12 - scale_suffix = "T" - elif model_params_count > 1e9 : - # Billions Of Parameters - scaled_model_params = model_params_count * 1e-9 - scale_suffix = "B" - elif model_params_count > 1e6 : - # Millions Of Parameters - scaled_model_params = model_params_count * 1e-6 - scale_suffix = "M" - else: - # Thousands Of Parameters - scaled_model_params = model_params_count * 1e-3 - scale_suffix = "K" - return f"{round(scaled_model_params)}{scale_suffix}" + if ".experts." in name: + if ".experts.0." in name: + expert_params += sum_weights_in_tensor + else: + shared_params += sum_weights_in_tensor + + total_params += sum_weights_in_tensor + + return total_params, shared_params, expert_params def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel: @@ -1231,34 +1249,24 @@ def load_vocab(self, vocab_types: list[str] | None, model_parent_path: Path) -> return vocab, special_vocab -def default_convention_outfile(file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> str: - quantization = { +def default_convention_outfile(file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> str: + name = metadata.name if metadata.name is not None else None + basename = metadata.basename if metadata.basename is not None else None + finetune = metadata.finetune if metadata.finetune is not None else None + version = metadata.version if metadata.version is not None else None + size_label = metadata.size_label if metadata.size_label is not None else gguf.size_label(*model_params_count, expert_count=expert_count or 0) + + output_type = { GGMLFileType.AllF32: "F32", GGMLFileType.MostlyF16: "F16", GGMLFileType.MostlyQ8_0: "Q8_0", }[file_type] - parameters = model_parameter_count_rounded_notation(model_params_count) - - expert_count = "" - if params.n_experts is not None: - expert_count = f"{params.n_experts}x" - - version = "" - if metadata is not None and metadata.version is not None: - version = f"-{metadata.version}" + return gguf.naming_convention(name, basename, finetune, version, size_label, output_type) - name = "ggml-model" - if metadata is not None and metadata.name is not None: - name = metadata.name - elif params.path_model is not None: - name = params.path_model.name - return f"{name}{version}-{expert_count}{parameters}-{quantization}" - - -def default_outfile(model_paths: list[Path], file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> Path: - default_filename = default_convention_outfile(file_type, params, model_params_count, metadata) +def default_outfile(model_paths: list[Path], file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> Path: + default_filename = default_convention_outfile(file_type, expert_count, model_params_count, metadata) ret = model_paths[0].parent / f"{default_filename}.gguf" if ret in model_paths: logger.error( @@ -1297,8 +1305,9 @@ def main(args_in: list[str] | None = None) -> None: parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides") parser.add_argument("--skip-unknown", action="store_true", help="skip unknown tensor names instead of failing") parser.add_argument("--verbose", action="store_true", help="increase output verbosity") - parser.add_argument("--metadata", type=Path, help="Specify the path for a metadata file") + parser.add_argument("--metadata", type=Path, help="Specify the path for an authorship metadata override file") parser.add_argument("--get-outfile", action="store_true", help="get calculated default outfile name") + parser.add_argument("--model-name", type=str, default=None, help="name of the model") args = parser.parse_args(args_in) @@ -1310,32 +1319,36 @@ def main(args_in: list[str] | None = None) -> None: else: logging.basicConfig(level=logging.INFO) - metadata = Metadata.load(args.metadata) + model_name = args.model_name + dir_model = args.model + + metadata = gguf.Metadata.load(args.metadata, dir_model, model_name) if args.get_outfile: - model_plus = load_some_model(args.model) + model_plus = load_some_model(dir_model) params = Params.load(model_plus) - model = convert_model_names(model_plus.model, params, args.skip_unknown) - model_params_count = model_parameter_count(model_plus.model) - ftype = pick_output_type(model, args.outtype) - print(f"{default_convention_outfile(ftype, params, model_params_count, metadata)}") # noqa: NP100 + model = convert_model_names(model_plus.model, params, args.skip_unknown) + model_params_count = per_model_weight_count_estimation(model_plus.model.items()) + ftype = pick_output_type(model, args.outtype) + + if (metadata is None or metadata.name is None) and params.path_model is not None: + metadata.name = params.path_model.name + + print(f"{default_convention_outfile(ftype, params.n_experts, model_params_count, metadata)}") # noqa: NP100 return if args.no_vocab and args.vocab_only: raise ValueError("--vocab-only does not make sense with --no-vocab") if args.dump_single: - model_plus = lazy_load_file(args.model) + model_plus = lazy_load_file(dir_model) do_dump_model(model_plus) return if not args.vocab_only: - model_plus = load_some_model(args.model) + model_plus = load_some_model(dir_model) else: - model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None) - - model_params_count = model_parameter_count(model_plus.model) - logger.info(f"model parameters count : {model_params_count} ({model_parameter_count_rounded_notation(model_params_count)})") + model_plus = ModelPlus(model = {}, paths = [dir_model / 'dummy'], format = 'none', vocab = None) if args.dump: do_dump_model(model_plus) @@ -1368,7 +1381,7 @@ def main(args_in: list[str] | None = None) -> None: logger.info(f"params = {params}") model_parent_path = model_plus.paths[0].parent - vocab_path = Path(args.vocab_dir or args.model or model_parent_path) + vocab_path = Path(args.vocab_dir or dir_model or model_parent_path) vocab_factory = VocabFactory(vocab_path) vocab_types = None if args.no_vocab else args.vocab_type.split(",") vocab, special_vocab = vocab_factory.load_vocab(vocab_types, model_parent_path) @@ -1399,13 +1412,21 @@ def main(args_in: list[str] | None = None) -> None: assert params is not None + if metadata.name is None and params.path_model is not None: + metadata.name = params.path_model.name + + model_params_count = per_model_weight_count_estimation(model_plus.model.items()) + logger.info(f"model parameters count : {model_params_count} ({gguf.model_weight_count_rounded_notation(model_params_count[0])})") + logger.info(f"Vocab info: {vocab}") logger.info(f"Special vocab info: {special_vocab}") model = model_plus.model model = convert_model_names(model, params, args.skip_unknown) ftype = pick_output_type(model, args.outtype) model = convert_to_output_type(model, ftype) - outfile = args.outfile or default_outfile(model_plus.paths, ftype, params, model_params_count, metadata) + outfile = args.outfile or default_outfile(model_plus.paths, ftype, params.n_experts, model_params_count, metadata=metadata) + + metadata.size_label = gguf.size_label(*model_params_count, expert_count=params.n_experts or 0) params.ftype = ftype logger.info(f"Writing {outfile}, format {ftype}") diff --git a/examples/deprecation-warning/README.md b/examples/deprecation-warning/README.md new file mode 100644 index 0000000000000..59918ec2bbf72 --- /dev/null +++ b/examples/deprecation-warning/README.md @@ -0,0 +1,49 @@ +# Migration notice for binary filenames + +> [!IMPORTANT] +[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809) + +This migration was important, but it is a breaking change that may not always be immediately obvious to users. + +Please update all scripts and workflows to use the new binary names. + +| Old Filename | New Filename | +| ---- | ---- | +| main | llama-cli | +| server | llama-server | +| llama-bench | llama-bench | +| embedding | llama-embedding | +| quantize | llama-quantize | +| tokenize | llama-tokenize | +| export-lora | llama-export-lora | +| libllava.a | libllava.a | +| baby-llama | llama-baby-llama | +| batched | llama-batched | +| batched-bench | llama-batched-bench | +| benchmark-matmult | llama-benchmark-matmult | +| convert-llama2c-to-ggml | llama-convert-llama2c-to-ggml | +| eval-callback | llama-eval-callback | +| gbnf-validator | llama-gbnf-validator | +| gguf | llama-gguf | +| gguf-split | llama-gguf-split | +| gritlm | llama-gritlm | +| imatrix | llama-imatrix | +| infill | llama-infill | +| llava-cli | llama-llava-cli | +| lookahead | llama-lookahead | +| lookup | llama-lookup | +| lookup-create | llama-lookup-create | +| lookup-merge | llama-lookup-merge | +| lookup-stats | llama-lookup-stats | +| parallel | llama-parallel | +| passkey | llama-passkey | +| perplexity | llama-perplexity | +| q8dot | llama-q8dot | +| quantize-stats | llama-quantize-stats | +| retrieval | llama-retrieval | +| save-load-state | llama-save-load-state | +| simple | llama-simple | +| speculative | llama-speculative | +| vdot | llama-vdot | +| tests/test-c.o | tests/test-c.o | + diff --git a/examples/deprecation-warning/deprecation-warning.cpp b/examples/deprecation-warning/deprecation-warning.cpp new file mode 100644 index 0000000000000..11b35d2c22500 --- /dev/null +++ b/examples/deprecation-warning/deprecation-warning.cpp @@ -0,0 +1,35 @@ +// Warns users that this filename was deprecated, and provides a link for more information. + +#include +#include +#include + +// Main +int main(int argc, char** argv) { + std::string filename = "main"; + if (argc >= 1) { + filename = argv[0]; + } + + // Get only the program name from the full path + auto pos = filename.find_last_of('/'); + if (pos != std::string::npos) { + filename = filename.substr(pos+1); + } + + // Append "llama-" to the beginning of filename to get the replacemnt filename + auto replacement_filename = "llama-" + filename; + + // The exception is if the filename is "main", then our replacement filename is "llama-cli" + if (filename == "main") { + replacement_filename = "llama-cli"; + } + + fprintf(stdout, "\n"); + fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str()); + fprintf(stdout, " Please use '%s' instead.\n", replacement_filename.c_str()); + fprintf(stdout, " See https://github.com/ggerganov/llama.cpp/tree/master/examples/deprecation-warning/README.md for more information.\n"); + fprintf(stdout, "\n"); + + return EXIT_FAILURE; +} diff --git a/examples/eval-callback/eval-callback.cpp b/examples/eval-callback/eval-callback.cpp index 64cd338c26351..c8a3016a4d80a 100644 --- a/examples/eval-callback/eval-callback.cpp +++ b/examples/eval-callback/eval-callback.cpp @@ -99,7 +99,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) { char src1_str[128] = {0}; if (src1) { - sprintf(src1_str, "%s{%s}", src1->name, ggml_ne_string(src1).c_str()); + snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, ggml_ne_string(src1).c_str()); } printf("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, diff --git a/examples/export-lora/README.md b/examples/export-lora/README.md index 1fb17feeca627..91c33c34acaa9 100644 --- a/examples/export-lora/README.md +++ b/examples/export-lora/README.md @@ -6,12 +6,11 @@ Apply LORA adapters to base model and export the resulting model. usage: llama-export-lora [options] options: - -h, --help show this help message and exit - -m FNAME, --model-base FNAME model path from which to load base model (default '') - -o FNAME, --model-out FNAME path to save exported model (default '') - -l FNAME, --lora FNAME apply LoRA adapter - -s FNAME S, --lora-scaled FNAME S apply LoRA adapter with user defined scaling S - -t N, --threads N number of threads to use during computation (default: 4) + -m, --model model path from which to load base model (default '') + --lora FNAME path to LoRA adapter (can be repeated to use multiple adapters) + --lora-scaled FNAME S path to LoRA adapter with user defined scaling S (can be repeated to use multiple adapters) + -t, --threads N number of threads to use during computation (default: 4) + -o, --output FNAME output file (default: 'ggml-lora-merged-f16.gguf') ``` For example: @@ -20,7 +19,15 @@ For example: ./bin/llama-export-lora \ -m open-llama-3b-v2-q8_0.gguf \ -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \ - -l lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.bin + --lora lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.gguf ``` -Multiple LORA adapters can be applied by passing multiple `-l FN` or `-s FN S` command line parameters. +Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters: + +```bash +./bin/llama-export-lora \ + -m your_base_model.gguf \ + -o your_merged_model.gguf \ + --lora-scaled lora_task_A.gguf 0.5 \ + --lora-scaled lora_task_B.gguf 0.5 +``` diff --git a/examples/export-lora/export-lora.cpp b/examples/export-lora/export-lora.cpp index 08413f57e4c3a..150f7e8d58f20 100644 --- a/examples/export-lora/export-lora.cpp +++ b/examples/export-lora/export-lora.cpp @@ -1,462 +1,420 @@ - #include "common.h" #include "ggml.h" #include "ggml-alloc.h" +#include #include #include #include +#include -struct lora_info { - std::string filename; - float scale; -}; - -struct export_lora_params { - std::string fn_model_base; - std::string fn_model_out; - std::vector lora; - int n_threads; -}; +static bool g_verbose = false; -struct lora_data { - struct lora_info info; - std::vector data; - struct ggml_context * ctx; +static std::string get_kv_str(struct gguf_context * ctx_gguf, const std::string & key){ + int id = gguf_find_key(ctx_gguf, key.c_str()); + return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id)); +} - uint32_t lora_r; - uint32_t lora_alpha; -}; +static float get_kv_f32(struct gguf_context * ctx_gguf, const std::string & key) { + int id = gguf_find_key(ctx_gguf, key.c_str()); + return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf, id); +} -struct llama_file { - // use FILE * so we don't have to re-open the file to mmap - FILE * fp; - size_t size; +static void zeros(std::ofstream & file, size_t n) { + char zero = 0; + for (size_t i = 0; i < n; ++i) { + file.write(&zero, 1); + } +} - llama_file(const char * fname, const char * mode) { - fp = std::fopen(fname, mode); - if (fp == NULL) { - size = 0; - } else { - seek(0, SEEK_END); - size = tell(); - seek(0, SEEK_SET); +static std::string ggml_ne_string(const ggml_tensor * t) { + std::string str; + for (int i = 0; i < GGML_MAX_DIMS; ++i) { + str += std::to_string(t->ne[i]); + if (i + 1 < GGML_MAX_DIMS) { + str += ", "; } } + return str; +} - size_t tell() const { -#ifdef _WIN32 - __int64 ret = _ftelli64(fp); -#else - long ret = std::ftell(fp); -#endif - GGML_ASSERT(ret != -1); // this really shouldn't fail - return (size_t) ret; +static struct gguf_context * load_gguf(std::string & fname, struct ggml_context ** ctx_ggml) { + struct gguf_init_params params = { + /*.no_alloc = */ true, + /*.ctx = */ ctx_ggml, + }; + struct gguf_context * ctx_gguf = gguf_init_from_file(fname.c_str(), params); + if (!ctx_gguf) { + throw std::runtime_error("failed to load input GGUF from " + fname); } + return ctx_gguf; +} - 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 - GGML_ASSERT(ret == 0); // same +static void replace_all(std::string & s, const std::string & search, const std::string & replace) { + std::string result; + for (size_t pos = 0; ; pos += search.length()) { + auto new_pos = s.find(search, pos); + if (new_pos == std::string::npos) { + result += s.substr(pos, s.size() - pos); + break; + } + result += s.substr(pos, new_pos - pos) + replace; + pos = new_pos; } + s = std::move(result); +} - 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)) { - die_fmt("read error: %s", strerror(errno)); - } - if (ret != 1) { - die("unexpectedly reached end of file"); +struct file_input { + struct ggml_context * ctx_meta = nullptr; + struct gguf_context * ctx_gguf = nullptr; + std::ifstream f_in; + std::map tensors; + float alpha; + float scale; + + file_input(std::string & fname, float scale): f_in(fname, std::ios::binary), scale(scale) { + if (!f_in.is_open()) { + throw std::runtime_error("failed to open input gguf from " + fname); } - } - std::uint32_t read_u32() { - std::uint32_t ret; - read_raw(&ret, sizeof(ret)); - return ret; + ctx_gguf = load_gguf(fname, &ctx_meta); + alpha = get_kv_f32(ctx_gguf, "adapter.lora.alpha"); + printf("%s: loaded gguf from %s\n", __func__, fname.c_str()); + + for (ggml_tensor * cur = ggml_get_first_tensor(ctx_meta); cur; cur = ggml_get_next_tensor(ctx_meta, cur)) { + std::string name(cur->name); + tensors[name] = cur; + if (g_verbose) { + printf("%s: %s\n", __func__, cur->name); + } + } } - std::string read_string(std::uint32_t len) { - std::vector chars(len); - read_raw(chars.data(), len); - return std::string(chars.data(), len); + ggml_tensor * get_tensor(std::string name) { + if (tensors.find(name) == tensors.end()) { + return nullptr; + } + return tensors[name]; } - void write_raw(const void * ptr, size_t size) { - if (size == 0) { - return; + void read_tensor_data(std::string name, std::vector & buf) { + if (tensors.find(name) == tensors.end()) { + throw std::runtime_error("cannot find tensor with name: " + name); } - errno = 0; - size_t ret = std::fwrite(ptr, size, 1, fp); - if (ret != 1) { - die_fmt("write error: %s", strerror(errno)); + auto len = ggml_nbytes(tensors[name]); + if (buf.size() < len) { + buf.resize(len); } + auto i_tensor_in = gguf_find_tensor(ctx_gguf, name.c_str()); // idx of tensor in the input file + auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor_in); + f_in.seekg(offset); + f_in.read((char* )buf.data(), len); } - void write_u32(std::uint32_t val) { - write_raw(&val, sizeof(val)); + ~file_input() { + gguf_free(ctx_gguf); + ggml_free(ctx_meta); } +}; - bool eof() { - return tell() >= size; - } +struct lora_merge_ctx { + // input base model + adapters + file_input base_model; + std::vector> adapters; - ~llama_file() { - if (fp) { - std::fclose(fp); + // for computing merged tensor + int n_threads; + ggml_backend_t backend = nullptr; + ggml_gallocr_t allocr = nullptr; + std::vector read_buf; + + // output file + struct gguf_context * ctx_out; + struct ggml_context * ctx_out_ggml; + std::ofstream fout; + + lora_merge_ctx( + std::string & base_fname, + std::vector> & lora_files, + std::string & outfile, + int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) { + fout.exceptions(std::ofstream::failbit); // fail fast on write errors + + if (gguf_find_key(base_model.ctx_gguf, LLM_KV_SPLIT_COUNT) >= 0) { + throw std::runtime_error("split model is not yet supported"); } - } -}; -static struct export_lora_params get_default_export_lora_params() { - struct export_lora_params result; - result.fn_model_base = ""; - result.fn_model_out = ""; - result.n_threads = GGML_DEFAULT_N_THREADS; - return result; -} + for (auto lora_inp : lora_files) { + auto fname = std::get<0>(lora_inp); + auto scale = std::get<1>(lora_inp); + std::unique_ptr adapter(new file_input(fname, scale)); + check_metadata_lora(adapter.get()); + adapters.push_back(std::move(adapter)); + } -static void export_lora_print_usage(int /*argc*/, char ** argv, const struct export_lora_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-base FNAME model path from which to load base model (default '%s')\n", params->fn_model_base.c_str()); - fprintf(stderr, " -o FNAME, --model-out FNAME path to save exported model (default '%s')\n", params->fn_model_out.c_str()); - fprintf(stderr, " -l FNAME, --lora FNAME apply LoRA adapter\n"); - fprintf(stderr, " -s FNAME S, --lora-scaled FNAME S apply LoRA adapter with user defined scaling S\n"); - fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params->n_threads); -} + ctx_out = gguf_init_empty(); + struct ggml_init_params params = { + /*.mem_size =*/ gguf_get_n_tensors(base_model.ctx_gguf)*ggml_tensor_overhead(), + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + ctx_out_ggml = ggml_init(params); + backend = ggml_backend_cpu_init(); + allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend)); + } -static bool export_lora_params_parse(int argc, char ** argv, struct export_lora_params * params) { - bool invalid_param = false; - std::string arg; - struct export_lora_params default_params = get_default_export_lora_params(); - const std::string arg_prefix = "--"; + void check_metadata_lora(file_input * adapter) { + auto general_type = get_kv_str(adapter->ctx_gguf, "general.type"); + if (general_type != "adapter") { + throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type); + } - for (int i = 1; i < argc; i++) { - arg = argv[i]; - if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) { - std::replace(arg.begin(), arg.end(), '_', '-'); + auto adapter_type = get_kv_str(adapter->ctx_gguf, "adapter.type"); + if (adapter_type != "lora") { + throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type); } - if (arg == "-m" || arg == "--model-base") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->fn_model_base = argv[i]; - } else if (arg == "-o" || arg == "--model-out") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->fn_model_out = argv[i]; - } else if (arg == "-l" || arg == "--lora") { - if (++i >= argc) { - invalid_param = true; - break; - } - struct lora_info lora; - lora.filename = argv[i]; - lora.scale = 1.0f; - params->lora.push_back(lora); - } else if (arg == "-s" || arg == "--lora-scaled") { - if (++i >= argc) { - invalid_param = true; - break; - } - struct lora_info lora; - lora.filename = argv[i]; - if (++i >= argc) { - invalid_param = true; - break; - } - lora.scale = std::stof(argv[i]); - params->lora.push_back(lora); - } else if (arg == "-t" || arg == "--threads") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_threads = std::stoi(argv[i]); - if (params->n_threads <= 0) { - params->n_threads = std::thread::hardware_concurrency(); - } - } else { - fprintf(stderr, "error: unknown argument: '%s'\n", arg.c_str()); - export_lora_print_usage(argc, argv, &default_params); - exit(1); + auto general_arch_base = get_kv_str(base_model.ctx_gguf, "general.architecture"); + auto general_arch_lora = get_kv_str(adapter->ctx_gguf, "general.architecture"); + if (general_arch_base != general_arch_lora) { + throw std::runtime_error("model arch and LoRA arch mismatch"); } } - if (params->fn_model_base == default_params.fn_model_base) { - fprintf(stderr, "error: please specify a filename for model-base.\n"); - export_lora_print_usage(argc, argv, &default_params); - exit(1); - } - if (params->fn_model_out == default_params.fn_model_out) { - fprintf(stderr, "error: please specify a filename for model-out.\n"); - export_lora_print_usage(argc, argv, &default_params); - exit(1); - } - if (invalid_param) { - fprintf(stderr, "error: invalid parameter for argument: '%s'\n", arg.c_str()); - export_lora_print_usage(argc, argv, &default_params); - exit(1); + ggml_type get_out_tensor_type(struct ggml_tensor * t) { + if (t->type == GGML_TYPE_F32) { + return GGML_TYPE_F32; + } else { + return GGML_TYPE_F16; + } } - return true; -} -static void free_lora(struct lora_data * lora) { - if (lora->ctx != NULL) { - ggml_free(lora->ctx); - } - delete lora; -} + void run_merge() { + // prepare metadata + gguf_set_kv(ctx_out, base_model.ctx_gguf); + // output is forced to f16 for now + gguf_set_val_u32(ctx_out, "general.file_type", LLAMA_FTYPE_MOSTLY_F16); + + // check if all lora adapters have the same tensors + // TODO: remove this when we can support merging subset of adapters. Ref: https://github.com/ggerganov/llama.cpp/pull/8607#discussion_r1686027777 + static const char * err_no_subset_adapter = "Input adapters do not have the same list of tensors. This is not yet supported. Please merge the adapter one-by-one instead of merging all at once."; + if (adapters.size() > 1) { + for (size_t i = 1; i < adapters.size(); ++i) { + if (adapters[0]->tensors.size() != adapters[i]->tensors.size()) { + throw std::runtime_error(err_no_subset_adapter); + } + for (auto & it : adapters[i]->tensors) { + if (adapters[0]->get_tensor(it.first) == nullptr) { + throw std::runtime_error(err_no_subset_adapter); + } + } + } + } -static struct lora_data * load_lora(struct lora_info * info) { - struct lora_data * result = new struct lora_data; - result->info = *info; - result->ctx = NULL; - result->lora_r = 1; - result->lora_alpha = 1; - - struct llama_file file(info->filename.c_str(), "rb"); - if (file.fp == NULL) { - fprintf(stderr, "warning: Could not open lora adapter '%s'. Ignoring this adapter.\n", - info->filename.c_str()); - free_lora(result); - return NULL; - } + // mapping base tensor to out tensor (same shape with base, but different type) + // if out_tensor == nullptr, we only copy it + std::vector> base_to_out_tensors; + for (auto & it : base_model.tensors) { + bool t_a = true; + bool t_b = true; + for (auto & adapter : adapters) { + t_a &= nullptr != adapter->get_tensor(it.first + ".lora_a"); + t_b &= nullptr != adapter->get_tensor(it.first + ".lora_b"); + } + auto base_tensor = it.second; + if (!t_a && !t_b) { + // only copy + struct ggml_tensor * cpy_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor); + ggml_set_name(cpy_tensor, base_tensor->name); + base_to_out_tensors.push_back(std::make_pair(cpy_tensor, nullptr)); + gguf_add_tensor(ctx_out, cpy_tensor); + } else if (t_a && t_b) { + // need merging + struct ggml_tensor * out_tensor = ggml_new_tensor( + ctx_out_ggml, get_out_tensor_type(base_tensor), GGML_MAX_DIMS, base_tensor->ne); + ggml_set_name(out_tensor, base_tensor->name); + base_to_out_tensors.push_back(std::make_pair(base_tensor, out_tensor)); + gguf_add_tensor(ctx_out, out_tensor); + } else { + throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b"); + } + } - struct ggml_init_params params_ggml; - params_ggml.mem_size = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE; - params_ggml.mem_buffer = NULL; - params_ggml.no_alloc = true; - result->ctx = ggml_init(params_ggml); + // placeholder for the meta data + { + size_t meta_size = gguf_get_meta_size(ctx_out); + zeros(fout, meta_size); + } - uint32_t magic = file.read_u32(); - if (magic != LLAMA_FILE_MAGIC_GGLA) { - die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str()); - } - uint32_t version = file.read_u32(); - if (version != 1) { - die_fmt("unexpected lora file version '%u' in '%s'", (unsigned) version, info->filename.c_str()); - } - result->lora_r = file.read_u32(); - result->lora_alpha = file.read_u32(); - // read tensor infos from file - std::vector name_buf; - std::vector tensors; - std::vector tensors_offset; - size_t total_nbytes_pad = 0; - while(!file.eof()) { - int64_t ne[4] = {1,1,1,1}; - uint32_t n_dims = file.read_u32(); - uint32_t namelen = file.read_u32(); - uint32_t type = file.read_u32(); - for (uint32_t k = 0; k < n_dims; ++k) { - ne[k] = (int64_t)file.read_u32(); + // process base model tensors + size_t n_merged = 0; + for (auto & it : base_to_out_tensors) { + if (it.second != nullptr) { + merge_tensor(it.first, it.second); + n_merged++; + } else { + copy_tensor(it.first); + } } - name_buf.clear(); - name_buf.resize(namelen + 1, '\0'); - file.read_raw(name_buf.data(), namelen); - file.seek((0-file.tell()) & 31, SEEK_CUR); - size_t offset = file.tell(); - struct ggml_tensor * tensor = ggml_new_tensor(result->ctx, (enum ggml_type) type, n_dims, ne); - ggml_set_name(tensor, name_buf.data()); - size_t nbytes = ggml_nbytes(tensor); - size_t nbytes_pad = ggml_nbytes_pad(tensor); - total_nbytes_pad += nbytes_pad; - tensors.push_back(tensor); - tensors_offset.push_back(offset); - file.seek(nbytes, SEEK_CUR); - } - // read tensor data - result->data.resize(total_nbytes_pad); - size_t data_offset = 0; - for (size_t i = 0; i < tensors.size(); ++i) { - struct ggml_tensor * tensor = tensors[i]; - size_t offset = tensors_offset[i]; - size_t nbytes = ggml_nbytes(tensor); - size_t nbytes_pad = ggml_nbytes_pad(tensor); - file.seek(offset, SEEK_SET); - tensor->data = result->data.data() + data_offset; - file.read_raw(tensor->data, nbytes); - data_offset += nbytes_pad; - } - return result; -} + // write output metadata + { + std::vector data(gguf_get_meta_size(ctx_out)); + gguf_get_meta_data(ctx_out, data.data()); + fout.seekp(0); + fout.write((const char *)data.data(), data.size()); + } -static struct ggml_cgraph * build_graph_lora( - struct ggml_context * ctx, - struct ggml_tensor * tensor, - struct ggml_tensor * lora_a, - struct ggml_tensor * lora_b, - float scaling -) { - struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b); - if (scaling != 1.0f) { - ab = ggml_scale(ctx, ab, scaling); + printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged); + printf("%s : wrote %ld tensors to output file\n", __func__, base_to_out_tensors.size()); } - struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab); - - struct ggml_cgraph * gf = ggml_new_graph(ctx); - ggml_build_forward_expand (gf, res); - return gf; -} -static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int n_threads) { - if (lora->ctx == NULL) { - return false; - } - std::string name = ggml_get_name(tensor); - std::string name_a = name + std::string(".loraA"); - std::string name_b = name + std::string(".loraB"); - struct ggml_tensor * lora_a = ggml_get_tensor(lora->ctx, name_a.c_str()); - struct ggml_tensor * lora_b = ggml_get_tensor(lora->ctx, name_b.c_str()); - if (lora_a == NULL || lora_b == NULL) { - return false; + void copy_tensor(struct ggml_tensor * base) { + printf("%s : %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str()); + size_t len = ggml_nbytes(base); + base_model.read_tensor_data(base->name, read_buf); + fout.write((char* )read_buf.data(), len); + zeros(fout, GGML_PAD(len, GGUF_DEFAULT_ALIGNMENT) - len); } - float scaling = lora->info.scale * (float)lora->lora_alpha / (float)lora->lora_r; - - struct ggml_init_params params; - params.mem_size = GGML_OBJECT_SIZE + ggml_graph_overhead() + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5; - params.mem_buffer = NULL; - params.no_alloc = true; - struct ggml_context * ctx = NULL; - struct ggml_gallocr * alloc = NULL; - struct ggml_cgraph * gf = NULL; - - ctx = ggml_init(params); - alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type()); - gf = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling); - - ggml_gallocr_alloc_graph(alloc, gf); - - struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads); - static std::vector data_work; - data_work.resize(cplan.work_size); - cplan.work_data = data_work.data(); + void merge_tensor(struct ggml_tensor * base, struct ggml_tensor * out) { + std::string name_base(base->name); + std::string name_lora_a = name_base + ".lora_a"; + std::string name_lora_b = name_base + ".lora_b"; + + printf("%s : %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str()); + + // context for input tensor + std::vector inp_a(adapters.size()); + std::vector inp_b(adapters.size()); + struct ggml_init_params params { + /*.mem_size =*/ ggml_tensor_overhead()*(2+adapters.size()*2), + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + struct ggml_context * ctx = ggml_init(params); + + // alloc tensors + struct ggml_tensor * inp_base = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, base->ne); + for (size_t i = 0; i < adapters.size(); ++i) { + auto t_a = adapters[i]->get_tensor(name_lora_a); + auto t_b = adapters[i]->get_tensor(name_lora_b); + inp_a[i] = ggml_dup_tensor(ctx, t_a); + inp_b[i] = ggml_dup_tensor(ctx, t_b); + } + ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx, backend); + + // load base tensor to backend buffer + base_model.read_tensor_data(name_base, read_buf); + if (base->type != GGML_TYPE_F32) { + // optionally dequantize it + printf("%s : + dequantize base tensor from %s to F32\n", __func__, ggml_type_name(base->type)); + auto nels = ggml_nelements(inp_base); + ggml_type_traits_t qtype = ggml_internal_get_type_traits(base->type); + std::vector dequant_buf(nels * sizeof(float)); + qtype.to_float(read_buf.data(), (float *)dequant_buf.data(), nels); + ggml_backend_tensor_set(inp_base, dequant_buf.data(), 0, dequant_buf.size()); + } else { + ggml_backend_tensor_set(inp_base, read_buf.data(), 0, ggml_nbytes(inp_base)); + } - ggml_graph_compute(gf, &cplan); + // load lora tensors to backend buffer + for (size_t i = 0; i < adapters.size(); ++i) { + adapters[i]->read_tensor_data(name_lora_a, read_buf); + ggml_backend_tensor_set(inp_a[i], read_buf.data(), 0, ggml_nbytes(inp_a[i])); + adapters[i]->read_tensor_data(name_lora_b, read_buf); + ggml_backend_tensor_set(inp_b[i], read_buf.data(), 0, ggml_nbytes(inp_b[i])); + } - ggml_gallocr_free(alloc); - ggml_free(ctx); - return true; -} + // build graph + struct ggml_cgraph * gf; + { + static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead(); + static std::vector buf(buf_size); + struct ggml_init_params params0 = { + /*.mem_size =*/ buf_size, + /*.mem_buffer =*/ buf.data(), + /*.no_alloc =*/ true, + }; + struct ggml_context * ctx0 = ggml_init(params0); + gf = ggml_new_graph(ctx0); + struct ggml_tensor * cur = inp_base; + for (size_t i = 0; i < adapters.size(); ++i) { + struct ggml_tensor * a_T = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32))); + struct ggml_tensor * delta = ggml_mul_mat(ctx0, a_T, ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32)); + // scale + const float alpha = adapters[i]->alpha; + const float rank = (float) inp_b[i]->ne[0]; + const float scale = alpha ? adapters[i]->scale * alpha / rank : adapters[i]->scale; + delta = ggml_scale(ctx0, delta, scale); + cur = ggml_add(ctx0, delta, cur); + printf("%s : + merging from adapter[%ld] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type)); + printf("%s : input_scale=%f calculated_scale=%f rank=%d\n", __func__, adapters[i]->scale, scale, (int) inp_b[i]->ne[0]); + } + cur = ggml_cast(ctx0, cur, out->type); + printf("%s : + output type is %s\n", __func__, ggml_type_name(out->type)); + ggml_build_forward_expand(gf, cur); + ggml_free(ctx0); + } -static void export_lora(struct export_lora_params * params) { - // load all loras - std::vector loras; - for (size_t i = 0; i < params->lora.size(); ++i) { - struct lora_data * lora = load_lora(¶ms->lora[i]); - if (lora != NULL) { - loras.push_back(lora); + // compute + { + ggml_gallocr_alloc_graph(allocr, gf); + ggml_backend_cpu_set_n_threads(backend, n_threads); + ggml_backend_graph_compute(backend, gf); } - } - if (loras.size() == 0) { - fprintf(stderr, "warning: no lora adapters will be applied.\n"); - } - // open input file - struct llama_file fin(params->fn_model_base.c_str(), "rb"); - if (!fin.fp) { - die_fmt("Could not open file '%s'\n", params->fn_model_base.c_str()); - } + // write data to output file + { + auto result = gf->nodes[gf->n_nodes - 1]; + size_t len = ggml_nbytes(result); + if (read_buf.size() < len) { + read_buf.resize(len); + } + ggml_backend_tensor_get(result, read_buf.data(), 0, len); + fout.write((char* )read_buf.data(), len); + zeros(fout, GGML_PAD(len, GGUF_DEFAULT_ALIGNMENT) - len); + } - // open base model gguf, read tensors without their data - struct ggml_context * ctx_in; - struct gguf_init_params params_gguf; - params_gguf.no_alloc = true; - params_gguf.ctx = &ctx_in; - struct gguf_context * gguf_in = gguf_init_from_file(params->fn_model_base.c_str(), params_gguf); - - // create new gguf - struct gguf_context * gguf_out = gguf_init_empty(); - - // copy meta data from base model: kv and tensors - gguf_set_kv(gguf_out, gguf_in); - int n_tensors = gguf_get_n_tensors(gguf_in); - for (int i=0; i < n_tensors; ++i) { - const char * name = gguf_get_tensor_name(gguf_in, i); - struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name); - gguf_add_tensor(gguf_out, tensor); + ggml_free(ctx); + ggml_backend_buffer_free(buffer); } - // create output file - struct llama_file fout(params->fn_model_out.c_str(), "wb"); - if (!fout.fp) { - die_fmt("Could not create file '%s'\n", params->fn_model_out.c_str()); + ~lora_merge_ctx() { + ggml_gallocr_free(allocr); + ggml_backend_free(backend); + gguf_free(ctx_out); + ggml_free(ctx_out_ggml); } +}; - // write gguf meta data - std::vector meta; - meta.resize(gguf_get_meta_size(gguf_out)); - gguf_get_meta_data(gguf_out, meta.data()); - fout.write_raw(meta.data(), meta.size()); - - std::vector data; - std::vector padding; - for (int i=0; i < n_tensors; ++i) { - const char * name = gguf_get_tensor_name(gguf_in, i); - struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name); - - // read tensor data - data.resize(ggml_nbytes(tensor)); - tensor->data = data.data(); - size_t offset = gguf_get_tensor_offset(gguf_in, i); - fin.seek(offset + meta.size(), SEEK_SET); - fin.read_raw(data.data(), data.size()); - - // apply all loras - for (size_t k = 0; k < loras.size(); ++k) { - apply_lora(tensor, loras[k], params->n_threads); - } - - // write tensor data + padding - padding.clear(); - padding.resize(GGML_PAD(data.size(), gguf_get_alignment(gguf_out)) - data.size(), 0); - - GGML_ASSERT(fout.tell() == offset + meta.size()); - // fout.seek(offset + meta.size(), SEEK_SET); - fout.write_raw(data.data(), data.size()); - fout.write_raw(padding.data(), padding.size()); +static void print_usage(int argc, char ** argv, const gpt_params & params) { + gpt_params_print_usage(argc, argv, params); - if (i % 2 == 0) { - printf("."); - } - } + printf("\nexample usage:\n"); + printf("\n %s -m base-model.gguf --lora lora-file.gguf -o merged-model-f16.gguf\n", argv[0]); + printf("\nNOTE: output model is F16\n"); printf("\n"); - - // close gguf - gguf_free(gguf_out); - gguf_free(gguf_in); - - // free loras - for (size_t i = 0; i < loras.size(); ++i) { - free_lora(loras[i]); - } } int main(int argc, char ** argv) { - struct export_lora_params params = get_default_export_lora_params(); + gpt_params params; - if (!export_lora_params_parse(argc, argv, ¶ms)) { + if (!gpt_params_parse(argc, argv, params)) { + print_usage(argc, argv, params); return 1; } - export_lora(¶ms); + g_verbose = (params.verbosity == 1); + try { + lora_merge_ctx ctx(params.model, params.lora_adapter, params.lora_outfile, params.n_threads); + ctx.run_merge(); + } catch (const std::exception & err) { + fprintf(stderr, "%s\n", err.what()); + exit(EXIT_FAILURE); + } + + printf("done, output file is %s\n", params.lora_outfile.c_str()); return 0; } diff --git a/examples/finetune/CMakeLists.txt b/examples/finetune/CMakeLists.txt deleted file mode 100644 index 64afe6ddc647a..0000000000000 --- a/examples/finetune/CMakeLists.txt +++ /dev/null @@ -1,5 +0,0 @@ -set(TARGET llama-finetune) -add_executable(${TARGET} finetune.cpp) -install(TARGETS ${TARGET} RUNTIME) -target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) -target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/finetune/README.md b/examples/finetune/README.md deleted file mode 100644 index 1c27df053bc04..0000000000000 --- a/examples/finetune/README.md +++ /dev/null @@ -1,90 +0,0 @@ -# finetune - -Basic usage instructions: - -```bash -# get training data -wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/shakespeare.txt - -# finetune LORA adapter -./bin/llama-finetune \ - --model-base open-llama-3b-v2-q8_0.gguf \ - --checkpoint-in chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf \ - --checkpoint-out chk-lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.gguf \ - --lora-out lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.bin \ - --train-data "shakespeare.txt" \ - --save-every 10 \ - --threads 6 --adam-iter 30 --batch 4 --ctx 64 \ - --use-checkpointing - -# predict -./bin/llama-cli -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin -``` - -**Only llama based models are supported!** The output files will be saved every N iterations (config with `--save-every N`). -The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output. -So in above example after 10 iterations these files will be written: -- chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf -- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf -- lora-open-llama-3b-v2-q8_0-shakespeare-10.bin -- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin - -After 10 more iterations: -- chk-lora-open-llama-3b-v2-q8_0-shakespeare-20.gguf -- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf -- lora-open-llama-3b-v2-q8_0-shakespeare-20.bin -- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin - -Checkpoint files (`--checkpoint-in FN`, `--checkpoint-out FN`) store the training process. When the input checkpoint file does not exist, it will begin finetuning a new randomly initialized adapter. - -llama.cpp compatible LORA adapters will be saved with filename specified by `--lora-out FN`. -These LORA adapters can then be used by `llama-cli` together with the base model, like in the 'predict' example command above. - -In `llama-cli` you can also load multiple LORA adapters, which will then be mixed together. - -For example if you have two LORA adapters `lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin` and `lora-open-llama-3b-v2-q8_0-bible-LATEST.bin`, you can mix them together like this: - -```bash -./bin/llama-cli -m open-llama-3b-v2-q8_0.gguf \ - --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin \ - --lora lora-open-llama-3b-v2-q8_0-bible-LATEST.bin -``` - -You can change how strong each LORA adapter is applied to the base model by using `--lora-scaled FN SCALE` instead of `--lora FN`. - -For example to apply 40% of the 'shakespeare' LORA adapter, 80% of the 'bible' LORA adapter and 100% of yet another one: - -```bash -./bin/llama-cli -m open-llama-3b-v2-q8_0.gguf \ - --lora-scaled lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin 0.4 \ - --lora-scaled lora-open-llama-3b-v2-q8_0-bible-LATEST.bin 0.8 \ - --lora lora-open-llama-3b-v2-q8_0-yet-another-one-LATEST.bin -``` - -The scale numbers don't need to add up to one, and you can also use numbers greater than 1 to further increase the influence of an adapter. But making the values too big will sometimes result in worse output. Play around to find good values. - -Gradient checkpointing reduces the memory requirements by ~50% but increases the runtime. -If you have enough RAM, you can make finetuning a bit faster by disabling checkpointing with `--no-checkpointing`. - -The default LORA rank can be specified with `--lora-r N`. -The LORA rank can be configured for each model tensor type separately with these command line options: - -```bash - --lora-r N LORA r: default rank. Also specifies resulting scaling together with lora-alpha. (default 4) - --rank-att-norm N LORA rank for attention norm tensor (default 1) - --rank-ffn-norm N LORA rank for feed-forward norm tensor (default 1) - --rank-out-norm N LORA rank for output norm tensor (default 1) - --rank-tok-embd N LORA rank for token embeddings tensor (default 4) - --rank-out N LORA rank for output tensor (default 4) - --rank-wq N LORA rank for wq tensor (default 4) - --rank-wk N LORA rank for wk tensor (default 4) - --rank-wv N LORA rank for wv tensor (default 4) - --rank-wo N LORA rank for wo tensor (default 4) - --rank-ffn_gate N LORA rank for ffn_gate tensor (default 4) - --rank-ffn_down N LORA rank for ffn_down tensor (default 4) - --rank-ffn_up N LORA rank for ffn_up tensor (default 4) -``` - -The LORA rank of 'norm' tensors should always be 1. - -To see all available options use `llama-finetune --help`. diff --git a/examples/finetune/convert_finetune_checkpoint_to_gguf.py b/examples/finetune/convert_finetune_checkpoint_to_gguf.py deleted file mode 100644 index 1b79d6995d92a..0000000000000 --- a/examples/finetune/convert_finetune_checkpoint_to_gguf.py +++ /dev/null @@ -1,487 +0,0 @@ -#!/usr/bin/env python3 -# finetune checkpoint --> gguf conversion - -import argparse -import gguf -import struct -import numpy as np -from pathlib import Path - -# gguf constants -LLM_KV_OPTIMIZER_TYPE = "optimizer.type" -LLM_KV_OPTIMIZER_TYPE_ADAM = "adam" -LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs" -LLM_KV_OPTIMIZER_FILE_VERSION = "optimizer.file_version" -LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT = "optimizer.convergence_past_count" -LLM_KV_OPTIMIZER_PARAMETER_COUNT = "optimizer.parameter_count" -LLM_KV_OPTIMIZER_ITERATION_COUNT = "optimizer.iteration_count" -LLM_KV_OPTIMIZER_JUST_INITIALIZED = "optimizer.just_initialized" -LLM_KV_OPTIMIZER_ADAM_BEST_LOSS = "optimizer.adam.best_loss" -LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS = "optimizer.adam.previous_loss" -LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT = "optimizer.adam.no_improvement_count" -LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count" -LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS = "optimizer.lbfgs.best_loss" -LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP = "optimizer.lbfgs.line_search_step" -LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J = "optimizer.lbfgs.line_search_j" -LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K = "optimizer.lbfgs.line_search_k" -LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END = "optimizer.lbfgs.line_search_end" -LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count" - -LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS = "optimizer.adam.first_moments" -LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS = "optimizer.adam.second_moments" -LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values" - -LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS = "optimizer.lbfgs.current_parameters" -LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters" -LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS = "optimizer.lbfgs.current_gradients" -LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS = "optimizer.lbfgs.previous_gradients" -LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION = "optimizer.lbfgs.search_direction" -LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES = "optimizer.lbfgs.past_loss_values" -LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA = "optimizer.lbfgs.memory_alpha" -LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS = "optimizer.lbfgs.memory_ys" -LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S = "optimizer.lbfgs.memory_s" -LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y = "optimizer.lbfgs.memory_y" - -LLM_KV_TRAINING_TYPE_TRAIN_MODEL = "train_model" -LLM_KV_TRAINING_TYPE_FINETUNE_LORA = "finetune_lora" -LLM_KV_TRAINING_TYPE = "training.type" -LLM_KV_TRAINING_FILE_VERSION = "training.file_version" -LLM_KV_TRAINING_ITERATION_COUNT = "training.iteration_count" -LLM_KV_TRAINING_SAMPLE_COUNT = "training.sample_count" -LLM_KV_TRAINING_TOKEN_COUNT = "training.token_count" - -LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD = "training.lora.rank.token_embd" -LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM = "training.lora.rank.output_norm" -LLM_KV_TRAINING_LORA_RANK_OUTPUT = "training.lora.rank.output" -LLM_KV_TRAINING_LORA_RANK_ATTN_NORM = "training.lora.rank.attn_norm" -LLM_KV_TRAINING_LORA_RANK_ATTN_Q = "training.lora.rank.attn_q" -LLM_KV_TRAINING_LORA_RANK_ATTN_K = "training.lora.rank.attn_k" -LLM_KV_TRAINING_LORA_RANK_ATTN_V = "training.lora.rank.attn_v" -LLM_KV_TRAINING_LORA_RANK_ATTN_OUT = "training.lora.rank.attn_output" -LLM_KV_TRAINING_LORA_RANK_FFN_NORM = "training.lora.rank.ffn_norm" -LLM_KV_TRAINING_LORA_RANK_FFN_GATE = "training.lora.rank.ffn_gate" -LLM_KV_TRAINING_LORA_RANK_FFN_DOWN = "training.lora.rank.ffn_down" -LLM_KV_TRAINING_LORA_RANK_FFN_UP = "training.lora.rank.ffn_up" - -class Tensor: - def __init__(self, dtype='f', ne=None): - if ne is None: - ne = [] - self.dtype = dtype - self.ne = ne - self.nbytes = 0 - if self.dtype == 'f': - if len(self.ne) == 0: - self.nbytes = 0 - else: - self.nbytes = int(np.prod(self.ne)) * 4 - else: - raise ValueError(f"Unhandled data type '{self.dtype}'") - - def load(self, data, offset): - nd = struct.unpack(' 0 else []) - - self.lbfgs_x = Tensor('f', [self.nx]) - self.lbfgs_xp = Tensor('f', [self.nx]) - self.lbfgs_g = Tensor('f', [self.nx]) - self.lbfgs_gp = Tensor('f', [self.nx]) - self.lbfgs_d = Tensor('f', [self.nx]) - self.lbfgs_pf = Tensor('f', [self.past] if self.past > 0 else []) - self.lbfgs_lmal = Tensor('f', [self.lbfgs_m]) - self.lbfgs_lmys = Tensor('f', [self.lbfgs_m]) - self.lbfgs_lms = Tensor('f', [self.nx, self.lbfgs_m]) - self.lbfgs_lmy = Tensor('f', [self.nx, self.lbfgs_m]) - - # forgot to save type in version 1: - # guess self.type from number of remaining bytes - size_type_0 = 12 + sum([t.max_storage_size() for t in - [self.adam_m, self.adam_v] - +([self.adam_pf] if (self.past > 0) else [])]) - size_type_1 = 24 + sum([t.max_storage_size() for t in - [self.lbfgs_x, self.lbfgs_xp, self.lbfgs_g, - self.lbfgs_gp, self.lbfgs_d, self.lbfgs_pf, - self.lbfgs_lmal, self.lbfgs_lmys, - self.lbfgs_lms, self.lbfgs_lmy] - +([self.lbfgs_pf] if (self.past > 0) else [])]) - # due to alignment padding the size might not by exact - # but the difference in size for both types is significant, - # so we can just use whichever is closest - remaining = len(data) - offset - if abs(remaining - size_type_0) < abs(remaining - size_type_1): - self.type = 0 - else: - self.type = 1 - - if self.type == 0: - offset = self.adam_m.load(data, offset) - offset = self.adam_v.load(data, offset) - offset = self.adam_pf.load(data,offset) - - self.adam_fx_best = struct.unpack(' 0: - self.adam_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES) - - elif self.type == 1: - gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS) - gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, self.lbfgs_m) - gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS, self.lbfgs_fx_best) - gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP, self.lbfgs_step) - gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J, self.lbfgs_j) - gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K, self.lbfgs_k) - gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END, self.lbfgs_end) - gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, self.lbfgs_n_no_improvement) - - self.lbfgs_x.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS) - self.lbfgs_xp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS) - self.lbfgs_g.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS) - self.lbfgs_gp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS) - self.lbfgs_d.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION) - if self.past > 0: - self.lbfgs_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES) - self.lbfgs_lmal.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA) - self.lbfgs_lmys.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS) - self.lbfgs_lms.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S) - self.lbfgs_lmy.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y) - else: - raise ValueError('Unknown optimizer type') - -class LoraParams: - def __init__(self): - pass - - def load(self, data, offset): - self.n_rank_attention_norm = struct.unpack(' -#include -#include -#include -#include - -#if defined(_MSC_VER) -#pragma warning(disable: 4244 4267) // possible loss of data -#endif - -struct my_llama_hparams { - uint32_t n_vocab = 32000; - uint32_t n_ctx = 512; - uint32_t n_embd = 4096; - uint32_t n_ff = 11008; - uint32_t n_head = 32; - uint32_t n_head_kv = 32; - uint32_t n_layer = 32; - - // float f_norm_eps = 1e-5f; // falcon - float f_norm_rms_eps = 1e-5f; // llama - - float rope_freq_base = 10000.0f; - float rope_freq_scale = 1.0f; - - uint32_t n_gqa() const { - return n_head/n_head_kv; - } - - uint32_t n_embd_head() const { - return n_embd/n_head; - } - - uint32_t n_embd_gqa() const { - return n_embd/n_gqa(); - } - - bool operator!=(const my_llama_hparams& other) const { - return memcmp(this, &other, sizeof(other)); - } -}; - -struct my_llama_layer { - // normalization - struct ggml_tensor * attention_norm; - - // attention - struct ggml_tensor * wq; - struct ggml_tensor * wk; - struct ggml_tensor * wv; - struct ggml_tensor * wo; - - // normalization - struct ggml_tensor * ffn_norm; - - // ff - struct ggml_tensor * ffn_gate; // w1 - struct ggml_tensor * ffn_down; // w2 - struct ggml_tensor * ffn_up; // w3 -}; - -struct my_llama_model { - struct my_llama_hparams hparams; - - struct ggml_tensor * tok_embeddings; - - struct ggml_tensor * norm; - struct ggml_tensor * output; - - std::vector layers; -}; - -struct my_llama_lora_hparams { - uint32_t lora_r = 1; - uint32_t lora_alpha = 1; - uint32_t n_rank_attention_norm = 1; - uint32_t n_rank_wq = 4; - uint32_t n_rank_wk = 4; - uint32_t n_rank_wv = 4; - uint32_t n_rank_wo = 4; - uint32_t n_rank_ffn_norm = 1; - uint32_t n_rank_ffn_gate = 4; - uint32_t n_rank_ffn_down = 4; - uint32_t n_rank_ffn_up = 4; - uint32_t n_rank_tok_embeddings = 4; - uint32_t n_rank_norm = 1; - uint32_t n_rank_output = 4; - - bool operator!=(const my_llama_lora_hparams& other) const { - return memcmp(this, &other, sizeof(other)); - } -}; - -struct my_llama_lora_layer { - // normalization - struct ggml_tensor * attention_norm_a; - struct ggml_tensor * attention_norm_b; - - // attention - struct ggml_tensor * wq_a; - struct ggml_tensor * wq_b; - struct ggml_tensor * wk_a; - struct ggml_tensor * wk_b; - struct ggml_tensor * wv_a; - struct ggml_tensor * wv_b; - struct ggml_tensor * wo_a; - struct ggml_tensor * wo_b; - - // normalization - struct ggml_tensor * ffn_norm_a; - struct ggml_tensor * ffn_norm_b; - - // ff - struct ggml_tensor * ffn_gate_a; - struct ggml_tensor * ffn_gate_b; - struct ggml_tensor * ffn_down_a; - struct ggml_tensor * ffn_down_b; - struct ggml_tensor * ffn_up_a; - struct ggml_tensor * ffn_up_b; -}; - -struct my_llama_lora { - struct ggml_context * ctx = NULL; - ggml_backend_buffer_t data; - - my_llama_lora_hparams hparams; - - struct ggml_tensor * tok_embeddings_a; - struct ggml_tensor * tok_embeddings_b; - - struct ggml_tensor * norm_a; - struct ggml_tensor * norm_b; - struct ggml_tensor * output_a; - struct ggml_tensor * output_b; - - std::vector layers; -}; - -// gguf constants -static const char * LLM_KV_TRAINING_TYPE_FINETUNE_LORA = "finetune_lora"; -static const char * LLM_KV_TRAINING_TYPE = "training.type"; - -static const char * LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD = "training.lora.rank.token_embd"; -static const char * LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM = "training.lora.rank.output_norm"; -static const char * LLM_KV_TRAINING_LORA_RANK_OUTPUT = "training.lora.rank.output"; -static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_NORM = "training.lora.rank.attn_norm"; -static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_Q = "training.lora.rank.attn_q"; -static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_K = "training.lora.rank.attn_k"; -static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_V = "training.lora.rank.attn_v"; -static const char * LLM_KV_TRAINING_LORA_RANK_ATTN_OUT = "training.lora.rank.attn_output"; -static const char * LLM_KV_TRAINING_LORA_RANK_FFN_NORM = "training.lora.rank.ffn_norm"; -static const char * LLM_KV_TRAINING_LORA_RANK_FFN_GATE = "training.lora.rank.ffn_gate"; -static const char * LLM_KV_TRAINING_LORA_RANK_FFN_DOWN = "training.lora.rank.ffn_down"; -static const char * LLM_KV_TRAINING_LORA_RANK_FFN_UP = "training.lora.rank.ffn_up"; - -// gguf constants (sync with gguf.py) - -static const char * LLM_KV_GENERAL_ARCHITECTURE = "general.architecture"; -static const char * LLM_KV_GENERAL_FILE_TYPE = "general.file_type"; - -static const char * LLM_KV_CONTEXT_LENGTH = "%s.context_length"; -static const char * LLM_KV_EMBEDDING_LENGTH = "%s.embedding_length"; -static const char * LLM_KV_BLOCK_COUNT = "%s.block_count"; -static const char * LLM_KV_FEED_FORWARD_LENGTH = "%s.feed_forward_length"; -static const char * LLM_KV_ATTENTION_HEAD_COUNT = "%s.attention.head_count"; -static const char * LLM_KV_ATTENTION_HEAD_COUNT_KV = "%s.attention.head_count_kv"; -static const char * LLM_KV_ATTENTION_LAYERNORM_RMS_EPS = "%s.attention.layer_norm_rms_epsilon"; -static const char * LLM_KV_ROPE_DIMENSION_COUNT = "%s.rope.dimension_count"; -static const char * LLM_KV_ROPE_FREQ_BASE = "%s.rope.freq_base"; // TODO load in llama.cpp -static const char * LLM_KV_ROPE_SCALE_LINEAR = "%s.rope.scale_linear"; - -static const char * LLM_TENSOR_TOKEN_EMBD = "token_embd"; -static const char * LLM_TENSOR_OUTPUT_NORM = "output_norm"; -static const char * LLM_TENSOR_OUTPUT = "output"; -static const char * LLM_TENSOR_ATTN_NORM = "blk.%d.attn_norm"; -static const char * LLM_TENSOR_ATTN_Q = "blk.%d.attn_q"; -static const char * LLM_TENSOR_ATTN_K = "blk.%d.attn_k"; -static const char * LLM_TENSOR_ATTN_V = "blk.%d.attn_v"; -static const char * LLM_TENSOR_ATTN_OUT = "blk.%d.attn_output"; -static const char * LLM_TENSOR_FFN_NORM = "blk.%d.ffn_norm"; -static const char * LLM_TENSOR_FFN_GATE = "blk.%d.ffn_gate"; -static const char * LLM_TENSOR_FFN_DOWN = "blk.%d.ffn_down"; -static const char * LLM_TENSOR_FFN_UP = "blk.%d.ffn_up"; - -static void print_params(struct my_llama_hparams * params) { - printf("%s: n_vocab : %u\n", __func__, params->n_vocab); - printf("%s: n_ctx : %u\n", __func__, params->n_ctx); - printf("%s: n_embd : %u\n", __func__, params->n_embd); - printf("%s: n_ff : %u\n", __func__, params->n_ff); - printf("%s: n_head : %u\n", __func__, params->n_head); - printf("%s: n_head_kv : %u\n", __func__, params->n_head_kv); - printf("%s: n_layer : %u\n", __func__, params->n_layer); - printf("%s: norm_rms_eps : %f\n", __func__, params->f_norm_rms_eps); - printf("%s: rope_freq_base : %f\n", __func__, params->rope_freq_base); - printf("%s: rope_freq_scale : %f\n", __func__, params->rope_freq_scale); -} - -static void print_lora_params(struct my_llama_lora_hparams * params) { - printf("%s: n_rank_attention_norm : %u\n", __func__, params->n_rank_attention_norm); - printf("%s: n_rank_wq : %u\n", __func__, params->n_rank_wq); - printf("%s: n_rank_wk : %u\n", __func__, params->n_rank_wk); - printf("%s: n_rank_wv : %u\n", __func__, params->n_rank_wv); - printf("%s: n_rank_wo : %u\n", __func__, params->n_rank_wo); - printf("%s: n_rank_ffn_norm : %u\n", __func__, params->n_rank_ffn_norm); - printf("%s: n_rank_ffn_gate : %u\n", __func__, params->n_rank_ffn_gate); - printf("%s: n_rank_ffn_down : %u\n", __func__, params->n_rank_ffn_down); - printf("%s: n_rank_ffn_up : %u\n", __func__, params->n_rank_ffn_up); - printf("%s: n_rank_tok_embeddings : %u\n", __func__, params->n_rank_tok_embeddings); - printf("%s: n_rank_norm : %u\n", __func__, params->n_rank_norm); - printf("%s: n_rank_output : %u\n", __func__, params->n_rank_output); -} - -#define GGUF_GET_KEY(ctx, dst, func, type, req, key) \ -{ \ - const std::string skey(key); \ - const int kid = gguf_find_key(ctx, skey.c_str()); \ - if (kid >= 0) { \ - enum gguf_type ktype = gguf_get_kv_type(ctx, kid); \ - if (ktype != (type)) { \ - die_fmt("key %s has wrong type: %s", skey.c_str(), gguf_type_name(ktype)); \ - } \ - (dst) = func(ctx, kid); \ - } else if (req) { \ - die_fmt("key not found in model: %s", skey.c_str()); \ - } \ -} - -static void load_model_hparams_gguf(struct gguf_context * ctx, struct my_llama_hparams * hparams, const char * expected_arch) { - std::string arch; - - GGUF_GET_KEY(ctx, arch, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_GENERAL_ARCHITECTURE); - if (expected_arch != NULL) { - if (arch != expected_arch) { - printf("%s: arch=%s expected_arch=%s\n", __func__, arch.c_str(), expected_arch); - } - GGML_ASSERT(arch == expected_arch); - } - - std::vector keybuf; - keybuf.resize(512); - auto kv = [&arch, &keybuf](const char * key) -> const char * { - snprintf(keybuf.data(), keybuf.size(), key, arch.c_str()); - return keybuf.data(); - }; - - GGUF_GET_KEY(ctx, hparams->n_embd, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_EMBEDDING_LENGTH)); - GGUF_GET_KEY(ctx, hparams->n_ctx, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_CONTEXT_LENGTH)); - GGUF_GET_KEY(ctx, hparams->n_ff, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_FEED_FORWARD_LENGTH)); - GGUF_GET_KEY(ctx, hparams->n_head, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_ATTENTION_HEAD_COUNT)); - GGUF_GET_KEY(ctx, hparams->n_layer, gguf_get_val_u32, GGUF_TYPE_UINT32, true, kv(LLM_KV_BLOCK_COUNT)); - - // n_head_kv is optional, default to n_head - hparams->n_head_kv = hparams->n_head; - GGUF_GET_KEY(ctx, hparams->n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV)); - - float rope_freq_scale = 1.0f; - GGUF_GET_KEY(ctx, hparams->f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS)); - GGUF_GET_KEY(ctx, hparams->rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE)); - GGUF_GET_KEY(ctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR)); - if (rope_freq_scale != 1.0f) { - hparams->rope_freq_scale = 1.0f / rope_freq_scale; - } -} - -static void init_model(struct llama_model * input, struct my_llama_model * model, const char * fn_model, uint32_t n_ctx) { - auto & hparams = model->hparams; - - std::vector tn_buf; - tn_buf.resize(GGML_MAX_NAME); - auto tn = [&tn_buf](const char * key) -> const char * { - snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", key); - return tn_buf.data(); - }; - auto tni = [&tn_buf](const char * key, int bid) -> const char * { - snprintf(tn_buf.data(), tn_buf.size(), key, bid); - std::string s = tn_buf.data(); - snprintf(tn_buf.data(), tn_buf.size(), "%s.weight", s.c_str()); - return tn_buf.data(); - }; - - - // get parameters directly from gguf file - { - struct gguf_init_params params = { - /*.no_alloc = */ false, - /*.ctx = */ NULL, - }; - struct gguf_context * mctx = gguf_init_from_file(fn_model, params); - - load_model_hparams_gguf(mctx, &hparams, "llama"); - - gguf_free(mctx); - } - hparams.n_vocab = llama_n_vocab(input); - hparams.n_ctx = n_ctx; - - // get tensors from llama_model (possibly mmapped) - model->tok_embeddings = llama_get_model_tensor(input, tn(LLM_TENSOR_TOKEN_EMBD)); - model->norm = llama_get_model_tensor(input, tn(LLM_TENSOR_OUTPUT_NORM)); - model->output = llama_get_model_tensor(input, tn(LLM_TENSOR_OUTPUT)); - - assert_shape_2d(model->tok_embeddings, hparams.n_embd, hparams.n_vocab); - assert_shape_1d(model->norm, hparams.n_embd); - assert_shape_2d(model->output, hparams.n_embd, hparams.n_vocab); - - model->layers.resize(hparams.n_layer); - for (uint32_t i = 0; i < hparams.n_layer; ++i) { - auto & layer = model->layers[i]; - - layer.attention_norm = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_NORM, i)); - layer.wq = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_Q, i)); - layer.wk = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_K, i)); - layer.wv = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_V, i)); - layer.wo = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_OUT, i)); - layer.ffn_norm = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_NORM, i)); - layer.ffn_gate = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i)); - layer.ffn_down = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i)); - layer.ffn_up = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i)); - - assert_shape_1d(layer.attention_norm, hparams.n_embd); - assert_shape_2d(layer.wq, hparams.n_embd, hparams.n_embd); - assert_shape_2d(layer.wk, hparams.n_embd, hparams.n_embd_gqa()); - assert_shape_2d(layer.wv, hparams.n_embd, hparams.n_embd_gqa()); - assert_shape_2d(layer.wo, hparams.n_embd, hparams.n_embd); - assert_shape_1d(layer.ffn_norm, hparams.n_embd); - assert_shape_2d(layer.ffn_gate, hparams.n_embd, hparams.n_ff); - assert_shape_2d(layer.ffn_down, hparams.n_ff, hparams.n_embd); - assert_shape_2d(layer.ffn_up, hparams.n_embd, hparams.n_ff); - } -} - -static void set_param_lora(struct my_llama_lora * lora) { - const uint32_t n_layer = lora->layers.size(); - - struct ggml_context* ctx = lora->ctx; - - ggml_set_param(ctx, lora->tok_embeddings_a); - ggml_set_param(ctx, lora->tok_embeddings_b); - ggml_set_param(ctx, lora->norm_a); - ggml_set_param(ctx, lora->norm_b); - ggml_set_param(ctx, lora->output_a); - ggml_set_param(ctx, lora->output_b); - - for (uint32_t i = 0; i < n_layer; ++i) { - auto & layer = lora->layers[i]; - - ggml_set_param(ctx, layer.attention_norm_a); - ggml_set_param(ctx, layer.attention_norm_b); - ggml_set_param(ctx, layer.wq_a); - ggml_set_param(ctx, layer.wq_b); - ggml_set_param(ctx, layer.wk_a); - ggml_set_param(ctx, layer.wk_b); - ggml_set_param(ctx, layer.wv_a); - ggml_set_param(ctx, layer.wv_b); - ggml_set_param(ctx, layer.wo_a); - ggml_set_param(ctx, layer.wo_b); - ggml_set_param(ctx, layer.ffn_norm_a); - ggml_set_param(ctx, layer.ffn_norm_b); - ggml_set_param(ctx, layer.ffn_gate_a); - ggml_set_param(ctx, layer.ffn_gate_b); - ggml_set_param(ctx, layer.ffn_down_a); - ggml_set_param(ctx, layer.ffn_down_b); - ggml_set_param(ctx, layer.ffn_up_a); - ggml_set_param(ctx, layer.ffn_up_b); - } -} - -static void init_lora(const struct my_llama_model * model, struct my_llama_lora * lora) { - const auto & lparams = lora->hparams; - - const uint32_t n_embd = model->hparams.n_embd; - const uint32_t n_embd_gqa = model->hparams.n_embd_gqa(); - const uint32_t n_layer = model->hparams.n_layer; - const uint32_t n_vocab = model->hparams.n_vocab; - const uint32_t n_ff = model->hparams.n_ff; - - std::vector tn_buf; - tn_buf.resize(GGML_MAX_NAME); - auto tn = [&tn_buf](const char * key, const char * suffix) -> const char * { - snprintf(tn_buf.data(), tn_buf.size(), "%s%s", key, suffix); - return tn_buf.data(); - }; - auto tni = [&tn_buf](const char * key, const char * suffix, int bid) -> const char * { - snprintf(tn_buf.data(), tn_buf.size(), key, bid); - std::string s = tn_buf.data(); - snprintf(tn_buf.data(), tn_buf.size(), "%s%s", s.c_str(), suffix); - return tn_buf.data(); - }; - - // context for lora tensors without their data - struct ggml_init_params ctx_lora_params; - ctx_lora_params.mem_size = ggml_tensor_overhead()*2*(6 + n_layer*18); - ctx_lora_params.mem_buffer = NULL; - ctx_lora_params.no_alloc = true; - - struct ggml_context * ctx = ggml_init(ctx_lora_params); - lora->ctx = ctx; - - lora->tok_embeddings_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_tok_embeddings, n_embd); - lora->tok_embeddings_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_tok_embeddings, n_vocab); - lora->norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_norm, n_embd); - lora->norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_norm, 1); - lora->output_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_output, n_embd); - lora->output_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_output, n_vocab); - - ggml_set_name(lora->tok_embeddings_a, tn(LLM_TENSOR_TOKEN_EMBD, ".weight.lora_a")); - ggml_set_name(lora->tok_embeddings_b, tn(LLM_TENSOR_TOKEN_EMBD, ".weight.lora_b")); - ggml_set_name(lora->norm_a, tn(LLM_TENSOR_OUTPUT_NORM, ".weight.lora_a")); - ggml_set_name(lora->norm_b, tn(LLM_TENSOR_OUTPUT_NORM, ".weight.lora_b")); - ggml_set_name(lora->output_a, tn(LLM_TENSOR_OUTPUT, ".weight.lora_a")); - ggml_set_name(lora->output_b, tn(LLM_TENSOR_OUTPUT, ".weight.lora_b")); - - lora->layers.resize(n_layer); - for (uint32_t i = 0; i < n_layer; ++i) { - auto & layer = lora->layers[i]; - - layer.attention_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_attention_norm, n_embd); - layer.attention_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_attention_norm, 1); - - layer.wq_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wq, n_embd); - layer.wq_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wq, n_embd); - layer.wk_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wk, n_embd); - layer.wk_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wk, n_embd_gqa); - layer.wv_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wv, n_embd); - layer.wv_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wv, n_embd_gqa); - layer.wo_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wo, n_embd); - layer.wo_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_wo, n_embd); - - layer.ffn_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, n_embd); - layer.ffn_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, 1); - - layer.ffn_gate_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_embd); - layer.ffn_gate_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_ff); - layer.ffn_down_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_ff); - layer.ffn_down_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_embd); - layer.ffn_up_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up, n_embd); - layer.ffn_up_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up, n_ff); - - ggml_set_name(layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_a", i)); - ggml_set_name(layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_b", i)); - ggml_set_name(layer.wq_a, tni(LLM_TENSOR_ATTN_Q, ".weight.lora_a", i)); - ggml_set_name(layer.wq_b, tni(LLM_TENSOR_ATTN_Q, ".weight.lora_b", i)); - ggml_set_name(layer.wk_a, tni(LLM_TENSOR_ATTN_K, ".weight.lora_a", i)); - ggml_set_name(layer.wk_b, tni(LLM_TENSOR_ATTN_K, ".weight.lora_b", i)); - ggml_set_name(layer.wv_a, tni(LLM_TENSOR_ATTN_V, ".weight.lora_a", i)); - ggml_set_name(layer.wv_b, tni(LLM_TENSOR_ATTN_V, ".weight.lora_b", i)); - ggml_set_name(layer.wo_a, tni(LLM_TENSOR_ATTN_OUT, ".weight.lora_a", i)); - ggml_set_name(layer.wo_b, tni(LLM_TENSOR_ATTN_OUT, ".weight.lora_b", i)); - ggml_set_name(layer.ffn_norm_a, tni(LLM_TENSOR_FFN_NORM, ".weight.lora_a", i)); - ggml_set_name(layer.ffn_norm_b, tni(LLM_TENSOR_FFN_NORM, ".weight.lora_b", i)); - ggml_set_name(layer.ffn_gate_a, tni(LLM_TENSOR_FFN_GATE, ".weight.lora_a", i)); - ggml_set_name(layer.ffn_gate_b, tni(LLM_TENSOR_FFN_GATE, ".weight.lora_b", i)); - ggml_set_name(layer.ffn_down_a, tni(LLM_TENSOR_FFN_DOWN, ".weight.lora_a", i)); - ggml_set_name(layer.ffn_down_b, tni(LLM_TENSOR_FFN_DOWN, ".weight.lora_b", i)); - ggml_set_name(layer.ffn_up_a, tni(LLM_TENSOR_FFN_UP, ".weight.lora_a", i)); - ggml_set_name(layer.ffn_up_b, tni(LLM_TENSOR_FFN_UP, ".weight.lora_b", i)); - } - - set_param_lora(lora); - - // allocate data for lora tensors - lora->data = ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_cpu_buffer_type()); -} - -static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, float std, float min, float max) { - const uint32_t n_layer = lora->layers.size(); - - struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max); - - randomize_tensor_normal(lora->tok_embeddings_a, rnd); - ggml_set_zero(lora->tok_embeddings_b); - randomize_tensor_normal(lora->norm_a, rnd); - ggml_set_zero(lora->norm_b); - randomize_tensor_normal(lora->output_a, rnd); - ggml_set_zero(lora->output_b); - - for (uint32_t i = 0; i < n_layer; ++i) { - auto & layer = lora->layers[i]; - randomize_tensor_normal(layer.attention_norm_a, rnd); - ggml_set_zero(layer.attention_norm_b); - - randomize_tensor_normal(layer.wq_a, rnd); - ggml_set_zero(layer.wq_b); - randomize_tensor_normal(layer.wk_a, rnd); - ggml_set_zero(layer.wk_b); - randomize_tensor_normal(layer.wv_a, rnd); - ggml_set_zero(layer.wv_b); - randomize_tensor_normal(layer.wo_a, rnd); - ggml_set_zero(layer.wo_b); - - randomize_tensor_normal(layer.ffn_norm_a, rnd); - ggml_set_zero(layer.ffn_norm_b); - - randomize_tensor_normal(layer.ffn_gate_a, rnd); - ggml_set_zero(layer.ffn_gate_b); - randomize_tensor_normal(layer.ffn_down_a, rnd); - ggml_set_zero(layer.ffn_down_b); - randomize_tensor_normal(layer.ffn_up_a, rnd); - ggml_set_zero(layer.ffn_up_b); - } - - free_random_normal_distribution(rnd); -} - -static struct ggml_tensor * llama_build_lora_finetune_graphs( - struct my_llama_model * model, - struct my_llama_lora * lora, - ggml_gallocr_t alloc, - struct ggml_context * ctx, - struct ggml_cgraph * gf, - struct ggml_cgraph * gb, - struct ggml_cgraph * gb_tmp, - struct ggml_tensor * * logits, - struct ggml_tensor * tokens_input, - struct ggml_tensor * targets, - const int n_tokens, - const int n_batch, - const bool enable_flash_attn, - const bool enable_checkpointing, - const bool measure_only) { - - ggml_set_scratch(ctx, { 0, 0, nullptr, }); - const int n_past = 0; - const int N = n_tokens; - const auto & hparams = model->hparams; - const int n_ctx = hparams.n_ctx; - const int n_vocab = hparams.n_vocab; - const int n_embd = hparams.n_embd; - const int n_layer = hparams.n_layer; - const int n_head = hparams.n_head; - const int n_head_kv = hparams.n_head_kv; - const int n_ff = hparams.n_ff; - const int n_rot = hparams.n_embd_head(); - const int n_embd_head = hparams.n_embd_head(); - const int n_embd_gqa = hparams.n_embd_gqa(); - - const float rms_norm_eps = hparams.f_norm_rms_eps; - const float rope_freq_base = hparams.rope_freq_base; - const float rope_freq_scale = hparams.rope_freq_scale; - - GGML_ASSERT((size_t) n_layer == lora->layers.size()); - - auto set_name = [](struct ggml_tensor * t, const char * n) { - ggml_set_name(t, n); - if (t->grad) { - ggml_format_name(t->grad, "%s->grad", n); - } - }; - - // KQ_pos - contains the positions - struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N); - ggml_set_input(KQ_pos); - - // rope has so much parameters that we make a custom function for it - auto rope = [ctx, KQ_pos, n_rot, n_ctx, rope_freq_base, rope_freq_scale] - (struct ggml_tensor * t) -> struct ggml_tensor * { - // not capturing these, to silcence warnings - const int rope_mode = 0; - - return ggml_rope_ext(ctx, - t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, - rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f - ); - }; - - set_name(tokens_input, "tokens_input"); - set_name(targets, "targets"); - - GGML_ASSERT(tokens_input->type == GGML_TYPE_I32); - - auto add_to_f32 = [] (struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) { - if (ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16 || a->type == GGML_TYPE_BF16) { - return ggml_add_cast(ctx, a, b, GGML_TYPE_F32); - } else if (a->type == GGML_TYPE_F32) { - return ggml_add(ctx, a, b); - } else { - die_fmt("%s: Finetuning on tensors with type '%s' is not yet supported.\n", - __func__, ggml_type_name(a->type)); - } - }; - - struct ggml_tensor * tok_embeddings = add_to_f32(ctx, model->tok_embeddings, ggml_mul_mat(ctx, lora->tok_embeddings_a, lora->tok_embeddings_b)); - struct ggml_tensor * norm = add_to_f32(ctx, model->norm, ggml_mul_mat(ctx, lora->norm_a, lora->norm_b)); - struct ggml_tensor * output = add_to_f32(ctx, model->output, ggml_mul_mat(ctx, lora->output_a, lora->output_b)); - - struct ggml_tensor * t00 = ggml_reshape_1d(ctx, tokens_input, N*n_batch); set_name(t00, "t00"); assert_shape_1d(t00, N*n_batch); - struct ggml_tensor * t01 = ggml_get_rows(ctx, tok_embeddings, t00); set_name(t01, "t01"); assert_shape_2d(t01, n_embd, N*n_batch); - - struct ggml_tensor * cur = t01; - - std::vector checkpoints; - if (enable_checkpointing) { - checkpoints.push_back(tokens_input); - checkpoints.push_back(targets); - checkpoints.push_back(t00); - checkpoints.push_back(t01); - } - - const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head); - - for (int il = 0; il < n_layer; ++il) { - struct my_llama_layer & layer = model->layers[il]; - struct my_llama_lora_layer & llayer = lora->layers[il]; - - struct ggml_tensor * attention_norm = add_to_f32(ctx, layer.attention_norm, ggml_mul_mat(ctx, llayer.attention_norm_a, llayer.attention_norm_b)); - struct ggml_tensor * ffn_norm = add_to_f32(ctx, layer.ffn_norm, ggml_mul_mat(ctx, llayer.ffn_norm_a, llayer.ffn_norm_b)); - struct ggml_tensor * wq = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b)); - struct ggml_tensor * wk = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b)); - struct ggml_tensor * wv = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b)); - struct ggml_tensor * wo = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b)); - struct ggml_tensor * ffn_gate = add_to_f32(ctx, layer.ffn_gate, ggml_mul_mat(ctx, llayer.ffn_gate_a, llayer.ffn_gate_b)); - struct ggml_tensor * ffn_down = add_to_f32(ctx, layer.ffn_down, ggml_mul_mat(ctx, llayer.ffn_down_a, llayer.ffn_down_b)); - struct ggml_tensor * ffn_up = add_to_f32(ctx, layer.ffn_up, ggml_mul_mat(ctx, llayer.ffn_up_a, llayer.ffn_up_b)); - - struct ggml_tensor * t02 = ggml_rms_norm (ctx, cur, rms_norm_eps); set_name(t02, "t02"); assert_shape_2d(t02, n_embd, N*n_batch); - struct ggml_tensor * t03 = ggml_repeat (ctx, attention_norm, t02); set_name(t03, "t03"); assert_shape_2d(t03, n_embd, N*n_batch); - struct ggml_tensor * t04 = ggml_mul (ctx, t03, t02); set_name(t04, "t04"); assert_shape_2d(t04, n_embd, N*n_batch); - struct ggml_tensor * t05 = ggml_mul_mat (ctx, wq, t04); set_name(t05, "t05"); assert_shape_2d(t05, n_embd, N*n_batch); - struct ggml_tensor * t06 = ggml_reshape_4d (ctx, t05, n_embd_head, n_head, N, n_batch); set_name(t06, "t06"); assert_shape_4d(t06, n_embd_head, n_head, N, n_batch); - struct ggml_tensor * t07 = rope (t06); set_name(t07, "t07"); assert_shape_4d(t07, n_embd_head, n_head, N, n_batch); - struct ggml_tensor * t08 = ggml_mul_mat (ctx, wk, t04); set_name(t08, "t08"); assert_shape_2d(t08, n_embd_gqa, N*n_batch); - struct ggml_tensor * t09 = ggml_reshape_4d (ctx, t08, n_embd_head, n_head_kv, N, n_batch); set_name(t09, "t09"); assert_shape_4d(t09, n_embd_head, n_head_kv, N, n_batch); - struct ggml_tensor * t10 = rope (t09); set_name(t10, "t10"); assert_shape_4d(t10, n_embd_head, n_head_kv, N, n_batch); - - struct ggml_tensor * t11; - if (ggml_is_quantized(wv->type)) { - struct ggml_tensor * t11_1 = ggml_mul_mat (ctx, wv, t04); set_name(t11_1, "t11_1"); assert_shape_2d(t11_1, n_embd_gqa, N*n_batch); - struct ggml_tensor * t11_2 = ggml_transpose(ctx, t11_1); set_name(t11_2, "t11_2"); assert_shape_2d(t11_2, N*n_batch, n_embd_gqa); - t11 = ggml_cont (ctx, t11_2); set_name(t11, "t11"); assert_shape_2d(t11, N*n_batch, n_embd_gqa); - } else { - t11 = ggml_mul_mat (ctx, t04, wv); set_name(t11, "t11"); assert_shape_2d(t11, N*n_batch, n_embd_gqa); - } - - struct ggml_tensor * t12 = ggml_reshape_4d (ctx, t11, N, n_batch, n_embd_head, n_head_kv); set_name(t12, "t12"); assert_shape_4d(t12, N, n_batch, n_embd_head, n_head_kv); - struct ggml_tensor * t13 = ggml_permute (ctx, t07, 0, 2, 1, 3); set_name(t13, "t13"); assert_shape_4d(t13, n_embd_head, N, n_head, n_batch); - struct ggml_tensor * t14 = ggml_permute (ctx, t10, 0, 2, 1, 3); set_name(t14, "t14"); assert_shape_4d(t14, n_embd_head, N, n_head_kv, n_batch); - struct ggml_tensor * t15 = ggml_permute (ctx, t12, 0, 3, 1, 2); set_name(t15, "t15"); assert_shape_4d(t15, N, n_embd_head, n_head_kv, n_batch); - struct ggml_tensor * t16; - if (enable_flash_attn) { - GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported"); - //t16 = ggml_flash_attn(ctx, t13, t14, t15, true); set_name(t16, "t16"); assert_shape_4d(t16, n_embd_head, N, n_head, n_batch); - } else { - struct ggml_tensor * t16_0 = ggml_mul_mat (ctx, t14, t13); set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch); - struct ggml_tensor * t16_1 = ggml_scale_inplace (ctx, t16_0, kv_scale); set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch); - struct ggml_tensor * t16_2 = ggml_diag_mask_inf_inplace(ctx, t16_1, n_past); set_name(t16_2, "t16_2"); assert_shape_4d(t16_2, N, N, n_head, n_batch); - struct ggml_tensor * t16_3 = ggml_soft_max_inplace (ctx, t16_2); set_name(t16_3, "t16_3"); assert_shape_4d(t16_3, N, N, n_head, n_batch); - t16 = ggml_mul_mat(ctx, t15, t16_3); set_name(t16, "t16"); assert_shape_4d(t16, n_embd_head, N, n_head, n_batch); - } - struct ggml_tensor * t17 = ggml_permute (ctx, t16, 0, 2, 1, 3); set_name(t17, "t17"); assert_shape_4d(t17, n_embd_head, n_head, N, n_batch); - struct ggml_tensor * t18 = ggml_cont (ctx, t17); set_name(t18, "t18"); assert_shape_4d(t18, n_embd_head, n_head, N, n_batch); - struct ggml_tensor * t19 = ggml_reshape_2d (ctx, t18, n_embd, N*n_batch); set_name(t19, "t19"); assert_shape_2d(t19, n_embd, N*n_batch); - struct ggml_tensor * t20 = ggml_mul_mat (ctx, wo, t19); set_name(t20, "t20"); assert_shape_2d(t20, n_embd, N*n_batch); - struct ggml_tensor * t21 = ggml_add (ctx, t20, cur); set_name(t21, "t21"); assert_shape_2d(t21, n_embd, N*n_batch); - struct ggml_tensor * t22 = ggml_rms_norm (ctx, t21, rms_norm_eps); set_name(t22, "t22"); assert_shape_2d(t22, n_embd, N*n_batch); - struct ggml_tensor * t23 = ggml_repeat (ctx, ffn_norm, t22); set_name(t23, "t23"); assert_shape_2d(t23, n_embd, N*n_batch); - struct ggml_tensor * t24 = ggml_mul (ctx, t23, t22); set_name(t24, "t24"); assert_shape_2d(t24, n_embd, N*n_batch); - struct ggml_tensor * t25 = ggml_mul_mat (ctx, ffn_up, t24); set_name(t25, "t25"); assert_shape_2d(t25, n_ff, N*n_batch); - struct ggml_tensor * t26 = ggml_mul_mat (ctx, ffn_gate, t24); set_name(t26, "t26"); assert_shape_2d(t26, n_ff, N*n_batch); - struct ggml_tensor * t27 = ggml_silu (ctx, t26); set_name(t27, "t27"); assert_shape_2d(t27, n_ff, N*n_batch); - struct ggml_tensor * t28 = ggml_mul (ctx, t27, t25); set_name(t28, "t28"); assert_shape_2d(t28, n_ff, N*n_batch); - struct ggml_tensor * t29 = ggml_mul_mat (ctx, ffn_down, t28); set_name(t29, "t29"); assert_shape_2d(t29, n_embd, N*n_batch); - struct ggml_tensor * t30 = ggml_add (ctx, t29, t21); set_name(t30, "t30"); assert_shape_2d(t30, n_embd, N*n_batch); - cur = t30; - if (enable_checkpointing) { - checkpoints.push_back(cur); - } - } - struct ggml_tensor * t31 = ggml_rms_norm (ctx, cur, rms_norm_eps); set_name(t31, "t31"); assert_shape_2d(t31, n_embd, N*n_batch); - struct ggml_tensor * t32 = ggml_repeat (ctx, norm, t31); set_name(t32, "t32"); assert_shape_2d(t32, n_embd, N*n_batch); - struct ggml_tensor * t33 = ggml_mul (ctx, t32, t31); set_name(t33, "t33"); assert_shape_2d(t33, n_embd, N*n_batch); - struct ggml_tensor * t34 = ggml_mul_mat (ctx, output, t33); set_name(t34, "t34"); assert_shape_2d(t34, n_vocab, N*n_batch); - struct ggml_tensor * t35 = ggml_reshape_3d (ctx, t34, n_vocab, N, n_batch); set_name(t35, "t35"); assert_shape_3d(t35, n_vocab, N, n_batch); - struct ggml_tensor * t36 = ggml_cross_entropy_loss(ctx, t35, targets); set_name(t36, "t36"); assert_shape_1d(t36, 1); - - if (enable_checkpointing) { - checkpoints.push_back(t31); - checkpoints.push_back(t32); - checkpoints.push_back(t33); - checkpoints.push_back(t34); - checkpoints.push_back(t35); - checkpoints.push_back(t36); - } - - ggml_build_forward_expand(gf, t36); - - if (enable_checkpointing) { - ggml_build_backward_gradient_checkpointing(ctx, gf, gb, gb_tmp, checkpoints.data(), (int) checkpoints.size()); - } else { - ggml_graph_cpy(gf, gb); - ggml_build_backward_expand(ctx, gf, gb, true); - } - - GGML_ASSERT(alloc != NULL); - - // make sure some tensors are not reallocated by inserting new temporary nodes depending on them - int n_leafs_before = gb->n_leafs; - int n_nodes_before = gb->n_nodes; - - // output tensors - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f)); - // input gradient - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f)); - GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL); - ggml_set_input(t36->grad); - // KQ_pos - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f)); - - // make sure base model tensors data cannot be used in viewable operations - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, 1.0f)); - for (int il = 0; il < n_layer; ++il) { - struct my_llama_layer & layer = model->layers[il]; - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_gate, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_down, 1.0f)); - ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_up, 1.0f)); - } - - // allocating checkpoints in one block to reduce memory fragmentation - // note: they will be freed in reverse order - for (unsigned int i = 0; i < checkpoints.size(); ++i) { - if (checkpoints[i]->data == NULL && checkpoints[i]->view_src == NULL) { - ggml_set_input(checkpoints[i]); - } - } - - if (measure_only) { - ggml_gallocr_reserve(alloc, gb); - } else { - ggml_gallocr_alloc_graph(alloc, gb); - - // set KQ_pos - { - int * data = (int *) KQ_pos->data; - for (int i = 0; i < N; ++i) { - data[i] = n_past + i; - } - } - } - - // remove the additional nodes and leafs - for (int i = n_leafs_before; i < gb->n_leafs; ++i) { - gb->leafs[i] = NULL; - } - for (int i = n_nodes_before; i < gb->n_nodes; ++i) { - gb->nodes[i] = NULL; - } - gb->n_leafs = n_leafs_before; - gb->n_nodes = n_nodes_before; - - *logits = t35; - return t36; -} - -static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct my_llama_lora * lora) { - // NOTE: gguf_context must be initialized with f_ggml_ctx and no_alloc=false, otherwise tensor data can not be read - - std::string arch; - - std::vector keybuf; - keybuf.resize(512); - - GGUF_GET_KEY(fctx, arch, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_GENERAL_ARCHITECTURE); - GGML_ASSERT(arch == "llama"); - - uint32_t ftype_u; - GGUF_GET_KEY(fctx, ftype_u, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_GENERAL_FILE_TYPE); - GGML_ASSERT((enum llama_ftype) ftype_u == LLAMA_FTYPE_ALL_F32); - - struct my_llama_hparams hparams; - load_model_hparams_gguf(fctx, &hparams, arch.c_str()); - - // parameters that define tensor shapes must match - GGML_ASSERT(hparams.n_embd == model->hparams.n_embd); - GGML_ASSERT(hparams.n_ff == model->hparams.n_ff); - GGML_ASSERT(hparams.n_head == model->hparams.n_head); - GGML_ASSERT(hparams.n_head_kv == model->hparams.n_head_kv); - GGML_ASSERT(hparams.n_layer == model->hparams.n_layer); - - GGUF_GET_KEY(fctx, lora->hparams.n_rank_tok_embeddings, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_norm, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_output, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_OUTPUT); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_attention_norm, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_NORM); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_wq, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_Q); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_wk, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_K); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_wv, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_V); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_wo, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_norm, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_NORM); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_gate, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_down, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN); - GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_up, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP); - - init_lora(model, lora); - - copy_tensor_by_name(lora->tok_embeddings_a, f_ggml_ctx, ggml_get_name(lora->tok_embeddings_a)); - copy_tensor_by_name(lora->tok_embeddings_b, f_ggml_ctx, ggml_get_name(lora->tok_embeddings_b)); - copy_tensor_by_name(lora->norm_a, f_ggml_ctx, ggml_get_name(lora->norm_a)); - copy_tensor_by_name(lora->norm_b, f_ggml_ctx, ggml_get_name(lora->norm_b)); - copy_tensor_by_name(lora->output_a, f_ggml_ctx, ggml_get_name(lora->output_a)); - copy_tensor_by_name(lora->output_b, f_ggml_ctx, ggml_get_name(lora->output_b)); - - for (uint32_t i = 0; i < lora->layers.size(); ++i) { - auto & layer = lora->layers[i]; - copy_tensor_by_name(layer.attention_norm_a, f_ggml_ctx, ggml_get_name(layer.attention_norm_a)); - copy_tensor_by_name(layer.attention_norm_b, f_ggml_ctx, ggml_get_name(layer.attention_norm_b)); - copy_tensor_by_name(layer.wq_a, f_ggml_ctx, ggml_get_name(layer.wq_a)); - copy_tensor_by_name(layer.wq_b, f_ggml_ctx, ggml_get_name(layer.wq_b)); - copy_tensor_by_name(layer.wk_a, f_ggml_ctx, ggml_get_name(layer.wk_a)); - copy_tensor_by_name(layer.wk_b, f_ggml_ctx, ggml_get_name(layer.wk_b)); - copy_tensor_by_name(layer.wv_a, f_ggml_ctx, ggml_get_name(layer.wv_a)); - copy_tensor_by_name(layer.wv_b, f_ggml_ctx, ggml_get_name(layer.wv_b)); - copy_tensor_by_name(layer.wo_a, f_ggml_ctx, ggml_get_name(layer.wo_a)); - copy_tensor_by_name(layer.wo_b, f_ggml_ctx, ggml_get_name(layer.wo_b)); - copy_tensor_by_name(layer.ffn_norm_a, f_ggml_ctx, ggml_get_name(layer.ffn_norm_a)); - copy_tensor_by_name(layer.ffn_norm_b, f_ggml_ctx, ggml_get_name(layer.ffn_norm_b)); - copy_tensor_by_name(layer.ffn_gate_a, f_ggml_ctx, ggml_get_name(layer.ffn_gate_a)); - copy_tensor_by_name(layer.ffn_gate_b, f_ggml_ctx, ggml_get_name(layer.ffn_gate_b)); - copy_tensor_by_name(layer.ffn_down_a, f_ggml_ctx, ggml_get_name(layer.ffn_down_a)); - copy_tensor_by_name(layer.ffn_down_b, f_ggml_ctx, ggml_get_name(layer.ffn_down_b)); - copy_tensor_by_name(layer.ffn_up_a, f_ggml_ctx, ggml_get_name(layer.ffn_up_a)); - copy_tensor_by_name(layer.ffn_up_b, f_ggml_ctx, ggml_get_name(layer.ffn_up_b)); - } -} - -static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_model * model, struct my_llama_lora * lora) { - const char * arch = "llama"; - enum llama_ftype ftype = LLAMA_FTYPE_ALL_F32; - - std::vector keybuf; - keybuf.resize(512); - auto kv = [arch, &keybuf](const char * key) -> const char * { - snprintf(keybuf.data(), keybuf.size(), key, arch); - return keybuf.data(); - }; - - gguf_set_val_str(fctx, LLM_KV_GENERAL_ARCHITECTURE, arch); - gguf_set_val_u32(fctx, LLM_KV_GENERAL_FILE_TYPE, ftype); - - gguf_set_val_u32(fctx, kv(LLM_KV_CONTEXT_LENGTH), model->hparams.n_ctx); - gguf_set_val_u32(fctx, kv(LLM_KV_EMBEDDING_LENGTH), model->hparams.n_embd); - gguf_set_val_u32(fctx, kv(LLM_KV_FEED_FORWARD_LENGTH), model->hparams.n_ff); - gguf_set_val_u32(fctx, kv(LLM_KV_ATTENTION_HEAD_COUNT), model->hparams.n_head); - gguf_set_val_u32(fctx, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV), model->hparams.n_head_kv); - gguf_set_val_u32(fctx, kv(LLM_KV_BLOCK_COUNT), model->hparams.n_layer); - gguf_set_val_u32(fctx, kv(LLM_KV_ROPE_DIMENSION_COUNT), model->hparams.n_embd_head()); - gguf_set_val_f32(fctx, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS), model->hparams.f_norm_rms_eps); - gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_FREQ_BASE), model->hparams.rope_freq_base); - gguf_set_val_f32(fctx, kv(LLM_KV_ROPE_SCALE_LINEAR), model->hparams.rope_freq_scale); - - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD, lora->hparams.n_rank_tok_embeddings); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM, lora->hparams.n_rank_norm); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_OUTPUT, lora->hparams.n_rank_output); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_NORM, lora->hparams.n_rank_attention_norm); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_Q, lora->hparams.n_rank_wq); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_K, lora->hparams.n_rank_wk); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_V, lora->hparams.n_rank_wv); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT, lora->hparams.n_rank_wo); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_NORM, lora->hparams.n_rank_ffn_norm); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE, lora->hparams.n_rank_ffn_gate); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN, lora->hparams.n_rank_ffn_down); - gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP, lora->hparams.n_rank_ffn_up); - - gguf_add_tensor(fctx, lora->tok_embeddings_a); - gguf_add_tensor(fctx, lora->tok_embeddings_b); - gguf_add_tensor(fctx, lora->norm_a); - gguf_add_tensor(fctx, lora->norm_b); - gguf_add_tensor(fctx, lora->output_a); - gguf_add_tensor(fctx, lora->output_b); - - for (uint32_t i = 0; i < lora->layers.size(); ++i) { - auto & layer = lora->layers[i]; - - gguf_add_tensor(fctx, layer.attention_norm_a); - gguf_add_tensor(fctx, layer.attention_norm_b); - gguf_add_tensor(fctx, layer.wq_a); - gguf_add_tensor(fctx, layer.wq_b); - gguf_add_tensor(fctx, layer.wk_a); - gguf_add_tensor(fctx, layer.wk_b); - gguf_add_tensor(fctx, layer.wv_a); - gguf_add_tensor(fctx, layer.wv_b); - gguf_add_tensor(fctx, layer.wo_a); - gguf_add_tensor(fctx, layer.wo_b); - gguf_add_tensor(fctx, layer.ffn_norm_a); - gguf_add_tensor(fctx, layer.ffn_norm_b); - gguf_add_tensor(fctx, layer.ffn_gate_a); - gguf_add_tensor(fctx, layer.ffn_gate_b); - gguf_add_tensor(fctx, layer.ffn_down_a); - gguf_add_tensor(fctx, layer.ffn_down_b); - gguf_add_tensor(fctx, layer.ffn_up_a); - gguf_add_tensor(fctx, layer.ffn_up_b); - } -} - -static void load_checkpoint_lora_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) { - std::string train_type = LLM_KV_TRAINING_TYPE_FINETUNE_LORA; - GGUF_GET_KEY(fctx, train_type, gguf_get_val_str, GGUF_TYPE_STRING, false, LLM_KV_TRAINING_TYPE); - GGML_ASSERT(train_type == LLM_KV_TRAINING_TYPE_FINETUNE_LORA); - - load_train_state_gguf(fctx, f_ggml_ctx, train); - load_llama_lora_gguf(fctx, f_ggml_ctx, model, lora); -} - -static void save_checkpoint_lora_gguf(struct gguf_context * fctx, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) { - gguf_set_val_str(fctx, LLM_KV_TRAINING_TYPE, LLM_KV_TRAINING_TYPE_FINETUNE_LORA); - save_llama_lora_gguf(fctx, model, lora); - save_train_state_gguf(fctx, train); -} - -static bool load_checkpoint_lora_file(const char * filename, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) { - struct ggml_context * f_ggml_ctx; - struct gguf_init_params params; - params.no_alloc = false; - params.ctx = &f_ggml_ctx; - struct gguf_context * fctx = gguf_init_from_file(filename, params); - if (fctx == NULL) { - return false; - } - - load_checkpoint_lora_gguf(fctx, f_ggml_ctx, model, lora, train); - - gguf_free(fctx); - return true; -} - -static void save_checkpoint_lora_file(const char * filename, struct my_llama_model * model, struct my_llama_lora * lora, struct train_state * train) { - printf("%s: saving to %s\n", __func__, filename); - struct gguf_context * fctx = gguf_init_empty(); - - save_checkpoint_lora_gguf(fctx, model, lora, train); - - // write file - const bool only_meta = false; - gguf_write_to_file(fctx, filename, only_meta); - gguf_free(fctx); -} - -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) { - size = 0; - } else { - 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 - GGML_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 - GGML_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)) { - die_fmt("read error: %s", strerror(errno)); - } - if (ret != 1) { - die("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) { - die_fmt("write error: %s", strerror(errno)); - } - } - - void write_u32(std::uint32_t val) { - write_raw(&val, sizeof(val)); - } - - ~llama_file() { - if (fp) { - std::fclose(fp); - } - } -}; - -static void write_tensor(struct llama_file * file, struct ggml_tensor * tensor, const char * name) { - if (tensor == NULL) { - file->write_u32(0); - file->write_u32(0); - file->write_u32(GGML_TYPE_F32); - file->seek((0-file->tell()) & 31, SEEK_CUR); - return; - } - if (name == NULL) { - name = ggml_get_name(tensor); - } - uint32_t name_len = strlen(name); - uint32_t nd = ggml_n_dims(tensor); - uint32_t ne[4] = { (uint32_t)tensor->ne[0], - (uint32_t)tensor->ne[1], - (uint32_t)tensor->ne[2], - (uint32_t)tensor->ne[3] }; - file->write_u32(nd); - file->write_u32(name_len); - file->write_u32(tensor->type); - file->write_raw(ne, sizeof(ne[0]) * nd); - file->write_raw(name, name_len); - file->seek((0-file->tell()) & 31, SEEK_CUR); - file->write_raw(tensor->data, ggml_nbytes(tensor)); -} - -static void save_as_llama_lora(const char * filename, struct my_llama_lora * lora) { - printf("%s: saving to %s\n", __func__, filename); - struct llama_file file(filename, "wb"); - if (file.fp == NULL) { - return; - } - - std::vector tn_buf; - tn_buf.resize(GGML_MAX_NAME); - - auto tn = [&tn_buf](const char * key, const char * suffix) -> const char * { - snprintf(tn_buf.data(), tn_buf.size(), "%s%s", key, suffix); - return tn_buf.data(); - }; - - auto tni = [&tn_buf](const char * key, int bid, const char * suffix) -> const char * { - snprintf(tn_buf.data(), tn_buf.size(), key, bid); - std::string s = tn_buf.data(); - snprintf(tn_buf.data(), tn_buf.size(), "%s%s", s.c_str(), suffix); - return tn_buf.data(); - }; - - // write_magic - file.write_u32(LLAMA_FILE_MAGIC_GGLA); // magic - file.write_u32(1); // version - // write_hparams - file.write_u32(lora->hparams.lora_r); - file.write_u32(lora->hparams.lora_alpha); - // write tensors - write_tensor(&file, lora->tok_embeddings_a, tn(LLM_TENSOR_TOKEN_EMBD, ".weight.loraA")); - write_tensor(&file, lora->tok_embeddings_b, tn(LLM_TENSOR_TOKEN_EMBD, ".weight.loraB")); - write_tensor(&file, lora->norm_a, tn(LLM_TENSOR_OUTPUT_NORM, ".weight.loraA")); - write_tensor(&file, lora->norm_b, tn(LLM_TENSOR_OUTPUT_NORM, ".weight.loraB")); - write_tensor(&file, lora->output_a, tn(LLM_TENSOR_OUTPUT, ".weight.loraA")); - write_tensor(&file, lora->output_b, tn(LLM_TENSOR_OUTPUT, ".weight.loraB")); - for (uint32_t i = 0; i < lora->layers.size(); ++i) { - auto & layer = lora->layers[i]; - write_tensor(&file, layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, i, ".weight.loraA")); - write_tensor(&file, layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, i, ".weight.loraB")); - write_tensor(&file, layer.wq_a, tni(LLM_TENSOR_ATTN_Q, i, ".weight.loraA")); - write_tensor(&file, layer.wq_b, tni(LLM_TENSOR_ATTN_Q, i, ".weight.loraB")); - write_tensor(&file, layer.wk_a, tni(LLM_TENSOR_ATTN_K, i, ".weight.loraA")); - write_tensor(&file, layer.wk_b, tni(LLM_TENSOR_ATTN_K, i, ".weight.loraB")); - write_tensor(&file, layer.wv_a, tni(LLM_TENSOR_ATTN_V, i, ".weight.loraA")); - write_tensor(&file, layer.wv_b, tni(LLM_TENSOR_ATTN_V, i, ".weight.loraB")); - write_tensor(&file, layer.wo_a, tni(LLM_TENSOR_ATTN_OUT, i, ".weight.loraA")); - write_tensor(&file, layer.wo_b, tni(LLM_TENSOR_ATTN_OUT, i, ".weight.loraB")); - write_tensor(&file, layer.ffn_norm_a, tni(LLM_TENSOR_FFN_NORM, i, ".weight.loraA")); - write_tensor(&file, layer.ffn_norm_b, tni(LLM_TENSOR_FFN_NORM, i, ".weight.loraB")); - write_tensor(&file, layer.ffn_gate_a, tni(LLM_TENSOR_FFN_GATE, i, ".weight.loraA")); - write_tensor(&file, layer.ffn_gate_b, tni(LLM_TENSOR_FFN_GATE, i, ".weight.loraB")); - write_tensor(&file, layer.ffn_down_a, tni(LLM_TENSOR_FFN_DOWN, i, ".weight.loraA")); - write_tensor(&file, layer.ffn_down_b, tni(LLM_TENSOR_FFN_DOWN, i, ".weight.loraB")); - write_tensor(&file, layer.ffn_up_a, tni(LLM_TENSOR_FFN_UP, i, ".weight.loraA")); - write_tensor(&file, layer.ffn_up_b, tni(LLM_TENSOR_FFN_UP, i, ".weight.loraB")); - } -} - -struct train_params { - struct train_params_common common; - - const char * fn_model_base; - const char * fn_lora_out; - - bool only_write_lora; - - float f_norm_rms_eps; - float rope_freq_base; - float rope_freq_scale; - - bool custom_f_norm_rms_eps; - bool custom_rope_freq_base; - bool custom_rope_freq_scale; - - int32_t lora_r; - int32_t lora_alpha; - bool custom_lora_alpha; - - uint32_t n_rank_attention_norm; - uint32_t n_rank_wq; - uint32_t n_rank_wk; - uint32_t n_rank_wv; - uint32_t n_rank_wo; - uint32_t n_rank_ffn_norm; - uint32_t n_rank_ffn_gate; - uint32_t n_rank_ffn_down; - uint32_t n_rank_ffn_up; - uint32_t n_rank_tok_embeddings; - uint32_t n_rank_norm; - uint32_t n_rank_output; - - bool custom_n_rank_attention_norm; - bool custom_n_rank_wq; - bool custom_n_rank_wk; - bool custom_n_rank_wv; - bool custom_n_rank_wo; - bool custom_n_rank_ffn_norm; - bool custom_n_rank_ffn_gate; - bool custom_n_rank_ffn_down; - bool custom_n_rank_ffn_up; - bool custom_n_rank_tok_embeddings; - bool custom_n_rank_norm; - bool custom_n_rank_output; -}; - -static struct train_params get_default_train_params() { - struct train_params params; - params.common = get_default_train_params_common(); - params.fn_model_base = ""; - params.fn_lora_out = "ggml-lora-ITERATION-f32.gguf"; - - params.only_write_lora = false; - - params.f_norm_rms_eps = 1e-5f; - params.rope_freq_base = 10000.0f; - params.rope_freq_scale = 1.0f; - - params.custom_f_norm_rms_eps = false; - params.custom_rope_freq_base = false; - params.custom_rope_freq_scale = false; - - params.lora_r = 4; - params.lora_alpha = 4; - params.custom_lora_alpha = false; - - params.n_rank_attention_norm = 1; - params.n_rank_wq = 4; - params.n_rank_wk = 4; - params.n_rank_wv = 4; - params.n_rank_wo = 4; - params.n_rank_ffn_norm = 1; - params.n_rank_ffn_gate = 4; - params.n_rank_ffn_down = 4; - params.n_rank_ffn_up = 4; - params.n_rank_tok_embeddings = 4; - params.n_rank_norm = 1; - params.n_rank_output = 4; - - params.custom_n_rank_attention_norm = false; - params.custom_n_rank_wq = false; - params.custom_n_rank_wk = false; - params.custom_n_rank_wv = false; - params.custom_n_rank_wo = false; - params.custom_n_rank_ffn_norm = false; - params.custom_n_rank_ffn_gate = false; - params.custom_n_rank_ffn_down = false; - params.custom_n_rank_ffn_up = false; - params.custom_n_rank_tok_embeddings = false; - params.custom_n_rank_norm = false; - params.custom_n_rank_output = false; - - return params; -} - -static void train_print_usage(int argc, char ** argv, const struct train_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, " --model-base FNAME model path from which to load base model (default '%s')\n", params->fn_model_base); - fprintf(stderr, " --lora-out FNAME path to save llama lora (default '%s')\n", params->fn_lora_out); - fprintf(stderr, " --only-write-lora only save llama lora, don't do any training. use this if you only want to convert a checkpoint to a lora adapter.\n"); - fprintf(stderr, " --norm-rms-eps F RMS-Norm epsilon value (default %f)\n", params->f_norm_rms_eps); - fprintf(stderr, " --rope-freq-base F Frequency base for ROPE (default %f)\n", params->rope_freq_base); - fprintf(stderr, " --rope-freq-scale F Frequency scale for ROPE (default %f)\n", params->rope_freq_scale); - fprintf(stderr, " --lora-alpha N LORA alpha : resulting LORA scaling is alpha/r. (default %d)\n", params->lora_alpha); - fprintf(stderr, " --lora-r N LORA r: default rank. Also specifies resulting scaling together with lora-alpha. (default %d)\n", params->lora_r); - fprintf(stderr, " --rank-att-norm N LORA rank for attention norm tensor, overrides default rank. Norm tensors should generally have rank 1.\n"); - fprintf(stderr, " --rank-ffn-norm N LORA rank for feed-forward norm tensor, overrides default rank. Norm tensors should generally have rank 1.\n"); - fprintf(stderr, " --rank-out-norm N LORA rank for output norm tensor, overrides default rank. Norm tensors should generally have rank 1.\n"); - fprintf(stderr, " --rank-tok-embd N LORA rank for token embeddings tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-out N LORA rank for output tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-wq N LORA rank for wq tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-wk N LORA rank for wk tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-wv N LORA rank for wv tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-wo N LORA rank for wo tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-ffn_gate N LORA rank for ffn_gate tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-ffn_down N LORA rank for ffn_down tensor, overrides default rank.\n"); - fprintf(stderr, " --rank-ffn_up N LORA rank for ffn_up tensor, overrides default rank.\n"); - - print_common_train_usage(argc, argv, ¶ms->common); -} - -static bool train_params_parse(int argc, char ** argv, struct train_params * params) { - bool invalid_param = false; - std::string arg; - struct train_params default_params = get_default_train_params(); - const std::string arg_prefix = "--"; - - for (int i = 1; i < argc; i++) { - arg = argv[i]; - if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) { - std::replace(arg.begin(), arg.end(), '_', '-'); - } - - if (consume_common_train_arg(argc, argv, &i, ¶ms->common, &invalid_param)) { - if (invalid_param) { - break; - } else if (params->common.print_usage) { - train_print_usage(argc, argv, &default_params); - exit(0); - } - } else if (arg == "--model-base") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->fn_model_base = argv[i]; - } else if (arg == "--lora-out") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->fn_lora_out = argv[i]; - } else if (arg == "--only-write-lora") { - params->only_write_lora = true; - } else if (arg == "--norm-rms-eps") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->f_norm_rms_eps = std::stof(argv[i]); - params->custom_f_norm_rms_eps = true; - } else if (arg == "--rope-freq-base") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->rope_freq_base = std::stof(argv[i]); - params->custom_rope_freq_base = true; - } else if (arg == "--rope-freq-scale") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->rope_freq_scale = std::stof(argv[i]); - params->custom_rope_freq_scale = true; - } else if (arg == "--lora-alpha") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->lora_alpha = std::stoi(argv[i]); - params->custom_lora_alpha = true; - } else if (arg == "--lora-r") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->lora_r = std::stoi(argv[i]); - } else if (arg == "--rank-att-norm") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_attention_norm = std::stoi(argv[i]); - params->custom_n_rank_attention_norm = true; - } else if (arg == "--rank-ffn-norm") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_ffn_norm = std::stoi(argv[i]); - params->custom_n_rank_ffn_norm = true; - } else if (arg == "--rank-out-norm") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_norm = std::stoi(argv[i]); - params->custom_n_rank_norm = true; - } else if (arg == "--rank-tok-embd") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_tok_embeddings = std::stoi(argv[i]); - params->custom_n_rank_tok_embeddings = true; - } else if (arg == "--rank-out") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_output = std::stoi(argv[i]); - params->custom_n_rank_output = true; - } else if (arg == "--rank-wq") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_wq = std::stoi(argv[i]); - params->custom_n_rank_wq = true; - } else if (arg == "--rank-wk") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_wk = std::stoi(argv[i]); - params->custom_n_rank_wk = true; - } else if (arg == "--rank-wv") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_wv = std::stoi(argv[i]); - params->custom_n_rank_wv = true; - } else if (arg == "--rank-wo") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_wo = std::stoi(argv[i]); - params->custom_n_rank_wo = true; - } else if (arg == "--rank-ffn_gate") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_ffn_gate = std::stoi(argv[i]); - params->custom_n_rank_ffn_gate = true; - } else if (arg == "--rank-ffn_down") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_ffn_down = std::stoi(argv[i]); - params->custom_n_rank_ffn_down = true; - } else if (arg == "--rank-ffn_up") { - if (++i >= argc) { - invalid_param = true; - break; - } - params->n_rank_ffn_up = std::stoi(argv[i]); - params->custom_n_rank_ffn_up = true; - } else { - fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); - train_print_usage(argc, argv, &default_params); - exit(1); - } - } - if (invalid_param) { - fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str()); - train_print_usage(argc, argv, &default_params); - exit(1); - } - finish_processing_train_args(¶ms->common); - return true; -} - -struct save_train_files_data { - const char * fn_checkpoint_out; - const char * fn_lora_out; - const char * pattern_fn_it; - const char * fn_latest; - struct my_llama_model * model; - struct my_llama_lora * lora; -}; - -static void save_train_files(void * vdata, struct train_state * train) { - struct save_train_files_data * data = (struct save_train_files_data *) vdata; - - int64_t iter = train->opt->iter; - - if (strlen(data->fn_checkpoint_out) > 0) { - save_checkpoint_lora_file(get_train_filename(data->fn_checkpoint_out, data->pattern_fn_it, data->fn_latest, iter).c_str(), data->model, data->lora, train); - save_checkpoint_lora_file(get_train_filename(data->fn_checkpoint_out, data->pattern_fn_it, data->fn_latest, -1 ).c_str(), data->model, data->lora, train); - } - if (strlen(data->fn_lora_out) > 0) { - save_as_llama_lora(get_train_filename(data->fn_lora_out, data->pattern_fn_it, data->fn_latest, iter).c_str(), data->lora); - save_as_llama_lora(get_train_filename(data->fn_lora_out, data->pattern_fn_it, data->fn_latest, -1 ).c_str(), data->lora); - } -} - -static int64_t get_parameter_count(struct my_llama_lora* lora) { - int64_t nx = 0; - nx += ggml_nelements(lora->tok_embeddings_a); - nx += ggml_nelements(lora->tok_embeddings_b); - nx += ggml_nelements(lora->norm_a); - nx += ggml_nelements(lora->norm_b); - nx += ggml_nelements(lora->output_a); - nx += ggml_nelements(lora->output_b); - - for (uint32_t i = 0; i < lora->layers.size(); ++i) { - auto & layer = lora->layers[i]; - nx += ggml_nelements(layer.attention_norm_a); - nx += ggml_nelements(layer.attention_norm_b); - nx += ggml_nelements(layer.wq_a); - nx += ggml_nelements(layer.wq_b); - nx += ggml_nelements(layer.wk_a); - nx += ggml_nelements(layer.wk_b); - nx += ggml_nelements(layer.wv_a); - nx += ggml_nelements(layer.wv_b); - nx += ggml_nelements(layer.wo_a); - nx += ggml_nelements(layer.wo_b); - nx += ggml_nelements(layer.ffn_norm_a); - nx += ggml_nelements(layer.ffn_norm_b); - nx += ggml_nelements(layer.ffn_gate_a); - nx += ggml_nelements(layer.ffn_gate_b); - nx += ggml_nelements(layer.ffn_down_a); - nx += ggml_nelements(layer.ffn_down_b); - nx += ggml_nelements(layer.ffn_up_a); - nx += ggml_nelements(layer.ffn_up_b); - } - return nx; -} - -int main(int argc, char ** argv) { - struct train_params params = get_default_train_params(); - - if (!train_params_parse(argc, argv, ¶ms)) { - return 1; - } - - if (params.common.seed == LLAMA_DEFAULT_SEED) { - params.common.seed = time(NULL); - } - printf("%s: seed: %u\n", __func__, params.common.seed); - srand(params.common.seed); - - struct llama_model_params llama_mparams = llama_model_default_params(); - llama_mparams.n_gpu_layers = params.common.n_gpu_layers; - llama_mparams.vocab_only = false; - - printf("%s: model base = '%s'\n", __func__, params.fn_model_base); - struct llama_model * lmodel = llama_load_model_from_file(params.fn_model_base, llama_mparams); - - struct llama_context_params llama_cparams = llama_context_default_params(); - struct llama_context * lctx = llama_new_context_with_model(lmodel, llama_cparams); - - struct my_llama_model model; - init_model(lmodel, &model, params.fn_model_base, params.common.n_ctx); - - struct my_llama_lora lora; - - struct train_state * train = init_train_state(); - struct ggml_opt_context * opt = train->opt; - - // set params from command line - if (params.custom_f_norm_rms_eps) { - model.hparams.f_norm_rms_eps = params.f_norm_rms_eps; - } - if (params.custom_rope_freq_base) { - model.hparams.rope_freq_base = params.rope_freq_base; - } - if (params.custom_rope_freq_scale) { - model.hparams.rope_freq_scale = params.rope_freq_scale; - } - lora.hparams.lora_r = params.lora_r; - lora.hparams.lora_alpha = params.custom_lora_alpha ? params.lora_alpha : params.lora_r; - uint32_t n_rank_attention_norm = params.custom_n_rank_attention_norm ? params.n_rank_attention_norm : 1; - uint32_t n_rank_wq = params.custom_n_rank_wq ? params.n_rank_wq : params.lora_r; - uint32_t n_rank_wk = params.custom_n_rank_wk ? params.n_rank_wk : params.lora_r; - uint32_t n_rank_wv = params.custom_n_rank_wv ? params.n_rank_wv : params.lora_r; - uint32_t n_rank_wo = params.custom_n_rank_wo ? params.n_rank_wo : params.lora_r; - uint32_t n_rank_ffn_norm = params.custom_n_rank_ffn_norm ? params.n_rank_ffn_norm : 1; - uint32_t n_rank_ffn_gate = params.custom_n_rank_ffn_gate ? params.n_rank_ffn_gate : params.lora_r; - uint32_t n_rank_ffn_down = params.custom_n_rank_ffn_down ? params.n_rank_ffn_down : params.lora_r; - uint32_t n_rank_ffn_up = params.custom_n_rank_ffn_up ? params.n_rank_ffn_up : params.lora_r; - uint32_t n_rank_tok_embeddings = params.custom_n_rank_tok_embeddings ? params.n_rank_tok_embeddings : params.lora_r; - uint32_t n_rank_norm = params.custom_n_rank_norm ? params.n_rank_norm : 1; - uint32_t n_rank_output = params.custom_n_rank_output ? params.n_rank_output : params.lora_r; - lora.hparams.n_rank_attention_norm = n_rank_attention_norm; - lora.hparams.n_rank_wq = n_rank_wq; - lora.hparams.n_rank_wk = n_rank_wk; - lora.hparams.n_rank_wv = n_rank_wv; - lora.hparams.n_rank_wo = n_rank_wo; - lora.hparams.n_rank_ffn_norm = n_rank_ffn_norm; - lora.hparams.n_rank_ffn_gate = n_rank_ffn_gate; - lora.hparams.n_rank_ffn_down = n_rank_ffn_down; - lora.hparams.n_rank_ffn_up = n_rank_ffn_up; - lora.hparams.n_rank_tok_embeddings = n_rank_tok_embeddings; - lora.hparams.n_rank_norm = n_rank_norm; - lora.hparams.n_rank_output = n_rank_output; - - // set opt params from command line - opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM); - opt->params.print_forward_graph = false; - opt->params.print_backward_graph = false; - opt->params.graph_size = LLAMA_TRAIN_MAX_NODES; - opt->params.n_threads = params.common.n_threads; - opt->params.past = params.common.opt_past; - opt->params.delta = params.common.opt_delta; - opt->params.max_no_improvement = params.common.opt_max_no_improvement; - opt->params.n_gradient_accumulation = params.common.n_gradient_accumulation; - opt->params.adam.n_iter = params.common.adam_n_iter; - opt->params.adam.sched = 1.0f; - opt->params.adam.alpha = params.common.adam_alpha; - opt->params.adam.decay = params.common.adam_decay; - opt->params.adam.decay_min_ndim = params.common.adam_decay_min_ndim; - opt->params.adam.beta1 = params.common.adam_beta1; - opt->params.adam.beta2 = params.common.adam_beta2; - opt->params.adam.gclip = params.common.adam_gclip; - opt->params.adam.eps_f = params.common.adam_eps_f; - - printf("%s: init model\n", __func__); - bool existed = load_checkpoint_lora_file(params.common.fn_checkpoint_in, &model, &lora, train); - - if (existed) { - // overwrite last n_ctx with user provided n_ctx - if (params.common.custom_n_ctx) { - model.hparams.n_ctx = params.common.n_ctx; - } - - const bool opt_param_count_changed = ( - (lora.hparams.n_rank_attention_norm != n_rank_attention_norm) - || (lora.hparams.n_rank_wq != n_rank_wq) - || (lora.hparams.n_rank_wk != n_rank_wk) - || (lora.hparams.n_rank_wv != n_rank_wv) - || (lora.hparams.n_rank_wo != n_rank_wo) - || (lora.hparams.n_rank_ffn_norm != n_rank_ffn_norm) - || (lora.hparams.n_rank_ffn_gate != n_rank_ffn_gate) - || (lora.hparams.n_rank_ffn_down != n_rank_ffn_down) - || (lora.hparams.n_rank_ffn_up != n_rank_ffn_up) - || (lora.hparams.n_rank_tok_embeddings != n_rank_tok_embeddings) - || (lora.hparams.n_rank_norm != n_rank_norm) - || (lora.hparams.n_rank_output != n_rank_output) - ); - - const bool opt_past_changed = opt->params.past != params.common.opt_past; - - if (opt_param_count_changed) { - print_lora_params(&lora.hparams); - die("Provided rank differs from checkpoint file. To use different rank start finetune from scratch with empty input checkpoint, e.g --checkpoint-in ''. Aborting."); - // need to discard previous optimizer gradient statistics and opt_init with new shapes - // TODO - } - if (opt_past_changed) { - die("Optimizer parameter '--opt-past N' differs from checkpoint file. To use different value finetune from scratch with empty input checkpoint, e.g --checkpoint-in ''. Aborting"); - // need to discard previous optimizer past function value statistics and opt_init with new shapes - // TODO - } - } else { // existed == false - init_lora(&model, &lora); - randomize_lora(&lora, params.common.seed, 0.0f, 1.0f, -1.0f, +1.0f); - if (!params.only_write_lora) { - ggml_opt_init(opt->ctx, opt, opt->params, get_parameter_count(&lora)); - } - } - opt->iter = train->train_its; - - print_params(&model.hparams); - print_lora_params(&lora.hparams); - printf("%s: total train_iterations %llu\n", __func__, (long long unsigned) train->train_its); - printf("%s: seen train_samples %llu\n", __func__, (long long unsigned) train->train_samples); - printf("%s: seen train_tokens %llu\n", __func__, (long long unsigned) train->train_tokens); - printf("%s: completed train_epochs %llu\n", __func__, (long long unsigned) train->train_epochs); - printf("%s: lora_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(lora.ctx) + ggml_backend_buffer_get_size(lora.data)), (float) (ggml_used_mem(lora.ctx) + ggml_backend_buffer_get_size(lora.data)) / (1024.0f*1024.0f)); - - if (params.only_write_lora) { - save_train_files_data save_data; - save_data.fn_checkpoint_out = ""; - save_data.fn_lora_out = params.fn_lora_out; - save_data.pattern_fn_it = params.common.pattern_fn_it; - save_data.fn_latest = params.common.fn_latest; - save_data.model = &model; - save_data.lora = &lora; - - save_train_files(&save_data, train); - - free_train_state(train); - ggml_free(lora.ctx); - llama_free(lctx); - llama_free_model(lmodel); - return 0; - } - - printf("%s: opt_size = %zu bytes (%.1f MB)\n", __func__, ggml_get_mem_size(opt->ctx), (float) ggml_get_mem_size(opt->ctx) / (1024.0f*1024.0f)); - printf("%s: opt iter %d\n", __func__, opt->iter); - - int n_tokens = model.hparams.n_ctx; - int n_vocab = model.hparams.n_vocab; - int n_batch = params.common.n_batch; - - // context for input tensors without their data - struct ggml_init_params ctx_input_params = { - ggml_tensor_overhead() * 2, // mem_size - NULL, // mem_buffer - true, // no_alloc - }; - struct ggml_context * ctx_input = ggml_init(ctx_input_params); - - // the input tensors - struct ggml_tensor * tokens_input = ggml_new_tensor_2d(ctx_input, GGML_TYPE_I32, n_tokens, n_batch); - struct ggml_tensor * target_probs = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab, n_tokens, n_batch); - - // allocate input tensors - // measure required memory for input tensors - ggml_backend_buffer_t input_data = ggml_backend_alloc_ctx_tensors_from_buft(ctx_input, ggml_backend_cpu_buffer_type()); - size_t max_input_size = ggml_backend_buffer_get_size(input_data); - printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f)); - - // context for compute tensors without their data - const size_t estimated_compute_size_wo_data = ( - 2*LLAMA_TRAIN_MAX_NODES*ggml_tensor_overhead() + - (params.common.use_checkpointing ? 3 : 2)*(GGML_OBJECT_SIZE+ggml_graph_overhead_custom(LLAMA_TRAIN_MAX_NODES, true)) - ); - struct ggml_init_params ctx_compute_params = { - estimated_compute_size_wo_data, // mem_size - NULL, // mem_buffer - true, // no_alloc - }; - struct ggml_context * ctx_compute = NULL; - - struct ggml_tensor * loss = NULL; - struct ggml_tensor * logits = NULL; - - struct ggml_cgraph * gf = NULL; - struct ggml_cgraph * gb = NULL; - struct ggml_cgraph * gb_tmp = NULL; - - // measure required memory for compute tensors - size_t best_compute_size = SIZE_MAX; - enum ggml_cgraph_eval_order best_order = GGML_CGRAPH_EVAL_ORDER_COUNT; - // find best evaluation order - for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) { - ctx_compute = ggml_init(ctx_compute_params); - ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type()); - gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); - gf->order = (enum ggml_cgraph_eval_order) order; - gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); - gb_tmp = params.common.use_checkpointing - ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true) - : NULL; - loss = llama_build_lora_finetune_graphs( - &model, &lora, alloc, ctx_compute, - gf, gb, gb_tmp, - &logits, tokens_input, target_probs, - n_tokens, n_batch, - params.common.use_flash, - params.common.use_checkpointing, - true - ); - size_t max_compute_size = ggml_gallocr_get_buffer_size(alloc, 0); // FIXME: this will still allocate the buffer - if (max_compute_size < best_compute_size) { - best_compute_size = max_compute_size; - best_order = gf->order; - } - ggml_gallocr_free(alloc); - ggml_free(ctx_compute); - } - size_t max_compute_size = best_compute_size; - printf("%s: compute_size = %zu bytes (%.1f MB)\n", __func__, max_compute_size, (float) max_compute_size / (1024.0f*1024.0f)); - printf("%s: evaluation order = %s\n", __func__, - (best_order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? "LEFT_TO_RIGHT" : - (best_order == GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT) ? "RIGHT_TO_LEFT" : - "invalid"); - - // allocate compute tensors - ctx_compute = ggml_init(ctx_compute_params); - ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type()); - gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); - gf->order = best_order; - gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); - gb_tmp = params.common.use_checkpointing - ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true) - : NULL; - loss = llama_build_lora_finetune_graphs( - &model, &lora, alloc, ctx_compute, - gf, gb, gb_tmp, - &logits, tokens_input, target_probs, - n_tokens, n_batch, - params.common.use_flash, - params.common.use_checkpointing, - false - ); - - // tokenize data - std::vector train_tokens; - std::vector train_samples_begin; - std::vector train_samples_size; - printf("%s: tokenize training data from %s\n", __func__, params.common.fn_train_data); - printf("%s: sample-start: %s\n", __func__, params.common.sample_start.c_str()); - printf("%s: include-sample-start: %s\n", __func__, params.common.include_sample_start ? "true" : "false"); - tokenize_file(lctx, - params.common.fn_train_data, - params.common.sample_start, - params.common.include_sample_start, - params.common.overlapping_samples, - n_tokens, - train_tokens, - train_samples_begin, - train_samples_size); - GGML_ASSERT(train_samples_begin.size() == train_samples_size.size()); - - printf("%s: number of training tokens: %zu\n", __func__, train_tokens.size()); - - std::vector token_noccurs; - token_noccurs.resize(model.hparams.n_vocab, 0); - for (unsigned int i = 0; i < train_tokens.size(); ++i) { - ++token_noccurs[train_tokens[i]]; - } - int n_unique_tokens = 0; - for (unsigned int i = 0; i < token_noccurs.size(); ++i) { - if (token_noccurs[i] == 0) continue; - ++n_unique_tokens; - } - printf("%s: number of unique tokens: %d\n", __func__, n_unique_tokens); - - size_t shuffle_samples_hash = compute_samples_hash(params.common.fn_train_data, train_samples_begin.data(), train_samples_size.data(), train_samples_size.size()); - const bool changed_train_data = (shuffle_samples_hash != train->shuffle_samples_hash) || (train->shuffle_sample_count != train_samples_size.size()); - if (changed_train_data) { - printf("%s: train data seems to have changed. restarting shuffled epoch.\n", __func__); - } - if (params.common.force_reshuffle) { - printf("%s: forced reshuffling of data. restarting with newly shuffled epoch.\n", __func__); - } - if ((train->shuffle_rng_state_current == "") || changed_train_data || params.common.force_reshuffle) { - train->shuffle_rng_state_current = mt19937_seed_to_state(params.common.seed); - train->shuffle_sample_count = train_samples_size.size(); - train->shuffle_next_sample = 0; - train->shuffle_samples_hash = shuffle_samples_hash; - } - std::vector train_shuffled_samples_offs; - std::vector train_shuffled_samples_begin; - std::vector train_shuffled_samples_size; - train_shuffled_samples_offs.resize(train_samples_begin.size()); - train_shuffled_samples_begin.resize(train_samples_begin.size()); - train_shuffled_samples_size.resize(train_samples_size.size()); - train->shuffle_rng_state_next = shuffle_samples( - train->shuffle_rng_state_current, - train_shuffled_samples_offs.data(), - train_shuffled_samples_begin.data(), - train_shuffled_samples_size.data(), - train_samples_begin.data(), - train_samples_size.data(), - train_samples_size.size()); - - printf("%s: begin training\n", __func__); - - save_train_files_data save_data; - save_data.fn_checkpoint_out = params.common.fn_checkpoint_out; - save_data.fn_lora_out = params.fn_lora_out; - save_data.pattern_fn_it = params.common.pattern_fn_it; - save_data.fn_latest = params.common.fn_latest; - save_data.model = &model; - save_data.lora = &lora; - - struct train_opt_callback_data opt_cb_data; - opt_cb_data.params = ¶ms.common; - opt_cb_data.train = train; - opt_cb_data.save_cb = &save_train_files; - opt_cb_data.save_data = &save_data; - opt_cb_data.lctx = lctx; - opt_cb_data.last_save_iter = opt->iter; - opt_cb_data.tokens_data = train_tokens.data(); - opt_cb_data.tokens_size = train_tokens.size(); - opt_cb_data.samples_begin = train_samples_begin.data(); - opt_cb_data.samples_size = train_samples_size.data(); - opt_cb_data.shuffled_samples_offs = train_shuffled_samples_offs.data(); - opt_cb_data.shuffled_samples_begin = train_shuffled_samples_begin.data(); - opt_cb_data.shuffled_samples_size = train_shuffled_samples_size.data(); - opt_cb_data.samples_count = train_samples_size.size(); - opt_cb_data.tokens_input = tokens_input; - opt_cb_data.target_probs = target_probs; - opt_cb_data.first_iter = opt->iter; - opt_cb_data.first_epoch = train->train_epochs; - opt_cb_data.iter_at_last_epoch = -1; - opt_cb_data.last_time = ggml_time_ms(); - opt_cb_data.millis_per_iter = 0.0; - - // measure required memory for work buffer - size_t max_work_size = ggml_graph_plan(gb, params.common.n_threads).work_size + GGML_OBJECT_SIZE; - printf("%s: work_size = %zu bytes (%.1f MB)\n", __func__, max_work_size, (float) max_work_size / (1024.0f*1024.0f)); - - // context for work buffer - struct ggml_init_params ctx_work_params = { - max_work_size, // mem_size - NULL, // mem_buffer - false, // no_alloc - }; - struct ggml_context * ctx_work = ggml_init(ctx_work_params); - - int64_t t0 = ggml_time_ms(); - - ggml_opt_resume_g(ctx_work, opt, loss, gf, gb, &train_opt_callback, (void *) &opt_cb_data); - - ggml_free(ctx_work); - ggml_free(ctx_compute); - ggml_free(ctx_input); - ggml_gallocr_free(alloc); - - - int64_t t1 = ggml_time_ms(); - printf("%s: total training time: ", __func__); - print_duration((double) (t1 - t0)); - printf("\n"); - - int new_iters = opt->iter - opt_cb_data.last_save_iter; - if (new_iters > 0) { - train->train_its += new_iters; - train->train_tokens += new_iters * opt->params.n_gradient_accumulation * n_batch * n_tokens; - - save_train_files(&save_data, train); - opt_cb_data.last_save_iter = opt->iter; - } - - ggml_free(opt->ctx); - free_train_state(train); - ggml_free(lora.ctx); - llama_free(lctx); - llama_free_model(lmodel); - return 0; -} diff --git a/examples/finetune/finetune.sh b/examples/finetune/finetune.sh deleted file mode 100644 index e3cc7f2712f17..0000000000000 --- a/examples/finetune/finetune.sh +++ /dev/null @@ -1,34 +0,0 @@ -#!/bin/bash -cd `dirname $0` -cd ../.. - -EXE="./llama-finetune" - -if [[ ! $LLAMA_MODEL_DIR ]]; then LLAMA_MODEL_DIR="./models"; fi -if [[ ! $LLAMA_TRAINING_DIR ]]; then LLAMA_TRAINING_DIR="."; fi - -# MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2-q8_0.gguf" # This is the model the readme uses. -MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2.gguf" # An f16 model. Note in this case with "-g", you get an f32-format .BIN file that isn't yet supported if you use it with "llama-cli --lora" with GPU inferencing. - -while getopts "dg" opt; do - case $opt in - d) - DEBUGGER="gdb --args" - ;; - g) - EXE="./build/bin/Release/finetune" - GPUARG="--gpu-layers 25" - ;; - esac -done - -$DEBUGGER $EXE \ - --model-base $MODEL \ - $GPUARG \ - --checkpoint-in chk-ol3b-shakespeare-LATEST.gguf \ - --checkpoint-out chk-ol3b-shakespeare-ITERATION.gguf \ - --lora-out lora-ol3b-shakespeare-ITERATION.bin \ - --train-data "$LLAMA_TRAINING_DIR\shakespeare.txt" \ - --save-every 10 \ - --threads 10 --adam-iter 30 --batch 4 --ctx 64 \ - --use-checkpointing diff --git a/examples/gbnf-validator/gbnf-validator.cpp b/examples/gbnf-validator/gbnf-validator.cpp index dd53ba9b1d551..48a705e15cea9 100644 --- a/examples/gbnf-validator/gbnf-validator.cpp +++ b/examples/gbnf-validator/gbnf-validator.cpp @@ -16,20 +16,25 @@ static bool llama_sample_grammar_string(struct llama_grammar * grammar, const st auto decoded = decode_utf8(input_str, {}); const auto & code_points = decoded.first; + const llama_grammar_rules & rules = llama_grammar_get_rules (grammar); + llama_grammar_stacks & cur_stacks = llama_grammar_get_stacks(grammar); + size_t pos = 0; for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) { - auto prev_stacks = grammar->stacks; - llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks); - if (grammar->stacks.empty()) { + const llama_grammar_stacks prev_stacks = llama_grammar_get_stacks(grammar); // copy + + llama_grammar_accept(rules, prev_stacks, *it, cur_stacks); + + if (cur_stacks.empty()) { error_pos = pos; error_msg = "Unexpected character '" + unicode_cpt_to_utf8(*it) + "'"; - grammar->stacks = prev_stacks; + cur_stacks = prev_stacks; return false; } ++pos; } - for (const auto & stack : grammar->stacks) { + for (const auto & stack : cur_stacks) { if (stack.empty()) { return true; } diff --git a/examples/gguf-hash/README.md b/examples/gguf-hash/README.md index 2320ffd8181b3..9871651e38ba8 100644 --- a/examples/gguf-hash/README.md +++ b/examples/gguf-hash/README.md @@ -201,6 +201,6 @@ Verification results for test.gguf.manifest - Success These micro c libraries dependencies was installed via the [clib c package manager](https://github.com/clibs) -- https://github.com/mofosyne/xxHash (From: https://github.com/Cyan4973/xxHash) +- https://github.com/Cyan4973/xxHash - https://github.com/clibs/sha1/ - https://github.com/jb55/sha256.c diff --git a/examples/gguf-hash/deps/xxhash/clib.json b/examples/gguf-hash/deps/xxhash/clib.json index 0048c74dc2741..242343c5d992d 100644 --- a/examples/gguf-hash/deps/xxhash/clib.json +++ b/examples/gguf-hash/deps/xxhash/clib.json @@ -1,7 +1,7 @@ { "name": "xxhash", "version": "0.8.2", - "repo": "mofosyne/xxhash", + "repo": "Cyan4973/xxhash", "description": "Extremely fast non-cryptographic hash algorithm", "keywords": ["xxhash", "hashing"], "license": "BSD-2-Clause", diff --git a/examples/gguf-hash/gguf-hash.cpp b/examples/gguf-hash/gguf-hash.cpp index c34728c3df4be..e96c75117f533 100644 --- a/examples/gguf-hash/gguf-hash.cpp +++ b/examples/gguf-hash/gguf-hash.cpp @@ -347,7 +347,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) { char hex_result[17]; for (int offset = 0; offset < 8; offset++) { unsigned int shift_bits_by = (8 * (8 - offset - 1)); - sprintf( ( hex_result + (2*offset)), "%02x", (unsigned char) (hash >> shift_bits_by)&0xff); + snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", (unsigned char) (hash >> shift_bits_by)&0xff); } if (hash_params.manifest_is_usable) { @@ -384,7 +384,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) { char hex_result[41] = {0}; for (int offset = 0; offset < 20; offset++) { - sprintf( ( hex_result + (2*offset)), "%02x", result[offset]&0xff); + snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff); } if (hash_params.manifest_is_usable) { @@ -421,7 +421,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) { char hex_result[SHA256_DIGEST_SIZE * 2 + 1] = {0}; for (int offset = 0; offset < SHA256_DIGEST_SIZE; offset++) { - sprintf( ( hex_result + (2*offset)), "%02x", result[offset]&0xff); + snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff); } if (hash_params.manifest_is_usable) { @@ -460,7 +460,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) { char hex_result[17]; for (int offset = 0; offset < 8; offset++) { unsigned int shift_bits_by = (8 * (8 - offset - 1)); - sprintf( ( hex_result + (2*offset)), "%02x", (unsigned char) (hash >> shift_bits_by)&0xff); + snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", (unsigned char) (hash >> shift_bits_by)&0xff); } if (hash_params.manifest_is_usable) { @@ -490,7 +490,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) { char hex_result[41]; for (int offset = 0; offset < 20; offset++) { - sprintf( ( hex_result + (2*offset)), "%02x", result[offset]&0xff); + snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff); } if (hash_params.manifest_is_usable) { @@ -520,7 +520,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) { char hex_result[SHA256_DIGEST_SIZE * 2 + 1] = {0}; for (int offset = 0; offset < SHA256_DIGEST_SIZE; offset++) { - sprintf( ( hex_result + (2*offset)), "%02x", result[offset]&0xff); + snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff); } if (hash_params.manifest_is_usable) { @@ -552,7 +552,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) { generate_uuidv5(result, uuid); char string_buffer[37] = {0}; - sprintf(string_buffer, "%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x", + snprintf(string_buffer, sizeof(string_buffer), "%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x", uuid[0], uuid[1], uuid[2], uuid[3], uuid[4], uuid[5], uuid[6], uuid[7], uuid[8], uuid[9], uuid[10], uuid[11], diff --git a/examples/gguf/gguf.cpp b/examples/gguf/gguf.cpp index 57514377107d7..7498f85efc4f9 100644 --- a/examples/gguf/gguf.cpp +++ b/examples/gguf/gguf.cpp @@ -92,6 +92,11 @@ static bool gguf_ex_read_0(const std::string & fname) { struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params); + if (!ctx) { + fprintf(stderr, "%s: failed to load '%s'\n", __func__, fname.c_str()); + return false; + } + printf("%s: version: %d\n", __func__, gguf_get_version(ctx)); printf("%s: alignment: %zu\n", __func__, gguf_get_alignment(ctx)); printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx)); diff --git a/examples/imatrix/README.md b/examples/imatrix/README.md index 29602881a0d21..bb5faec94c20a 100644 --- a/examples/imatrix/README.md +++ b/examples/imatrix/README.md @@ -1,6 +1,6 @@ # llama.cpp/examples/imatrix -Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantum models. +Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantized models. More information is available here: https://github.com/ggerganov/llama.cpp/pull/4861 ## Usage diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index d641a9f12b388..a6497b6e0bf82 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -23,6 +23,10 @@ #include "ggml-cuda.h" #include "ggml-sycl.h" +#ifdef GGML_USE_CANN +#include "ggml-cann.h" +#endif + // utils static uint64_t get_time_ns() { using clock = std::chrono::high_resolution_clock; @@ -120,6 +124,17 @@ static std::string get_gpu_info() { id += "/"; } } +#endif +#ifdef GGML_USE_CANN + uint32_t count = ggml_backend_cann_get_device_count(); + for (uint32_t i = 0; i < count; i++) { + char buf[128]; + ggml_backend_cann_get_device_description(i, buf, sizeof(buf)); + id += buf; + if (i < count - 1) { + id += "/"; + } + } #endif // TODO: other backends return id; diff --git a/examples/llama.android/llama/src/main/cpp/llama-android.cpp b/examples/llama.android/llama/src/main/cpp/llama-android.cpp index 92a6b16b1a00c..2aafe23167557 100644 --- a/examples/llama.android/llama/src/main/cpp/llama-android.cpp +++ b/examples/llama.android/llama/src/main/cpp/llama-android.cpp @@ -409,7 +409,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop( const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value); if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) { - return env->NewStringUTF(""); + return nullptr; } auto new_token_chars = llama_token_to_piece(context, new_token_id); diff --git a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift index 2a3f9f75890fc..58c32ca533bb1 100644 --- a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift +++ b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift @@ -26,11 +26,12 @@ actor LlamaContext { private var context: OpaquePointer private var batch: llama_batch private var tokens_list: [llama_token] + var is_done: Bool = false /// This variable is used to store temporarily invalid cchars private var temporary_invalid_cchars: [CChar] - var n_len: Int32 = 64 + var n_len: Int32 = 1024 var n_cur: Int32 = 0 var n_decode: Int32 = 0 @@ -160,6 +161,7 @@ actor LlamaContext { if llama_token_is_eog(model, new_token_id) || n_cur == n_len { print("\n") + is_done = true let new_token_str = String(cString: temporary_invalid_cchars + [0]) temporary_invalid_cchars.removeAll() return new_token_str diff --git a/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift b/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift index 2c1e3f61bad20..b8f6a31d582cd 100644 --- a/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift +++ b/examples/llama.swiftui/llama.swiftui/Models/LlamaState.swift @@ -132,7 +132,7 @@ class LlamaState: ObservableObject { messageLog += "\(text)" Task.detached { - while await llamaContext.n_cur < llamaContext.n_len { + while await !llamaContext.is_done { let result = await llamaContext.completion_loop() await MainActor.run { self.messageLog += "\(result)" diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index d6882eec31e1a..d23e282fbe655 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -16,6 +16,10 @@ #include "ggml-metal.h" #endif +#ifdef GGML_USE_CANN +#include "ggml-cann.h" +#endif + #define STB_IMAGE_IMPLEMENTATION #include "stb_image.h" @@ -1001,6 +1005,11 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { LOG_TEE("%s: CLIP using Metal backend\n", __func__); #endif +#ifdef GGML_USE_CANN + new_clip->backend = ggml_backend_cann_init(0); + LOG_TEE("%s: CLIP using CANN backend\n", __func__); +#endif + if (!new_clip->backend) { new_clip->backend = ggml_backend_cpu_init(); diff --git a/examples/lookup/lookup-stats.cpp b/examples/lookup/lookup-stats.cpp index 0b171c87273d1..2fe67100e6c03 100644 --- a/examples/lookup/lookup-stats.cpp +++ b/examples/lookup/lookup-stats.cpp @@ -31,7 +31,6 @@ int main(int argc, char ** argv){ // load the model std::tie(model, ctx) = llama_init_from_gpt_params(params); - GGML_ASSERT(llama_n_vocab(model) < (1 << 16)); // tokenize the prompt std::vector inp; @@ -65,7 +64,7 @@ int main(int argc, char ** argv){ } const int n_input = inp.size(); - const int n_ctx = params.n_ctx; + const int n_ctx = llama_n_ctx(ctx); int n_drafted = 0; int n_accept = 0; diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp index 80ecd925d5962..bb571bac4d778 100644 --- a/examples/lookup/lookup.cpp +++ b/examples/lookup/lookup.cpp @@ -39,7 +39,6 @@ int main(int argc, char ** argv){ // load the model std::tie(model, ctx) = llama_init_from_gpt_params(params); - GGML_ASSERT(llama_n_vocab(model) < (1 << 16)); // tokenize the prompt std::vector inp; diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 4ef55c1e6b3cf..61e960ea2abe6 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -124,6 +124,7 @@ static std::string chat_add_and_format(struct llama_model * model, std::vectorchat_template, chat_msgs, new_msg, role == "user"); chat_msgs.push_back({role, content}); + LOG("formatted: %s\n", formatted.c_str()); return formatted; } @@ -289,8 +290,13 @@ int main(int argc, char ** argv) { // Should not run without any tokens if (embd_inp.empty()) { - embd_inp.push_back(llama_token_bos(model)); - LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str()); + if (add_bos) { + embd_inp.push_back(llama_token_bos(model)); + LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str()); + } else { + LOG_TEE("error: input is empty\n"); + return -1; + } } // Tokenize negative prompt diff --git a/examples/pydantic_models_to_grammar.py b/examples/pydantic_models_to_grammar.py index d8145710ce250..93e5dcb6c3855 100644 --- a/examples/pydantic_models_to_grammar.py +++ b/examples/pydantic_models_to_grammar.py @@ -6,7 +6,7 @@ from copy import copy from enum import Enum from inspect import getdoc, isclass -from typing import TYPE_CHECKING, Any, Callable, List, Optional, Union, get_args, get_origin +from typing import TYPE_CHECKING, Any, Callable, List, Optional, Union, get_args, get_origin, get_type_hints from docstring_parser import parse from pydantic import BaseModel, create_model @@ -53,35 +53,38 @@ class PydanticDataType(Enum): def map_pydantic_type_to_gbnf(pydantic_type: type[Any]) -> str: - if isclass(pydantic_type) and issubclass(pydantic_type, str): + origin_type = get_origin(pydantic_type) + origin_type = pydantic_type if origin_type is None else origin_type + + if isclass(origin_type) and issubclass(origin_type, str): return PydanticDataType.STRING.value - elif isclass(pydantic_type) and issubclass(pydantic_type, bool): + elif isclass(origin_type) and issubclass(origin_type, bool): return PydanticDataType.BOOLEAN.value - elif isclass(pydantic_type) and issubclass(pydantic_type, int): + elif isclass(origin_type) and issubclass(origin_type, int): return PydanticDataType.INTEGER.value - elif isclass(pydantic_type) and issubclass(pydantic_type, float): + elif isclass(origin_type) and issubclass(origin_type, float): return PydanticDataType.FLOAT.value - elif isclass(pydantic_type) and issubclass(pydantic_type, Enum): + elif isclass(origin_type) and issubclass(origin_type, Enum): return PydanticDataType.ENUM.value - elif isclass(pydantic_type) and issubclass(pydantic_type, BaseModel): - return format_model_and_field_name(pydantic_type.__name__) - elif get_origin(pydantic_type) is list: + elif isclass(origin_type) and issubclass(origin_type, BaseModel): + return format_model_and_field_name(origin_type.__name__) + elif origin_type is list: element_type = get_args(pydantic_type)[0] return f"{map_pydantic_type_to_gbnf(element_type)}-list" - elif get_origin(pydantic_type) is set: + elif origin_type is set: element_type = get_args(pydantic_type)[0] return f"{map_pydantic_type_to_gbnf(element_type)}-set" - elif get_origin(pydantic_type) is Union: + elif origin_type is Union: union_types = get_args(pydantic_type) union_rules = [map_pydantic_type_to_gbnf(ut) for ut in union_types] return f"union-{'-or-'.join(union_rules)}" - elif get_origin(pydantic_type) is Optional: + elif origin_type is Optional: element_type = get_args(pydantic_type)[0] return f"optional-{map_pydantic_type_to_gbnf(element_type)}" - elif isclass(pydantic_type): - return f"{PydanticDataType.CUSTOM_CLASS.value}-{format_model_and_field_name(pydantic_type.__name__)}" - elif get_origin(pydantic_type) is dict: + elif isclass(origin_type): + return f"{PydanticDataType.CUSTOM_CLASS.value}-{format_model_and_field_name(origin_type.__name__)}" + elif origin_type is dict: key_type, value_type = get_args(pydantic_type) return f"custom-dict-key-type-{format_model_and_field_name(map_pydantic_type_to_gbnf(key_type))}-value-type-{format_model_and_field_name(map_pydantic_type_to_gbnf(value_type))}" else: @@ -118,7 +121,7 @@ def get_members_structure(cls, rule_name): # Modify this comprehension members = [ f' "\\"{name}\\"" ":" {map_pydantic_type_to_gbnf(param_type)}' - for name, param_type in cls.__annotations__.items() + for name, param_type in get_type_hints(cls).items() if name != "self" ] @@ -297,17 +300,20 @@ def generate_gbnf_rule_for_type( field_name = format_model_and_field_name(field_name) gbnf_type = map_pydantic_type_to_gbnf(field_type) - if isclass(field_type) and issubclass(field_type, BaseModel): + origin_type = get_origin(field_type) + origin_type = field_type if origin_type is None else origin_type + + if isclass(origin_type) and issubclass(origin_type, BaseModel): nested_model_name = format_model_and_field_name(field_type.__name__) nested_model_rules, _ = generate_gbnf_grammar(field_type, processed_models, created_rules) rules.extend(nested_model_rules) gbnf_type, rules = nested_model_name, rules - elif isclass(field_type) and issubclass(field_type, Enum): + elif isclass(origin_type) and issubclass(origin_type, Enum): enum_values = [f'"\\"{e.value}\\""' for e in field_type] # Adding escaped quotes enum_rule = f"{model_name}-{field_name} ::= {' | '.join(enum_values)}" rules.append(enum_rule) gbnf_type, rules = model_name + "-" + field_name, rules - elif get_origin(field_type) == list: # Array + elif origin_type is list: # Array element_type = get_args(field_type)[0] element_rule_name, additional_rules = generate_gbnf_rule_for_type( model_name, f"{field_name}-element", element_type, is_optional, processed_models, created_rules @@ -317,7 +323,7 @@ def generate_gbnf_rule_for_type( rules.append(array_rule) gbnf_type, rules = model_name + "-" + field_name, rules - elif get_origin(field_type) == set or field_type == set: # Array + elif origin_type is set: # Array element_type = get_args(field_type)[0] element_rule_name, additional_rules = generate_gbnf_rule_for_type( model_name, f"{field_name}-element", element_type, is_optional, processed_models, created_rules @@ -371,7 +377,7 @@ def generate_gbnf_rule_for_type( gbnf_type = f"{model_name}-{field_name}-optional" else: gbnf_type = f"{model_name}-{field_name}-union" - elif isclass(field_type) and issubclass(field_type, str): + elif isclass(origin_type) and issubclass(origin_type, str): if field_info and hasattr(field_info, "json_schema_extra") and field_info.json_schema_extra is not None: triple_quoted_string = field_info.json_schema_extra.get("triple_quoted_string", False) markdown_string = field_info.json_schema_extra.get("markdown_code_block", False) @@ -387,8 +393,8 @@ def generate_gbnf_rule_for_type( gbnf_type = PydanticDataType.STRING.value elif ( - isclass(field_type) - and issubclass(field_type, float) + isclass(origin_type) + and issubclass(origin_type, float) and field_info and hasattr(field_info, "json_schema_extra") and field_info.json_schema_extra is not None @@ -413,8 +419,8 @@ def generate_gbnf_rule_for_type( ) elif ( - isclass(field_type) - and issubclass(field_type, int) + isclass(origin_type) + and issubclass(origin_type, int) and field_info and hasattr(field_info, "json_schema_extra") and field_info.json_schema_extra is not None @@ -462,7 +468,7 @@ def generate_gbnf_grammar(model: type[BaseModel], processed_models: set[type[Bas if not issubclass(model, BaseModel): # For non-Pydantic classes, generate model_fields from __annotations__ or __init__ if hasattr(model, "__annotations__") and model.__annotations__: - model_fields = {name: (typ, ...) for name, typ in model.__annotations__.items()} # pyright: ignore[reportGeneralTypeIssues] + model_fields = {name: (typ, ...) for name, typ in get_type_hints(model).items()} else: init_signature = inspect.signature(model.__init__) parameters = init_signature.parameters @@ -470,7 +476,7 @@ def generate_gbnf_grammar(model: type[BaseModel], processed_models: set[type[Bas name != "self"} else: # For Pydantic models, use model_fields and check for ellipsis (required fields) - model_fields = model.__annotations__ + model_fields = get_type_hints(model) model_rule_parts = [] nested_rules = [] @@ -706,7 +712,7 @@ def generate_markdown_documentation( else: documentation += f" Fields:\n" # noqa: F541 if isclass(model) and issubclass(model, BaseModel): - for name, field_type in model.__annotations__.items(): + for name, field_type in get_type_hints(model).items(): # if name == "markdown_code_block": # continue if get_origin(field_type) == list: @@ -754,14 +760,17 @@ def generate_field_markdown( field_info = model.model_fields.get(field_name) field_description = field_info.description if field_info and field_info.description else "" - if get_origin(field_type) == list: + origin_type = get_origin(field_type) + origin_type = field_type if origin_type is None else origin_type + + if origin_type == list: element_type = get_args(field_type)[0] field_text = f"{indent}{field_name} ({format_model_and_field_name(field_type.__name__)} of {format_model_and_field_name(element_type.__name__)})" if field_description != "": field_text += ":\n" else: field_text += "\n" - elif get_origin(field_type) == Union: + elif origin_type == Union: element_types = get_args(field_type) types = [] for element_type in element_types: @@ -792,9 +801,9 @@ def generate_field_markdown( example_text = f"'{field_example}'" if isinstance(field_example, str) else field_example field_text += f"{indent} Example: {example_text}\n" - if isclass(field_type) and issubclass(field_type, BaseModel): + if isclass(origin_type) and issubclass(origin_type, BaseModel): field_text += f"{indent} Details:\n" - for name, type_ in field_type.__annotations__.items(): + for name, type_ in get_type_hints(field_type).items(): field_text += generate_field_markdown(name, type_, field_type, depth + 2) return field_text @@ -855,7 +864,7 @@ def generate_text_documentation( if isclass(model) and issubclass(model, BaseModel): documentation_fields = "" - for name, field_type in model.__annotations__.items(): + for name, field_type in get_type_hints(model).items(): # if name == "markdown_code_block": # continue if get_origin(field_type) == list: @@ -948,7 +957,7 @@ def generate_field_text( if isclass(field_type) and issubclass(field_type, BaseModel): field_text += f"{indent} Details:\n" - for name, type_ in field_type.__annotations__.items(): + for name, type_ in get_type_hints(field_type).items(): field_text += generate_field_text(name, type_, field_type, depth + 2) return field_text diff --git a/examples/pydantic_models_to_grammar_examples.py b/examples/pydantic_models_to_grammar_examples.py old mode 100644 new mode 100755 index 8e7f46cf99a43..eb000d5ccba24 --- a/examples/pydantic_models_to_grammar_examples.py +++ b/examples/pydantic_models_to_grammar_examples.py @@ -1,8 +1,15 @@ -# Function calling example using pydantic models. +#!/usr/bin/env python3 + +"""Function calling example using pydantic models.""" + from __future__ import annotations +import argparse import datetime import json +import logging +import textwrap +import sys from enum import Enum from typing import Optional, Union @@ -12,28 +19,54 @@ create_dynamic_model_from_function, generate_gbnf_grammar_and_documentation) -# Function to get completion on the llama.cpp server with grammar. -def create_completion(prompt, grammar): - headers = {"Content-Type": "application/json"} - data = {"prompt": prompt, "grammar": grammar} +def create_completion(host, prompt, gbnf_grammar): + """Calls the /completion API on llama-server. - response = requests.post("http://127.0.0.1:8080/completion", headers=headers, json=data) - data = response.json() - - print(data["content"]) - return data["content"] + See + https://github.com/ggerganov/llama.cpp/tree/HEAD/examples/server#api-endpoints + """ + print(f" Request:\n Grammar:\n{textwrap.indent(gbnf_grammar, ' ')}\n Prompt:\n{textwrap.indent(prompt.rstrip(), ' ')}") + headers = {"Content-Type": "application/json"} + data = {"prompt": prompt, "grammar": gbnf_grammar} + result = requests.post(f"http://{host}/completion", headers=headers, json=data).json() + assert data.get("error") is None, data + logging.info("Result: %s", result) + content = result["content"] + print(f" Model: {result['model']}") + print(f" Result:\n{textwrap.indent(json.dumps(json.loads(content), indent=2), ' ')}") + return content # A function for the agent to send a message to the user. class SendMessageToUser(BaseModel): - """ - Send a message to the User. - """ + """Send a message to the User.""" chain_of_thought: str = Field(..., description="Your chain of thought while sending the message.") message: str = Field(..., description="Message you want to send to the user.") def run(self): - print(self.message) + print(f"SendMessageToUser: {self.message}") + + +def example_rce(host): + """Minimal test case where the LLM call an arbitrary python function.""" + print("- example_rce") + tools = [SendMessageToUser] + gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation( + pydantic_model_list=tools, outer_object_name="function", + outer_object_content="function_parameters", model_prefix="Function", fields_prefix="Parameters") + system_message = "You are an advanced AI, tasked to assist the user by calling functions in JSON format. The following are the available functions and their parameters and types:\n\n" + documentation + user_message = "What is 42 * 42?" + prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant" + text = create_completion(host, prompt, gbnf_grammar) + json_data = json.loads(text) + tools_map = {tool.__name__:tool for tool in tools} + # This finds "SendMessageToUser": + tool = tools_map.get(json_data["function"]) + if not tool: + print(f"Error: unknown tool {json_data['function']}") + return 1 + tool(**json_data["function_parameters"]).run() + return 0 # Enum for the calculator tool. @@ -44,11 +77,11 @@ class MathOperation(Enum): DIVIDE = "divide" -# Simple pydantic calculator tool for the agent that can add, subtract, multiply, and divide. Docstring and description of fields will be used in system prompt. +# Simple pydantic calculator tool for the agent that can add, subtract, +# multiply, and divide. Docstring and description of fields will be used in +# system prompt. class Calculator(BaseModel): - """ - Perform a math operation on two numbers. - """ + """Perform a math operation on two numbers.""" number_one: Union[int, float] = Field(..., description="First number.") operation: MathOperation = Field(..., description="Math operation to perform.") number_two: Union[int, float] = Field(..., description="Second number.") @@ -66,55 +99,61 @@ def run(self): raise ValueError("Unknown operation.") -# Here the grammar gets generated by passing the available function models to generate_gbnf_grammar_and_documentation function. This also generates a documentation usable by the LLM. -# pydantic_model_list is the list of pydanitc models -# outer_object_name is an optional name for an outer object around the actual model object. Like a "function" object with "function_parameters" which contains the actual model object. If None, no outer object will be generated -# outer_object_content is the name of outer object content. -# model_prefix is the optional prefix for models in the documentation. (Default="Output Model") -# fields_prefix is the prefix for the model fields in the documentation. (Default="Output Fields") -gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation( - pydantic_model_list=[SendMessageToUser, Calculator], outer_object_name="function", - outer_object_content="function_parameters", model_prefix="Function", fields_prefix="Parameters") +def example_calculator(host): + """Have the LLM ask to get a calculation done. -print(gbnf_grammar) -print(documentation) + Here the grammar gets generated by passing the available function models to + generate_gbnf_grammar_and_documentation function. This also generates a + documentation usable by the LLM. -system_message = "You are an advanced AI, tasked to assist the user by calling functions in JSON format. The following are the available functions and their parameters and types:\n\n" + documentation + pydantic_model_list is the list of pydantic models outer_object_name is an + optional name for an outer object around the actual model object. Like a + "function" object with "function_parameters" which contains the actual model + object. If None, no outer object will be generated outer_object_content is + the name of outer object content. -user_message = "What is 42 * 42?" -prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant" - -text = create_completion(prompt=prompt, grammar=gbnf_grammar) -# This should output something like this: -# { -# "function": "calculator", -# "function_parameters": { -# "number_one": 42, -# "operation": "multiply", -# "number_two": 42 -# } -# } -function_dictionary = json.loads(text) -if function_dictionary["function"] == "calculator": - function_parameters = {**function_dictionary["function_parameters"]} - - print(Calculator(**function_parameters).run()) - # This should output: 1764 + model_prefix is the optional prefix for models in the documentation. (Default="Output Model") + fields_prefix is the prefix for the model fields in the documentation. (Default="Output Fields") + """ + print("- example_calculator") + tools = [SendMessageToUser, Calculator] + gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation( + pydantic_model_list=tools, outer_object_name="function", + outer_object_content="function_parameters", model_prefix="Function", fields_prefix="Parameters") + system_message = "You are an advanced AI, tasked to assist the user by calling functions in JSON format. The following are the available functions and their parameters and types:\n\n" + documentation + user_message1 = "What is 42 * 42?" + prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message1}<|im_end|>\n<|im_start|>assistant" + text = create_completion(host, prompt, gbnf_grammar) + json_data = json.loads(text) + expected = { + "function": "Calculator", + "function_parameters": { + "number_one": 42, + "operation": "multiply", + "number_two": 42 + } + } + if json_data != expected: + print(" Result is not as expected!") + tools_map = {tool.__name__:tool for tool in tools} + # This finds "Calculator": + tool = tools_map.get(json_data["function"]) + if not tool: + print(f"Error: unknown tool {json_data['function']}") + return 1 + result = tool(**json_data["function_parameters"]).run() + print(f" Call {json_data['function']} gave result {result}") + return 0 -# A example structured output based on pydantic models. The LLM will create an entry for a Book database out of an unstructured text. class Category(Enum): - """ - The category of the book. - """ + """The category of the book.""" Fiction = "Fiction" NonFiction = "Non-Fiction" class Book(BaseModel): - """ - Represents an entry about a book. - """ + """Represents an entry about a book.""" title: str = Field(..., description="Title of the book.") author: str = Field(..., description="Author of the book.") published_year: Optional[int] = Field(..., description="Publishing year of the book.") @@ -123,33 +162,42 @@ class Book(BaseModel): summary: str = Field(..., description="Summary of the book.") -# We need no additional parameters other than our list of pydantic models. -gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation([Book]) - -system_message = "You are an advanced AI, tasked to create a dataset entry in JSON for a Book. The following is the expected output model:\n\n" + documentation - -text = """The Feynman Lectures on Physics is a physics textbook based on some lectures by Richard Feynman, a Nobel laureate who has sometimes been called "The Great Explainer". The lectures were presented before undergraduate students at the California Institute of Technology (Caltech), during 1961–1963. The book's co-authors are Feynman, Robert B. Leighton, and Matthew Sands.""" -prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant" - -text = create_completion(prompt=prompt, grammar=gbnf_grammar) +def example_struct(host): + """A example structured output based on pydantic models. -json_data = json.loads(text) + The LLM will create an entry for a Book database out of an unstructured + text. We need no additional parameters other than our list of pydantic + models. + """ + print("- example_struct") + tools = [Book] + gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation(pydantic_model_list=tools) + system_message = "You are an advanced AI, tasked to create a dataset entry in JSON for a Book. The following is the expected output model:\n\n" + documentation + text = """The Feynman Lectures on Physics is a physics textbook based on some lectures by Richard Feynman, a Nobel laureate who has sometimes been called "The Great Explainer". The lectures were presented before undergraduate students at the California Institute of Technology (Caltech), during 1961–1963. The book's co-authors are Feynman, Robert B. Leighton, and Matthew Sands.""" + prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant" + text = create_completion(host, prompt, gbnf_grammar) + json_data = json.loads(text) + # In this case, there's no function nor function_parameters. + # Here the result will vary based on the LLM used. + keys = sorted(["title", "author", "published_year", "keywords", "category", "summary"]) + if keys != sorted(json_data.keys()): + print(f"Unexpected result: {sorted(json_data.keys())}") + return 1 + book = Book(**json_data) + print(f" As a Book object: %s" % book) + return 0 -print(Book(**json_data)) -# An example for parallel function calling with a Python function, a pydantic function model and an OpenAI like function definition. def get_current_datetime(output_format: Optional[str] = None): - """ - Get the current date and time in the given format. + """Get the current date and time in the given format. + Args: output_format: formatting string for the date and time, defaults to '%Y-%m-%d %H:%M:%S' """ - if output_format is None: - output_format = '%Y-%m-%d %H:%M:%S' - return datetime.datetime.now().strftime(output_format) + return datetime.datetime.now().strftime(output_format or "%Y-%m-%d %H:%M:%S") -# Example function to get the weather +# Example function to get the weather. def get_current_weather(location, unit): """Get the current weather in a given location""" if "London" in location: @@ -158,68 +206,107 @@ def get_current_weather(location, unit): return json.dumps({"location": "New York", "temperature": "24", "unit": unit.value}) elif "North Pole" in location: return json.dumps({"location": "North Pole", "temperature": "-42", "unit": unit.value}) - else: - return json.dumps({"location": location, "temperature": "unknown"}) - - -# Here is a function definition in OpenAI style -current_weather_tool = { - "type": "function", - "function": { - "name": "get_current_weather", - "description": "Get the current weather in a given location", - "parameters": { - "type": "object", - "properties": { - "location": { - "type": "string", - "description": "The city and state, e.g. San Francisco, CA", + return json.dumps({"location": location, "temperature": "unknown"}) + + +def example_concurrent(host): + """An example for parallel function calling with a Python function, a pydantic + function model and an OpenAI like function definition. + """ + print("- example_concurrent") + # Function definition in OpenAI style. + current_weather_tool = { + "type": "function", + "function": { + "name": "get_current_weather", + "description": "Get the current weather in a given location", + "parameters": { + "type": "object", + "properties": { + "location": { + "type": "string", + "description": "The city and state, e.g. San Francisco, CA", + }, + "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}, }, - "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}, + "required": ["location"], }, - "required": ["location"], }, - }, -} - -# Convert OpenAI function definition into pydantic model -current_weather_tool_model = convert_dictionary_to_pydantic_model(current_weather_tool) -# Add the actual function to a pydantic model -current_weather_tool_model = add_run_method_to_dynamic_model(current_weather_tool_model, get_current_weather) - -# Convert normal Python function to a pydantic model -current_datetime_model = create_dynamic_model_from_function(get_current_datetime) - -tool_list = [SendMessageToUser, Calculator, current_datetime_model, current_weather_tool_model] - - -gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation( - pydantic_model_list=tool_list, outer_object_name="function", - outer_object_content="params", model_prefix="Function", fields_prefix="Parameters", list_of_outputs=True) - -system_message = "You are an advanced AI assistant. You are interacting with the user and with your environment by calling functions. You call functions by writing JSON objects, which represent specific function calls.\nBelow is a list of your available function calls:\n\n" + documentation - - -text = """Get the date and time, get the current weather in celsius in London and solve the following calculation: 42 * 42""" -prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant" - -text = create_completion(prompt=prompt, grammar=gbnf_grammar) - -json_data = json.loads(text) - -print(json_data) -# Should output something like this: -# [{'function': 'get_current_datetime', 'params': {'output_format': '%Y-%m-%d %H:%M:%S'}}, {'function': 'get_current_weather', 'params': {'location': 'London', 'unit': 'celsius'}}, {'function': 'Calculator', 'params': {'number_one': 42, 'operation': 'multiply', 'number_two': 42}}] - - -for call in json_data: - if call["function"] == "Calculator": - print(Calculator(**call["params"]).run()) - elif call["function"] == "get_current_datetime": - print(current_datetime_model(**call["params"]).run()) # pyright: ignore[reportAttributeAccessIssue] - elif call["function"] == "get_current_weather": - print(current_weather_tool_model(**call["params"]).run()) # pyright: ignore[reportAttributeAccessIssue] -# Should output something like this: -# 2024-01-14 13:36:06 -# {"location": "London", "temperature": "42", "unit": "celsius"} -# 1764 + } + # Convert OpenAI function definition into pydantic model. + current_weather_tool_model = convert_dictionary_to_pydantic_model(current_weather_tool) + # Add the actual function to a pydantic model. + current_weather_tool_model = add_run_method_to_dynamic_model(current_weather_tool_model, get_current_weather) + + # Convert normal Python function to a pydantic model. + current_datetime_model = create_dynamic_model_from_function(get_current_datetime) + + tools = [SendMessageToUser, Calculator, current_datetime_model, current_weather_tool_model] + gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation( + pydantic_model_list=tools, outer_object_name="function", + outer_object_content="params", model_prefix="Function", fields_prefix="Parameters", list_of_outputs=True) + system_message = "You are an advanced AI assistant. You are interacting with the user and with your environment by calling functions. You call functions by writing JSON objects, which represent specific function calls.\nBelow is a list of your available function calls:\n\n" + documentation + text = """Get the date and time, get the current weather in celsius in London and solve the following calculation: 42 * 42""" + prompt = f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant" + text = create_completion(host, prompt, gbnf_grammar) + json_data = json.loads(text) + expected = [ + { + "function": "get_current_datetime", + "params": { + "output_format": "%Y-%m-%d %H:%M:%S" + } + }, + { + "function": "get_current_weather", + "params": { + "location": "London", + "unit": "celsius" + } + }, + { + "function": "Calculator", + "params": { + "number_one": 42, + "operation": "multiply", + "number_two": 42 + } + } + ] + res = 0 + if json_data != expected: + print(" Result is not as expected!") + print(" This can happen on highly quantized models") + res = 1 + tools_map = {tool.__name__:tool for tool in tools} + for call in json_data: + tool = tools_map.get(call["function"]) + if not tool: + print(f"Error: unknown tool {call['function']}") + return 1 + result = tool(**call["params"]).run() + print(f" Call {call['function']} returned {result}") + # Should output something like this: + # Call get_current_datetime returned 2024-07-15 09:50:38 + # Call get_current_weather returned {"location": "London", "temperature": "42", "unit": "celsius"} + # Call Calculator returned 1764 + return res + + +def main(): + parser = argparse.ArgumentParser(description=sys.modules[__name__].__doc__) + parser.add_argument("--host", default="localhost:8080", help="llama.cpp server") + parser.add_argument("-v", "--verbose", action="store_true", help="enables logging") + args = parser.parse_args() + logging.basicConfig(level=logging.INFO if args.verbose else logging.ERROR) + ret = 0 + # Comment out below to only run the example you want. + ret = ret or example_rce(args.host) + ret = ret or example_calculator(args.host) + ret = ret or example_struct(args.host) + ret = ret or example_concurrent(args.host) + return ret + + +if __name__ == "__main__": + sys.exit(main()) diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp index 746df8446b85e..68cf8d3595e87 100644 --- a/examples/quantize-stats/quantize-stats.cpp +++ b/examples/quantize-stats/quantize-stats.cpp @@ -154,7 +154,7 @@ static void test_roundtrip_on_chunk( } if (use_reference) { - qfns.from_float_reference(input_scratch, quantized_scratch, chunk_size); + qfns.from_float_ref(input_scratch, quantized_scratch, chunk_size); } else { qfns.from_float(input_scratch, quantized_scratch, chunk_size); } diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 76e2052d55d79..8d764725878b9 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -16,41 +16,44 @@ struct quant_option { }; static const std::vector QUANT_OPTIONS = { - { "Q4_0", LLAMA_FTYPE_MOSTLY_Q4_0, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, - { "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 4.78G, +0.4511 ppl @ Llama-3-8B", }, - { "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 5.21G, +0.1316 ppl @ Llama-3-8B", }, - { "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 5.65G, +0.1062 ppl @ Llama-3-8B", }, - { "IQ2_XXS",LLAMA_FTYPE_MOSTLY_IQ2_XXS," 2.06 bpw quantization", }, - { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization", }, - { "IQ2_S", LLAMA_FTYPE_MOSTLY_IQ2_S, " 2.5 bpw quantization", }, - { "IQ2_M", LLAMA_FTYPE_MOSTLY_IQ2_M, " 2.7 bpw quantization", }, - { "IQ1_S", LLAMA_FTYPE_MOSTLY_IQ1_S, " 1.56 bpw quantization", }, - { "IQ1_M", LLAMA_FTYPE_MOSTLY_IQ1_M, " 1.75 bpw quantization", }, - { "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.96G, +3.5199 ppl @ Llama-3-8B", }, - { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.96G, +3.1836 ppl @ Llama-3-8B", }, - { "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization", }, - { "IQ3_S", LLAMA_FTYPE_MOSTLY_IQ3_S, " 3.44 bpw quantization", }, - { "IQ3_M", LLAMA_FTYPE_MOSTLY_IQ3_M, " 3.66 bpw quantization mix", }, - { "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" }, - { "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw quantization", }, - { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 3.41G, +1.6321 ppl @ Llama-3-8B", }, - { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.74G, +0.6569 ppl @ Llama-3-8B", }, - { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 4.03G, +0.5562 ppl @ Llama-3-8B", }, - { "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.50 bpw non-linear quantization", }, - { "IQ4_XS", LLAMA_FTYPE_MOSTLY_IQ4_XS, " 4.25 bpw non-linear quantization", }, - { "Q4_K", LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", }, - { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 4.37G, +0.2689 ppl @ Llama-3-8B", }, - { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 4.58G, +0.1754 ppl @ Llama-3-8B", }, - { "Q5_K", LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", }, - { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 5.21G, +0.1049 ppl @ Llama-3-8B", }, - { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 5.33G, +0.0569 ppl @ Llama-3-8B", }, - { "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 6.14G, +0.0217 ppl @ Llama-3-8B", }, - { "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 7.96G, +0.0026 ppl @ Llama-3-8B", }, - { "F16", LLAMA_FTYPE_MOSTLY_F16, "14.00G, +0.0020 ppl @ Mistral-7B", }, - { "BF16", LLAMA_FTYPE_MOSTLY_BF16, "14.00G, -0.0050 ppl @ Mistral-7B", }, - { "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", }, + { "Q4_0", LLAMA_FTYPE_MOSTLY_Q4_0, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, + { "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 4.78G, +0.4511 ppl @ Llama-3-8B", }, + { "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 5.21G, +0.1316 ppl @ Llama-3-8B", }, + { "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 5.65G, +0.1062 ppl @ Llama-3-8B", }, + { "IQ2_XXS", LLAMA_FTYPE_MOSTLY_IQ2_XXS, " 2.06 bpw quantization", }, + { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization", }, + { "IQ2_S", LLAMA_FTYPE_MOSTLY_IQ2_S, " 2.5 bpw quantization", }, + { "IQ2_M", LLAMA_FTYPE_MOSTLY_IQ2_M, " 2.7 bpw quantization", }, + { "IQ1_S", LLAMA_FTYPE_MOSTLY_IQ1_S, " 1.56 bpw quantization", }, + { "IQ1_M", LLAMA_FTYPE_MOSTLY_IQ1_M, " 1.75 bpw quantization", }, + { "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.96G, +3.5199 ppl @ Llama-3-8B", }, + { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.96G, +3.1836 ppl @ Llama-3-8B", }, + { "IQ3_XXS", LLAMA_FTYPE_MOSTLY_IQ3_XXS, " 3.06 bpw quantization", }, + { "IQ3_S", LLAMA_FTYPE_MOSTLY_IQ3_S, " 3.44 bpw quantization", }, + { "IQ3_M", LLAMA_FTYPE_MOSTLY_IQ3_M, " 3.66 bpw quantization mix", }, + { "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" }, + { "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw quantization", }, + { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 3.41G, +1.6321 ppl @ Llama-3-8B", }, + { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.74G, +0.6569 ppl @ Llama-3-8B", }, + { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 4.03G, +0.5562 ppl @ Llama-3-8B", }, + { "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.50 bpw non-linear quantization", }, + { "IQ4_XS", LLAMA_FTYPE_MOSTLY_IQ4_XS, " 4.25 bpw non-linear quantization", }, + { "Q4_K", LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", }, + { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 4.37G, +0.2689 ppl @ Llama-3-8B", }, + { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 4.58G, +0.1754 ppl @ Llama-3-8B", }, + { "Q5_K", LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", }, + { "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 5.21G, +0.1049 ppl @ Llama-3-8B", }, + { "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 5.33G, +0.0569 ppl @ Llama-3-8B", }, + { "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 6.14G, +0.0217 ppl @ Llama-3-8B", }, + { "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 7.96G, +0.0026 ppl @ Llama-3-8B", }, + { "Q4_0_4_4", LLAMA_FTYPE_MOSTLY_Q4_0_4_4, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, + { "Q4_0_4_8", LLAMA_FTYPE_MOSTLY_Q4_0_4_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, + { "Q4_0_8_8", LLAMA_FTYPE_MOSTLY_Q4_0_8_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", }, + { "F16", LLAMA_FTYPE_MOSTLY_F16, "14.00G, +0.0020 ppl @ Mistral-7B", }, + { "BF16", LLAMA_FTYPE_MOSTLY_BF16, "14.00G, -0.0050 ppl @ Mistral-7B", }, + { "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", }, // Note: Ensure COPY comes after F32 to avoid ftype 0 from matching. - { "COPY", LLAMA_FTYPE_ALL_F32, "only copy tensors, no quantizing", }, + { "COPY", LLAMA_FTYPE_ALL_F32, "only copy tensors, no quantizing", }, }; static const char * const LLM_KV_QUANTIZE_IMATRIX_FILE = "quantize.imatrix.file"; diff --git a/examples/server/README.md b/examples/server/README.md index cb45ee06dd0a0..33a2b95cc0edb 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -5,7 +5,7 @@ Fast, lightweight, pure C/C++ HTTP server based on [httplib](https://github.com/ Set of LLM REST APIs and a simple web front end to interact with llama.cpp. **Features:** - * LLM inference of F16 and quantum models on GPU and CPU + * LLM inference of F16 and quantized models on GPU and CPU * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes * Parallel decoding with multi-user support * Continuous batching @@ -15,69 +15,281 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp. The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216). -**Command line options:** - -- `-v`, `--verbose`: Enable verbose server output. When using the `/completion` endpoint, this includes the tokenized prompt, the full request and the full response. -- `-t N`, `--threads N`: Set the number of threads to use by CPU layers during generation. Not used by model layers that are offloaded to GPU. This option has no effect when using the maximum number of GPU layers. Default: `std::thread::hardware_concurrency()` (number of CPU cores). -- `-tb N, --threads-batch N`: Set the number of threads to use by CPU layers during batch and prompt processing (>= 32 tokens). This option has no effect if a GPU is available. Default: `--threads`. -- `--threads-http N`: Number of threads in the http server pool to process requests. Default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)` -- `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`). -- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file. Default: unused -- `-hfr REPO, --hf-repo REPO`: Hugging Face model repository. Default: unused -- `-hff FILE, --hf-file FILE`: Hugging Face model file. Default: unused -- `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses. -- `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is `512`, but LLaMA models were built with a context of `2048`, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of `4096`. -- `-ngl N`, `--n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance. -- `-mg i, --main-gpu i`: When using multiple GPUs, this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default, GPU `0` is used. -- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs, this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default, the data is split in proportion to VRAM, but this may not be optimal for performance. -- `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `2048` -- `-ub N`, `--ubatch-size N`: Physical maximum batch size. Default: `512` -- `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped. -- `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed. -- `--numa STRATEGY`: Attempt one of the below optimization strategies that may help on some NUMA systems -- `--numa distribute`: Spread execution evenly over all nodes -- `--numa isolate`: Only spawn threads on CPUs on the node that execution started on -- `--numa numactl`: Use the CPU map provided by numactl. If run without this previously, it is recommended to drop the system page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/1437 -- `--numa`: Attempt optimizations that may help on some NUMA systems. -- `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains. -- `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation. -- `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600` -- `--host`: Set the hostname or ip address to listen. Default `127.0.0.1` -- `--port`: Set the port to listen. Default: `8080` -- `--path`: Path from which to serve static files. Default: disabled -- `--api-key`: Set an api key for request authorization. By default, the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys. -- `--api-key-file`: Path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`s. -- `--embeddings`: Enable embedding vector output and the OAI compatible endpoint /v1/embeddings. Physical batch size (`--ubatch-size`) must be carefully defined. Default: disabled -- `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`. Values > 1 will allow for higher throughput with multiple parallel requests but the results will **not** be deterministic due to differences in rounding error. -- `-cb`, `--cont-batching`: Enable continuous batching (a.k.a dynamic batching). Default: disabled -- `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load a system prompt (initial prompt of all slots). This is useful for chat applications. [See more](#change-system-prompt-on-runtime) -- `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA. -- `--grp-attn-n`: Set the group attention factor to extend context size through self-extend. Used together with group attention width `--grp-attn-w`. Default: `1`, which is disabled. -- `--grp-attn-w`: Set the group attention width to extend context size through self-extend. Used together with group attention factor `--grp-attn-n`. Default: `512` -- `-n N, --n-predict N`: Set the maximum tokens to predict. Default: `-1` -- `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included. -- `--metrics`: enable prometheus `/metrics` compatible endpoint. Default: disabled -- `--slot-save-path PATH`: Specifies the path where the state of slots (the prompt cache) can be stored. If not provided, the slot management endpoints will be disabled. -- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name. Default: template taken from model's metadata. We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) -- `--log-disable`: Output logs to stdout only, not to `llama.log`. Default: enabled -- `--log-format FORMAT`: Define the log output to FORMAT: json or text Default: `json` -- `--rope-scaling` : RoPE scaling method. Defaults to linear unless otherwise specified by the model. Options are `none`, `linear`, `yarn` -- `--rope-freq-base N` : RoPE frequency base (default: loaded from model) -- `--rope-freq-scale N`: RoPE frequency scaling factor, expands context by a factor of 1/N (e.g. 0.25) -- `--yarn-ext-factor N` : YaRN: extrapolation mix factor (Default: 1.0, 0.0 = full interpolation) -- `--yarn-attn-factor N` : YaRN: scale sqrt(t) or attention magnitude (default: 1.0) -- `--yarn-beta-slow N`: YaRN: High correction dim or alpha (default: 1.0) -- `--yarn-beta-fast N`: YaRN: low correction dim or beta (default: 32.0) -- `--pooling` : Pooling type for embeddings, use model default if unspecified. Options are `none`, `mean`, `cls` -- `-dt N`, `--defrag-thold N`: KV cache defragmentation threshold (default: -1.0, < 0 = disabled) -- `-fa`, `--flash-attn` : enable flash attention (default: disabled). -- `-ctk TYPE`, `--cache-type-k TYPE` : KV cache data type for K (default: `f16`, options `f32`, `f16`, `q8_0`, `q4_0`, `q4_1`, `iq4_nl`, `q5_0`, or `q5_1`) -- `-ctv TYPE`, `--cache-type-v TYPE` : KV cache type for V (default `f16`, see `-ctk` for options) -- `--spm-infill` : Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. - -**If compiled with `LLAMA_SERVER_SSL=ON`** -- `--ssl-key-file FNAME`: path to file a PEM-encoded SSL private key -- `--ssl-cert-file FNAME`: path to file a PEM-encoded SSL certificate +## Usage + +``` +usage: ./llama-server [options] + +general: + + -h, --help, --usage print usage and exit + --version show version and build info + -v, --verbose print verbose information + --verbosity N set specific verbosity level (default: 0) + --verbose-prompt print a verbose prompt before generation (default: false) + --no-display-prompt don't print prompt at generation (default: false) + -co, --color colorise output to distinguish prompt and user input from generations (default: false) + -s, --seed SEED RNG seed (default: -1, use random seed for < 0) + -t, --threads N number of threads to use during generation (default: 8) + -tb, --threads-batch N number of threads to use during batch and prompt processing (default: same as --threads) + -td, --threads-draft N number of threads to use during generation (default: same as --threads) + -tbd, --threads-batch-draft N number of threads to use during batch and prompt processing (default: same as --threads-draft) + --draft N number of tokens to draft for speculative decoding (default: 5) + -ps, --p-split N speculative decoding split probability (default: 0.1) + -lcs, --lookup-cache-static FNAME + path to static lookup cache to use for lookup decoding (not updated by generation) + -lcd, --lookup-cache-dynamic FNAME + path to dynamic lookup cache to use for lookup decoding (updated by generation) + -c, --ctx-size N size of the prompt context (default: 0, 0 = loaded from model) + -n, --predict N number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled) + -b, --batch-size N logical maximum batch size (default: 2048) + -ub, --ubatch-size N physical maximum batch size (default: 512) + --keep N number of tokens to keep from the initial prompt (default: 0, -1 = all) + --chunks N max number of chunks to process (default: -1, -1 = all) + -fa, --flash-attn enable Flash Attention (default: disabled) + -p, --prompt PROMPT prompt to start generation with + in conversation mode, this will be used as system prompt + (default: '') + -f, --file FNAME a file containing the prompt (default: none) + --in-file FNAME an input file (repeat to specify multiple files) + -bf, --binary-file FNAME binary file containing the prompt (default: none) + -e, --escape process escapes sequences (\n, \r, \t, \', \", \\) (default: true) + --no-escape do not process escape sequences + -ptc, --print-token-count N print token count every N tokens (default: -1) + --prompt-cache FNAME file to cache prompt state for faster startup (default: none) + --prompt-cache-all if specified, saves user input and generations to cache as well + not supported with --interactive or other interactive options + --prompt-cache-ro if specified, uses the prompt cache but does not update it + -r, --reverse-prompt PROMPT halt generation at PROMPT, return control in interactive mode + can be specified more than once for multiple prompts + -sp, --special special tokens output enabled (default: false) + -cnv, --conversation run in conversation mode, does not print special tokens and suffix/prefix + if suffix/prefix are not specified, default chat template will be used + (default: false) + -i, --interactive run in interactive mode (default: false) + -if, --interactive-first run in interactive mode and wait for input right away (default: false) + -mli, --multiline-input allows you to write or paste multiple lines without ending each in '\' + --in-prefix-bos prefix BOS to user inputs, preceding the `--in-prefix` string + --in-prefix STRING string to prefix user inputs with (default: empty) + --in-suffix STRING string to suffix after user inputs with (default: empty) + --spm-infill use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) + +sampling: + + --samplers SAMPLERS samplers that will be used for generation in the order, separated by ';' + (default: top_k;tfs_z;typical_p;top_p;min_p;temperature) + --sampling-seq SEQUENCE simplified sequence for samplers that will be used (default: kfypmt) + --ignore-eos ignore end of stream token and continue generating (implies --logit-bias EOS-inf) + --penalize-nl penalize newline tokens (default: false) + --temp N temperature (default: 0.8) + --top-k N top-k sampling (default: 40, 0 = disabled) + --top-p N top-p sampling (default: 0.9, 1.0 = disabled) + --min-p N min-p sampling (default: 0.1, 0.0 = disabled) + --tfs N tail free sampling, parameter z (default: 1.0, 1.0 = disabled) + --typical N locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) + --repeat-last-n N last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) + --repeat-penalty N penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) + --presence-penalty N repeat alpha presence penalty (default: 0.0, 0.0 = disabled) + --frequency-penalty N repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) + --dynatemp-range N dynamic temperature range (default: 0.0, 0.0 = disabled) + --dynatemp-exp N dynamic temperature exponent (default: 1.0) + --mirostat N use Mirostat sampling. + Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used. + (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) + --mirostat-lr N Mirostat learning rate, parameter eta (default: 0.1) + --mirostat-ent N Mirostat target entropy, parameter tau (default: 5.0) + -l TOKEN_ID(+/-)BIAS modifies the likelihood of token appearing in the completion, + i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello', + or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' + --cfg-negative-prompt PROMPT + negative prompt to use for guidance (default: '') + --cfg-negative-prompt-file FNAME + negative prompt file to use for guidance + --cfg-scale N strength of guidance (default: 1.0, 1.0 = disable) + --chat-template JINJA_TEMPLATE + set custom jinja chat template (default: template taken from model's metadata) + if suffix/prefix are specified, template will be disabled + only commonly used templates are accepted: + https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template + +grammar: + + --grammar GRAMMAR BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') + --grammar-file FNAME file to read grammar from + -j, --json-schema SCHEMA JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object + For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead + +embedding: + + --pooling {none,mean,cls,last} + pooling type for embeddings, use model default if unspecified + --attention {causal,non-causal} + attention type for embeddings, use model default if unspecified + +context hacking: + + --rope-scaling {none,linear,yarn} + RoPE frequency scaling method, defaults to linear unless specified by the model + --rope-scale N RoPE context scaling factor, expands context by a factor of N + --rope-freq-base N RoPE base frequency, used by NTK-aware scaling (default: loaded from model) + --rope-freq-scale N RoPE frequency scaling factor, expands context by a factor of 1/N + --yarn-orig-ctx N YaRN: original context size of model (default: 0 = model training context size) + --yarn-ext-factor N YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation) + --yarn-attn-factor N YaRN: scale sqrt(t) or attention magnitude (default: 1.0) + --yarn-beta-slow N YaRN: high correction dim or alpha (default: 1.0) + --yarn-beta-fast N YaRN: low correction dim or beta (default: 32.0) + -gan, --grp-attn-n N group-attention factor (default: 1) + -gaw, --grp-attn-w N group-attention width (default: 512.0) + -dkvc, --dump-kv-cache verbose print of the KV cache + -nkvo, --no-kv-offload disable KV offload + -ctk, --cache-type-k TYPE KV cache data type for K (default: f16) + -ctv, --cache-type-v TYPE KV cache data type for V (default: f16) + +perplexity: + + --all-logits return logits for all tokens in the batch (default: false) + --hellaswag compute HellaSwag score over random tasks from datafile supplied with -f + --hellaswag-tasks N number of tasks to use when computing the HellaSwag score (default: 400) + --winogrande compute Winogrande score over random tasks from datafile supplied with -f + --winogrande-tasks N number of tasks to use when computing the Winogrande score (default: 0) + --multiple-choice compute multiple choice score over random tasks from datafile supplied with -f + --multiple-choice-tasks N + number of tasks to use when computing the multiple choice score (default: 0) + --kl-divergence computes KL-divergence to logits provided via --kl-divergence-base + --ppl-stride N stride for perplexity calculation (default: 0) + --ppl-output-type {0,1} output type for perplexity calculation (default: 0) + +parallel: + + -dt, --defrag-thold N KV cache defragmentation threshold (default: -1.0, < 0 - disabled) + -np, --parallel N number of parallel sequences to decode (default: 1) + -ns, --sequences N number of sequences to decode (default: 1) + -cb, --cont-batching enable continuous batching (a.k.a dynamic batching) (default: enabled) + +multi-modality: + + --mmproj FILE path to a multimodal projector file for LLaVA. see examples/llava/README.md + --image FILE path to an image file. use with multimodal models. Specify multiple times for batching + +backend: + + --rpc SERVERS comma separated list of RPC servers + --mlock force system to keep model in RAM rather than swapping or compressing + --no-mmap do not memory-map model (slower load but may reduce pageouts if not using mlock) + --numa TYPE attempt optimizations that help on some NUMA systems + - distribute: spread execution evenly over all nodes + - isolate: only spawn threads on CPUs on the node that execution started on + - numactl: use the CPU map provided by numactl + if run without this previously, it is recommended to drop the system page cache before using this + see https://github.com/ggerganov/llama.cpp/issues/1437 + +model: + + --check-tensors check model tensor data for invalid values (default: false) + --override-kv KEY=TYPE:VALUE + advanced option to override model metadata by key. may be specified multiple times. + types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false + --lora FNAME apply LoRA adapter (implies --no-mmap) + --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap) + --lora-base FNAME optional model to use as a base for the layers modified by the LoRA adapter + --control-vector FNAME add a control vector + note: this argument can be repeated to add multiple control vectors + --control-vector-scaled FNAME SCALE + add a control vector with user defined scaling SCALE + note: this argument can be repeated to add multiple scaled control vectors + --control-vector-layer-range START END + layer range to apply the control vector(s) to, start and end inclusive + -m, --model FNAME model path (default: models/$filename with filename from --hf-file + or --model-url if set, otherwise models/7B/ggml-model-f16.gguf) + -md, --model-draft FNAME draft model for speculative decoding (default: unused) + -mu, --model-url MODEL_URL model download url (default: unused) + -hfr, --hf-repo REPO Hugging Face model repository (default: unused) + -hff, --hf-file FILE Hugging Face model file (default: unused) + -hft, --hf-token TOKEN Hugging Face access token (default: value from HF_TOKEN environment variable) + +retrieval: + + --context-file FNAME file to load context from (repeat to specify multiple files) + --chunk-size N minimum length of embedded text chunks (default: 64) + --chunk-separator STRING + separator between chunks (default: ' + ') + +passkey: + + --junk N number of times to repeat the junk text (default: 250) + --pos N position of the passkey in the junk text (default: -1) + +imatrix: + + -o, --output FNAME output file (default: 'imatrix.dat') + --output-frequency N output the imatrix every N iterations (default: 10) + --save-frequency N save an imatrix copy every N iterations (default: 0) + --process-output collect data for the output tensor (default: false) + --no-ppl do not compute perplexity (default: true) + --chunk N start processing the input from chunk N (default: 0) + +bench: + + -pps is the prompt shared across parallel sequences (default: false) + -npp n0,n1,... number of prompt tokens + -ntg n0,n1,... number of text generation tokens + -npl n0,n1,... number of parallel prompts + +embedding: + + --embd-normalize normalisation for embendings (default: 2) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm) + --embd-output-format empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix + --embd-separator separator of embendings (default \n) for example "<#sep#>" + +server: + + --host HOST ip address to listen (default: 127.0.0.1) + --port PORT port to listen (default: 8080) + --path PATH path to serve static files from (default: ) + --embedding(s) enable embedding endpoint (default: disabled) + --api-key KEY API key to use for authentication (default: none) + --api-key-file FNAME path to file containing API keys (default: none) + --ssl-key-file FNAME path to file a PEM-encoded SSL private key + --ssl-cert-file FNAME path to file a PEM-encoded SSL certificate + --timeout N server read/write timeout in seconds (default: 600) + --threads-http N number of threads used to process HTTP requests (default: -1) + --system-prompt-file FNAME + set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications + --log-format {text,json} + log output format: json or text (default: json) + --metrics enable prometheus compatible metrics endpoint (default: disabled) + --no-slots disables slots monitoring endpoint (default: enabled) + --slot-save-path PATH path to save slot kv cache (default: disabled) + --chat-template JINJA_TEMPLATE + set custom jinja chat template (default: template taken from model's metadata) + only commonly used templates are accepted: + https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template + -sps, --slot-prompt-similarity SIMILARITY + how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled) + + +logging: + + --simple-io use basic IO for better compatibility in subprocesses and limited consoles + -ld, --logdir LOGDIR path under which to save YAML logs (no logging if unset) + --log-test Run simple logging test + --log-disable Disable trace logs + --log-enable Enable trace logs + --log-file FNAME Specify a log filename (without extension) + --log-new Create a separate new log file on start. Each log file will have unique name: "..log" + --log-append Don't truncate the old log file. + +cvector: + + -o, --output FNAME output file (default: 'control_vector.gguf') + --positive-file FNAME positive prompts file, one prompt per line (default: 'examples/cvector-generator/positive.txt') + --negative-file FNAME negative prompts file, one prompt per line (default: 'examples/cvector-generator/negative.txt') + --pca-batch N batch size used for PCA. Larger batch runs faster, but uses more memory (default: 100) + --pca-iter N number of iterations used for PCA (default: 1000) + --method {pca,mean} dimensionality reduction method to be used (default: pca) +``` + ## Build @@ -232,7 +444,7 @@ node index.js `n_predict`: Set the maximum number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. Default: `-1`, where `-1` is infinity. - `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded. + `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded. The number excludes the BOS token. By default, this value is set to `0`, meaning no tokens are kept. Use `-1` to retain all tokens from the prompt. `stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`. diff --git a/examples/server/public/completion.js b/examples/server/public/completion.js index 987b9a3b891d0..36818f7644aa1 100644 --- a/examples/server/public/completion.js +++ b/examples/server/public/completion.js @@ -21,7 +21,7 @@ let generation_settings = null; // export async function* llama(prompt, params = {}, config = {}) { let controller = config.controller; - const api_url = config.api_url || ""; + const api_url = config.api_url?.replace(/\/+$/, '') || ""; if (!controller) { controller = new AbortController(); @@ -196,7 +196,7 @@ export const llamaComplete = async (params, controller, callback) => { // Get the model info from the server. This is useful for getting the context window and so on. export const llamaModelInfo = async (config = {}) => { if (!generation_settings) { - const api_url = config.api_url || ""; + const api_url = config.api_url?.replace(/\/+$/, '') || ""; const props = await fetch(`${api_url}/props`).then(r => r.json()); generation_settings = props.default_generation_settings; } diff --git a/examples/server/public/index-new.html b/examples/server/public/index-new.html index 5513e9121a40f..c87dd8f1e1d32 100644 --- a/examples/server/public/index-new.html +++ b/examples/server/public/index-new.html @@ -14,10 +14,10 @@