root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-27 03:44:35 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

Merge branch 'master' into gg/flash-attn

Browse Source
This commit is contained in:
Georgi Gerganov 2024-02-19 12:58:18 +02:00
commit 31109ca00a
No known key found for this signature in database
GPG Key ID: BF970631944C16B7
87 changed files with 5115 additions and 1531 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
.github/workflows/build.yml vendored
View File

@ -37,6 +37,8 @@ jobs:
- name: Build
id: make_build
env:
LLAMA_FATAL_WARNINGS: 1
run: |
CC=gcc-8 make -j $(nproc)
@ -65,7 +67,7 @@ jobs:
run: |
mkdir build
cd build
cmake ..
cmake .. -DLLAMA_FATAL_WARNINGS=ON
cmake --build . --config Release -j $(nproc)
- name: Test
@ -100,7 +102,7 @@ jobs:
run: |
mkdir build
cd build
cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
- name: Test
@ -244,6 +246,8 @@ jobs:
- name: Build
id: make_build
env:
LLAMA_FATAL_WARNINGS: 1
run: |
LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
@ -277,7 +281,7 @@ jobs:
sysctl -a
mkdir build
cd build
cmake -DLLAMA_METAL=OFF ..
cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF ..
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test

2
.gitignore vendored
View File

@ -23,11 +23,13 @@
.clang-tidy
.vs/
.vscode/
.idea/
lcov-report/
gcovr-report/
build*
cmake-build-*
out/
tmp/

424
CMakeLists.txt
View File

@ -55,6 +55,9 @@ option(LLAMA_ALL_WARNINGS "llama: enable all compiler warnings"
option(LLAMA_ALL_WARNINGS_3RD_PARTY "llama: enable all compiler warnings in 3rd party libs" OFF)
option(LLAMA_GPROF "llama: enable gprof" OFF)
# build
option(LLAMA_FATAL_WARNINGS "llama: enable -Werror flag" OFF)
# sanitizers
option(LLAMA_SANITIZE_THREAD "llama: enable thread sanitizer" OFF)
option(LLAMA_SANITIZE_ADDRESS "llama: enable address sanitizer" OFF)
@ -112,17 +115,14 @@ option(LLAMA_MPI "llama: use MPI"
option(LLAMA_QKK_64 "llama: use super-block size of 64 for k-quants" OFF)
option(LLAMA_SYCL "llama: use SYCL" OFF)
option(LLAMA_SYCL_F16 "llama: use 16 bit floats for sycl calculations" OFF)
option(LLAMA_CPU_HBM "llama: use memkind for CPU HBM" OFF)
option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_SERVER "llama: build server example" ON)
# add perf arguments
option(LLAMA_PERF "llama: enable perf" OFF)
if (LLAMA_PERF)
add_definitions(-DGGML_PERF)
endif()
# Required for relocatable CMake package
include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
@ -130,6 +130,7 @@ include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
#
# Compile flags
#
if (LLAMA_SYCL)
set(CMAKE_CXX_STANDARD 17)
else()
@ -140,9 +141,18 @@ set(CMAKE_CXX_STANDARD_REQUIRED true)
set(CMAKE_C_STANDARD 11)
set(CMAKE_C_STANDARD_REQUIRED true)
set(THREADS_PREFER_PTHREAD_FLAG ON)
find_package(Threads REQUIRED)
include(CheckCXXCompilerFlag)
if (LLAMA_FATAL_WARNINGS)
if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
add_compile_options(-Werror)
elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
add_compile_options(/WX)
endif()
endif()
# enable libstdc++ assertions for debug builds
if (CMAKE_SYSTEM_NAME MATCHES "Linux")
add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
@ -151,17 +161,17 @@ endif()
if (NOT MSVC)
if (LLAMA_SANITIZE_THREAD)
add_compile_options(-fsanitize=thread)
link_libraries(-fsanitize=thread)
link_libraries (-fsanitize=thread)
endif()
if (LLAMA_SANITIZE_ADDRESS)
add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
link_libraries(-fsanitize=address)
link_libraries (-fsanitize=address)
endif()
if (LLAMA_SANITIZE_UNDEFINED)
add_compile_options(-fsanitize=undefined)
link_libraries(-fsanitize=undefined)
link_libraries (-fsanitize=undefined)
endif()
endif()
@ -298,14 +308,17 @@ if (LLAMA_BLAS)
endif()
message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
add_compile_options(${BLAS_LINKER_FLAGS})
add_compile_definitions(GGML_USE_OPENBLAS)
if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${LLAMA_BLAS_VENDOR} MATCHES "Generic" OR ${LLAMA_BLAS_VENDOR} MATCHES "Intel"))
add_compile_definitions(GGML_BLAS_USE_MKL)
endif()
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES})
set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES})
set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
else()
message(WARNING "BLAS not found, please refer to "
"https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
@ -330,9 +343,6 @@ if (LLAMA_CUBLAS)
set(GGML_SOURCES_CUDA ggml-cuda.cu)
add_compile_definitions(GGML_USE_CUBLAS)
# if (LLAMA_CUDA_CUBLAS)
# add_compile_definitions(GGML_CUDA_CUBLAS)
# endif()
if (LLAMA_CUDA_FORCE_DMMV)
add_compile_definitions(GGML_CUDA_FORCE_DMMV)
endif()
@ -387,15 +397,20 @@ if (LLAMA_MPI)
find_package(MPI)
if (MPI_C_FOUND)
message(STATUS "MPI found")
set(GGML_HEADERS_MPI ggml-mpi.h)
set(GGML_SOURCES_MPI ggml-mpi.c ggml-mpi.h)
set(GGML_SOURCES_MPI ggml-mpi.c)
add_compile_definitions(GGML_USE_MPI)
add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
if (NOT MSVC)
add_compile_options(-Wno-cast-qual)
endif()
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${MPI_C_LIBRARIES})
set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
# Even if you're only using the C header, C++ programs may bring in MPI
# C++ functions, so more linkage is needed
if (MPI_CXX_FOUND)
@ -427,31 +442,28 @@ if (LLAMA_VULKAN)
if (Vulkan_FOUND)
message(STATUS "Vulkan found")
add_library(ggml-vulkan OBJECT ggml-vulkan.cpp ggml-vulkan.h)
if (BUILD_SHARED_LIBS)
set_target_properties(ggml-vulkan PROPERTIES POSITION_INDEPENDENT_CODE ON)
endif()
target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
set(GGML_HEADERS_VULKAN ggml-vulkan.h)
set(GGML_SOURCES_VULKAN ggml-vulkan.cpp)
add_compile_definitions(GGML_USE_VULKAN)
if (LLAMA_VULKAN_CHECK_RESULTS)
target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_CHECK_RESULTS)
add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
endif()
if (LLAMA_VULKAN_DEBUG)
target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_DEBUG)
add_compile_definitions(GGML_VULKAN_DEBUG)
endif()
if (LLAMA_VULKAN_VALIDATE)
target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_VALIDATE)
add_compile_definitions(GGML_VULKAN_VALIDATE)
endif()
if (LLAMA_VULKAN_RUN_TESTS)
target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_RUN_TESTS)
add_compile_definitions(GGML_VULKAN_RUN_TESTS)
endif()
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-vulkan)
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} Vulkan::Vulkan)
else()
message(WARNING "Vulkan not found")
endif()
@ -463,43 +475,45 @@ if (LLAMA_HIPBLAS)
if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
endif()
if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
endif()
find_package(hip)
find_package(hipblas)
find_package(rocblas)
find_package(hip REQUIRED)
find_package(hipblas REQUIRED)
find_package(rocblas REQUIRED)
if (${hipblas_FOUND} AND ${hip_FOUND})
message(STATUS "HIP and hipBLAS found")
add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
if (LLAMA_HIP_UMA)
add_compile_definitions(GGML_HIP_UMA)
endif()
add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
if (BUILD_SHARED_LIBS)
set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON)
endif()
if (LLAMA_CUDA_FORCE_DMMV)
target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
endif()
if (LLAMA_CUDA_FORCE_MMQ)
target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_MMQ)
endif()
target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
message(STATUS "HIP and hipBLAS found")
if (LLAMA_STATIC)
message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
endif()
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-rocm)
else()
message(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
set(GGML_HEADERS_ROCM ggml-cuda.h)
set(GGML_SOURCES_ROCM ggml-cuda.cu)
add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
if (LLAMA_HIP_UMA)
add_compile_definitions(GGML_HIP_UMA)
endif()
if (LLAMA_CUDA_FORCE_DMMV)
add_compile_definitions(GGML_CUDA_FORCE_DMMV)
endif()
if (LLAMA_CUDA_FORCE_MMQ)
add_compile_definitions(GGML_CUDA_FORCE_MMQ)
endif()
add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
if (LLAMA_STATIC)
message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
endif()
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
endif()
if (LLAMA_SYCL)
@ -509,10 +523,14 @@ if (LLAMA_SYCL)
#todo: AOT
find_package(IntelSYCL REQUIRED)
message(STATUS "SYCL found")
add_compile_definitions(GGML_USE_SYCL)
if (LLAMA_SYCL_F16)
add_compile_definitions(GGML_SYCL_F16)
endif()
add_compile_definitions(GGML_USE_SYCL)
add_compile_options(-I./) #include DPCT
add_compile_options(-I/${SYCL_INCLUDE_DIR})
@ -521,7 +539,7 @@ if (LLAMA_SYCL)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
set(GGML_HEADERS_SYCL ggml.h ggml-sycl.h)
set(GGML_HEADERS_SYCL ggml-sycl.h)
set(GGML_SOURCES_SYCL ggml-sycl.cpp)
if (WIN32)
@ -540,61 +558,61 @@ if (LLAMA_KOMPUTE)
endif()
function(compile_shader)
set(options)
set(oneValueArgs)
set(multiValueArgs SOURCES)
cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
foreach(source ${compile_shader_SOURCES})
get_filename_component(filename ${source} NAME)
set(spv_file ${filename}.spv)
add_custom_command(
OUTPUT ${spv_file}
DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
COMMENT "Compiling ${source} to ${spv_file}"
)
set(options)
set(oneValueArgs)
set(multiValueArgs SOURCES)
cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
foreach(source ${compile_shader_SOURCES})
get_filename_component(filename ${source} NAME)
set(spv_file ${filename}.spv)
add_custom_command(
OUTPUT ${spv_file}
DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
COMMENT "Compiling ${source} to ${spv_file}"
)
get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
set(FILE_NAME "shader${RAW_FILE_NAME}")
string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
if(CMAKE_GENERATOR MATCHES "Visual Studio")
add_custom_command(
OUTPUT ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
DEPENDS ${spv_file} xxd
COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
)
else()
add_custom_command(
OUTPUT ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
DEPENDS ${spv_file} xxd
COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
)
endif()
endforeach()
get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
set(FILE_NAME "shader${RAW_FILE_NAME}")
string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
if(CMAKE_GENERATOR MATCHES "Visual Studio")
add_custom_command(
OUTPUT ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
DEPENDS ${spv_file} xxd
COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
)
else()
add_custom_command(
OUTPUT ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
DEPENDS ${spv_file} xxd
COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
)
endif()
endforeach()
endfunction()
if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
@ -604,66 +622,66 @@ if (LLAMA_KOMPUTE)
# Compile our shaders
compile_shader(SOURCES
kompute-shaders/op_scale.comp
kompute-shaders/op_scale_8.comp
kompute-shaders/op_add.comp
kompute-shaders/op_addrow.comp
kompute-shaders/op_mul.comp
kompute-shaders/op_silu.comp
kompute-shaders/op_relu.comp
kompute-shaders/op_gelu.comp
kompute-shaders/op_softmax.comp
kompute-shaders/op_norm.comp
kompute-shaders/op_rmsnorm.comp
kompute-shaders/op_diagmask.comp
kompute-shaders/op_mul_mat_mat_f32.comp
kompute-shaders/op_mul_mat_f16.comp
kompute-shaders/op_mul_mat_q8_0.comp
kompute-shaders/op_mul_mat_q4_0.comp
kompute-shaders/op_mul_mat_q4_1.comp
kompute-shaders/op_mul_mat_q6_k.comp
kompute-shaders/op_getrows_f16.comp
kompute-shaders/op_getrows_q4_0.comp
kompute-shaders/op_getrows_q4_1.comp
kompute-shaders/op_getrows_q6_k.comp
kompute-shaders/op_rope_f16.comp
kompute-shaders/op_rope_f32.comp
kompute-shaders/op_cpy_f16_f16.comp
kompute-shaders/op_cpy_f16_f32.comp
kompute-shaders/op_cpy_f32_f16.comp
kompute-shaders/op_cpy_f32_f32.comp
kompute-shaders/op_scale.comp
kompute-shaders/op_scale_8.comp
kompute-shaders/op_add.comp
kompute-shaders/op_addrow.comp
kompute-shaders/op_mul.comp
kompute-shaders/op_silu.comp
kompute-shaders/op_relu.comp
kompute-shaders/op_gelu.comp
kompute-shaders/op_softmax.comp
kompute-shaders/op_norm.comp
kompute-shaders/op_rmsnorm.comp
kompute-shaders/op_diagmask.comp
kompute-shaders/op_mul_mat_mat_f32.comp
kompute-shaders/op_mul_mat_f16.comp
kompute-shaders/op_mul_mat_q8_0.comp
kompute-shaders/op_mul_mat_q4_0.comp
kompute-shaders/op_mul_mat_q4_1.comp
kompute-shaders/op_mul_mat_q6_k.comp
kompute-shaders/op_getrows_f16.comp
kompute-shaders/op_getrows_q4_0.comp
kompute-shaders/op_getrows_q4_1.comp
kompute-shaders/op_getrows_q6_k.comp
kompute-shaders/op_rope_f16.comp
kompute-shaders/op_rope_f32.comp
kompute-shaders/op_cpy_f16_f16.comp
kompute-shaders/op_cpy_f16_f32.comp
kompute-shaders/op_cpy_f32_f16.comp
kompute-shaders/op_cpy_f32_f32.comp
)
# Create a custom target for our generated shaders
add_custom_target(generated_shaders DEPENDS
shaderop_scale.h
shaderop_scale_8.h
shaderop_add.h
shaderop_addrow.h
shaderop_mul.h
shaderop_silu.h
shaderop_relu.h
shaderop_gelu.h
shaderop_softmax.h
shaderop_norm.h
shaderop_rmsnorm.h
shaderop_diagmask.h
shaderop_mul_mat_mat_f32.h
shaderop_mul_mat_f16.h
shaderop_mul_mat_q8_0.h
shaderop_mul_mat_q4_0.h
shaderop_mul_mat_q4_1.h
shaderop_mul_mat_q6_k.h
shaderop_getrows_f16.h
shaderop_getrows_q4_0.h
shaderop_getrows_q4_1.h
shaderop_getrows_q6_k.h
shaderop_rope_f16.h
shaderop_rope_f32.h
shaderop_cpy_f16_f16.h
shaderop_cpy_f16_f32.h
shaderop_cpy_f32_f16.h
shaderop_cpy_f32_f32.h
shaderop_scale.h
shaderop_scale_8.h
shaderop_add.h
shaderop_addrow.h
shaderop_mul.h
shaderop_silu.h
shaderop_relu.h
shaderop_gelu.h
shaderop_softmax.h
shaderop_norm.h
shaderop_rmsnorm.h
shaderop_diagmask.h
shaderop_mul_mat_mat_f32.h
shaderop_mul_mat_f16.h
shaderop_mul_mat_q8_0.h
shaderop_mul_mat_q4_0.h
shaderop_mul_mat_q4_1.h
shaderop_mul_mat_q6_k.h
shaderop_getrows_f16.h
shaderop_getrows_q4_0.h
shaderop_getrows_q4_1.h
shaderop_getrows_q6_k.h
shaderop_rope_f16.h
shaderop_rope_f32.h
shaderop_cpy_f16_f16.h
shaderop_cpy_f16_f32.h
shaderop_cpy_f32_f16.h
shaderop_cpy_f32_f32.h
)
# Create a custom command that depends on the generated_shaders
@ -676,8 +694,10 @@ if (LLAMA_KOMPUTE)
# Add the stamp to the main sources to ensure dependency tracking
set(GGML_SOURCES_KOMPUTE ggml-kompute.cpp ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
set(GGML_HEADERS_KOMPUTE ggml-kompute.h ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
set(GGML_HEADERS_KOMPUTE ggml-kompute.h ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
add_compile_definitions(GGML_USE_KOMPUTE)
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} kompute)
set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${CMAKE_BINARY_DIR})
else()
@ -685,6 +705,18 @@ if (LLAMA_KOMPUTE)
endif()
endif()
if (LLAMA_CPU_HBM)
find_library(memkind memkind REQUIRED)
add_compile_definitions(GGML_USE_CPU_HBM)
target_link_libraries(ggml PUBLIC memkind)
endif()
if (LLAMA_PERF)
add_compile_definitions(GGML_PERF)
endif()
function(get_flags CCID CCVER)
set(C_FLAGS "")
set(CXX_FLAGS "")
@ -709,13 +741,6 @@ function(get_flags CCID CCVER)
if (CCVER VERSION_GREATER_EQUAL 8.1.0)
list(APPEND CXX_FLAGS -Wextra-semi)
endif()
elseif (CCID MATCHES "Intel")
if (NOT LLAMA_SYCL)
# enable max optimization level when using Intel compiler
set(C_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
add_link_options(-fuse-ld=lld -static-intel)
endif()
endif()
set(GF_C_FLAGS ${C_FLAGS} PARENT_SCOPE)
@ -746,10 +771,7 @@ endif()
set(CUDA_CXX_FLAGS "")
if (LLAMA_CUBLAS)
set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
if (NOT MSVC)
list(APPEND CUDA_FLAGS -Wno-pedantic)
endif()
set(CUDA_FLAGS -use_fast_math)
if (LLAMA_ALL_WARNINGS AND NOT MSVC)
set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
@ -782,7 +804,11 @@ if (LLAMA_CUBLAS)
message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
get_flags(${CUDA_CCID} ${CUDA_CCVER})
list(APPEND CUDA_CXX_FLAGS ${GF_CXX_FLAGS}) # This is passed to -Xcompiler later
list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS}) # This is passed to -Xcompiler later
endif()
if (NOT MSVC)
list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
endif()
endif()
@ -821,6 +847,7 @@ execute_process(
ERROR_VARIABLE output
OUTPUT_QUIET
)
if (output MATCHES "dyld-1015\.7")
add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
endif()
@ -830,10 +857,10 @@ endif()
# feel free to update the Makefile for your architecture and send a pull request or issue
message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
if (MSVC)
string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
else ()
set(CMAKE_GENERATOR_PLATFORM_LWR "")
set(CMAKE_GENERATOR_PLATFORM_LWR "")
endif ()
if (NOT MSVC)
@ -855,11 +882,21 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
message(STATUS "ARM detected")
if (MSVC)
add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
add_compile_definitions(__ARM_NEON)
add_compile_definitions(__ARM_FEATURE_FMA)
add_compile_definitions(__ARM_FEATURE_DOTPROD)
# add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) # MSVC doesn't support vdupq_n_f16, vld1q_f16, vst1q_f16
add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
if (GGML_COMPILER_SUPPORT_DOTPROD)
add_compile_definitions(__ARM_FEATURE_DOTPROD)
endif ()
check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
endif ()
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
else()
check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
@ -1017,11 +1054,6 @@ endif()
# ggml
if (GGML_USE_CPU_HBM)
add_definitions(-DGGML_USE_CPU_HBM)
find_library(memkind memkind REQUIRED)
endif()
add_library(ggml OBJECT
ggml.c
ggml.h
@ -1038,16 +1070,17 @@ add_library(ggml OBJECT
${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA}
${GGML_SOURCES_SYCL} ${GGML_HEADERS_SYCL}
${GGML_SOURCES_KOMPUTE} ${GGML_HEADERS_KOMPUTE}
${GGML_SOURCES_VULKAN} ${GGML_HEADERS_VULKAN}
${GGML_SOURCES_ROCM} ${GGML_HEADERS_ROCM}
)
target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
target_compile_features(ggml PUBLIC c_std_11) # don't bump
target_compile_features (ggml PUBLIC c_std_11) # don't bump
target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
if (GGML_USE_CPU_HBM)
target_link_libraries(ggml PUBLIC memkind)
endif()
add_library(ggml_static STATIC $<TARGET_OBJECTS:ggml>)
if (BUILD_SHARED_LIBS)
set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
add_library(ggml_shared SHARED $<TARGET_OBJECTS:ggml>)
@ -1063,7 +1096,8 @@ add_library(llama
)
target_include_directories(llama PUBLIC .)
target_compile_features(llama PUBLIC cxx_std_11) # don't bump
target_compile_features (llama PUBLIC cxx_std_11) # don't bump
target_link_libraries(llama PRIVATE
ggml
${LLAMA_EXTRA_LIBS}
@ -1114,7 +1148,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama)
set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
"${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}"
"${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}"
"${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")

21
Makefile
View File

@ -215,6 +215,11 @@ MK_CFLAGS += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
-Werror=implicit-function-declaration
MK_CXXFLAGS += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
ifeq ($(LLAMA_FATAL_WARNINGS),1)
MK_CFLAGS += -Werror
MK_CXXFLAGS += -Werror
endif
# this version of Apple ld64 is buggy
ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
@ -439,7 +444,7 @@ ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
ifdef JETSON_EOL_MODULE_DETECT
$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
else
$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
$(NVCC) $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
endif # JETSON_EOL_MODULE_DETECT
endif # LLAMA_CUBLAS
@ -550,7 +555,7 @@ override LDFLAGS := $(MK_LDFLAGS) $(LDFLAGS)
ifdef LLAMA_CUBLAS
GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
include scripts/get-flags.mk
CUDA_CXXFLAGS := $(GF_CXXFLAGS)
CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
endif
#
@ -569,6 +574,14 @@ $(info I CC: $(shell $(CC) --version | head -n 1))
$(info I CXX: $(shell $(CXX) --version | head -n 1))
ifdef LLAMA_CUBLAS
$(info I NVCC: $(shell $(NVCC) --version | tail -n 1))
CUDA_VERSION := $(shell nvcc --version | grep -oP 'release (\K[0-9]+\.[0-9])')
ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
ifndef CUDA_DOCKER_ARCH
ifndef CUDA_POWER_ARCH
$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via CUDA_DOCKER_ARCH)
endif # CUDA_POWER_ARCH
endif # CUDA_DOCKER_ARCH
endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
endif # LLAMA_CUBLAS
$(info )
@ -854,3 +867,7 @@ tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o te
tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
tests/test-chat-template: tests/test-chat-template.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

6
README.md
View File

@ -61,7 +61,7 @@ variety of hardware - locally and in the cloud.
- Plain C/C++ implementation without any dependencies
- Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
- AVX, AVX2 and AVX512 support for x86 architectures
- 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
- 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
- Vulkan, SYCL, and (partial) OpenCL backend support
- CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
@ -768,7 +768,7 @@ The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 thread
#### How to run
1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
3. Output:
```
@ -958,7 +958,7 @@ We have three Docker images available for this project:
1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executabhle file. (platforms: `linux/amd64`, `linux/arm64`)
3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
Additionally, there the following images, similar to the above:

56
ci/run.sh
View File

@ -33,7 +33,7 @@ sd=`dirname $0`
cd $sd/../
SRC=`pwd`
CMAKE_EXTRA=""
CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
if [ ! -z ${GG_BUILD_METAL} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON"
@ -219,7 +219,7 @@ function gg_run_open_llama_3b_v2 {
gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
@ -401,7 +401,7 @@ function gg_run_open_llama_7b_v2 {
gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
path_models="../models-mnt/open-llama/7B-v2"
@ -568,6 +568,54 @@ function gg_sum_open_llama_7b_v2 {
#gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
}
# bge-small
function gg_run_embd_bge_small {
cd ${SRC}
gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/tokenizer.model
gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/tokenizer_config.json
gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/special_tokens_map.json
gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/pytorch_model.bin
gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/sentence_bert_config.json
gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/vocab.txt
gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/modules.json
gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
gg_wget models-mnt/bge-small/1_Pooling https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/1_Pooling/config.json
path_models="../models-mnt/bge-small"
rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
set -e
(time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
(time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log
python3 ../convert-hf-to-gguf.py ${path_models}
model_f16="${path_models}/ggml-model-f16.gguf"
model_q8_0="${path_models}/ggml-model-q8_0.gguf"
./bin/quantize ${model_f16} ${model_q8_0} q8_0
(time ./bin/embedding --model ${model_f16} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
set +e
}
function gg_sum_embd_bge_small {
gg_printf '### %s\n\n' "${ci}"
gg_printf 'BGE Small (BERT):\n'
gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
}
## main
if [ -z ${GG_BUILD_LOW_PERF} ]; then
@ -591,6 +639,8 @@ test $ret -eq 0 && gg_run ctest_debug
test $ret -eq 0 && gg_run ctest_release
if [ -z ${GG_BUILD_LOW_PERF} ]; then
test $ret -eq 0 && gg_run embd_bge_small
if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
if [ -z ${GG_BUILD_CUDA} ]; then
test $ret -eq 0 && gg_run open_llama_3b_v2

87
common/common.cpp
View File

@ -341,7 +341,7 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
break;
}
const auto sampler_names = string_split(argv[i], ';');
sparams.samplers_sequence = sampler_types_from_names(sampler_names);
sparams.samplers_sequence = sampler_types_from_names(sampler_names, true);
} else if (arg == "--sampling-seq") {
if (++i >= argc) {
invalid_param = true;
@ -671,7 +671,15 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
} else if (arg == "--no-mmap") {
params.use_mmap = false;
} else if (arg == "--numa") {
params.numa = true;
if (++i >= argc) {
invalid_param = true;
break;
}
std::string value(argv[i]);
/**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
else { invalid_param = true; break; }
} else if (arg == "--verbose-prompt") {
params.verbose_prompt = true;
} else if (arg == "--no-display-prompt") {
@ -935,7 +943,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
printf(" -tb N, --threads-batch N\n");
printf(" number of threads to use during batch and prompt processing (default: same as --threads)\n");
printf(" -td N, --threads-draft N");
printf(" number of threads to use during generation (default: same as --threads)");
printf(" number of threads to use during generation (default: same as --threads)\n");
printf(" -tbd N, --threads-batch-draft N\n");
printf(" number of threads to use during batch and prompt processing (default: same as --threads-draft)\n");
printf(" -p PROMPT, --prompt PROMPT\n");
@ -956,7 +964,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
printf(" -n N, --n-predict N number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
printf(" -c N, --ctx-size N size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
printf(" -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch);
printf(" --samplers samplers that will be used for generation in the order, separated by \';\' (default: %s)\n", sampler_type_names.c_str());
printf(" --samplers samplers that will be used for generation in the order, separated by \';\'\n");
printf(" (default: %s)\n", sampler_type_names.c_str());
printf(" --sampling-seq simplified sequence for samplers that will be used (default: %s)\n", sampler_type_chars.c_str());
printf(" --top-k N top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
printf(" --top-p N top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
@ -1005,7 +1014,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
printf(" --winogrande-tasks N number of tasks to use when computing the Winogrande score (default: %zu)\n", params.winogrande_tasks);
printf(" --multiple-choice compute multiple choice score over random tasks from datafile supplied with -f\n");
printf(" --multiple-choice-tasks N number of tasks to use when computing the multiple choice score (default: %zu)\n", params.winogrande_tasks);
printf(" --kl-divergence computes KL-divergence to logits provided via --kl-divergence-base");
printf(" --kl-divergence computes KL-divergence to logits provided via --kl-divergence-base\n");
printf(" --keep N number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
printf(" --draft N number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft);
printf(" --chunks N max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
@ -1022,7 +1031,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
if (llama_supports_mmap()) {
printf(" --no-mmap do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
}
printf(" --numa attempt optimizations that help on some NUMA systems\n");
printf(" --numa TYPE attempt optimizations that help on some NUMA systems\n");
printf(" - distribute: spread execution evenly over all nodes\n");
printf(" - isolate: only spawn threads on CPUs on the node that execution started on\n");
printf(" - numactl: use the CPU map provided by numactl\n");
printf(" if run without this previously, it is recommended to drop the system page cache before using this\n");
printf(" see https://github.com/ggerganov/llama.cpp/issues/1437\n");
if (llama_supports_gpu_offload()) {
@ -1122,34 +1134,50 @@ std::vector<std::string> string_split(std::string input, char separator) {
return parts;
}
std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names) {
std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
{"top_k", llama_sampler_type::TOP_K},
{"top_p", llama_sampler_type::TOP_P},
{"typical_p", llama_sampler_type::TYPICAL_P},
{"min_p", llama_sampler_type::MIN_P},
{"tfs_z", llama_sampler_type::TFS_Z},
{"temperature", llama_sampler_type::TEMPERATURE}
};
// since samplers names are written multiple ways
// make it ready for both system names and input names
std::unordered_map<std::string, llama_sampler_type> sampler_name_map {
{"top_k", llama_sampler_type::TOP_K},
std::unordered_map<std::string, llama_sampler_type> sampler_alt_name_map {
{"top-k", llama_sampler_type::TOP_K},
{"top_p", llama_sampler_type::TOP_P},
{"top-p", llama_sampler_type::TOP_P},
{"nucleus", llama_sampler_type::TOP_P},
{"typical_p", llama_sampler_type::TYPICAL_P},
{"typical-p", llama_sampler_type::TYPICAL_P},
{"typical", llama_sampler_type::TYPICAL_P},
{"min_p", llama_sampler_type::MIN_P},
{"min-p", llama_sampler_type::MIN_P},
{"tfs_z", llama_sampler_type::TFS_Z},
{"tfs-z", llama_sampler_type::TFS_Z},
{"tfs", llama_sampler_type::TFS_Z},
{"temp", llama_sampler_type::TEMP},
{"temperature", llama_sampler_type::TEMP}
{"temp", llama_sampler_type::TEMPERATURE}
};
std::vector<llama_sampler_type> sampler_types;
sampler_types.reserve(names.size());
for (const auto& name : names) {
const auto sampler_item = sampler_name_map.find(name);
if (sampler_item != sampler_name_map.end()) {
for (const auto & name : names)
{
auto sampler_item = sampler_canonical_name_map.find(name);
if (sampler_item != sampler_canonical_name_map.end())
{
sampler_types.push_back(sampler_item->second);
}
else
{
if (allow_alt_names)
{
sampler_item = sampler_alt_name_map.find(name);
if (sampler_item != sampler_alt_name_map.end())
{
sampler_types.push_back(sampler_item->second);
}
}
}
}
return sampler_types;
}
@ -1161,7 +1189,7 @@ std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & nam
{'y', llama_sampler_type::TYPICAL_P},
{'m', llama_sampler_type::MIN_P},
{'f', llama_sampler_type::TFS_Z},
{'t', llama_sampler_type::TEMP}
{'t', llama_sampler_type::TEMPERATURE}
};
std::vector<llama_sampler_type> sampler_types;
@ -1177,12 +1205,12 @@ std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & nam
std::string sampler_type_to_name_string(llama_sampler_type sampler_type) {
switch (sampler_type) {
case llama_sampler_type::TOP_K: return "top_k";
case llama_sampler_type::TFS_Z: return "tfs_z";
case llama_sampler_type::TYPICAL_P: return "typical_p";
case llama_sampler_type::TOP_P: return "top_p";
case llama_sampler_type::MIN_P: return "min_p";
case llama_sampler_type::TEMP: return "temp";
case llama_sampler_type::TOP_K: return "top_k";
case llama_sampler_type::TFS_Z: return "tfs_z";
case llama_sampler_type::TYPICAL_P: return "typical_p";
case llama_sampler_type::TOP_P: return "top_p";
case llama_sampler_type::MIN_P: return "min_p";
case llama_sampler_type::TEMPERATURE: return "temperature";
default : return "";
}
}
@ -1676,6 +1704,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
}
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);
fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
@ -1689,7 +1718,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
fprintf(stream, "no_mul_mat_q: %s # default: false\n", !params.mul_mat_q ? "true" : "false");
fprintf(stream, "no_penalize_nl: %s # default: false\n", !sparams.penalize_nl ? "true" : "false");
fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false");
fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);
@ -1714,7 +1742,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
fprintf(stream, "rope_freq_base: %f # default: 10000.0\n", params.rope_freq_base);
fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
fprintf(stream, "seed: %d # default: -1 (random seed)\n", params.seed);
fprintf(stream, "seed: %u # default: -1 (random seed)\n", params.seed);
fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
@ -1723,7 +1751,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
dump_vector_float_yaml(stream, "tensor_split", tensor_split_vector);
fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency());
fprintf(stream, "threads: %d # default: %u\n", params.n_threads, std::thread::hardware_concurrency());
fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
@ -1774,7 +1802,8 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
if (cs_curr[j] < 0) { continue; }
if (seqs.find(cs_curr[j]) == seqs.end()) {
if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
seqs[cs_curr[j]] = seqs.size();
const size_t sz = seqs.size();
seqs[cs_curr[j]] = sz;
}
}
if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }

4
common/common.h
View File

@ -76,6 +76,7 @@ struct gpt_params {
float yarn_beta_slow = 1.0f; // YaRN high correction dim
int32_t yarn_orig_ctx = 0; // YaRN original context length
int32_t rope_scaling_type = LLAMA_ROPE_SCALING_UNSPECIFIED;
ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
// // sampling parameters
struct llama_sampling_params sparams;
@ -134,7 +135,6 @@ struct gpt_params {
bool logits_all = false; // return logits for all tokens in the batch
bool use_mmap = true; // use mmap for faster loads
bool use_mlock = false; // use mlock to keep model in memory
bool numa = false; // attempt optimizations that help on some NUMA systems
bool verbose_prompt = false; // print prompt tokens before generation
bool display_prompt = true; // print prompt before generation
bool infill = false; // use infill mode
@ -165,7 +165,7 @@ void process_escapes(std::string& input);
// String utils
//
std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names);
std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
std::vector<std::string> string_split(std::string input, char separator);
std::string sampler_type_to_name_string(llama_sampler_type sampler_type);

6
common/sampling.cpp
View File

@ -121,7 +121,7 @@ static void sampler_queue(
struct llama_context * ctx_main,
const llama_sampling_params & params,
llama_token_data_array & cur_p,
size_t & min_keep) {
size_t min_keep) {
const float temp = params.temp;
const float dynatemp_range = params.dynatemp_range;
const float dynatemp_exponent = params.dynatemp_exponent;
@ -139,7 +139,7 @@ static void sampler_queue(
case llama_sampler_type::TYPICAL_P: llama_sample_typical (ctx_main, &cur_p, typical_p, min_keep); break;
case llama_sampler_type::TOP_P : llama_sample_top_p (ctx_main, &cur_p, top_p, min_keep); break;
case llama_sampler_type::MIN_P : llama_sample_min_p (ctx_main, &cur_p, min_p, min_keep); break;
case llama_sampler_type::TEMP:
case llama_sampler_type::TEMPERATURE:
if (dynatemp_range > 0) {
float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
float dynatemp_max = std::max(0.0f, temp + dynatemp_range);
@ -249,7 +249,7 @@ static llama_token llama_sampling_sample_impl(
id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
} else {
// temperature sampling
size_t min_keep = std::max(1, params.n_probs);
size_t min_keep = std::max(1, params.min_keep);
sampler_queue(ctx_main, params, cur_p, min_keep);

15
common/sampling.h
View File

@ -10,18 +10,19 @@
// sampler types
enum class llama_sampler_type : char {
TOP_K = 'k',
TOP_P = 'p',
MIN_P = 'm',
TFS_Z = 'f',
TYPICAL_P = 'y',
TEMP = 't'
TOP_K = 'k',
TOP_P = 'p',
MIN_P = 'm',
TFS_Z = 'f',
TYPICAL_P = 'y',
TEMPERATURE = 't'
};
// sampling parameters
typedef struct llama_sampling_params {
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled
@ -45,7 +46,7 @@ typedef struct llama_sampling_params {
llama_sampler_type::TYPICAL_P,
llama_sampler_type::TOP_P,
llama_sampler_type::MIN_P,
llama_sampler_type::TEMP
llama_sampler_type::TEMPERATURE
};
std::string grammar; // optional BNF-like grammar to constrain sampling

140
convert-hf-to-gguf.py
View File

@ -10,7 +10,7 @@ import re
import sys
from enum import IntEnum
from pathlib import Path
from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast
from typing import TYPE_CHECKING, Any, ContextManager, Iterator, Sequence, cast
import numpy as np
import torch
@ -25,15 +25,6 @@ import gguf
from convert import HfVocab
# check for any of the given keys in the dictionary and return the value of the first key found
def get_key_opts(d, keys):
for k in keys:
if k in d:
return d[k]
print(f"Could not find any of {keys}")
sys.exit()
###### MODEL DEFINITIONS ######
class SentencePieceTokenTypes(IntEnum):
@ -58,6 +49,15 @@ class Model:
self.hparams = Model.load_hparams(self.dir_model)
self.model_arch = self._get_model_architecture()
self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
def find_hparam(self, keys: Sequence[str], optional: bool = False) -> Any:
key = next((k for k in keys if k in self.hparams), None)
if key is not None:
return self.hparams[key]
if optional:
return None
raise KeyError(f"could not find any of: {keys}")
def set_vocab(self):
self._set_vocab_gpt2()
@ -79,28 +79,33 @@ class Model:
def set_gguf_parameters(self):
self.gguf_writer.add_name(self.dir_model.name)
self.gguf_writer.add_block_count(self.hparams.get(
"n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")),
))
if (n_ctx := self.hparams.get("max_position_embeddings")) is not None:
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:
self.gguf_writer.add_context_length(n_ctx)
if (n_embd := self.hparams.get("hidden_size")) is not None:
self.gguf_writer.add_embedding_length(n_embd)
if (n_ff := self.hparams.get("intermediate_size")) is not None:
n_embd = self.find_hparam(["hidden_size", "n_embd"])
self.gguf_writer.add_embedding_length(n_embd)
if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
self.gguf_writer.add_feed_forward_length(n_ff)
if (n_head := self.hparams.get("num_attention_heads")) is not None:
self.gguf_writer.add_head_count(n_head)
n_head = self.find_hparam(["num_attention_heads", "n_head"])
self.gguf_writer.add_head_count(n_head)
if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
self.gguf_writer.add_head_count_kv(n_head_kv)
if (n_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
self.gguf_writer.add_layer_norm_rms_eps(n_rms_eps)
if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon"], optional=True)) is not None:
self.gguf_writer.add_layer_norm_eps(f_norm_eps)
if (n_experts := self.hparams.get("num_local_experts")) is not None:
self.gguf_writer.add_expert_count(n_experts)
if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
self.gguf_writer.add_expert_used_count(n_experts_used)
self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
self.gguf_writer.add_file_type(self.ftype)
def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
@ -211,6 +216,8 @@ class Model:
return MiniCPMModel
if model_architecture == "BertModel":
return BertModel
if model_architecture == "NomicBertModel":
return NomicBertModel
return Model
def _is_model_safetensors(self) -> bool:
@ -268,6 +275,8 @@ class Model:
return gguf.MODEL_ARCH.MINICPM
if arch == "BertModel":
return gguf.MODEL_ARCH.BERT
if arch == "NomicBertModel":
return gguf.MODEL_ARCH.NOMIC_BERT
raise NotImplementedError(f'Architecture "{arch}" not supported!')
@ -1297,21 +1306,21 @@ class GPT2Model(Model):
class Phi2Model(Model):
def set_gguf_parameters(self):
block_count = get_key_opts(self.hparams, ["num_hidden_layers", "n_layer"])
block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"])
n_embd = get_key_opts(self.hparams, ["hidden_size", "n_embd"])
n_head = get_key_opts(self.hparams, ["num_attention_heads", "n_head"])
rot_pct = self.find_hparam(["partial_rotary_factor"])
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(get_key_opts(self.hparams, ["n_positions", "max_position_embeddings"]))
self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
self.gguf_writer.add_embedding_length(n_embd)
self.gguf_writer.add_feed_forward_length(4 * n_embd)
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(n_head)
self.gguf_writer.add_head_count_kv(n_head)
self.gguf_writer.add_layer_norm_eps(get_key_opts(self.hparams, ["layer_norm_epsilon", "layer_norm_eps"]))
self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_epsilon", "layer_norm_eps"]))
self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_add_bos_token(False)
@ -1636,19 +1645,34 @@ in chat mode so that the conversation can end normally.")
class BertModel(Model):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.block_count = self.hparams["num_hidden_layers"]
self.vocab_size = None
def set_gguf_parameters(self):
# TODO(cebtenzzre): merge with parent class
self.gguf_writer.add_name(self.dir_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_feed_forward_length(self.hparams["intermediate_size"])
self.gguf_writer.add_block_count(self.block_count)
self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
super().set_gguf_parameters()
self.gguf_writer.add_causal_attention(False)
self.gguf_writer.add_file_type(self.ftype)
# get pooling path
with open(self.dir_model / "modules.json", encoding="utf-8") as f:
modules = json.load(f)
pooling_path = None
for mod in modules:
if mod["type"] == "sentence_transformers.models.Pooling":
pooling_path = mod["path"]
break
# get pooling type
pooling_type = gguf.PoolingType.NONE
if pooling_path is not None:
with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
pooling = json.load(f)
if pooling["pooling_mode_mean_tokens"]:
pooling_type = gguf.PoolingType.MEAN
elif pooling["pooling_mode_cls_token"]:
pooling_type = gguf.PoolingType.CLS
else:
raise NotImplementedError("Only MEAN and CLS pooling types supported")
self.gguf_writer.add_pooling_type(pooling_type.value)
def set_vocab(self):
path = self.dir_model
@ -1658,6 +1682,7 @@ class BertModel(Model):
vocab = HfVocab(path, added_tokens_path)
tokens, scores, toktypes = zip(*vocab.all_tokens())
assert len(tokens) == vocab.vocab_size
self.vocab_size = vocab.vocab_size
# we need this to validate the size of the token_type embeddings
# though currently we are passing all zeros to the token_type embeddings
@ -1671,7 +1696,7 @@ class BertModel(Model):
if tok.startswith(b"##"):
return tok[2:]
return b"\xe2\x96\x81" + tok
tokens = [phantom(t, y) for t, y in zip(tokens, toktypes)]
tokens = tuple(phantom(t, y) for t, y in zip(tokens, toktypes))
# set up bos and eos tokens (cls and sep)
self.gguf_writer.add_bos_token_id(vocab.tokenizer.cls_token_id)
@ -1723,6 +1748,43 @@ class BertModel(Model):
self.gguf_writer.add_tensor(new_name, data)
class NomicBertModel(BertModel):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# the HF config claims n_ctx=8192, but it uses RoPE scaling
self.hparams["n_ctx"] = 2048
# SwigLU activation
assert self.hparams["activation_function"] == "swiglu"
# this doesn't do anything in the HF version
assert self.hparams["causal"] is False
# no bias tensors
assert self.hparams["qkv_proj_bias"] is False
assert self.hparams["mlp_fc1_bias"] is False
assert self.hparams["mlp_fc2_bias"] is False
# norm at end of layer
assert self.hparams["prenorm"] is False
# standard RoPE
assert self.hparams["rotary_emb_fraction"] == 1.0
assert self.hparams["rotary_emb_interleaved"] is False
assert self.hparams["rotary_emb_scale_base"] is None
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
def get_tensors(self):
assert self.vocab_size is not None
for name, data in super().get_tensors():
# Nomic Embed's token embeddings tensor is padded, but llama.cpp wants tensor sizes to match exactly.
if name == 'embeddings.word_embeddings.weight' and data.shape[1] != self.vocab_size:
rounded_vocab_size = (self.vocab_size + 63) // 64 * 64
assert data.shape == (rounded_vocab_size, self.hparams["n_embd"])
data = data[:self.vocab_size, :]
yield name, data
###### CONVERSION LOGIC ######

37
convert.py
View File

@ -1173,7 +1173,7 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM
for (name, tensor) in model.items()}
def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> LazyModel:
tmap = gguf.TensorNameMap(ARCH, params.n_layer)
should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
@ -1199,7 +1199,11 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
for name, lazy_tensor in model.items():
tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
if name_new is None:
raise Exception(f"Unexpected tensor name: {name}")
if skip_unknown:
print(f"Unexpected tensor name: {name} - skipping")
continue
else:
raise Exception(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
if tensor_type in should_skip:
print(f"skipping tensor {name_new}")
@ -1377,19 +1381,20 @@ def main(args_in: list[str] | None = None) -> None:
output_choices.append("q8_0")
vocab_types = ["spm", "bpe", "hfft"]
parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file")
parser.add_argument("--awq-path", type=Path, help="Path to scale awq cache file", default=None)
parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model")
parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file")
parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
parser.add_argument("--outtype", choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file")
parser.add_argument("--vocab-type", choices=vocab_types, help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm")
parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input")
parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
parser.add_argument("--ctx", type=int, help="model training context (default: based on input)")
parser.add_argument("--concurrency", type=int, help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
parser.add_argument("--big-endian", action="store_true", help="model is executed on big endian machine")
parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides")
parser.add_argument("--awq-path", type=Path, help="Path to scale awq cache file", default=None)
parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model")
parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file")
parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
parser.add_argument("--outtype", choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file")
parser.add_argument("--vocab-type", choices=vocab_types, help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm")
parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input")
parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
parser.add_argument("--ctx", type=int, help="model training context (default: based on input)")
parser.add_argument("--concurrency", type=int, help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
parser.add_argument("--big-endian", action="store_true", help="model is executed on big endian machine")
parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides")
parser.add_argument("--skip-unknown", action="store_true", help="skip unknown tensor names instead of failing")
args = parser.parse_args(args_in)
if args.awq_path:
@ -1461,7 +1466,7 @@ def main(args_in: list[str] | None = None) -> None:
print(f"Special vocab info: {special_vocab}")
model = model_plus.model
model = convert_model_names(model, params)
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)

1
examples/CMakeLists.txt
View File

@ -38,6 +38,7 @@ else()
add_subdirectory(speculative)
add_subdirectory(lookahead)
add_subdirectory(lookup)
add_subdirectory(gguf)
add_subdirectory(train-text-from-scratch)
add_subdirectory(imatrix)
if (LLAMA_BUILD_SERVER)

22
examples/baby-llama/baby-llama.cpp
View File

@ -1533,16 +1533,17 @@ int main(int argc, char ** argv) {
int n_past = 0;
ggml_cgraph gf = {};
struct ggml_cgraph * gf = NULL;
gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
get_example_targets_batch(ctx0, 64*ex+0, tokens_input, targets);
struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, &gf, tokens_input, n_tokens, n_past, n_batch);
struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, gf, tokens_input, n_tokens, n_past, n_batch);
// struct ggml_tensor * e = cross_entropy_loss(ctx0, targets, logits);
struct ggml_tensor * e = square_error_loss(ctx0, targets, logits);
ggml_build_forward_expand(&gf, e);
ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
ggml_build_forward_expand(gf, e);
ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
float error_before_opt = ggml_get_f32_1d(e, 0);
@ -1552,8 +1553,8 @@ int main(int argc, char ** argv) {
opt_params_lbfgs.lbfgs.n_iter = 16;
ggml_opt(ctx0, opt_params_lbfgs, e);
//
ggml_build_forward_expand(&gf, e);
ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
ggml_build_forward_expand(gf, e);
ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
float error_after_opt = ggml_get_f32_1d(e, 0);
@ -1600,13 +1601,14 @@ int main(int argc, char ** argv) {
};
struct ggml_context * ctx0 = ggml_init(params);
ggml_cgraph gf = {};
struct ggml_cgraph * gf = NULL;
gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
int n_past = 0;
struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, &gf, tokens_input, sample_ctx, n_past);
struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, gf, tokens_input, sample_ctx, n_past);
ggml_build_forward_expand(&gf, logits);
ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
ggml_build_forward_expand(gf, logits);
ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
struct ggml_tensor * best_samples = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sample_ctx);
struct ggml_tensor * probs = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_vocab, sample_ctx);

5
examples/batched-bench/batched-bench.cpp
View File

@ -82,7 +82,8 @@ int main(int argc, char ** argv) {
// init LLM
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
// initialize the model
@ -158,7 +159,7 @@ int main(int argc, char ** argv) {
}
LOG_TEE("\n");
LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %d, n_threads_batch = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
LOG_TEE("\n");
LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");

2
examples/batched.swift/Sources/main.swift
View File

@ -17,7 +17,7 @@ let n_parallel: Int = arguments.count > 3 && Int(arguments[3]) != nil ? Int(argu
let n_len: Int = 32
// init LLM
llama_backend_init(false)
llama_backend_init()
defer {
llama_backend_free()
}

5
examples/batched/batched.cpp
View File

@ -50,7 +50,8 @@ int main(int argc, char ** argv) {
// init LLM
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
// initialize the model
@ -91,7 +92,7 @@ int main(int argc, char ** argv) {
const int n_ctx = llama_n_ctx(ctx);
LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %d, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
// make sure the KV cache is big enough to hold all the prompt and generated tokens
if (n_kv_req > n_ctx) {

3
examples/beam-search/beam-search.cpp
View File

@ -119,7 +119,8 @@ int main(int argc, char ** argv)
// Init LLM :
//---------------------------------
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model;
llama_context * ctx;

38
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
View File

@ -325,14 +325,14 @@ struct train_params {
};
static void print_params(struct my_llama_hparams * params) {
printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
printf("%s: n_ctx: %d\n", __func__, params->n_ctx);
printf("%s: n_embd: %d\n", __func__, params->n_embd);
printf("%s: n_mult: %d\n", __func__, params->n_mult);
printf("%s: n_head: %d\n", __func__, params->n_head);
printf("%s: n_ff: %d\n", __func__, params->n_ff);
printf("%s: n_layer: %d\n", __func__, params->n_layer);
printf("%s: n_rot: %d\n", __func__, params->n_rot);
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_mult: %u\n", __func__, params->n_mult);
printf("%s: n_head: %u\n", __func__, params->n_head);
printf("%s: n_ff: %u\n", __func__, params->n_ff);
printf("%s: n_layer: %u\n", __func__, params->n_layer);
printf("%s: n_rot: %u\n", __func__, params->n_rot);
}
static void init_model(struct my_llama_model * model) {
@ -350,25 +350,25 @@ static void init_model(struct my_llama_model * model) {
model->train_tokens = 0;
model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
printf("[%s:GG] Allocating [%d] x [%d] = [%d] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
printf("[%s:GG] Allocating [%u] x [%u] = [%u] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
model->norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
printf("[%s:GG] Allocating [%d] float space for model->norm\n",__func__,n_embd);
printf("[%s:GG] Allocating [%u] float space for model->norm\n",__func__,n_embd);
model->output = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
// printing the per-layer allocations here so we dont print in the for loop.
printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wq for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wk for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wv for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wo for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wq for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wk for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wv for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wo for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
printf("[%s:GG] Allocating [%d] float space for layer.ffn_norm for [%d] layers\n",__func__,n_embd, n_layer);
printf("[%s:GG] Allocating [%u] float space for layer.ffn_norm for [%u] layers\n",__func__,n_embd, n_layer);
printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w1 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w2 for [%d] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w3 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w1 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w2 for [%u] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w3 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
ggml_set_name(model->norm, "norm.weight");

149
examples/embedding/embedding.cpp
View File

@ -7,6 +7,51 @@
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
static std::vector<std::string> split_lines(const std::string & s) {
std::string line;
std::vector<std::string> lines;
std::stringstream ss(s);
while (std::getline(ss, line)) {
lines.push_back(line);
}
return lines;
}
static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
for (size_t i = 0; i < tokens.size(); i++) {
llama_batch_add(batch, tokens[i], i, { seq_id }, false);
}
}
static void normalize(float * vec, float * out, int n) {
float norm = 0;
for (int i = 0; i < n; i++) {
norm += vec[i] * vec[i];
}
norm = sqrt(norm);
for (int i = 0; i < n; i++) {
out[i] = vec[i] / norm;
}
}
static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
// clear previous kv_cache values (irrelevant for embeddings)
llama_kv_cache_clear(ctx);
// run model
fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
if (llama_decode(ctx, batch) < 0) {
fprintf(stderr, "%s : failed to decode\n", __func__);
}
// normalize on copy
for (int k = 0; k < n_seq; k++) {
float * emb = llama_get_embeddings_ith(ctx, k);
float * out = output + k * n_embd;
normalize(emb, out, n_embd);
}
}
int main(int argc, char ** argv) {
gpt_params params;
@ -29,7 +74,8 @@ int main(int argc, char ** argv) {
params.prompt = gpt_random_prompt(rng);
}
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model;
llama_context * ctx;
@ -55,59 +101,84 @@ int main(int argc, char ** argv) {
fprintf(stderr, "%s\n", get_system_info(params).c_str());
}
int n_past = 0;
// split the prompt into lines
std::vector<std::string> prompts = split_lines(params.prompt);
// tokenize the prompt
auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
// max batch size
const uint64_t n_batch = params.n_batch;
GGML_ASSERT(params.n_batch == params.n_ctx);
// tokenize the prompts and trim
std::vector<std::vector<int32_t>> inputs;
for (const auto & prompt : prompts) {
auto inp = ::llama_tokenize(ctx, prompt, true);
if (inp.size() > n_batch) {
inp.resize(n_batch);
}
inputs.push_back(inp);
}
// tokenization stats
if (params.verbose_prompt) {
fprintf(stderr, "\n");
fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
for (int i = 0; i < (int) embd_inp.size(); i++) {
fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
for (int i = 0; i < (int) inputs.size(); i++) {
fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
for (int j = 0; j < (int) inputs[i].size(); j++) {
fprintf(stderr, "%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
}
fprintf(stderr, "\n\n");
}
fprintf(stderr, "\n");
}
if (embd_inp.size() > (size_t)n_ctx) {
fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
__func__, embd_inp.size(), n_ctx);
return 1;
}
while (!embd_inp.empty()) {
int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
if (llama_decode(ctx, llama_batch_get_one(embd_inp.data(), n_tokens, n_past, 0))) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return 1;
}
n_past += n_tokens;
embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_tokens);
}
// initialize batch
const int n_prompts = prompts.size();
struct llama_batch batch = llama_batch_init(n_batch, 0, n_prompts);
// allocate output
const int n_embd = llama_n_embd(model);
auto * embeddings = llama_get_embeddings(ctx);
std::vector<float> embeddings(n_prompts * n_embd, 0);
float * emb = embeddings.data();
// l2-normalize embeddings
float norm = 0;
for (int i = 0; i < n_embd; i++) {
norm += embeddings[i] * embeddings[i];
}
norm = sqrt(norm);
for (int i = 0; i < n_embd; i++) {
embeddings[i] /= norm;
// break into batches
int p = 0; // number of prompts processed already
int s = 0; // number of prompts in current batch
for (int k = 0; k < n_prompts; k++) {
// clamp to n_batch tokens
auto & inp = inputs[k];
const uint64_t n_toks = inp.size();
// encode if at capacity
if (batch.n_tokens + n_toks > n_batch) {
float * out = emb + p * n_embd;
batch_decode(ctx, batch, out, s, n_embd);
llama_batch_clear(batch);
p += s;
s = 0;
}
// add to batch
batch_add_seq(batch, inp, s);
s += 1;
}
for (int i = 0; i < n_embd; i++) {
printf("%f ", embeddings[i]);
}
printf("\n");
// final batch
float * out = emb + p * n_embd;
batch_decode(ctx, batch, out, s, n_embd);
// print first 3 embeddings
for (int j = 0; j < std::min(3, n_prompts); j++) {
fprintf(stderr, "embedding %d: ", j);
for (int i = 0; i < n_embd; i++) {
fprintf(stderr, "%f ", emb[j * n_embd + i]);
}
fprintf(stderr, "\n\n");
}
fprintf(stderr, "\n");
// clean up
llama_print_timings(ctx);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
return 0;

2
examples/export-lora/export-lora.cpp
View File

@ -7,8 +7,6 @@
#include <string>
#include <thread>
static const size_t tensor_alignment = 32;
struct lora_info {
std::string filename;
float scale;

6
examples/finetune/README.md
View File

@ -80,9 +80,9 @@ The LORA rank can be configured for each model tensor type separately with these
--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-w1 N LORA rank for w1 tensor (default 4)
--rank-w2 N LORA rank for w2 tensor (default 4)
--rank-w3 N LORA rank for w3 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.

242
examples/finetune/finetune.cpp
View File

@ -60,9 +60,9 @@ struct my_llama_layer {
struct ggml_tensor * ffn_norm;
// ff
struct ggml_tensor * w1;
struct ggml_tensor * w2;
struct ggml_tensor * w3;
struct ggml_tensor * ffn_gate; // w1
struct ggml_tensor * ffn_down; // w2
struct ggml_tensor * ffn_up; // w3
};
struct my_llama_model {
@ -85,9 +85,9 @@ struct my_llama_lora_hparams {
uint32_t n_rank_wv = 4;
uint32_t n_rank_wo = 4;
uint32_t n_rank_ffn_norm = 1;
uint32_t n_rank_w1 = 4;
uint32_t n_rank_w2 = 4;
uint32_t n_rank_w3 = 4;
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;
@ -117,12 +117,12 @@ struct my_llama_lora_layer {
struct ggml_tensor * ffn_norm_b;
// ff
struct ggml_tensor * w1_a;
struct ggml_tensor * w1_b;
struct ggml_tensor * w2_a;
struct ggml_tensor * w2_b;
struct ggml_tensor * w3_a;
struct ggml_tensor * w3_b;
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 {
@ -208,9 +208,9 @@ static void print_lora_params(struct my_llama_lora_hparams * params) {
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_w1 : %u\n", __func__, params->n_rank_w1);
printf("%s: n_rank_w2 : %u\n", __func__, params->n_rank_w2);
printf("%s: n_rank_w3 : %u\n", __func__, params->n_rank_w3);
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);
@ -319,9 +319,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
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.w1 = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
layer.w2 = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
layer.w3 = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, 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);
@ -329,9 +329,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
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.w1, hparams.n_embd, hparams.n_ff);
assert_shape_2d(layer.w2, hparams.n_ff, hparams.n_embd);
assert_shape_2d(layer.w3, hparams.n_embd, hparams.n_ff);
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);
}
}
@ -362,12 +362,12 @@ static void set_param_lora(struct my_llama_lora * lora) {
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.w1_a);
ggml_set_param(ctx, layer.w1_b);
ggml_set_param(ctx, layer.w2_a);
ggml_set_param(ctx, layer.w2_b);
ggml_set_param(ctx, layer.w3_a);
ggml_set_param(ctx, layer.w3_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);
}
}
@ -435,12 +435,12 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
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.w1_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_embd);
layer.w1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_ff);
layer.w2_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_ff);
layer.w2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_embd);
layer.w3_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_embd);
layer.w3_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_ff);
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));
@ -454,12 +454,12 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
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.w1_a, tni(LLM_TENSOR_FFN_GATE, ".weight.lora_a", i));
ggml_set_name(layer.w1_b, tni(LLM_TENSOR_FFN_GATE, ".weight.lora_b", i));
ggml_set_name(layer.w2_a, tni(LLM_TENSOR_FFN_DOWN, ".weight.lora_a", i));
ggml_set_name(layer.w2_b, tni(LLM_TENSOR_FFN_DOWN, ".weight.lora_b", i));
ggml_set_name(layer.w3_a, tni(LLM_TENSOR_FFN_UP, ".weight.lora_a", i));
ggml_set_name(layer.w3_b, tni(LLM_TENSOR_FFN_UP, ".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);
@ -497,12 +497,12 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl
randomize_tensor_normal(layer.ffn_norm_a, rnd);
ggml_set_zero(layer.ffn_norm_b);
randomize_tensor_normal(layer.w1_a, rnd);
ggml_set_zero(layer.w1_b);
randomize_tensor_normal(layer.w2_a, rnd);
ggml_set_zero(layer.w2_b);
randomize_tensor_normal(layer.w3_a, rnd);
ggml_set_zero(layer.w3_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);
@ -610,13 +610,13 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
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 * w1 = add_to_f32(ctx, layer.w1, ggml_mul_mat(ctx, llayer.w1_a, llayer.w1_b));
struct ggml_tensor * w2 = add_to_f32(ctx, layer.w2, ggml_mul_mat(ctx, llayer.w2_a, llayer.w2_b));
struct ggml_tensor * w3 = add_to_f32(ctx, layer.w3, ggml_mul_mat(ctx, llayer.w3_a, llayer.w3_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);
@ -659,11 +659,11 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
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, w3, t24); set_name(t25, "t25"); assert_shape_2d(t25, n_ff, N*n_batch);
struct ggml_tensor * t26 = ggml_mul_mat (ctx, w1, t24); set_name(t26, "t26"); assert_shape_2d(t26, n_ff, 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, w2, t28); set_name(t29, "t29"); assert_shape_2d(t29, n_embd, 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) {
@ -723,9 +723,9 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
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.w1, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, 1.0f));
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, 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
@ -798,9 +798,9 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
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_w1, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
GGUF_GET_KEY(fctx, lora->hparams.n_rank_w2, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
GGUF_GET_KEY(fctx, lora->hparams.n_rank_w3, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
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);
@ -825,12 +825,12 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
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.w1_a, f_ggml_ctx, ggml_get_name(layer.w1_a));
copy_tensor_by_name(layer.w1_b, f_ggml_ctx, ggml_get_name(layer.w1_b));
copy_tensor_by_name(layer.w2_a, f_ggml_ctx, ggml_get_name(layer.w2_a));
copy_tensor_by_name(layer.w2_b, f_ggml_ctx, ggml_get_name(layer.w2_b));
copy_tensor_by_name(layer.w3_a, f_ggml_ctx, ggml_get_name(layer.w3_a));
copy_tensor_by_name(layer.w3_b, f_ggml_ctx, ggml_get_name(layer.w3_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));
}
}
@ -868,9 +868,9 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
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_w1);
gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN, lora->hparams.n_rank_w2);
gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP, lora->hparams.n_rank_w3);
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);
@ -894,12 +894,12 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
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.w1_a);
gguf_add_tensor(fctx, layer.w1_b);
gguf_add_tensor(fctx, layer.w2_a);
gguf_add_tensor(fctx, layer.w2_b);
gguf_add_tensor(fctx, layer.w3_a);
gguf_add_tensor(fctx, layer.w3_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);
}
}
@ -1104,12 +1104,12 @@ static void save_as_llama_lora(const char * filename, struct my_llama_lora * lor
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.w1_a, tni(LLM_TENSOR_FFN_GATE, i, ".weight.loraA"));
write_tensor(&file, layer.w1_b, tni(LLM_TENSOR_FFN_GATE, i, ".weight.loraB"));
write_tensor(&file, layer.w2_a, tni(LLM_TENSOR_FFN_DOWN, i, ".weight.loraA"));
write_tensor(&file, layer.w2_b, tni(LLM_TENSOR_FFN_DOWN, i, ".weight.loraB"));
write_tensor(&file, layer.w3_a, tni(LLM_TENSOR_FFN_UP, i, ".weight.loraA"));
write_tensor(&file, layer.w3_b, tni(LLM_TENSOR_FFN_UP, 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"));
}
}
@ -1139,9 +1139,9 @@ struct train_params {
uint32_t n_rank_wv;
uint32_t n_rank_wo;
uint32_t n_rank_ffn_norm;
uint32_t n_rank_w1;
uint32_t n_rank_w2;
uint32_t n_rank_w3;
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;
@ -1152,9 +1152,9 @@ struct train_params {
bool custom_n_rank_wv;
bool custom_n_rank_wo;
bool custom_n_rank_ffn_norm;
bool custom_n_rank_w1;
bool custom_n_rank_w2;
bool custom_n_rank_w3;
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;
@ -1186,9 +1186,9 @@ static struct train_params get_default_train_params() {
params.n_rank_wv = 4;
params.n_rank_wo = 4;
params.n_rank_ffn_norm = 1;
params.n_rank_w1 = 4;
params.n_rank_w2 = 4;
params.n_rank_w3 = 4;
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;
@ -1199,9 +1199,9 @@ static struct train_params get_default_train_params() {
params.custom_n_rank_wv = false;
params.custom_n_rank_wo = false;
params.custom_n_rank_ffn_norm = false;
params.custom_n_rank_w1 = false;
params.custom_n_rank_w2 = false;
params.custom_n_rank_w3 = 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;
@ -1232,9 +1232,9 @@ static void train_print_usage(int argc, char ** argv, const struct train_params
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-w1 N LORA rank for w1 tensor, overrides default rank.\n");
fprintf(stderr, " --rank-w2 N LORA rank for w2 tensor, overrides default rank.\n");
fprintf(stderr, " --rank-w3 N LORA rank for w3 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, &params->common);
}
@ -1369,27 +1369,27 @@ static bool train_params_parse(int argc, char ** argv, struct train_params * par
}
params->n_rank_wo = std::stoi(argv[i]);
params->custom_n_rank_wo = true;
} else if (arg == "--rank-w1") {
} else if (arg == "--rank-ffn_gate") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->n_rank_w1 = std::stoi(argv[i]);
params->custom_n_rank_w1 = true;
} else if (arg == "--rank-w2") {
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_w2 = std::stoi(argv[i]);
params->custom_n_rank_w2 = true;
} else if (arg == "--rank-w3") {
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_w3 = std::stoi(argv[i]);
params->custom_n_rank_w3 = true;
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);
@ -1452,12 +1452,12 @@ static int64_t get_parameter_count(struct my_llama_lora* lora) {
nx += ggml_nelements(layer.wo_b);
nx += ggml_nelements(layer.ffn_norm_a);
nx += ggml_nelements(layer.ffn_norm_b);
nx += ggml_nelements(layer.w1_a);
nx += ggml_nelements(layer.w1_b);
nx += ggml_nelements(layer.w2_a);
nx += ggml_nelements(layer.w2_b);
nx += ggml_nelements(layer.w3_a);
nx += ggml_nelements(layer.w3_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;
}
@ -1511,9 +1511,9 @@ int main(int argc, char ** argv) {
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_w1 = params.custom_n_rank_w1 ? params.n_rank_w1 : params.lora_r;
uint32_t n_rank_w2 = params.custom_n_rank_w2 ? params.n_rank_w2 : params.lora_r;
uint32_t n_rank_w3 = params.custom_n_rank_w3 ? params.n_rank_w3 : params.lora_r;
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;
@ -1523,9 +1523,9 @@ int main(int argc, char ** argv) {
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_w1 = n_rank_w1;
lora.hparams.n_rank_w2 = n_rank_w2;
lora.hparams.n_rank_w3 = n_rank_w3;
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;
@ -1566,9 +1566,9 @@ int main(int argc, char ** argv) {
|| (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_w1 != n_rank_w1)
|| (lora.hparams.n_rank_w2 != n_rank_w2)
|| (lora.hparams.n_rank_w3 != n_rank_w3)
|| (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)

3
examples/imatrix/imatrix.cpp
View File

@ -568,7 +568,8 @@ int main(int argc, char ** argv) {
params.prompt = gpt_random_prompt(rng);
}
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
llama_model_params mparams = llama_model_params_from_gpt_params(params);

3
examples/infill/infill.cpp
View File

@ -202,7 +202,8 @@ int main(int argc, char ** argv) {
std::mt19937 rng(params.seed);
LOG("%s: llama backend init\n", __func__);
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model;
llama_context * ctx;

3
examples/llama-bench/llama-bench.cpp
View File

@ -1151,8 +1151,7 @@ int main(int argc, char ** argv) {
if (!params.verbose) {
llama_log_set(llama_null_log_callback, NULL);
}
bool numa = false;
llama_backend_init(numa);
llama_backend_init();
// initialize printer
std::unique_ptr<printer> p;

4
examples/llama.android/app/src/main/cpp/llama-android.cpp
View File

@ -274,8 +274,8 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb
extern "C"
JNIEXPORT void JNICALL
Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject, jboolean numa) {
llama_backend_init(numa);
Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) {
llama_backend_init();
}
extern "C"

2
examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
View File

@ -51,7 +51,7 @@ actor LlamaContext {
}
static func create_context(path: String) throws -> LlamaContext {
llama_backend_init(false)
llama_backend_init()
var model_params = llama_model_default_params()
#if targetEnvironment(simulator)

54
examples/llava/README.md
View File

@ -1,10 +1,12 @@
# LLaVA
Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants.
Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants,
as well as llava-1.6 [llava-v1.6](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2) variants.
The pre-converted [7b](https://huggingface.co/mys/ggml_llava-v1.5-7b)
and [13b](https://huggingface.co/mys/ggml_llava-v1.5-13b)
models are available.
For llava-1.6 a variety of prepared gguf models are available as well [7b-34b](https://huggingface.co/cmp-nct/llava-1.6-gguf)
After API is confirmed, more models will be supported / uploaded.
@ -18,10 +20,11 @@ After building, run: `./llava-cli` to see the usage. For example:
```
**note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
**note**: For GPU offloading ensure to use the `-ngl` flag just like usual
## Model conversion
## LLaVA 1.5
- Clone `llava-v15-7b` and `clip-vit-large-patch14-336` locally:
- Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
```sh
git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
@ -50,13 +53,54 @@ python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-pa
5. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
```sh
python ./convert.py ../llava-v1.5-7b
python ./convert.py ../llava-v1.5-7b --skip-unknown
```
Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` directory.
## LLaVA 1.6 gguf conversion
1) Backup your pth/safetensor model files as llava-surgery modifies them
2) Use `python llava-surgery-v2.py -C -m /path/to/hf-model` which also supports llava-1.5 variants pytorch as well as safetensor models:
- you will find a llava.projector and a llava.clip file in your model directory
3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory (https://huggingface.co/cmp-nct/llava-1.6-gguf/blob/main/config_vit.json) and rename it to config.json.
4) Create the visual gguf model: `python ./examples/llava/convert-image-encoder-to-gguf.py -m ../path/to/vit --llava-projector ../path/to/llava.projector --output-dir ../path/to/output --clip-model-is-vision`
- This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
5) Everything else as usual: convert.py the hf model, quantize as needed
**note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
**note** llava-1.6 greatly benefits from batched prompt processing (defaults work)
## llava-cli templating and llava-1.6 prompting
llava-1.5 models all use the same vicuna prompt, here you can just add your image question like `-p "Provide a full description."`
For llava-1.5 models which are not vicuna (mistral and Yi) you need to adapt system prompt as well as user prompt, for this purpose llava-cli has a basic templating system:
**For Mistral and using llava-cli binary:**
Add this: `-p "<image>\nUSER:\nProvide a full description.\nASSISTANT:\n"`
The mistral template for llava-1.6 seems to be no system print and a USER/ASSISTANT role
**For the 34B this should work:**
Add this: `-e -p <|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nProvide a full description.<|im_end|><|im_start|>assistant\n`
## How to know if you are running in llava-1.5 or llava-1.6 mode
When running llava-cli you will see a visual information right before the prompt is being processed:
**Llava-1.5:**
`encode_image_with_clip: image embedding created: 576 tokens`
**Llava-1.6 (anything above 576):**
`encode_image_with_clip: image embedding created: 2880 tokens`
Alternatively just pay notice to how many "tokens" have been used for your prompt, it will also show 1000+ tokens for llava-1.6
## TODO
- [ ] Support non-CPU backend for the image encoding part.
- [x] Support non-CPU backend for the image encoding part.
- [ ] Support different sampling methods.
- [ ] Support more model variants.

778
examples/llava/clip.cpp
View File

@ -1,7 +1,7 @@
// NOTE: This is modified from clip.cpp only for LLaVA,
// so there might be still unnecessary artifacts hanging around
// I'll gradually clean and extend it
// Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
#include "clip.h"
#include "ggml.h"
#include "ggml-alloc.h"
@ -30,6 +30,26 @@
#include <vector>
#include <sstream>
#include <cinttypes>
#include <limits>
//#define CLIP_DEBUG_FUNCTIONS
// RGB uint8 image
struct clip_image_u8 {
int nx;
int ny;
std::vector<uint8_t> buf;
};
// RGB float32 image (NHWC)
// Memory layout: RGBRGBRGB...
struct clip_image_f32 {
int nx;
int ny;
std::vector<float> buf;
};
static std::string format(const char * fmt, ...) {
va_list ap;
@ -50,50 +70,56 @@ static std::string format(const char * fmt, ...) {
// key constants
//
#define KEY_FTYPE "general.file_type"
#define KEY_NAME "general.name"
#define KEY_DESCRIPTION "general.description"
#define KEY_HAS_TEXT_ENC "clip.has_text_encoder"
#define KEY_HAS_VIS_ENC "clip.has_vision_encoder"
#define KEY_FTYPE "general.file_type"
#define KEY_NAME "general.name"
#define KEY_DESCRIPTION "general.description"
#define KEY_HAS_TEXT_ENC "clip.has_text_encoder"
#define KEY_HAS_VIS_ENC "clip.has_vision_encoder"
#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector"
#define KEY_USE_GELU "clip.use_gelu"
#define KEY_N_EMBD "clip.%s.embedding_length"
#define KEY_N_FF "clip.%s.feed_forward_length"
#define KEY_N_BLOCK "clip.%s.block_count"
#define KEY_N_HEAD "clip.%s.attention.head_count"
#define KEY_USE_GELU "clip.use_gelu"
#define KEY_N_EMBD "clip.%s.embedding_length"
#define KEY_N_FF "clip.%s.feed_forward_length"
#define KEY_N_BLOCK "clip.%s.block_count"
#define KEY_N_HEAD "clip.%s.attention.head_count"
#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon"
#define KEY_PROJ_DIM "clip.%s.projection_dim"
#define KEY_TOKENS "tokenizer.ggml.tokens"
#define KEY_N_POSITIONS "clip.text.context_length"
#define KEY_IMAGE_SIZE "clip.vision.image_size"
#define KEY_PATCH_SIZE "clip.vision.patch_size"
#define KEY_IMAGE_MEAN "clip.vision.image_mean"
#define KEY_IMAGE_STD "clip.vision.image_std"
#define KEY_PROJ_TYPE "clip.projector_type"
#define KEY_PROJ_DIM "clip.%s.projection_dim"
#define KEY_TOKENS "tokenizer.ggml.tokens"
#define KEY_N_POSITIONS "clip.text.context_length"
#define KEY_IMAGE_SIZE "clip.vision.image_size"
#define KEY_PATCH_SIZE "clip.vision.patch_size"
#define KEY_IMAGE_MEAN "clip.vision.image_mean"
#define KEY_IMAGE_STD "clip.vision.image_std"
#define KEY_PROJ_TYPE "clip.projector_type"
#define KEY_MM_PATCH_MERGE_TYPE "clip.vision.mm_patch_merge_type"
#define KEY_IMAGE_GRID_PINPOINTS "clip.vision.image_grid_pinpoints"
#define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
//
// tensor name constants
//
#define TN_TOKEN_EMBD "%s.token_embd.weight"
#define TN_POS_EMBD "%s.position_embd.weight"
#define TN_CLASS_EMBD "v.class_embd"
#define TN_PATCH_EMBD "v.patch_embd.weight"
#define TN_ATTN_K "%s.blk.%d.attn_k.%s"
#define TN_ATTN_Q "%s.blk.%d.attn_q.%s"
#define TN_ATTN_V "%s.blk.%d.attn_v.%s"
#define TN_ATTN_OUTPUT "%s.blk.%d.attn_out.%s"
#define TN_FFN_DOWN "%s.blk.%d.ffn_down.%s"
#define TN_FFN_UP "%s.blk.%d.ffn_up.%s"
#define TN_LN_1 "%s.blk.%d.ln1.%s"
#define TN_LN_2 "%s.blk.%d.ln2.%s"
#define TN_LN_PRE "%s.pre_ln.%s"
#define TN_LN_POST "%s.post_ln.%s"
#define TN_TEXT_PROJ "text_projection.weight"
#define TN_VIS_PROJ "visual_projection.weight"
#define TN_LLAVA_PROJ "mm.%d.%s"
#define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s"
#define TN_TOKEN_EMBD "%s.token_embd.weight"
#define TN_POS_EMBD "%s.position_embd.weight"
#define TN_CLASS_EMBD "v.class_embd"
#define TN_PATCH_EMBD "v.patch_embd.weight"
#define TN_ATTN_K "%s.blk.%d.attn_k.%s"
#define TN_ATTN_Q "%s.blk.%d.attn_q.%s"
#define TN_ATTN_V "%s.blk.%d.attn_v.%s"
#define TN_ATTN_OUTPUT "%s.blk.%d.attn_out.%s"
#define TN_FFN_DOWN "%s.blk.%d.ffn_down.%s"
#define TN_FFN_UP "%s.blk.%d.ffn_up.%s"
#define TN_LN_1 "%s.blk.%d.ln1.%s"
#define TN_LN_2 "%s.blk.%d.ln2.%s"
#define TN_LN_PRE "%s.pre_ln.%s"
#define TN_LN_POST "%s.post_ln.%s"
#define TN_TEXT_PROJ "text_projection.weight"
#define TN_VIS_PROJ "visual_projection.weight"
#define TN_LLAVA_PROJ "mm.%d.%s"
#define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s"
#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
#define TN_IMAGE_NEWLINE "model.image_newline"
enum projector_type {
@ -104,8 +130,8 @@ enum projector_type {
};
static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
{ PROJECTOR_TYPE_MLP, "mlp" },
{ PROJECTOR_TYPE_LDP, "ldp" },
{ PROJECTOR_TYPE_MLP, "mlp" },
{ PROJECTOR_TYPE_LDP, "ldp" },
};
@ -165,7 +191,6 @@ static std::string gguf_data_to_str(enum gguf_type type, const void * data, int
}
}
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()) {
@ -217,7 +242,7 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
}
}
static void print_tensor_info(const ggml_tensor* tensor, const char* prefix = "") {
static void print_tensor_info(const ggml_tensor * tensor, const char * prefix = "") {
size_t tensor_size = ggml_nbytes(tensor);
printf("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n",
prefix, ggml_n_dims(tensor), tensor->name, tensor_size,
@ -233,31 +258,136 @@ static projector_type clip_projector_type_from_string(const std::string & name)
return PROJECTOR_TYPE_UNKNOWN;
}
//
// image data
//
#ifdef CLIP_DEBUG_FUNCTIONS
static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::string& filename) {
std::ofstream file(filename, std::ios::binary);
if (!file.is_open()) {
std::cerr << "Failed to open file for writing: " << filename << std::endl;
return;
}
// RGB uint8 image
struct clip_image_u8 {
int nx;
int ny;
// PPM header: P6 format, width, height, and max color value
file << "P6\n" << img.nx << " " << img.ny << "\n255\n";
std::vector<uint8_t> buf;
};
// Write pixel data
for (size_t i = 0; i < img.buf.size(); i += 3) {
// PPM expects binary data in RGB format, which matches our image buffer
file.write(reinterpret_cast<const char*>(&img.buf[i]), 3);
}
// RGB float32 image (NHWC)
// Memory layout: RGBRGBRGB...
struct clip_image_f32 {
int nx;
int ny;
file.close();
}
static void clip_image_save_to_bmp(const clip_image_u8& img, const std::string& filename) {
std::ofstream file(filename, std::ios::binary);
if (!file.is_open()) {
std::cerr << "Failed to open file for writing: " << filename << std::endl;
return;
}
int fileSize = 54 + 3 * img.nx * img.ny; // File header + info header + pixel data
int bytesPerPixel = 3;
int widthInBytes = img.nx * bytesPerPixel;
int paddingAmount = (4 - (widthInBytes % 4)) % 4;
int stride = widthInBytes + paddingAmount;
// Bitmap file header
unsigned char fileHeader[14] = {
'B','M', // Signature
0,0,0,0, // Image file size in bytes
0,0,0,0, // Reserved
54,0,0,0 // Start of pixel array
};
// Total file size
fileSize = 54 + (stride * img.ny);
fileHeader[2] = (unsigned char)(fileSize);
fileHeader[3] = (unsigned char)(fileSize >> 8);
fileHeader[4] = (unsigned char)(fileSize >> 16);
fileHeader[5] = (unsigned char)(fileSize >> 24);
// Bitmap information header (BITMAPINFOHEADER)
unsigned char infoHeader[40] = {
40,0,0,0, // Size of this header (40 bytes)
0,0,0,0, // Image width
0,0,0,0, // Image height
1,0, // Number of color planes
24,0, // Bits per pixel
0,0,0,0, // No compression
0,0,0,0, // Image size (can be 0 for no compression)
0,0,0,0, // X pixels per meter (not specified)
0,0,0,0, // Y pixels per meter (not specified)
0,0,0,0, // Total colors (color table not used)
0,0,0,0 // Important colors (all are important)
};
// Width and height in the information header
infoHeader[4] = (unsigned char)(img.nx);
infoHeader[5] = (unsigned char)(img.nx >> 8);
infoHeader[6] = (unsigned char)(img.nx >> 16);
infoHeader[7] = (unsigned char)(img.nx >> 24);
infoHeader[8] = (unsigned char)(img.ny);
infoHeader[9] = (unsigned char)(img.ny >> 8);
infoHeader[10] = (unsigned char)(img.ny >> 16);
infoHeader[11] = (unsigned char)(img.ny >> 24);
// Write file headers
file.write(reinterpret_cast<char*>(fileHeader), sizeof(fileHeader));
file.write(reinterpret_cast<char*>(infoHeader), sizeof(infoHeader));
// Pixel data
std::vector<unsigned char> padding(3, 0); // Max padding size to be added to each row
for (int y = img.ny - 1; y >= 0; --y) { // BMP files are stored bottom-to-top
for (int x = 0; x < img.nx; ++x) {
// Each pixel
size_t pixelIndex = (y * img.nx + x) * 3;
unsigned char pixel[3] = {
img.buf[pixelIndex + 2], // BMP stores pixels in BGR format
img.buf[pixelIndex + 1],
img.buf[pixelIndex]
};
file.write(reinterpret_cast<char*>(pixel), 3);
}
// Write padding for the row
file.write(reinterpret_cast<char*>(padding.data()), paddingAmount);
}
file.close();
}
// debug function to convert f32 to u8
static void clip_image_convert_f32_to_u8(const clip_image_f32& src, clip_image_u8& dst) {
dst.nx = src.nx;
dst.ny = src.ny;
dst.buf.resize(3 * src.nx * src.ny);
for (size_t i = 0; i < src.buf.size(); ++i) {
dst.buf[i] = static_cast<uint8_t>(std::min(std::max(int(src.buf[i] * 255.0f), 0), 255));
}
}
#endif
std::vector<float> buf;
};
//
// clip layers
//
struct clip_hparams {
int32_t image_size;
int32_t patch_size;
int32_t hidden_size;
int32_t n_intermediate;
int32_t projection_dim;
int32_t n_head;
int32_t n_layer;
float eps;
char mm_patch_merge_type[32] = "flat"; // spatial_unpad or flat (default)
int32_t image_grid_pinpoints[32];
int32_t image_crop_resolution;
};
struct clip_layer {
// attention
struct ggml_tensor * k_w;
@ -287,7 +417,7 @@ struct clip_layer {
};
struct clip_vision_model {
struct clip_vision_hparams hparams;
struct clip_hparams hparams;
// embeddings
struct ggml_tensor * class_embedding;
@ -310,6 +440,8 @@ struct clip_vision_model {
struct ggml_tensor * mm_2_w = NULL;
struct ggml_tensor * mm_2_b = NULL;
struct ggml_tensor * image_newline = NULL;
// Yi type models with mlp+normalization projection
struct ggml_tensor * mm_1_w = NULL; // Yi type models have 0, 1, 3, 4
struct ggml_tensor * mm_1_b = NULL;
@ -364,9 +496,10 @@ struct clip_ctx {
std::vector<uint8_t> buf_compute_meta;
// memory buffers to evaluate the model
ggml_backend_buffer_t params_buffer = NULL;
ggml_backend_buffer_t params_buffer = NULL;
ggml_backend_buffer_t compute_buffer = NULL;
ggml_backend_t backend = NULL;
ggml_backend_t backend = NULL;
ggml_gallocr_t compute_alloc = NULL;
};
@ -379,18 +512,19 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
const auto & model = ctx->vision_model;
const auto & hparams = model.hparams;
const int image_size = hparams.image_size;
const int patch_size = hparams.patch_size;
const int num_patches = ((image_size / patch_size) * (image_size / patch_size));
const int num_positions = num_patches + 1;
const int hidden_size = hparams.hidden_size;
const int n_head = hparams.n_head;
const int d_head = hidden_size / n_head;
const int n_layer = hparams.n_layer;
//const int n_intermediate = hparams.n_intermediate;
//const int projection_dim = hparams.projection_dim;
const float eps = hparams.eps;
int batch_size = imgs->size;
const int image_size = hparams.image_size;
const int patch_size = hparams.patch_size;
const int num_patches = ((image_size / patch_size) * (image_size / patch_size));
const int num_patches_per_side = image_size / patch_size; GGML_UNUSED(num_patches_per_side);
const int num_positions = num_patches + 1;
const int hidden_size = hparams.hidden_size;
const int n_head = hparams.n_head;
const int d_head = hidden_size / n_head;
const int n_layer = hparams.n_layer;
const float eps = hparams.eps;
const int batch_size = imgs->size;
if (ctx->has_llava_projector) {
GGML_ASSERT(batch_size == 1);
}
@ -540,7 +674,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
embeddings = ggml_gelu(ctx0, embeddings);
embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
@ -791,10 +924,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
if (idx != -1) {
const std::string proj_type = gguf_get_val_str(ctx, idx);
new_clip->proj_type = clip_projector_type_from_string(proj_type);
}
else {
} else {
new_clip->proj_type = PROJECTOR_TYPE_MLP;
}
if (new_clip->proj_type == PROJECTOR_TYPE_MLP) {
if (gguf_find_tensor(ctx, format(TN_LLAVA_PROJ, 3, "weight").c_str()) != -1) {
new_clip->proj_type = PROJECTOR_TYPE_MLP_NORM;
@ -920,11 +1053,41 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
hparams.projection_dim = get_u32(ctx, format(KEY_PROJ_DIM, "vision"));
hparams.eps = get_f32(ctx, format(KEY_LAYER_NORM_EPS, "vision"));
try {
int idx = get_key_idx(ctx, KEY_IMAGE_GRID_PINPOINTS);
int n = gguf_get_arr_n(ctx, idx);
const int32_t * pinpoints = (const int32_t *)gguf_get_arr_data(ctx, idx);
for (int i = 0; i < 32 && i < n && pinpoints[i] != 0; ++i) {
hparams.image_grid_pinpoints[i] = pinpoints[i];
}
if (n < 32)
hparams.image_grid_pinpoints[n] = 0;
} catch (std::runtime_error & e) {
hparams.image_grid_pinpoints[0]=0;
}
try {
int idx = get_key_idx(ctx, KEY_MM_PATCH_MERGE_TYPE);
strcpy(hparams.mm_patch_merge_type, gguf_get_val_str(ctx, idx));
} catch (std::runtime_error & e) {
strcpy(hparams.mm_patch_merge_type, "flat");
}
try {
hparams.image_crop_resolution = get_u32(ctx, KEY_IMAGE_CROP_RESOLUTION); // llava-1.6
} catch(const std::exception& e) {
hparams.image_crop_resolution = hparams.image_size;
}
int idx_mean = get_key_idx(ctx, KEY_IMAGE_MEAN);
int idx_std = get_key_idx(ctx, KEY_IMAGE_STD);
const float * mean_data = (const float *)gguf_get_arr_data(ctx, idx_mean);
const float * std_data = (const float *)gguf_get_arr_data(ctx, idx_std);
for (int i = 0; i < 3; ++i) {
new_clip->image_mean[i] = *((const float *)gguf_get_arr_data(ctx, idx_mean));
new_clip->image_std[i] = *((const float *)gguf_get_arr_data(ctx, idx_std));
new_clip->image_mean[i] = mean_data[i];
new_clip->image_std[i] = std_data[i];
}
if (verbosity >= 2) {
@ -936,13 +1099,27 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
printf("v_projection_dim %d\n", hparams.projection_dim);
printf("v_n_head %d\n", hparams.n_head);
printf("v_n_layer %d\n", hparams.n_layer);
printf("v_eps %f\n", hparams.eps);
printf("v_image_mean %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
printf("v_image_std %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
printf("v_image_grid_pinpoints: ");
for (int i = 0; i < 32 && (hparams.image_grid_pinpoints[i] != 0); ++i) {
printf("%d ", hparams.image_grid_pinpoints[i]);
}
printf("\n");
printf("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
}
vision_model.patch_embeddings = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
vision_model.class_embedding = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD);
vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
vision_model.pre_ln_w = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
vision_model.pre_ln_b = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
try {
vision_model.patch_embeddings = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
vision_model.class_embedding = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD);
vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
vision_model.pre_ln_w = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
vision_model.pre_ln_b = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
} catch(const std::exception& e) {
fprintf(stderr, "%s: failed to load vision model tensors\n", __func__);
}
// LLaVA projection
if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) {
@ -968,40 +1145,43 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
vision_model.mm_4_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "weight"));
vision_model.mm_4_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "bias"));
} catch (std::runtime_error & e) { }
}
else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) {
try {
vision_model.image_newline = get_tensor(new_clip->ctx_data, TN_IMAGE_NEWLINE);
// fprintf(stderr, "%s: image_newline tensor (llava-1.6) found\n", __func__);
} catch (std::runtime_error & e) { }
} else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) {
// MobileVLM projection
vision_model.mm_model_mlp_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "weight"));
vision_model.mm_model_mlp_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "bias"));
vision_model.mm_model_mlp_3_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "weight"));
vision_model.mm_model_mlp_3_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "bias"));
vision_model.mm_model_block_1_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "0.weight"));
vision_model.mm_model_block_1_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.weight"));
vision_model.mm_model_block_1_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.bias"));
vision_model.mm_model_mlp_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "weight"));
vision_model.mm_model_mlp_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "bias"));
vision_model.mm_model_mlp_3_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "weight"));
vision_model.mm_model_mlp_3_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "bias"));
vision_model.mm_model_block_1_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "0.weight"));
vision_model.mm_model_block_1_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.weight"));
vision_model.mm_model_block_1_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.bias"));
vision_model.mm_model_block_1_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.weight"));
vision_model.mm_model_block_1_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.bias"));
vision_model.mm_model_block_1_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.weight"));
vision_model.mm_model_block_1_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.bias"));
vision_model.mm_model_block_1_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "0.weight"));
vision_model.mm_model_block_1_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.weight"));
vision_model.mm_model_block_1_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.bias"));
vision_model.mm_model_block_2_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "0.weight"));
vision_model.mm_model_block_2_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.weight"));
vision_model.mm_model_block_2_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.bias"));
vision_model.mm_model_block_1_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "0.weight"));
vision_model.mm_model_block_1_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.weight"));
vision_model.mm_model_block_1_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.bias"));
vision_model.mm_model_block_2_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "0.weight"));
vision_model.mm_model_block_2_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.weight"));
vision_model.mm_model_block_2_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.bias"));
vision_model.mm_model_block_2_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.weight"));
vision_model.mm_model_block_2_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.bias"));
vision_model.mm_model_block_2_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.weight"));
vision_model.mm_model_block_2_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.bias"));
vision_model.mm_model_block_2_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight"));
vision_model.mm_model_block_2_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight"));
vision_model.mm_model_block_2_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias"));
}
else {
vision_model.mm_model_block_2_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight"));
vision_model.mm_model_block_2_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight"));
vision_model.mm_model_block_2_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias"));
} else {
std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
}
vision_model.layers.resize(hparams.n_layer);
for (int il = 0; il < hparams.n_layer; ++il) {
auto & layer = vision_model.layers[il];
layer.k_w = get_tensor(new_clip->ctx_data, format(TN_ATTN_K, "v", il, "weight"));
@ -1050,8 +1230,20 @@ struct clip_image_f32 * clip_image_f32_init() {
return new clip_image_f32();
}
void clip_image_u8_free (struct clip_image_u8 * img) { delete img; }
void clip_image_u8_free(struct clip_image_u8 * img) { delete img; }
void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
void clip_image_u8_batch_free(struct clip_image_u8_batch & batch) {
if (batch.size > 0) {
delete[] batch.data;
batch.size = 0;
}
}
void clip_image_f32_batch_free(struct clip_image_f32_batch & batch) {
if (batch.size > 0) {
delete[] batch.data;
batch.size = 0;
}
}
static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) {
img->nx = nx;
@ -1084,24 +1276,252 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length
return true;
}
// normalize: x = (x - mean) / std
// TODO: implement bicubic interpolation instead of linear.
bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32 * res, const bool pad2square) {
// Linear interpolation between two points
inline float lerp(float s, float e, float t) {
return s + (e - s) * t;
}
// Bilinear resize function
static void bilinear_resize(const clip_image_u8& src, clip_image_u8& dst, int target_width, int target_height) {
dst.nx = target_width;
dst.ny = target_height;
dst.buf.resize(3 * target_width * target_height);
float x_ratio = static_cast<float>(src.nx - 1) / target_width;
float y_ratio = static_cast<float>(src.ny - 1) / target_height;
for (int y = 0; y < target_height; y++) {
for (int x = 0; x < target_width; x++) {
float px = x_ratio * x;
float py = y_ratio * y;
int x_floor = static_cast<int>(px);
int y_floor = static_cast<int>(py);
float x_lerp = px - x_floor;
float y_lerp = py - y_floor;
for (int c = 0; c < 3; c++) {
float top = lerp(
static_cast<float>(src.buf[3 * (y_floor * src.nx + x_floor) + c]),
static_cast<float>(src.buf[3 * (y_floor * src.nx + (x_floor + 1)) + c]),
x_lerp
);
float bottom = lerp(
static_cast<float>(src.buf[3 * ((y_floor + 1) * src.nx + x_floor) + c]),
static_cast<float>(src.buf[3 * ((y_floor + 1) * src.nx + (x_floor + 1)) + c]),
x_lerp
);
dst.buf[3 * (y * target_width + x) + c] = static_cast<uint8_t>(lerp(top, bottom, y_lerp));
}
}
}
}
// Normalize image to float32 - careful with pytorch .to(model.device, dtype=torch.float16) - this sometimes reduces precision (32>16>32), sometimes not
static void normalize_image_u8_to_f32(const clip_image_u8* src, clip_image_f32* dst, const float mean[3], const float std[3]) {
dst->nx = src->nx;
dst->ny = src->ny;
dst->buf.resize(src->buf.size());
for (size_t i = 0; i < src->buf.size(); ++i) {
int c = i % 3; // rgb
dst->buf[i] = (static_cast<float>(src->buf[i]) / 255.0f - mean[c]) / std[c];
}
}
inline float clip(float x, float lower, float upper) {
return std::max(lower, std::min(x, upper));
}
static bool bicubic_resize(const clip_image_u8 &img, clip_image_u8 &dst, int target_width, int target_height) {
const int nx = img.nx;
const int ny = img.ny;
dst.nx = target_width;
dst.ny = target_height;
dst.buf.resize(3 * target_width * target_height);
float Cc;
float C[5];
float d0, d2, d3, a0, a1, a2, a3;
int i, j, k, jj;
int x, y;
float dx, dy;
float tx, ty;
tx = (float)nx / (float)target_width;
ty = (float)ny / (float)target_height;
// Bicubic interpolation; adapted from ViT.cpp, inspired from :
// -> https://github.com/yglukhov/bicubic-interpolation-image-processing/blob/master/libimage.c#L36
// -> https://en.wikipedia.org/wiki/Bicubic_interpolation
for (i = 0; i < target_height; i++) {
for (j = 0; j < target_width; j++) {
x = (int)(tx * j);
y = (int)(ty * i);
dx = tx * j - x;
dy = ty * i - y;
for (k = 0; k < 3; k++) {
for (jj = 0; jj <= 3; jj++) {
d0 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x - 1, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
d2 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x + 1, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
d3 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x + 2, 0, nx - 1)) * 3 + k] - img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
a0 = img.buf[(clip(y - 1 + jj, 0, ny - 1) * nx + clip(x, 0, nx - 1)) * 3 + k];
a1 = -1.0 / 3 * d0 + d2 - 1.0 / 6 * d3;
a2 = 1.0 / 2 * d0 + 1.0 / 2 * d2;
a3 = -1.0 / 6 * d0 - 1.0 / 2 * d2 + 1.0 / 6 * d3;
C[jj] = a0 + a1 * dx + a2 * dx * dx + a3 * dx * dx * dx;
d0 = C[0] - C[1];
d2 = C[2] - C[1];
d3 = C[3] - C[1];
a0 = C[1];
a1 = -1.0 / 3 * d0 + d2 - 1.0 / 6 * d3;
a2 = 1.0 / 2 * d0 + 1.0 / 2 * d2;
a3 = -1.0 / 6 * d0 - 1.0 / 2 * d2 + 1.0 / 6 * d3;
Cc = a0 + a1 * dy + a2 * dy * dy + a3 * dy * dy * dy;
const uint8_t Cc2 = std::min(std::max(std::round(Cc), 0.0f), 255.0f);
dst.buf[(i * target_width + j) * 3 + k] = float(Cc2);
}
}
}
}
return true;
}
// llava-1.6 type of resize_and_pad (black)
static void resize_and_pad_image(const clip_image_u8& image, clip_image_u8 &image_output, const std::pair<int, int>& target_resolution) {
int target_width = target_resolution.first;
int target_height = target_resolution.second;
float scale_w = static_cast<float>(target_width) / image.nx;
float scale_h = static_cast<float>(target_height) / image.ny;
int new_width, new_height;
if (scale_w < scale_h) {
new_width = target_width;
new_height = std::min(static_cast<int>(std::ceil(image.ny * scale_w)), target_height);
} else {
new_height = target_height;
new_width = std::min(static_cast<int>(std::ceil(image.nx * scale_h)), target_width);
}
clip_image_u8 resized_image;
// bilinear_resize(image, resized_image, new_width, new_height);
bicubic_resize(image, resized_image, new_width, new_height);
clip_image_u8 padded_image;
padded_image.nx = target_width;
padded_image.ny = target_height;
padded_image.buf.resize(3 * target_width * target_height, 0); // Initialize with black
// Calculate padding offsets
int pad_x = (target_width - new_width) / 2;
int pad_y = (target_height - new_height) / 2;
// Copy the resized image into the center of the padded buffer
for (int y = 0; y < new_height; ++y) {
for (int x = 0; x < new_width; ++x) {
for (int c = 0; c < 3; ++c) {
padded_image.buf[3 * ((y + pad_y) * target_width + (x + pad_x)) + c] = resized_image.buf[3 * (y * new_width + x) + c];
}
}
}
image_output = std::move(padded_image);
}
/**
* Selects the best resolution from a list of possible resolutions based on the original size.
*
* @param original_size The original size of the image in the format (width, height).
* @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
* @return The best fit resolution in the format (width, height).
*/
static std::pair<int, int> select_best_resolution(const std::pair<int, int> & original_size, const std::vector<std::pair<int, int>> & possible_resolutions) {
int original_width = original_size.first;
int original_height = original_size.second;
std::pair<int, int> best_fit;
int max_effective_resolution = 0;
int min_wasted_resolution = std::numeric_limits<int>::max();
for (const auto& resolution : possible_resolutions) {
int width = resolution.first;
int height = resolution.second;
float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
int downscaled_width = static_cast<int>(original_width * scale);
int downscaled_height = static_cast<int>(original_height * scale);
int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
int wasted_resolution = (width * height) - effective_resolution;
// fprintf(stderr, "resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
max_effective_resolution = effective_resolution;
min_wasted_resolution = wasted_resolution;
best_fit = resolution;
}
}
return best_fit;
}
static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & image, int patch_size) {
std::vector<clip_image_u8*> patches;
int width = image.nx;
int height = image.ny;
for (int i = 0; i < height; i += patch_size) {
for (int j = 0; j < width; j += patch_size) {
clip_image_u8 *patch = clip_image_u8_init();
patch->nx = std::min(patch_size, width - j);
patch->ny = std::min(patch_size, height - i);
patch->buf.resize(3 * patch->nx * patch->ny);
for (int y = 0; y < patch->ny; ++y) {
for (int x = 0; x < patch->nx; ++x) {
for (int c = 0; c < 3; ++c) {
patch->buf[3 * (y * patch->nx + x) + c] = image.buf[3 * ((i + y) * width + (j + x)) + c];
}
}
}
patches.push_back(patch);
}
}
return patches;
}
// returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
// res_imgs memory is being allocated here, previous allocations will be freed if found
bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs) {
bool pad_to_square = true;
if (!ctx->has_vision_encoder) {
printf("This gguf file seems to have no vision encoder\n");
return false;
}
auto & params = ctx->vision_model.hparams;
// The model config actually contains all we need to decide on how to preprocess, here we automatically switch to the new llava-1.6 preprocessing
if (strcmp(params.mm_patch_merge_type, "spatial_unpad") == 0) {
pad_to_square = false;
}
// free the previous res_imgs if any set
if (res_imgs.size > 0) {
clip_image_f32_batch_free(res_imgs);
}
res_imgs.data = nullptr;
res_imgs.size = 0;
// the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
// see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
clip_image_u8 * temp = clip_image_u8_init(); // we will keep the input image data here temporarily
if (pad2square && img->nx != img->ny) {
if (pad_to_square && img->nx != img->ny) {
int longer_side = std::max(img->nx, img->ny);
temp->nx = longer_side;
temp->ny = longer_side;
temp->buf.resize(3 * longer_side * longer_side);
const uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA
const uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA (this is the mean rgb color * 255)
// fill with background color
for (size_t i = 0; i < temp->buf.size(); i++) {
@ -1119,18 +1539,63 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
}
}
} else {
temp->nx = img->nx;
temp->ny = img->ny;
temp->buf.resize(img->buf.size());
memcpy(temp->buf.data(), img->buf.data(), temp->buf.size());
if (params.image_grid_pinpoints[0] != 0) {
// "spatial_unpad" with "anyres" processing for llava-1.6
std::vector<std::pair<int, int>> possible_resolutions;
for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
}
std::pair<int, int> best_resolution = select_best_resolution({img->nx, img->ny}, possible_resolutions);
// clip_image_save_to_bmp(*img, "input.bmp");
resize_and_pad_image(*img, *temp, best_resolution); // we do not pad with mean-bg color anymore in llava-1.6
// clip_image_save_to_bmp(*temp, "resized.bmp");
// visually verify normalized image:
// normalize_image_u8_to_f32(*temp, *res, ctx->image_mean, ctx->image_std);
// {
// clip_image_u8 * temp2 = clip_image_u8_init();
// clip_image_convert_f32_to_u8(*res, *temp2);
// clip_image_save_to_bmp(*temp2, "resized_normalized_f32.bmp");
// clip_image_u8_free(temp2);
// }
std::vector<clip_image_u8 *> patches = divide_to_patches_u8(*temp, params.image_size); // prepare spatial sorted main patches of image_size each (336 in llava-1.6)
clip_image_u8 *image_original_resize = clip_image_u8_init();
// bilinear_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square
bicubic_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square
patches.insert(patches.begin(), image_original_resize);
// clip_image_f32_batch_init(patches.size());
res_imgs.size = patches.size();
res_imgs.data = new clip_image_f32[res_imgs.size];
int num=0;
for (auto& patch : patches) {
normalize_image_u8_to_f32(patch, &res_imgs.data[num], ctx->image_mean, ctx->image_std);
num++;
}
for (size_t i = 0; i < patches.size(); i++) {
// printf("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
clip_image_u8_free(patches[i]);
}
clip_image_u8_free(temp);
return true;
} else {
temp->nx = img->nx;
temp->ny = img->ny;
temp->buf.resize(img->buf.size());
memcpy(temp->buf.data(), img->buf.data(), temp->buf.size());
}
}
const int nx = temp->nx;
const int ny = temp->ny;
// clip_image_save_to_bmp(*temp, "resized_vanilla.bmp");
const int nx2 = ctx->vision_model.hparams.image_size;
const int ny2 = ctx->vision_model.hparams.image_size;
clip_image_f32 * res = clip_image_f32_init();
res->nx = nx2;
res->ny = ny2;
res->buf.resize(3 * nx2 * ny2);
@ -1184,9 +1649,26 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
}
clip_image_u8_free(temp);
// {
// clip_image_u8 * temp2 = clip_image_u8_init();
// clip_image_convert_f32_to_u8(*res, *temp2);
// clip_image_save_to_bmp(*temp2, "resized_normalized_f32_vanilla.bmp");
// clip_image_u8_free(temp2);
// }
// res_imgs.push_back(res);
res_imgs.size = 1;
res_imgs.data = new clip_image_f32[res_imgs.size];
res_imgs.data[0] = *res;
clip_image_f32_free(res);
return true;
}
ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) {
return ctx->vision_model.image_newline;
}
void clip_free(clip_ctx * ctx) {
ggml_free(ctx->ctx_data);
gguf_free(ctx->ctx_gguf);
@ -1194,6 +1676,42 @@ void clip_free(clip_ctx * ctx) {
delete ctx;
}
size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
}
int32_t clip_image_size(const struct clip_ctx * ctx) {
return ctx->vision_model.hparams.image_size;
}
int32_t clip_patch_size(const struct clip_ctx * ctx) {
return ctx->vision_model.hparams.patch_size;
}
int32_t clip_hidden_size(const struct clip_ctx * ctx) {
return ctx->vision_model.hparams.hidden_size;
}
const char * clip_patch_merge_type(const struct clip_ctx * ctx) {
return ctx->vision_model.hparams.mm_patch_merge_type;
}
const int32_t * clip_image_grid(const struct clip_ctx * ctx) {
return ctx->vision_model.hparams.image_grid_pinpoints;
}
int clip_n_patches(const struct clip_ctx * ctx) {
const auto & params = ctx->vision_model.hparams;
int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
n_patches /= 4;
}
return n_patches;
}
bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
if (!ctx->has_vision_encoder) {
printf("This gguf file seems to have no vision encoder\n");
@ -1213,7 +1731,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
}
int batch_size = imgs->size;
if(ctx->has_llava_projector) {
if (ctx->has_llava_projector) {
GGML_ASSERT(batch_size == 1); // TODO: support multiple images
}
@ -1224,9 +1742,10 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
// set inputs
const auto & model = ctx->vision_model;
const auto & hparams = model.hparams;
const int image_size = hparams.image_size;
const int patch_size = hparams.patch_size;
const int num_patches = ((image_size / patch_size) * (image_size / patch_size));
const int image_size = hparams.image_size;
const int patch_size = hparams.patch_size;
const int num_patches = ((image_size / patch_size) * (image_size / patch_size));
const int num_positions = num_patches + 1;
{
@ -1301,11 +1820,11 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
// copy the embeddings to the location passed by the user
ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
return true;
}
bool clip_model_quantize(const char * fname_inp, const char * fname_out, const int itype) {
ggml_type type = GGML_TYPE_Q4_1;
assert(itype < GGML_TYPE_COUNT);
@ -1494,26 +2013,13 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
return ctx->vision_model.mm_model_block_1_block_2_1_b->ne[0];
}
else if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
return ctx->vision_model.mm_2_b->ne[0];
} else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
}
if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
return ctx->vision_model.mm_3_b->ne[0];
}
else {
std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
}
}
int clip_n_patches(const struct clip_ctx * ctx) {
auto & params = ctx->vision_model.hparams;
int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
n_patches /= 4;
}
return n_patches;
}
size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
}

49
examples/llava/clip.h
View File

@ -24,25 +24,7 @@ struct clip_ctx;
extern "C" {
#endif
struct clip_vision_hparams {
int32_t image_size;
int32_t patch_size;
int32_t hidden_size;
int32_t n_intermediate;
int32_t projection_dim;
int32_t n_head;
int32_t n_layer;
float eps;
};
CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity);
CLIP_API void clip_free(struct clip_ctx * ctx);
CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
CLIP_API int clip_n_patches (const struct clip_ctx * ctx);
CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
struct clip_ctx;
struct clip_image_u8_batch {
struct clip_image_u8 * data;
@ -54,18 +36,43 @@ struct clip_image_f32_batch {
size_t size;
};
CLIP_API struct clip_ctx * clip_model_load (const char * fname, int verbosity);
CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
CLIP_API void clip_free(struct clip_ctx * ctx);
CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
// TODO: should be enum, not string
CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
CLIP_API int clip_n_patches (const struct clip_ctx * ctx);
CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
CLIP_API struct clip_image_u8 * clip_image_u8_init ();
CLIP_API struct clip_image_f32 * clip_image_f32_init();
CLIP_API void clip_image_u8_free (struct clip_image_u8 * img);
CLIP_API void clip_image_u8_free (struct clip_image_u8 * img);
CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch & batch);
CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch & batch);
CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
/** interpret bytes as an image file with length bytes_length, and use the result to populate img */
CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
CLIP_API bool clip_image_preprocess (struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, bool pad2square);
/** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs );
CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
CLIP_API bool clip_image_encode (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);

66
examples/llava/convert-image-encoder-to-gguf.py
View File

@ -78,18 +78,19 @@ ap.add_argument("--text-only", action="store_true", required=False,
help="Save a text-only model. It can't be used to encode images")
ap.add_argument("--vision-only", action="store_true", required=False,
help="Save a vision-only model. It can't be used to encode texts")
ap.add_argument("--clip_model_is_vision", action="store_true", required=False,
ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
help="The clip model is from openclip (for ViT-SO400M type))")
ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp", choices=["mlp", "ldp"], default="mlp")
ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values")
ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values")
ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
default_image_mean = [0.48145466, 0.4578275, 0.40821073]
default_image_std = [0.26862954, 0.26130258, 0.27577711]
ap.add_argument('--image_mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
ap.add_argument('--image_std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
# with proper
args = ap.parse_args()
@ -105,7 +106,7 @@ if args.use_f32:
# output in the same directory as the model if output_dir is None
dir_model = args.model_dir
if args.clip_model_is_vision:
if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
vocab = None
tokens = None
else:
@ -133,7 +134,7 @@ ftype = 1
if args.use_f32:
ftype = 0
if args.clip_model_is_vision:
if args.clip_model_is_vision or args.clip_model_is_openclip:
model = CLIPVisionModel.from_pretrained(dir_model)
processor = None
else:
@ -202,6 +203,57 @@ if has_vision_encoder:
fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"])
block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
# /**
# "image_grid_pinpoints": [
# [
# 336,
# 672
# ],
# [
# 672,
# 336
# ],
# [
# 672,
# 672
# ],
# [
# 1008,
# 336
# ],
# [
# 336,
# 1008
# ]
# ],
# Flattened:
# [
# 336, 672,
# 672, 336,
# 672, 672,
# 1008, 336,
# 336, 1008
# ]
# *
# */
if "image_grid_pinpoints" in v_hparams:
# flatten it
image_grid_pinpoints = []
for pinpoint in v_hparams["image_grid_pinpoints"]:
for p in pinpoint:
image_grid_pinpoints.append(p)
fout.add_array("clip.vision.image_grid_pinpoints", image_grid_pinpoints)
if "image_crop_resolution" in v_hparams:
fout.add_uint32("clip.vision.image_crop_resolution", v_hparams["image_crop_resolution"])
if "image_aspect_ratio" in v_hparams:
fout.add_string("clip.vision.image_aspect_ratio", v_hparams["image_aspect_ratio"])
if "image_split_resolution" in v_hparams:
fout.add_uint32("clip.vision.image_split_resolution", v_hparams["image_split_resolution"])
if "mm_patch_merge_type" in v_hparams:
fout.add_string("clip.vision.mm_patch_merge_type", v_hparams["mm_patch_merge_type"])
if "mm_projector_type" in v_hparams:
fout.add_string("clip.vision.mm_projector_type", v_hparams["mm_projector_type"])
if processor is not None:
image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean

29
examples/llava/llava-cli.cpp
View File

@ -155,11 +155,29 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
system_prompt = prompt.substr(0, image_pos);
user_prompt = prompt.substr(image_pos + std::string("<image>").length());
printf("system_prompt: %s\n", system_prompt.c_str());
if (params->verbose_prompt) {
auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
printf("user_prompt: %s\n", user_prompt.c_str());
if (params->verbose_prompt) {
auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
} else {
// llava-1.5 native mode
system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
user_prompt = prompt + "\nASSISTANT:";
if (params->verbose_prompt) {
auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
}
eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos);
@ -171,13 +189,17 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
fprintf(stderr, "\n");
struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
std::string response = "";
for (int i = 0; i < max_tgt_len; i++) {
const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0) break;
if (strstr(tmp, "###")) break; // Yi-VL behavior
printf("%s", tmp);
if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
fflush(stdout);
}
@ -196,7 +218,8 @@ static struct llava_context * llava_init(gpt_params * params) {
auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
llama_backend_init(params->numa);
llama_backend_init();
llama_numa_init(params->numa);
llama_model_params model_params = llama_model_params_from_gpt_params(*params);

167
examples/llava/llava-surgery-v2.py Normal file
View File

@ -0,0 +1,167 @@
import argparse
import glob
import os
import torch
from safetensors.torch import load as safe_load, save as safe_save, safe_open, save_file
# Function to determine if file is a SafeTensor file
def is_safetensor_file(file_path):
return file_path.endswith('.safetensors')
# Unified loading function
def load_model(file_path):
if is_safetensor_file(file_path):
tensors = {}
with safe_open(file_path, framework="pt", device="cpu") as f:
for key in f.keys():
tensors[key] = f.get_tensor(key).clone()
# output shape
print(f"{key} : {tensors[key].shape}")
return tensors, 'safetensor'
else:
return torch.load(file_path, map_location=torch.device('cpu')), 'pytorch'
# Unified saving function
def save_model(model, file_path, file_type):
if file_type == 'safetensor':
# safe_save(model, file_path)
save_file(model, file_path)
else:
torch.save(model, file_path)
# Adapted function to clean vision tower from checkpoint
def clean_vision_tower_from_checkpoint(checkpoint_path):
checkpoint, file_type = load_model(checkpoint_path)
# file_type = 'pytorch'
model_path = os.path.dirname(checkpoint_path)
print(f"Searching for vision tower tensors in {checkpoint_path}")
clip_tensors = [k for k, v in checkpoint.items() if (k.startswith("model.vision_tower") or k.startswith("vit."))]
if len(clip_tensors) > 0:
print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}")
# Adapted for file type
clip_path = os.path.join(model_path, "llava.clip")
if os.path.exists(clip_path):
print(f"Loading existing llava.clip from {clip_path}")
existing_clip, _ = load_model(clip_path)
else:
print(f"Creating new llava.clip at {clip_path}")
existing_clip = {}
# Update existing_clip with new tensors, avoid duplicates
for name in clip_tensors:
simple_name = name[name.index('vision_model.'):] if 'vision_model.' in name else name
print(f"Adding {simple_name} to llava.clip")
if simple_name not in existing_clip:
existing_clip[simple_name] = checkpoint[name]
# Save the updated clip tensors back to llava.clip
save_model(existing_clip, clip_path, 'pytorch')
# Remove the tensors from the original checkpoint
for name in clip_tensors:
del checkpoint[name]
# Save the updated checkpoint
checkpoint_path = checkpoint_path
save_model(checkpoint, checkpoint_path, file_type)
return True
return False
def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector):
newline_checkpoint_path = None
projector_checkpoint_path = None
for path in checkpoint_paths:
checkpoint, _ = load_model(path)
if newline_criteria(checkpoint) and newline_checkpoint_path is None:
newline_checkpoint_path = path
if projector(checkpoint):
projector_checkpoint_path = path
return newline_checkpoint_path, projector_checkpoint_path
def newline_criteria(checkpoint):
return any(k.startswith("model.image_newline") for k in checkpoint.keys())
def proj_criteria(checkpoint):
return any(k.startswith("model.mm_projector") or k.startswith("vision_proj.") for k in checkpoint.keys())
# Command-line interface setup
ap = argparse.ArgumentParser()
ap.add_argument("-m", "--model", required=True, help="Path to LLaVA v1.5+ model")
ap.add_argument("-C", "--clean-vision-tower", action="store_true", help="Remove any vision tower from the model files")
args = ap.parse_args()
if args.clean_vision_tower:
# Generalized to handle both PyTorch and SafeTensors models
model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
# checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and path.startswith('pytorch')) or (path.endswith('.safetensors') and path.startswith('model'))]
checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
for projector_checkpoint_path in checkpoint_paths:
print(f"Cleaning {projector_checkpoint_path}")
if not clean_vision_tower_from_checkpoint(projector_checkpoint_path):
print(f"No vision tower found in {projector_checkpoint_path}")
# we break once none is found, so far all models append them at the end
# break
print("Done! All vision tower tensors are removed from the model files and stored in llava.clip file.")
# Now we look for the projector in the last checkpoint
model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
# last_checkpoint_path = checkpoint_paths[0]
# first_checkpoint_path = checkpoint_paths[-1]
newline_checkpoint_path, projector_checkpoint_path = find_relevant_checkpoints(checkpoint_paths, newline_criteria, proj_criteria)
print(f"Taking projector from {projector_checkpoint_path}")
first_mm_tensors = []
first_checkpoint = None
if newline_checkpoint_path is not None:
print(f"Taking newline from {newline_checkpoint_path}")
first_checkpoint, file_type = load_model(newline_checkpoint_path)
first_mm_tensors = [k for k, v in first_checkpoint.items() if k.startswith("model.image_newline")]
# Load the checkpoint
mm_tensors = []
last_checkpoint = None
if projector_checkpoint_path is not None:
last_checkpoint, file_type = load_model(projector_checkpoint_path)
mm_tensors = [k for k, v in last_checkpoint.items() if k.startswith("model.mm_projector") or k.startswith("vision_proj.")]
if len(mm_tensors) == 0:
if last_checkpoint is not None:
for k, v in last_checkpoint.items():
print(k)
print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint)} tensors.")
print("No tensors found. Is this a LLaVA model?")
exit()
print(f"Found {len(mm_tensors)} tensors to extract.")
print(f"Found additional {len(first_mm_tensors)} tensors to extract.")
# projector = {name: checkpoint.[name].float() for name in mm_tensors}
projector = {}
for name in mm_tensors:
projector[name] = last_checkpoint[name].float()
for name in first_mm_tensors:
projector[name] = first_checkpoint[name].float()
if len(projector) > 0:
save_model(projector, f"{args.model}/llava.projector", 'pytorch')
for name in mm_tensors:
del last_checkpoint[name]
for name in first_mm_tensors:
del first_checkpoint[name]
if len(mm_tensors) > 0:
save_model(last_checkpoint, projector_checkpoint_path, file_type)
if len(first_mm_tensors) > 0:
save_model(first_checkpoint, newline_checkpoint_path, file_type)
print("Done!")
print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

8
examples/llava/llava-surgery.py
View File

@ -19,19 +19,12 @@ mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_project
projector = {name: checkpoint[name].float() for name in mm_tensors}
torch.save(projector, f"{args.model}/llava.projector")
# remove these tensors from the checkpoint and save it again
for name in mm_tensors:
del checkpoint[name]
# BakLLaVA models contain CLIP tensors in it
clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
if len(clip_tensors) > 0:
clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
torch.save(clip, f"{args.model}/llava.clip")
# remove these tensors
for name in clip_tensors:
del checkpoint[name]
# added tokens should be removed to be able to convert Mistral models
if os.path.exists(f"{args.model}/added_tokens.json"):
@ -39,7 +32,6 @@ if len(clip_tensors) > 0:
f.write("{}\n")
torch.save(checkpoint, path)
print("Done!")
print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")

297
examples/llava/llava.cpp
View File

@ -2,32 +2,296 @@
#include "common.h"
#include "llama.h"
#include "llava.h"
#include "base64.hpp"
#include <cstdio>
#include <cstdlib>
#include <vector>
#include <numeric>
// RGB uint8 image
struct clip_image_u8 {
int nx;
int ny;
std::vector<uint8_t> buf;
};
// RGB float32 image (NHWC)
// Memory layout: RGBRGBRGB...
struct clip_image_f32 {
int nx;
int ny;
std::vector<float> buf;
};
struct clip_image_grid_shape {
int first;
int second;
};
/**
* Selects the best resolution from a list of possible resolutions based on the original size.
*
* @param original_size The original size of the image in the format (width, height).
* @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
* @return The best fit resolution in the format (width, height).
*/
static std::pair<int, int> select_best_resolution(const std::pair<int, int>& original_size, const std::vector<std::pair<int, int>>& possible_resolutions) {
int original_width = original_size.first;
int original_height = original_size.second;
std::pair<int, int> best_fit;
int max_effective_resolution = 0;
int min_wasted_resolution = std::numeric_limits<int>::max();
for (const auto& resolution : possible_resolutions) {
int width = resolution.first;
int height = resolution.second;
float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
int downscaled_width = static_cast<int>(original_width * scale);
int downscaled_height = static_cast<int>(original_height * scale);
int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
int wasted_resolution = (width * height) - effective_resolution;
// fprintf(stderr, "resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
max_effective_resolution = effective_resolution;
min_wasted_resolution = wasted_resolution;
best_fit = resolution;
}
}
return best_fit;
}
/**
* @brief Get the anyres image grid shape object
*
* @param image_size
* @param grid_pinpoints
* @param image_patch_size
* @return <int, int>
*/
static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<int, int> & image_size, const std::vector<std::pair<int, int>> & grid_pinpoints, int image_patch_size) {
/**
Conversion from gguf flat array to vector:
std::vector<std::pair<int, int>> possible_resolutions;
for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
}
*/
auto best_resolution = select_best_resolution(image_size, grid_pinpoints);
return {best_resolution.first / image_patch_size, best_resolution.second / image_patch_size};
}
// Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
struct {
struct ggml_tensor * newline;
struct ggml_context * ctx;
} model;
const int32_t image_size = clip_image_size(ctx_clip);
const int32_t patch_size = clip_patch_size(ctx_clip);
int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
int num_patches_width = grid_shape.first; // grid 1-4
int num_patches_height = grid_shape.second; // grid 1-4
const size_t num_images = num_patches_width * num_patches_height + 1;
// TODO: size calculation is not calculated - it's only tens of MB
size_t ctx_size = 0;
{
ctx_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features
ctx_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32);
}
struct ggml_init_params params {
/*.mem_size =*/ ctx_size,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ false, // NOTE: this should be false when using the legacy API
};
// Python reference code for full unpad:
/*
base_image_feature = image_feature[0]
image_feature = image_feature[1:]
image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
image_feature = image_feature.flatten(1, 2).flatten(2, 3)
image_feature = unpad_image(image_feature, image_sizes[image_idx])
image_feature = torch.cat((
image_feature,
self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1)
), dim=-1)
image_feature = image_feature.flatten(1, 2).transpose(0, 1)
image_feature = torch.cat((base_image_feature, image_feature), dim=0)
*/
// We now have two options: unpad or no unpad. Unpad removes tokens for faster llm eval.
// In terms of result quality it appears to make no difference, so we'll start with the easier approach given 5D tensors are not supported in ggml yet.
// Without unpad we have to split the sub-image embeddings into patches of 24 features each and permute them.
// Once all images are processed to prepended the base_image_features without any changes.
// Pytorch reference simplified, modified for ggml compatibility - confirmed identical output in python (for a 2x2 grid image (676x676 scaling))
/*
image_feature = image_feature.view(2, 2, 24, 24, 4096)
image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
image_feature = image_feature.view(2, 24, 2, 24, 4096)
image_feature = image_feature.flatten(0, 3)
// Reshape to 4D tensor by merging the last two dimensions
image_feature = image_feature.view(2, 2, 24, 24*4096)
image_feature = image_feature.permute(0, 2, 1, 3).contiguous()
image_feature = image_feature.view(-1, 4096)
*/
model.ctx = ggml_init(params);
ggml_tensor * newline_tmp = clip_get_newline_tensor(ctx_clip);
model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
if (newline_tmp->backend != GGML_BACKEND_CPU) {
if (newline_tmp->buffer == NULL) {
printf("newline_tmp tensor buffer is NULL\n");
}
ggml_backend_tensor_get(newline_tmp, model.newline->data, 0, ggml_nbytes(newline_tmp));
} else {
model.newline->data = newline_tmp->data;
if (model.newline->data == NULL) {
printf("newline_tmp tensor data is NULL\n");
}
}
struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
// ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
// fill it with the image embeddings, ignoring the base
for (size_t i = 1; i < num_images; i++) {
size_t offset = (i-1) * clip_embd_nbytes(ctx_clip);
memcpy((uint8_t *)(image_features->data) + offset, image_embd_v[i], clip_embd_nbytes(ctx_clip));
}
struct ggml_cgraph * gf = ggml_new_graph(model.ctx);
size_t size_ele = ggml_type_size(GGML_TYPE_F32);
struct ggml_tensor *image_features_patchview = ggml_view_4d(model.ctx, image_features,
num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
num_patches_per_side,
num_patches_width,
num_patches_height,
size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side,
size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side * num_patches_width, 0);
// ggml_tensor_printf(image_features_patchview,"image_features_patchview",__LINE__,false,false);
struct ggml_tensor *permuted_cont = ggml_cont(model.ctx, ggml_permute(model.ctx, image_features_patchview, 0, 2, 1, 3));
/**
At the end of each row we have to add the row_end embeddings, which are the same as the newline embeddings
image_feature = torch.cat((
image_feature,
self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
), dim=-1)
*
*/
// ggml_tensor_printf(permuted_cont,"permuted_cont",__LINE__,false,false);
struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side, size_ele * clip_n_mmproj_embd(ctx_clip), 0);
// ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
ggml_build_forward_expand(gf, flatten);
ggml_graph_compute_with_ctx(model.ctx, gf, 1);
struct ggml_tensor* result = gf->nodes[gf->n_nodes - 1];
memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
// append without newline tokens (default behavior in llava_arch when not using unpad ):
memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
*n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
// Debug: Test single segments
// Current findings: sending base image, sending a segment embedding all works similar to python
// However, permuted embeddings do not work yet (stride issue?)
// memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as context
// memcpy(image_embd_out, (float*)prepared_cont->data, clip_embd_nbytes(ctx_clip)); // main image as context
// *n_img_pos_out=576;
ggml_free(model.ctx);
return true;
}
#include "base64.hpp"
static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
clip_image_f32 * img_res = clip_image_f32_init();
if (!clip_image_preprocess(ctx_clip, img, img_res, /*pad2square =*/ true)) {
// std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
clip_image_f32_batch img_res_v;
img_res_v.size = 0;
img_res_v.data = nullptr;
if (!clip_image_preprocess(ctx_clip, img, img_res_v)) {
fprintf(stderr, "%s: unable to preprocess image\n", __func__);
clip_image_f32_free(img_res);
delete[] img_res_v.data;
return false;
}
*n_img_pos = clip_n_patches(ctx_clip);
const int64_t t_img_enc_start_us = ggml_time_us();
bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
clip_image_f32_free(img_res);
if (!encoded) {
fprintf(stderr, "Unable to encode image\n");
return false;
const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
// flat / default llava-1.5 type embedding
*n_img_pos = clip_n_patches(ctx_clip);
bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
delete[] img_res_v.data;
if (!encoded) {
fprintf(stderr, "Unable to encode image\n");
return false;
}
} else {
// spatial_unpad llava-1.6 type embedding
// TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
std::vector<float *> image_embd_v;
image_embd_v.resize(img_res_v.size);
for (size_t i = 0; i < img_res_v.size; i++) {
image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
if (!encoded) {
fprintf(stderr, "Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
return false;
}
}
const int64_t t_img_enc_batch_us = ggml_time_us();
printf("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
const int32_t * image_grid = clip_image_grid(ctx_clip);
std::vector<std::pair<int, int>> grid_pinpoints;
for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) {
grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
}
// free all img_res_v - not needed anymore
delete[] img_res_v.data;
img_res_v.size = 0;
img_res_v.data = nullptr;
const int32_t image_size = clip_image_size(ctx_clip);
struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
int n_img_pos_out;
clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
*n_img_pos = n_img_pos_out;
for (size_t i = 0; i < image_embd_v.size(); i++) {
free(image_embd_v[i]);
}
image_embd_v.clear();
// debug image/segment/normalization content:
// clip_image_u8 * tmp = clip_image_u8_init();
// clip_image_convert_f32_to_u8(*image_feature, *tmp);
// clip_image_save_to_bmp(*tmp, "image_feature.bmp");
}
printf("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
const int64_t t_img_enc_end_us = ggml_time_us();
float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
@ -48,10 +312,9 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
}
static bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip));
float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
if (!image_embd) {
fprintf(stderr, "Unable to allocate memory for image embeddings\n");
free(image_embd);
return false;
}
@ -85,7 +348,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
return true;
}
LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
clip_image_u8 * img = clip_image_u8_init();
if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
clip_image_u8_free(img);
@ -142,7 +405,7 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
return true;
}
LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
unsigned char* image_bytes;
long image_bytes_length;
auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
@ -151,13 +414,13 @@ LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct
return NULL;
}
auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
llava_image_embed *embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
free(image_bytes);
return embed;
}
LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed) {
void llava_image_embed_free(struct llava_image_embed * embed) {
free(embed->embed);
free(embed);
}

2
examples/llava/llava.h
View File

@ -3,7 +3,6 @@
#include "ggml.h"
#ifdef LLAMA_SHARED
# if defined(_WIN32) && !defined(__MINGW32__)
# ifdef LLAMA_BUILD
@ -42,7 +41,6 @@ LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
/** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);
#ifdef __cplusplus
}
#endif

3
examples/lookahead/lookahead.cpp
View File

@ -54,7 +54,8 @@ int main(int argc, char ** argv) {
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model = NULL;
llama_context * ctx = NULL;

3
examples/lookup/lookup.cpp
View File

@ -31,7 +31,8 @@ int main(int argc, char ** argv){
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model = NULL;
llama_context * ctx = NULL;

6
examples/main/README.md
View File

@ -283,7 +283,11 @@ These options help improve the performance and memory usage of the LLaMA models.
### NUMA support
- `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
- `--numa distribute`: Pin an equal proportion of the threads to the cores on each NUMA node. This will spread the load amongst all cores on the system, utilitizing all memory channels at the expense of potentially requiring memory to travel over the slow links between nodes.
- `--numa isolate`: Pin all threads to the NUMA node that the program starts on. This limits the number of cores and amount of memory that can be used, but guarantees all memory access remains local to the NUMA node.
- `--numa numactl`: Pin threads to the CPUMAP that is passed to the program by starting it with the numactl utility. This is the most flexible mode, and allow arbitraty core usage patterns, for example a map that uses all the cores on one NUMA nodes, and just enough cores on a second node to saturate the inter-node memory bus.
These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
### Memory Float 32

3
examples/main/main.cpp
View File

@ -185,7 +185,8 @@ int main(int argc, char ** argv) {
}
LOG("%s: llama backend init\n", __func__);
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model;
llama_context * ctx;

3
examples/parallel/parallel.cpp
View File

@ -122,7 +122,8 @@ int main(int argc, char ** argv) {
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model = NULL;
llama_context * ctx = NULL;

3
examples/passkey/passkey.cpp
View File

@ -71,7 +71,8 @@ int main(int argc, char ** argv) {
// init LLM
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
// initialize the model

9
examples/perplexity/perplexity.cpp
View File

@ -309,7 +309,7 @@ static void process_logits(int n_vocab, const float * logits, const int * tokens
}
static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
// Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
// Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
// Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
// Output: `perplexity: 13.5106 [114/114]`
// BOS tokens will be added for each chunk before eval
@ -447,7 +447,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
return perplexity_v2(ctx, params);
}
// Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
// Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
// Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
// Output: `perplexity: 13.5106 [114/114]`
// BOS tokens will be added for each chunk before eval
@ -1623,7 +1623,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
uint32_t n_ctx;
in.read((char *)&n_ctx, sizeof(n_ctx));
if (n_ctx > llama_n_ctx(ctx)) {
fprintf(stderr, "%s: %s has been computed with %d, while the current context is %d. Increase it with -c and retry\n",
fprintf(stderr, "%s: %s has been computed with %u, while the current context is %d. Increase it with -c and retry\n",
__func__, params.logits_file.c_str(), n_ctx, params.n_ctx);
}
@ -1809,7 +1809,8 @@ int main(int argc, char ** argv) {
params.prompt = gpt_random_prompt(rng);
}
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model;
llama_context * ctx;

8
examples/quantize/quantize.cpp
View File

@ -23,6 +23,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
{ "IQ2_XXS",LLAMA_FTYPE_MOSTLY_IQ2_XXS," 2.06 bpw quantization", },
{ "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization", },
{ "IQ1_S", LLAMA_FTYPE_MOSTLY_IQ1_S, " 1.56 bpw quantization", },
{ "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
{ "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
{ "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization", },
@ -237,7 +238,7 @@ int main(int argc, char ** argv) {
params.imatrix = &imatrix_data;
}
llama_backend_init(false);
llama_backend_init();
// parse command line arguments
const std::string fname_inp = argv[arg_idx];
@ -287,9 +288,10 @@ int main(int argc, char ** argv) {
}
}
if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) && imatrix_data.empty()) {
if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && imatrix_data.empty()) {
fprintf(stderr, "\n===============================================================================================\n");
fprintf(stderr, "Please do not use IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
fprintf(stderr, "Please do not use IQ1_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
fprintf(stderr, "===============================================================================================\n\n\n");
return 1;
}

77
examples/server/README.md
View File

@ -16,6 +16,13 @@ Command line options:
- `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
- `--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 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 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.
@ -32,6 +39,8 @@ Command line options:
- `--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(default: 1=disabled), used together with group attention width `--grp-attn-w`
- `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
- `-n, --n-predict`: 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.
## Build
@ -128,6 +137,7 @@ node index.js
- `{"status": "loading model"}` if the model is still being loaded.
- `{"status": "error"}` if the model failed to load.
- `{"status": "ok"}` if the model is successfully loaded and the server is ready for further requests mentioned below.
- `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slot are currently available
- **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
@ -189,6 +199,8 @@ node index.js
`n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
`min_keep`: If greater than 0, force samplers to return N possible tokens at minimum (default: 0)
`image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
`slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
@ -197,6 +209,8 @@ node index.js
`system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
`samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. (default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values)
### Result JSON
- Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
@ -370,6 +384,69 @@ Notice that each `probs` is an array of length `n_probs`.
}'
```
- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
### Result JSON
```json
[
{
"dynatemp_exponent": 1.0,
"dynatemp_range": 0.0,
"frequency_penalty": 0.0,
"grammar": "",
"id": 0,
"ignore_eos": false,
"logit_bias": [],
"min_p": 0.05000000074505806,
"mirostat": 0,
"mirostat_eta": 0.10000000149011612,
"mirostat_tau": 5.0,
"model": "llama-2-7b-32k-instruct.Q2_K.gguf",
"n_ctx": 2048,
"n_keep": 0,
"n_predict": 100000,
"n_probs": 0,
"next_token": {
"has_next_token": true,
"n_remain": -1,
"num_tokens_predicted": 0,
"stopped_eos": false,
"stopped_limit": false,
"stopped_word": false,
"stopping_word": ""
},
"penalize_nl": true,
"penalty_prompt_tokens": [],
"presence_penalty": 0.0,
"prompt": "Say hello to llama.cpp",
"repeat_last_n": 64,
"repeat_penalty": 1.100000023841858,
"samplers": [
"top_k",
"tfs_z",
"typical_p",
"top_p",
"min_p",
"temperature"
],
"seed": 42,
"state": 1,
"stop": [
"\n"
],
"stream": false,
"task_id": 0,
"temperature": 0.0,
"tfs_z": 1.0,
"top_k": 40,
"top_p": 0.949999988079071,
"typical_p": 1.0,
"use_penalty_prompt_tokens": false
}
]
```
## More examples
### Change system prompt on runtime

4
examples/server/public/index.html
View File

@ -234,6 +234,7 @@
mirostat_eta: 0.1, // learning rate
grammar: '',
n_probs: 0, // no completion_probabilities,
min_keep: 0, // min probs from each sampler,
image_data: [],
cache_prompt: true,
api_key: ''
@ -791,6 +792,9 @@
<fieldset>
${IntField({ label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs })}
</fieldset>
<fieldset>
${IntField({ label: "Min Probabilities from each Sampler", max: 10, min: 0, name: "min_keep", value: params.value.min_keep })}
</fieldset>
<fieldset>
<label for="api_key">API Key</label>
<input type="text" name="api_key" value="${params.value.api_key}" placeholder="Enter API key" oninput=${updateParams} />

219
examples/server/server.cpp
View File

@ -28,6 +28,7 @@
#include <chrono>
#include <condition_variable>
#include <atomic>
#include <signal.h>
using json = nlohmann::json;
@ -40,6 +41,7 @@ struct server_params
int32_t port = 8080;
int32_t read_timeout = 600;
int32_t write_timeout = 600;
bool slots_endpoint = true;
};
bool server_verbose = false;
@ -158,6 +160,7 @@ struct llama_client_slot
int32_t n_decoded = 0;
int32_t n_remaining = -1;
int32_t i_batch = -1;
int32_t n_predict = -1;
int32_t num_prompt_tokens = 0;
int32_t num_prompt_tokens_processed = 0;
@ -409,6 +412,7 @@ struct llama_server_context
slot.id = i;
slot.n_ctx = n_ctx_slot;
slot.n_predict = params.n_predict;
LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot);
@ -436,10 +440,6 @@ struct llama_server_context
default_generation_settings_for_props["seed"] = -1;
batch = llama_batch_init(n_ctx, 0, params.n_parallel);
// empty system prompt
system_prompt = "";
system_tokens.clear();
}
std::vector<llama_token> tokenize(const json & json_prompt, bool add_bos) const
@ -548,6 +548,16 @@ struct llama_server_context
slot->params.seed = json_value(data, "seed", default_params.seed);
slot->sparams.grammar = json_value(data, "grammar", default_sparams.grammar);
slot->sparams.n_probs = json_value(data, "n_probs", default_sparams.n_probs);
slot->sparams.min_keep = json_value(data, "min_keep", default_sparams.min_keep);
if (slot->n_predict > 0 && slot->params.n_predict > slot->n_predict) {
// Might be better to reject the request with a 400 ?
LOG_WARNING("Max tokens to predict exceeds server configuration", {
{"params.n_predict", slot->params.n_predict},
{"slot.n_predict", slot->n_predict},
});
slot->params.n_predict = slot->n_predict;
}
// infill
if (data.count("input_prefix") != 0)
@ -676,6 +686,24 @@ struct llama_server_context
}
}
const auto &samplers_sequence = data.find("samplers");
if (samplers_sequence != data.end() && samplers_sequence->is_array())
{
std::vector<std::string> sampler_names;
for (const auto &sampler_name : *samplers_sequence)
{
if (sampler_name.is_string())
{
sampler_names.emplace_back(sampler_name);
}
}
slot->sparams.samplers_sequence = sampler_types_from_names(sampler_names, false);
}
else
{
slot->sparams.samplers_sequence = default_sparams.samplers_sequence;
}
if (multimodal)
{
const auto &images_data = data.find("image_data");
@ -765,27 +793,30 @@ struct llama_server_context
}
void update_system_prompt() {
system_tokens = ::llama_tokenize(ctx, system_prompt, add_bos_token);
llama_batch_clear(batch);
kv_cache_clear();
system_tokens.clear();
for (int i = 0; i < (int) system_tokens.size(); ++i)
{
llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
}
if (!system_prompt.empty()) {
system_tokens = ::llama_tokenize(ctx, system_prompt, add_bos_token);
if (llama_decode(ctx, batch) != 0)
{
LOG_TEE("%s: llama_decode() failed\n", __func__);
return;
}
llama_batch_clear(batch);
// assign the system KV cache to all parallel sequences
for (int32_t i = 1; i < params.n_parallel; ++i)
{
llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size());
for (int i = 0; i < (int)system_tokens.size(); ++i)
{
llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
}
if (llama_decode(ctx, batch) != 0)
{
LOG_TEE("%s: llama_decode() failed\n", __func__);
return;
}
// assign the system KV cache to all parallel sequences
for (int32_t i = 1; i < params.n_parallel; ++i)
{
llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size());
}
}
LOG_TEE("system prompt updated\n");
@ -807,10 +838,8 @@ struct llama_server_context
name_user = sys_props.value("anti_prompt", "");
name_assistant = sys_props.value("assistant_name", "");
if (slots.size() > 0)
{
notify_system_prompt_changed();
}
notify_system_prompt_changed();
}
static size_t find_stopping_strings(const std::string &text, const size_t last_token_size,
@ -968,18 +997,31 @@ struct llama_server_context
{
continue;
}
clip_image_f32 * img_res = clip_image_f32_init();
if (!clip_image_preprocess(clp_ctx, img.img_data, img_res, /*pad2square =*/ true))
clip_image_f32_batch img_res_v;
img_res_v.size = 0;
img_res_v.data = nullptr;
if (!clip_image_preprocess(clp_ctx, img.img_data, img_res_v))
{
LOG_TEE("Error processing the given image");
clip_free(clp_ctx);
clip_image_f32_batch_free(img_res_v);
return false;
}
if (img_res_v.size == 0)
{
LOG_TEE("Error processing the given image");
return false;
}
// note: assumes only one image was returned by clip_image_preprocess
clip_image_f32 * img_res = img_res_v.data;
img.image_tokens = clip_n_patches(clp_ctx);
img.image_embedding = (float *)malloc(clip_embd_nbytes(clp_ctx));
if (!img.image_embedding)
{
LOG_TEE("Unable to allocate memory for image embeddings\n");
clip_image_f32_batch_free(img_res_v);
clip_free(clp_ctx);
return false;
}
@ -987,9 +1029,12 @@ struct llama_server_context
if (!clip_image_encode(clp_ctx, params.n_threads, img_res, img.image_embedding))
{
LOG_TEE("Unable to encode image\n");
clip_image_f32_batch_free(img_res_v);
return false;
}
clip_image_f32_free(img_res);
clip_image_f32_batch_free(img_res_v);
img.request_encode_image = false;
}
@ -1013,8 +1058,15 @@ struct llama_server_context
const auto eos_bias = slot.sparams.logit_bias.find(llama_token_eos(model));
const bool ignore_eos = eos_bias != slot.sparams.logit_bias.end() &&
eos_bias->second < 0.0f && std::isinf(eos_bias->second);
std::vector<std::string> samplers_sequence;
for (const auto &sampler_type : slot.sparams.samplers_sequence)
{
samplers_sequence.emplace_back(sampler_type_to_name_string(sampler_type));
}
return json {
{"n_ctx", slot.n_ctx},
{"n_predict", slot.n_predict},
{"model", params.model_alias},
{"seed", slot.params.seed},
{"temperature", slot.sparams.temp},
@ -1042,7 +1094,9 @@ struct llama_server_context
{"stream", slot.params.stream},
{"logit_bias", slot.sparams.logit_bias},
{"n_probs", slot.sparams.n_probs},
{"min_keep", slot.sparams.min_keep},
{"grammar", slot.sparams.grammar},
{"samplers", samplers_sequence}
};
}
@ -1839,7 +1893,10 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
{
printf(" --no-mmap do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
}
printf(" --numa attempt optimizations that help on some NUMA systems\n");
printf(" --numa TYPE attempt optimizations that help on some NUMA systems\n");
printf(" - distribute: spread execution evenly over all nodes\n");
printf(" - isolate: only spawn threads on CPUs on the node that execution started on\n");
printf(" - numactl: use the CPU map provided my numactl\n");
if (llama_supports_gpu_offload()) {
printf(" -ngl N, --n-gpu-layers N\n");
printf(" number of layers to store in VRAM\n");
@ -1872,14 +1929,16 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
printf(" set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n");
printf(" --mmproj MMPROJ_FILE path to a multimodal projector file for LLaVA.\n");
printf(" --log-disable disables logging to a file.\n");
printf(" --slots-endpoint-disable disables slots monitoring endpoint.\n");
printf("\n");
printf(" -n, --n-predict maximum tokens to predict (default: %d)\n", params.n_predict);
printf(" --override-kv KEY=TYPE:VALUE\n");
printf(" advanced option to override model metadata by key. may be specified multiple times.\n");
printf(" types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
printf(" -gan N, --grp-attn-n N set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`");
printf(" -gaw N, --grp-attn-w N set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`");
printf(" --chat-template FORMAT_NAME");
printf(" set chat template, possible valus is: llama2, chatml (default %s)", sparams.chat_template.c_str());
printf(" set chat template, possible value is: llama2, chatml (default %s)", sparams.chat_template.c_str());
printf("\n");
}
@ -2248,9 +2307,17 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
{
params.use_mmap = false;
}
else if (arg == "--numa")
{
params.numa = true;
else if (arg == "--numa") {
if (++i >= argc) {
invalid_param = true;
break;
} else {
std::string value(argv[i]);
/**/ if (value == "distribute" || value == "" ) { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
else { invalid_param = true; break; }
}
}
else if (arg == "--embedding")
{
@ -2311,6 +2378,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
log_set_target(stdout);
LOG_INFO("logging to file is disabled.", {});
}
else if (arg == "--slots-endpoint-disable")
{
sparams.slots_endpoint = false;
}
else if (arg == "--chat-template")
{
if (++i >= argc)
@ -2462,6 +2533,9 @@ static void append_to_generated_text_from_generated_token_probs(llama_server_con
}
}
std::function<void(int)> shutdown_handler;
inline void signal_handler(int signal) { shutdown_handler(signal); }
int main(int argc, char **argv)
{
#if SERVER_VERBOSE != 1
@ -2481,7 +2555,8 @@ int main(int argc, char **argv)
params.model_alias = params.model;
}
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
LOG_INFO("build info", {{"build", LLAMA_BUILD_NUMBER},
{"commit", LLAMA_COMMIT}});
@ -2511,8 +2586,35 @@ int main(int argc, char **argv)
server_state current_state = state.load();
switch(current_state) {
case SERVER_STATE_READY:
res.set_content(R"({"status": "ok"})", "application/json");
res.status = 200; // HTTP OK
if (llama.all_slots_are_idle) {
res.set_content(R"({"status": "ok"})", "application/json");
res.status = 200; // HTTP OK
} else {
int available_slots = 0;
int processing_slots = 0;
for (llama_client_slot & slot : llama.slots) {
if (slot.available()) {
available_slots++;
} else {
processing_slots++;
}
}
if (available_slots > 0) {
json health = {
{"status", "ok"},
{"slots_idle", available_slots},
{"slots_processing", processing_slots}};
res.set_content(health.dump(), "application/json");
res.status = 200; // HTTP OK
} else {
json health = {
{"status", "no slot available"},
{"slots_idle", available_slots},
{"slots_processing", processing_slots}};
res.set_content(health.dump(), "application/json");
res.status = 503; // HTTP Service Unavailable
}
}
break;
case SERVER_STATE_LOADING_MODEL:
res.set_content(R"({"status": "loading model"})", "application/json");
@ -2525,6 +2627,32 @@ int main(int argc, char **argv)
}
});
if (sparams.slots_endpoint) {
svr.Get("/slots", [&](const httplib::Request&, httplib::Response& res) {
json slots;
for (llama_client_slot & slot : llama.slots) {
json slot_data = llama.get_formated_generation(slot);
slot_data["id"] = slot.id;
slot_data["task_id"] = slot.task_id;
slot_data["state"] = slot.state;
slot_data["prompt"] = slot.prompt;
slot_data["next_token"] = {
{"has_next_token", slot.has_next_token},
{"n_remain", slot.n_remaining},
{"num_tokens_predicted", slot.n_decoded},
{"stopped_eos", slot.stopped_eos},
{"stopped_word", slot.stopped_word},
{"stopped_limit", slot.stopped_limit},
{"stopping_word", slot.stopping_word},
};
slots.push_back(slot_data);
}
res.set_content(slots.dump(), "application/json");
res.status = 200; // HTTP OK
});
}
svr.set_logger(log_server_request);
svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep)
@ -3078,8 +3206,25 @@ int main(int argc, char **argv)
std::placeholders::_2,
std::placeholders::_3
));
llama.queue_tasks.start_loop();
shutdown_handler = [&](int) {
llama.queue_tasks.terminate();
};
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
struct sigaction sigint_action;
sigint_action.sa_handler = signal_handler;
sigemptyset (&sigint_action.sa_mask);
sigint_action.sa_flags = 0;
sigaction(SIGINT, &sigint_action, NULL);
#elif defined (_WIN32)
auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
};
SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
#endif
llama.queue_tasks.start_loop();
svr.stop();
t.join();
llama_backend_free();

20
examples/server/utils.hpp
View File

@ -220,6 +220,7 @@ inline std::string format_chatml(std::vector<json> messages)
struct llama_server_queue {
int id = 0;
std::mutex mutex_tasks;
bool running;
// queues
std::vector<task_server> queue_tasks;
std::vector<task_server> queue_tasks_deferred;
@ -278,9 +279,18 @@ struct llama_server_queue {
queue_tasks_deferred.clear();
}
// Start the main loop. This call is blocking
[[noreturn]]
// end the start_loop routine
void terminate() {
{
std::unique_lock<std::mutex> lock(mutex_tasks);
running = false;
}
condition_tasks.notify_all();
}
// Start the main loop.
void start_loop() {
running = true;
while (true) {
// new task arrived
LOG_VERBOSE("have new task", {});
@ -324,8 +334,12 @@ struct llama_server_queue {
{
std::unique_lock<std::mutex> lock(mutex_tasks);
if (queue_tasks.empty()) {
if (!running) {
LOG_VERBOSE("ending start_loop", {});
return;
}
condition_tasks.wait(lock, [&]{
return !queue_tasks.empty();
return (!queue_tasks.empty() || !running);
});
}
}

3
examples/simple/simple.cpp
View File

@ -31,7 +31,8 @@ int main(int argc, char ** argv) {
// init LLM
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
// initialize the model

3
examples/speculative/speculative.cpp
View File

@ -50,7 +50,8 @@ int main(int argc, char ** argv) {
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init(params.numa);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model_tgt = NULL;
llama_model * model_dft = NULL;

2
examples/tokenize/tokenize.cpp
View File

@ -17,7 +17,7 @@ int main(int argc, char ** argv) {
const bool printing_ids = argc > 3 && std::string(argv[3]) == "--ids";
llama_backend_init(false);
llama_backend_init();
llama_model_params model_params = llama_model_default_params();
model_params.vocab_only = true;

68
examples/train-text-from-scratch/train-text-from-scratch.cpp
View File

@ -50,9 +50,9 @@ struct my_llama_layer {
struct ggml_tensor * ffn_norm;
// ff
struct ggml_tensor * w1;
struct ggml_tensor * w2;
struct ggml_tensor * w3;
struct ggml_tensor * ffn_gate; // w1
struct ggml_tensor * ffn_down; // w2
struct ggml_tensor * ffn_up; // w3
};
struct my_llama_model {
@ -111,13 +111,13 @@ 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: %d\n", __func__, params->n_vocab);
printf("%s: n_ctx: %d\n", __func__, params->n_ctx);
printf("%s: n_embd: %d\n", __func__, params->n_embd);
printf("%s: n_head: %d\n", __func__, params->n_head);
printf("%s: n_ff: %d\n", __func__, params->n_ff);
printf("%s: n_layer: %d\n", __func__, params->n_layer);
printf("%s: n_rot: %d\n", __func__, params->n_rot);
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_head: %u\n", __func__, params->n_head);
printf("%s: n_ff: %u\n", __func__, params->n_ff);
printf("%s: n_layer: %u\n", __func__, params->n_layer);
printf("%s: n_rot: %u\n", __func__, params->n_rot);
}
static void set_param_model(struct my_llama_model * model) {
@ -140,9 +140,9 @@ static void set_param_model(struct my_llama_model * model) {
ggml_set_param(ctx, layer.wv);
ggml_set_param(ctx, layer.wo);
ggml_set_param(ctx, layer.ffn_norm);
ggml_set_param(ctx, layer.w1);
ggml_set_param(ctx, layer.w2);
ggml_set_param(ctx, layer.w3);
ggml_set_param(ctx, layer.ffn_gate);
ggml_set_param(ctx, layer.ffn_down);
ggml_set_param(ctx, layer.ffn_up);
}
}
@ -198,9 +198,9 @@ static void init_model(struct my_llama_model * model) {
layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.w1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_ff, n_embd);
layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
layer.ffn_gate = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
layer.ffn_down = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_ff, n_embd);
layer.ffn_up = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ff);
ggml_set_name(layer.attention_norm, tni(LLM_TENSOR_ATTN_NORM, i));
@ -211,9 +211,9 @@ static void init_model(struct my_llama_model * model) {
ggml_set_name(layer.ffn_norm, tni(LLM_TENSOR_FFN_NORM, i));
ggml_set_name(layer.w1, tni(LLM_TENSOR_FFN_GATE, i));
ggml_set_name(layer.w2, tni(LLM_TENSOR_FFN_DOWN, i));
ggml_set_name(layer.w3, tni(LLM_TENSOR_FFN_UP, i));
ggml_set_name(layer.ffn_gate, tni(LLM_TENSOR_FFN_GATE, i));
ggml_set_name(layer.ffn_down, tni(LLM_TENSOR_FFN_DOWN, i));
ggml_set_name(layer.ffn_up, tni(LLM_TENSOR_FFN_UP, i));
}
set_param_model(model);
@ -244,9 +244,9 @@ static void randomize_model(struct my_llama_model * model, int seed, float mean,
randomize_tensor_normal(layer.ffn_norm, rnd);
randomize_tensor_normal(layer.w1, rnd);
randomize_tensor_normal(layer.w2, rnd);
randomize_tensor_normal(layer.w3, rnd);
randomize_tensor_normal(layer.ffn_gate, rnd);
randomize_tensor_normal(layer.ffn_down, rnd);
randomize_tensor_normal(layer.ffn_up, rnd);
}
free_random_normal_distribution(rnd);
@ -356,11 +356,11 @@ static struct ggml_tensor * llama_build_train_graphs(
struct ggml_tensor * t22 = ggml_rms_norm (ctx, t21, f_norm_rms_eps); set_name(t22, "t22"); assert_shape_2d(t22, n_embd, N*n_batch);
struct ggml_tensor * t23 = ggml_repeat (ctx, layer.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, layer.w3, t24); set_name(t25, "t25"); assert_shape_2d(t25, n_ff, N*n_batch);
struct ggml_tensor * t26 = ggml_mul_mat (ctx, layer.w1, t24); set_name(t26, "t26"); assert_shape_2d(t26, n_ff, N*n_batch);
struct ggml_tensor * t25 = ggml_mul_mat (ctx, layer.ffn_up, t24); set_name(t25, "t25"); assert_shape_2d(t25, n_ff, N*n_batch);
struct ggml_tensor * t26 = ggml_mul_mat (ctx, layer.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, layer.w2, t28); set_name(t29, "t29"); assert_shape_2d(t29, n_embd, N*n_batch);
struct ggml_tensor * t29 = ggml_mul_mat (ctx, layer.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;
checkpoints.push_back(cur);
@ -521,9 +521,9 @@ static void load_llama_model_gguf(struct gguf_context * fctx, struct ggml_contex
copy_tensor_by_name(layer.wv, f_ggml_ctx, tni(LLM_TENSOR_ATTN_V, i));
copy_tensor_by_name(layer.wo, f_ggml_ctx, tni(LLM_TENSOR_ATTN_OUT, i));
copy_tensor_by_name(layer.ffn_norm, f_ggml_ctx, tni(LLM_TENSOR_FFN_NORM, i));
copy_tensor_by_name(layer.w1, f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i));
copy_tensor_by_name(layer.w2, f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i));
copy_tensor_by_name(layer.w3, f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i));
copy_tensor_by_name(layer.ffn_gate, f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i));
copy_tensor_by_name(layer.ffn_down, f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i));
copy_tensor_by_name(layer.ffn_up, f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i));
}
}
@ -664,9 +664,9 @@ static void save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vo
gguf_add_tensor(fctx, layer.wv);
gguf_add_tensor(fctx, layer.wo);
gguf_add_tensor(fctx, layer.ffn_norm);
gguf_add_tensor(fctx, layer.w1);
gguf_add_tensor(fctx, layer.w2);
gguf_add_tensor(fctx, layer.w3);
gguf_add_tensor(fctx, layer.ffn_gate);
gguf_add_tensor(fctx, layer.ffn_down);
gguf_add_tensor(fctx, layer.ffn_up);
}
}
@ -915,9 +915,9 @@ static int64_t get_parameter_count(struct my_llama_model* model) {
nx += ggml_nelements(layer.wv);
nx += ggml_nelements(layer.wo);
nx += ggml_nelements(layer.ffn_norm);
nx += ggml_nelements(layer.w1);
nx += ggml_nelements(layer.w2);
nx += ggml_nelements(layer.w3);
nx += ggml_nelements(layer.ffn_gate);
nx += ggml_nelements(layer.ffn_down);
nx += ggml_nelements(layer.ffn_up);
}
return nx;
}

6
flake.lock
View File

@ -20,11 +20,11 @@
},
"nixpkgs": {
"locked": {
"lastModified": 1707268954,
"narHash": "sha256-2en1kvde3cJVc3ZnTy8QeD2oKcseLFjYPLKhIGDanQ0=",
"lastModified": 1708118438,
"narHash": "sha256-kk9/0nuVgA220FcqH/D2xaN6uGyHp/zoxPNUmPCMmEE=",
"owner": "NixOS",
"repo": "nixpkgs",
"rev": "f8e2ebd66d097614d51a56a755450d4ae1632df1",
"rev": "5863c27340ba4de8f83e7e3c023b9599c3cb3c80",
"type": "github"
},
"original": {

6
ggml-alloc.c
View File

@ -551,7 +551,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
}
for (int j = 0; j < GGML_MAX_SRC; j++) {
if (graph->nodes[i]->src[j] == NULL) {
break;
continue;
}
if (graph->nodes[i]->src[j]->flags & GGML_TENSOR_FLAG_INPUT) {
ggml_gallocr_allocate_node(galloc, graph->nodes[i]->src[j], get_node_buffer_id(node_buffer_ids, i));
@ -787,7 +787,7 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
break;
continue;
}
if (!ggml_gallocr_node_needs_realloc(galloc, src, node_alloc, &node_alloc->src[j])) {
#ifndef NDEBUG
@ -833,7 +833,7 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
break;
continue;
}
ggml_gallocr_init_tensor(galloc, src, node_alloc, &node_alloc->src[j]);
}

27
ggml-backend.c
View File

@ -219,6 +219,10 @@ GGML_CALL void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void *
GGML_ASSERT(buf != NULL && "tensor buffer not set");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
if (!size) {
return;
}
tensor->buffer->iface.set_tensor(buf, tensor, data, offset, size);
}
@ -229,6 +233,10 @@ GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void *
GGML_ASSERT(tensor->buffer != NULL && "tensor buffer not set");
GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
if (!size) {
return;
}
tensor->buffer->iface.get_tensor(buf, tensor, data, offset, size);
}
@ -748,7 +756,7 @@ GGML_CALL static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, str
GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
switch (op->op) {
case GGML_OP_CPY:
return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS; // missing type_traits.from_float
return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS && op->type != GGML_TYPE_IQ1_S; // missing type_traits.from_float
case GGML_OP_MUL_MAT:
return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
default:
@ -998,6 +1006,7 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, gg
}
}
GGML_ASSERT(false && "tensor buffer type not supported by any backend");
return -1; // silence warning
}
#if 0
@ -1032,7 +1041,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
for (int i = 0; i < GGML_MAX_SRC; i++) {
const struct ggml_tensor * src = tensor->src[i];
if (src == NULL) {
break;
continue;
}
if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
int src_backend = ggml_backend_sched_backend_from_buffer(sched, src->buffer);
@ -1079,7 +1088,7 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
break;
continue;
}
ggml_backend_t src_backend = tensor_backend(src);
fprintf(stderr, " %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
@ -1135,7 +1144,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
break;
continue;
}
if (tensor_backend_id(src) == -1) {
tensor_backend_id(src) = ggml_backend_sched_backend_id_from_cur(sched, src);
@ -1247,7 +1256,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
break;
continue;
}
int src_backend_id = tensor_backend_id(src);
if (src_backend_id == -1) {
@ -1306,7 +1315,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
break;
continue;
}
int src_backend_id = tensor_backend_id(src);
assert(src_backend_id != -1); // all inputs should be assigned by now
@ -1353,7 +1362,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
for (int j = 0; j < GGML_MAX_SRC; j++) {
struct ggml_tensor * src = node->src[j];
if (src == NULL) {
break;
continue;
}
ggml_backend_t src_backend = tensor_backend(src);
if (src_backend != tensor_backend /* && src_backend != NULL */) {
@ -1659,7 +1668,7 @@ static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set,
for (int i = 0; i < GGML_MAX_SRC; i++) {
struct ggml_tensor * s = src->src[i];
if (s == NULL) {
break;
continue;
}
dst->src[i] = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, s);
}
@ -1688,7 +1697,7 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
for (int i = 0; i < GGML_MAX_SRC; i++) {
struct ggml_tensor * s = src->src[i];
if (s == NULL) {
break;
continue;
}
graph_copy_init_tensor(hash_set, node_copies, node_init, s);
}

488
ggml-cuda.cu
View File

@ -518,6 +518,15 @@ typedef struct {
} block_iq3_xxs;
static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
#define QR1_S 8
#define QI1_S (QK_K / (4*QR1_S))
typedef struct {
half d;
uint8_t qs[QK_K/8];
uint8_t scales[QK_K/16];
} block_iq1_s;
static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
#define WARP_SIZE 32
#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
@ -1708,6 +1717,137 @@ static const __device__ uint32_t iq3xxs_grid[256] = {
0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
};
static const __device__ uint64_t iq1s_grid[512] = {
0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01,
0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100,
0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00,
0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101,
0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100,
0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00,
0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff,
0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000,
0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000,
0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001,
0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff,
0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01,
0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001,
0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00,
0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001,
0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100,
0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000,
0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000,
0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000,
0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff,
0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff,
0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01,
0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100,
0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff,
0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000,
0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101,
0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff,
0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff,
0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001,
0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01,
0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101,
0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100,
0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00,
0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001,
0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff,
0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000,
0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000,
0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100,
0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100,
0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01,
0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff,
0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101,
0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000,
0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff,
0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000,
0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff,
0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00,
0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101,
0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000,
0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000,
0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000,
0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100,
0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000,
0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff,
0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000,
0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000,
0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000,
0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff,
0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000,
0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001,
0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100,
0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000,
0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00,
0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100,
0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000,
0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001,
0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00,
0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff,
0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100,
0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff,
0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000,
0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff,
0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff,
0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00,
0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001,
0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001,
0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01,
0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000,
0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101,
0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00,
0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100,
0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101,
0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101,
0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000,
0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff,
0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff,
0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101,
0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff,
0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101,
0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001,
0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff,
0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff,
0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01,
0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff,
0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100,
0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001,
0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00,
0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff,
0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff,
0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000,
0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000,
0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101,
0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001,
0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000,
0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101,
0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000,
0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001,
0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000,
0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100,
0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000,
0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000,
0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100,
0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff,
0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff,
0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00,
0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101,
0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000,
0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00,
0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000,
0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff,
0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101,
0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff,
0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00,
0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff,
};
static const __device__ uint8_t ksigns_iq2xs[128] = {
0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15,
144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159,
@ -1850,6 +1990,29 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
}
template<typename dst_t>
static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
const int i = blockIdx.x;
const block_iq1_s * x = (const block_iq1_s *) vx;
const int tid = threadIdx.x;
#if QK_K == 256
const int il = tid/8; // 0...3
const int ib = tid%8; // 0...7
dst_t * y = yy + i*QK_K + 32*ib + 8*il;
const int i8 = 4*ib+il;
uint8_t h = x[i].scales[i8/2] >> 4*(i8%2);
const int8_t * grid = (const int8_t *)(iq1s_grid + (x[i].qs[i8] | ((h & 8) << 5)));
const float d = (float)x[i].d * (2*(h & 7) + 1);
for (int j = 0; j < 8; ++j) y[j] = d * grid[j];
#else
assert(false);
#endif
}
static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
@ -4549,6 +4712,49 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
#endif
}
static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if QK_K == 256
const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
const int ib32 = iqs;
int sumi1 = 0, sumi2 = 0, sumi3 = 0, sumi4 = 0;
const uint8_t h1 = bq1->scales[2*ib32+0];
const uint8_t h2 = bq1->scales[2*ib32+1];
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const int * q8 = (const int *)bq8_1[ib32].qs;
const int * grid1 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+0] | ((h1 & 0x08) << 5)));
const int * grid2 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+1] | ((h1 & 0x80) << 1)));
const int * grid3 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+2] | ((h2 & 0x08) << 5)));
const int * grid4 = (const int *)(iq1s_grid + (bq1->qs[4*ib32+3] | ((h2 & 0x80) << 1)));
for (int j = 0; j < 2; ++j) {
sumi1 = __dp4a(q8[j+0], grid1[j], sumi1);
sumi2 = __dp4a(q8[j+2], grid2[j], sumi2);
sumi3 = __dp4a(q8[j+4], grid3[j], sumi3);
sumi4 = __dp4a(q8[j+6], grid4[j], sumi4);
}
#else
const int8_t * q8 = bq8_1[ib32].qs;
const int8_t * grid1 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+0] | ((h1 & 0x08) << 5)));
const int8_t * grid2 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+1] | ((h1 & 0x80) << 1)));
const int8_t * grid3 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+2] | ((h2 & 0x08) << 5)));
const int8_t * grid4 = (const int8_t *)(iq1s_grid + (bq1->qs[4*ib32+3] | ((h2 & 0x80) << 1)));
for (int j = 0; j < 8; ++j) {
sumi1 += q8[j+ 0] * grid1[j];
sumi2 += q8[j+ 8] * grid2[j];
sumi3 += q8[j+16] * grid3[j];
sumi4 += q8[j+24] * grid4[j];
}
#endif
const float d = (float)bq1->d * __low2float(bq8_1[ib32].ds);
return d * (sumi1 * (2*(h1 & 7) + 1) + sumi2 * (2*((h1 >> 4) & 7) + 1) +
sumi3 * (2*(h2 & 7) + 1) + sumi4 * (2*((h2 >> 4) & 7) + 1));
#else
assert(false);
return 0.f;
#endif
}
template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x, int mmq_y, int nwarps,
allocate_tiles_cuda_t allocate_tiles, load_tiles_cuda_t load_tiles, int vdr, vec_dot_q_mul_mat_cuda_t vec_dot>
static __device__ __forceinline__ void mul_mat_q(
@ -5983,149 +6189,31 @@ static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int
dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
}
template <bool vals_smem, int ncols_template, int block_size_template, bool need_check>
static __global__ void soft_max_f16(const float * x, const half * y, float * dst, const int ncols_par, const int nrows_y, const float scale) {
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
const int ncols_data = ncols_template == 0 ? ncols_par : ncols_template;
const int ncols_smem = GGML_PAD(ncols_data, 2*WARP_SIZE)/2;
const int tid = threadIdx.x;
const int rowx = blockIdx.x;
const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
const int warp_id = threadIdx.x / WARP_SIZE;
const int lane_id = threadIdx.x % WARP_SIZE;
extern __shared__ half data_soft_max_f16[];
half * buf_iw = data_soft_max_f16 + 0; // shared memory buffer for inter-warp communication
// (shared memory) buffer to cache values between iterations:
half2 * vals = vals_smem ? (half2 *) (buf_iw + WARP_SIZE) : (half2 *) (dst + rowx*ncols_data);
// if the buffer is larger than max. shared memory per block, use dst as temp. buffer instead
// in that case col_smem == col_data must be enforced to avoid race conditions
half2 max_val = make_half2(-INFINITY, -INFINITY);
#pragma unroll
for (int col0 = 0; col0 < ncols_smem; col0 += block_size) {
const int col_data = 2*col0 + 2*WARP_SIZE*warp_id + lane_id;
const int col_smem = vals_smem ? col0 + tid : col_data;
const int ix = rowx*ncols_data + col_data;
const int iy = rowy*ncols_data + col_data;
half2 val;
if (need_check && col_data + 0 >= ncols_data) {
val.x = -INFINITY;
} else {
val.x = x[ix + 0]*scale + (y ? __half2float(y[iy + 0]) : 0.0f);
}
if (need_check && col_data + WARP_SIZE >= ncols_data) {
val.y = -INFINITY;
} else {
val.y = x[ix + WARP_SIZE]*scale + (y ? __half2float(y[iy + WARP_SIZE]) : 0.0f);
}
if (!need_check || col_smem < (vals_smem ? ncols_smem : ncols_data)) {
vals[col_smem] = val;
}
max_val = __hmax2(max_val, val);
}
// find the max value in the block
max_val = warp_reduce_max(max_val);
if (block_size > WARP_SIZE) {
if (warp_id == 0) {
buf_iw[lane_id] = -INFINITY;
}
__syncthreads();
if (lane_id == 0) {
buf_iw[warp_id] = __hmax(max_val.x, max_val.y);
}
__syncthreads();
max_val = __half2half2(buf_iw[lane_id]);
max_val = warp_reduce_max(max_val);
} else {
max_val = __half2half2(__hmax(max_val.x, max_val.y));
}
half2 tmp = make_half2(0.0f, 0.0f); // partial sums
#pragma unroll
for (int col0 = 0; col0 < ncols_smem; col0 += block_size) {
const int col_smem = vals_smem ? col0 + tid : 2*col0 + 2*warp_id*WARP_SIZE + lane_id;
if (ncols_template == 0 && col_smem >= (vals_smem ? ncols_smem : ncols_data)) {
break;
}
const half2 val = h2exp(vals[col_smem] - max_val);
tmp += val;
vals[col_smem] = val;
}
// find the sum of exps in the block
tmp = warp_reduce_sum(tmp);
if (block_size > WARP_SIZE) {
if (warp_id == 0) {
buf_iw[lane_id] = 0.0f;
}
__syncthreads();
if (lane_id == 0) {
buf_iw[warp_id] = tmp.x + tmp.y;
}
__syncthreads();
tmp = __half2half2(buf_iw[lane_id]);
tmp = warp_reduce_sum(tmp);
} else {
tmp = __half2half2(tmp.x + tmp.y);
}
const half2 inv_sum = make_half2(1.0f, 1.0f) / tmp;
#pragma unroll
for (int col0 = 0; col0 < ncols_smem; col0 += block_size) {
const int col_data = 2*col0 + 2*WARP_SIZE*warp_id + lane_id;
const int col_smem = vals_smem ? col0 + tid : col_data;
const int idst = rowx*ncols_data + col_data;
const half2 result = vals[col_smem] * inv_sum;
if (need_check && col_data + 0 >= ncols_data) {
return;
}
dst[idst] = result.x;
if (need_check && col_data + WARP_SIZE >= ncols_data) {
return;
}
dst[idst + WARP_SIZE] = result.y;
}
#else
(void) x; (void) y; (void) dst; (void) ncols_par; (void) nrows_y; (void) scale;
NO_DEVICE_CODE;
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX
}
template <bool vals_smem, int ncols_template, int block_size_template>
static __global__ void soft_max_f32(const float * x, const half * y, float * dst, const int ncols_par, const int nrows_y, const float scale) {
static __global__ void soft_max_f32(const float * x, const half * mask, const half * pos, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
const int tid = threadIdx.x;
const int rowx = blockIdx.x;
const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
const int warp_id = threadIdx.x / WARP_SIZE;
const int lane_id = threadIdx.x % WARP_SIZE;
float slope = 0.0f;
// ALiBi
if (max_bias > 0.0f) {
const int h = rowx/nrows_y; // head index
const float base = h < n_head_log2 ? m0 : m1;
const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
slope = powf(base, exp);
}
extern __shared__ float data_soft_max_f32[];
float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
// shared memory buffer to cache values between iterations:
@ -6144,7 +6232,8 @@ static __global__ void soft_max_f32(const float * x, const half * y, float * dst
const int ix = rowx*ncols + col;
const int iy = rowy*ncols + col;
const float val = x[ix]*scale + (y ? __half2float(y[iy]) : 0.0f);
const float val = x[ix]*scale + (mask ? __half2float(mask[iy]) : 0.0f) + (pos ? __half2float(slope*pos[col]) : 0.0f);
vals[col] = val;
max_val = max(max_val, val);
}
@ -7235,6 +7324,12 @@ static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int k,
dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
}
template<typename dst_t>
static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
const int nb = k / QK_K;
dequantize_block_iq1_s<<<nb, 32, 0, stream>>>(vx, y);
}
template <typename src_t, typename dst_t>
static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
@ -7274,6 +7369,8 @@ static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
return dequantize_row_iq2_xs_cuda;
case GGML_TYPE_IQ3_XXS:
return dequantize_row_iq3_xxs_cuda;
case GGML_TYPE_IQ1_S:
return dequantize_row_iq1_s_cuda;
case GGML_TYPE_F32:
return convert_unary_cuda<float>;
default:
@ -7309,6 +7406,8 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
return dequantize_row_iq2_xs_cuda;
case GGML_TYPE_IQ3_XXS:
return dequantize_row_iq3_xxs_cuda;
case GGML_TYPE_IQ1_S:
return dequantize_row_iq1_s_cuda;
case GGML_TYPE_F16:
return convert_unary_cuda<half>;
default:
@ -8146,89 +8245,53 @@ static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols
diag_mask_inf_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x, rows_per_channel, n_past);
}
static void soft_max_f16_cuda(const float * x, const half * y, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, cudaStream_t stream) {
int nth = WARP_SIZE;
while (nth < ncols_x/2 && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
const dim3 block_dims(nth, 1, 1);
const dim3 block_nums(nrows_x, 1, 1);
const size_t shmem = (GGML_PAD(ncols_x, 2*WARP_SIZE) + WARP_SIZE)*sizeof(half);
static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
if (shmem <= g_device_caps[g_main_device].smpb) {
switch (ncols_x) {
case 32:
soft_max_f16<true, 32, 32, true><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
break;
case 64:
soft_max_f16<true, 64, 32, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
break;
case 128:
soft_max_f16<true, 128, 64, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
break;
case 256:
soft_max_f16<true, 256, 128, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
break;
case 512:
soft_max_f16<true, 512, 256, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
break;
case 1024:
soft_max_f16<true, 1024, 512, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
break;
case 2048:
soft_max_f16<true, 2048, 1024, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
break;
case 4096:
soft_max_f16<true, 4096, 1024, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
break;
default:
soft_max_f16<true, 0, 0, true><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
break;
}
} else {
const size_t shmem_low = WARP_SIZE*sizeof(half);
soft_max_f16<false, 0, 0, true><<<block_nums, block_dims, shmem_low, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
}
}
static void soft_max_f32_cuda(const float * x, const half * y, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, cudaStream_t stream) {
static void soft_max_f32_cuda(const float * x, const half * mask, const half * pos, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
int nth = WARP_SIZE;
while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
const dim3 block_dims(nth, 1, 1);
const dim3 block_nums(nrows_x, 1, 1);
const size_t shmem = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
const uint32_t n_head_kv = nrows_x/nrows_y;
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
if (shmem < g_device_caps[g_main_device].smpb) {
switch (ncols_x) {
case 32:
soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 64:
soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 128:
soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 256:
soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 512:
soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 1024:
soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 2048:
soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
case 4096:
soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
default:
soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
break;
}
} else {
const size_t shmem_low = WARP_SIZE*sizeof(float);
soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
}
}
@ -8507,6 +8570,7 @@ GGML_CALL void ggml_init_cublas() {
if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
initialized = true;
g_cublas_loaded = false;
fprintf(stderr, "%s: no " GGML_CUDA_NAME " devices found, " GGML_CUDA_NAME " will be disabled\n", __func__);
return;
}
@ -9094,6 +9158,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
return max_compute_capability >= CC_RDNA2 ? 128 : 64;
default:
GGML_ASSERT(false);
@ -9117,6 +9182,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
return max_compute_capability >= CC_VOLTA ? 128 : 64;
case GGML_TYPE_Q6_K:
return 64;
@ -9214,6 +9280,10 @@ static void ggml_cuda_op_mul_mat_vec_q(
mul_mat_vec_q_cuda<QK_K, QI3_XXS, block_iq3_xxs, 1, vec_dot_iq3_xxs_q8_1>
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
break;
case GGML_TYPE_IQ1_S:
mul_mat_vec_q_cuda<QK_K, QI1_S, block_iq1_s, 1, vec_dot_iq1_s_q8_1>
(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
break;
default:
GGML_ASSERT(false);
break;
@ -9653,30 +9723,36 @@ static void ggml_cuda_op_soft_max(
GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16); // src1 contains mask and it is optional
const int64_t ne00 = src0->ne[0];
const int64_t ne00 = src0->ne[0];
const int64_t nrows_x = ggml_nrows(src0);
const int64_t nrows_y = src1 ? src0->ne[1] : 1; // note: using number of queries since mask can be padded!
const int64_t nrows_y = src0->ne[1];
float scale = 1.0f;
memcpy(&scale, dst->op_params, sizeof(float));
float scale = 1.0f;
float max_bias = 0.0f;
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION >= CUDART_HMAX
#ifdef GGML_CUDA_F16
const bool use_f16_soft_max = true;
#else
const bool use_f16_soft_max = false;
#endif // GGML_CUDA_F16
#else
const bool use_f16_soft_max = false;
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && CUDART_VERSION >= CUDART_HMAX
memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
if (use_f16_soft_max) {
soft_max_f16_cuda(src0_dd, src1 ? (const half *) src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream);
} else {
soft_max_f32_cuda(src0_dd, src1 ? (const half *) src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream);
// positions tensor
half * src2_dd = nullptr;
cuda_pool_alloc<half> src2_f;
ggml_tensor * src2 = dst->src[2];
const bool use_src2 = src2 != nullptr;
if (use_src2) {
const bool src2_on_device = src2->backend == GGML_BACKEND_GPU;
if (src2_on_device) {
ggml_tensor_extra_gpu * src2_extra = (ggml_tensor_extra_gpu *) src2->extra;
src2_dd = (half *) src2_extra->data_device[g_main_device];
} else {
src2_dd = src2_f.alloc(ggml_nelements(src2));
CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src2_dd, src2, 0, 0, 0, 1, main_stream));
}
}
(void) dst;
soft_max_f32_cuda(src0_dd, src1 ? (const half *) src1_dd : nullptr, src2_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream);
}
static void ggml_cuda_op_scale(
@ -12139,7 +12215,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
return false;
}
ggml_type a_type = a->type;
if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS) {
if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS || a_type == GGML_TYPE_IQ1_S) {
if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
return false;
}

71
ggml-metal.m
View File

@ -61,6 +61,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,
GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
GGML_METAL_KERNEL_TYPE_RMS_NORM,
GGML_METAL_KERNEL_TYPE_GROUP_NORM,
@ -83,6 +84,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
//GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,
@ -101,6 +103,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,
@ -116,6 +119,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
@ -131,6 +135,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_ROPE_F32,
GGML_METAL_KERNEL_TYPE_ROPE_F16,
GGML_METAL_KERNEL_TYPE_ALIBI_F32,
@ -182,7 +187,7 @@ struct ggml_metal_context {
// MSL code
// TODO: move the contents here when ready
// for now it is easier to work in a separate file
//static NSString * const msl_library_source = @"see metal.metal";
// static NSString * const msl_library_source = @"see metal.metal";
// Here to assist with NSBundle Path Hack
@interface GGMLMetalClass : NSObject
@ -442,6 +447,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS, get_rows_iq2_xxs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS, get_rows_iq2_xs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS, get_rows_iq3_xxs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S, get_rows_iq1_s, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, ctx->support_simdgroup_reduction);
@ -464,6 +470,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, ctx->support_simdgroup_reduction);
@ -482,6 +489,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, ctx->support_simdgroup_mm);
@ -497,6 +505,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, ctx->support_simdgroup_mm);
@ -512,6 +521,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32, rope_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16, rope_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32, alibi_f32, true);
@ -744,6 +754,7 @@ static bool ggml_metal_graph_compute(
size_t offs_src0 = 0;
size_t offs_src1 = 0;
size_t offs_src2 = 0;
size_t offs_dst = 0;
id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
@ -762,6 +773,7 @@ static bool ggml_metal_graph_compute(
struct ggml_tensor * src0 = gf->nodes[i]->src[0];
struct ggml_tensor * src1 = gf->nodes[i]->src[1];
struct ggml_tensor * src2 = gf->nodes[i]->src[2];
struct ggml_tensor * dst = gf->nodes[i];
switch (dst->op) {
@ -823,6 +835,7 @@ static bool ggml_metal_graph_compute(
id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil;
id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil;
id<MTLBuffer> id_src2 = src2 ? ggml_metal_get_buffer(src2, &offs_src2) : nil;
id<MTLBuffer> id_dst = dst ? ggml_metal_get_buffer(dst, &offs_dst) : nil;
//GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
@ -1206,7 +1219,16 @@ static bool ggml_metal_graph_compute(
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX].pipeline;
}
const float scale = ((float *) dst->op_params)[0];
const float scale = ((float *) dst->op_params)[0];
const float max_bias = ((float *) dst->op_params)[1];
const int64_t nrows_x = ggml_nrows(src0);
const int64_t nrows_y = src0->ne[1];
const uint32_t n_head_kv = nrows_x/nrows_y;
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@ -1215,11 +1237,20 @@ static bool ggml_metal_graph_compute(
} else {
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&scale length:sizeof(scale) atIndex:6];
if (id_src2) {
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
} else {
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:2];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:3];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:4];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:5];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:6];
[encoder setBytes:&scale length:sizeof(scale) atIndex:7];
[encoder setBytes:&max_bias length:sizeof(max_bias) atIndex:8];
[encoder setBytes:&m0 length:sizeof(m0) atIndex:9];
[encoder setBytes:&m1 length:sizeof(m1) atIndex:10];
[encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:11];
[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
@ -1315,6 +1346,7 @@ static bool ggml_metal_graph_compute(
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32 ].pipeline; break;
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
}
@ -1449,6 +1481,12 @@ static bool ggml_metal_graph_compute(
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
} break;
case GGML_TYPE_IQ1_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
} break;
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
@ -1483,7 +1521,7 @@ static bool ggml_metal_graph_compute(
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) {
src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_IQ1_S) { // || src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
@ -1532,8 +1570,6 @@ static bool ggml_metal_graph_compute(
// max size of the src1ids array in the kernel stack
GGML_ASSERT(ne11 <= 512);
struct ggml_tensor * src2 = gf->nodes[i]->src[2];
const int64_t ne20 = src2 ? src2->ne[0] : 0;
const int64_t ne21 = src2 ? src2->ne[1] : 0;
const int64_t ne22 = src2 ? src2->ne[2] : 0;
@ -1591,6 +1627,7 @@ static bool ggml_metal_graph_compute(
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32 ].pipeline; break;
default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
}
@ -1728,6 +1765,12 @@ static bool ggml_metal_graph_compute(
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline;
} break;
case GGML_TYPE_IQ1_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline;
} break;
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
@ -1778,7 +1821,7 @@ static bool ggml_metal_graph_compute(
if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1 ||
src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1 || src2t == GGML_TYPE_Q8_0 ||
src2t == GGML_TYPE_Q2_K) { // || src2t == GGML_TYPE_Q4_K) {
src2t == GGML_TYPE_Q2_K || src2t == GGML_TYPE_IQ1_S) { // || src2t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_IQ2_XS) {
@ -1832,6 +1875,7 @@ static bool ggml_metal_graph_compute(
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S ].pipeline; break;
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32 ].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
}
@ -2202,18 +2246,13 @@ static bool ggml_metal_graph_compute(
GGML_ASSERT(ne00 % 4 == 0);
GGML_ASSERT(src0->type == GGML_TYPE_F32);
struct ggml_tensor * src2 = gf->nodes[i]->src[2];
struct ggml_tensor * src3 = gf->nodes[i]->src[3];
GGML_ASSERT(ggml_are_same_shape(src1, src2));
GGML_ASSERT(src3);
size_t offs_src2 = 0;
size_t offs_src3 = 0;
GGML_ASSERT(src2);
id<MTLBuffer> id_src2 = ggml_metal_get_buffer(src2, &offs_src2);
id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16);

402
ggml-metal.metal
View File

@ -351,12 +351,17 @@ kernel void kernel_sum_rows(
kernel void kernel_soft_max(
device const char * src0,
device const char * src1,
device const char * src2,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant float & scale,
threadgroup float * buf [[threadgroup(0)]],
constant float & max_bias,
constant float & m0,
constant float & m1,
constant uint32_t & n_head_log2,
threadgroup float * buf [[threadgroup(0)]],
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]],
@ -368,13 +373,26 @@ kernel void kernel_soft_max(
device const float * psrc0 = (device const float *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
device const half * pmask = src1 != src0 ? (device const half *) src1 + i01*ne00 : nullptr;
device const half * ppos = src2 != src0 ? (device const half *) src2 : nullptr;
device float * pdst = (device float *) dst + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
float slope = 0.0f;
// ALiBi
if (max_bias > 0.0f) {
const int64_t h = i02;
const float base = h < n_head_log2 ? m0 : m1;
const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
slope = pow(base, exp);
}
// parallel max
float lmax = -INFINITY;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f));
lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
}
// find the max value in the block
@ -399,7 +417,7 @@ kernel void kernel_soft_max(
// parallel sum
float lsum = 0.0f;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f)) - max_val);
const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
lsum += exp_psrc0;
pdst[i00] = exp_psrc0;
}
@ -437,11 +455,16 @@ kernel void kernel_soft_max(
kernel void kernel_soft_max_4(
device const char * src0,
device const char * src1,
device const char * src2,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant float & scale,
constant float & max_bias,
constant float & m0,
constant float & m1,
constant uint32_t & n_head_log2,
threadgroup float * buf [[threadgroup(0)]],
uint tgpig[[threadgroup_position_in_grid]],
uint tpitg[[thread_position_in_threadgroup]],
@ -454,13 +477,25 @@ kernel void kernel_soft_max_4(
device const float4 * psrc4 = (device const float4 *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4;
device const half4 * pmask = src1 != src0 ? (device const half4 *) src1 + i01*ne00/4 : nullptr;
device const half4 * ppos = src2 != src0 ? (device const half4 *) src2 : nullptr;
device float4 * pdst4 = (device float4 *) dst + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4;
float slope = 0.0f;
if (max_bias > 0.0f) {
const int64_t h = i02;
const float base = h < n_head_log2 ? m0 : m1;
const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
slope = pow(base, exp);
}
// parallel max
float4 lmax4 = -INFINITY;
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
lmax4 = fmax(lmax4, psrc4[i00]*scale + (float4) (pmask ? pmask[i00] : 0.0f));
lmax4 = fmax(lmax4, psrc4[i00]*scale + (float4)((pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)));
}
const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
@ -486,7 +521,7 @@ kernel void kernel_soft_max_4(
// parallel sum
float4 lsum4 = 0.0f;
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
const float4 exp_psrc4 = exp((psrc4[i00]*scale + (float4) (pmask ? pmask[i00] : 0.0f)) - max_val);
const float4 exp_psrc4 = exp((psrc4[i00]*scale + (float4)((pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f))) - max_val);
lsum4 += exp_psrc4;
pdst4[i00] = exp_psrc4;
}
@ -2895,6 +2930,13 @@ typedef struct {
} block_iq3_xxs;
// 98 bytes / block for QK_K = 256, so 3.0625 bpw
typedef struct {
half d;
uint8_t qs[QK_K/8];
uint8_t scales[QK_K/16];
} block_iq1_s;
//====================================== dot products =========================
void kernel_mul_mv_q2_K_f32_impl(
@ -4152,6 +4194,137 @@ constexpr constant static uint32_t iq3xxs_grid[256] = {
0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
};
#define NGRID_IQ1S 512
constexpr constant static uint64_t iq1s_grid[NGRID_IQ1S] = {
0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01,
0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100,
0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00,
0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101,
0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100,
0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00,
0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff,
0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000,
0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000,
0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001,
0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff,
0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01,
0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001,
0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00,
0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001,
0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100,
0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000,
0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000,
0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000,
0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff,
0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff,
0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01,
0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100,
0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff,
0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000,
0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101,
0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff,
0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff,
0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001,
0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01,
0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101,
0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100,
0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00,
0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001,
0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff,
0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000,
0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000,
0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100,
0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100,
0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01,
0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff,
0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101,
0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000,
0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff,
0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000,
0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff,
0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00,
0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101,
0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000,
0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000,
0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000,
0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100,
0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000,
0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff,
0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000,
0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000,
0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000,
0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff,
0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000,
0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001,
0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100,
0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000,
0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00,
0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100,
0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000,
0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001,
0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00,
0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff,
0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100,
0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff,
0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000,
0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff,
0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff,
0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00,
0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001,
0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001,
0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01,
0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000,
0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101,
0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00,
0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100,
0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101,
0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101,
0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000,
0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff,
0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff,
0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101,
0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff,
0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101,
0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001,
0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff,
0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff,
0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01,
0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff,
0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100,
0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001,
0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00,
0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff,
0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff,
0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000,
0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000,
0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101,
0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001,
0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000,
0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101,
0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000,
0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001,
0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000,
0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100,
0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000,
0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000,
0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100,
0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff,
0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff,
0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00,
0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101,
0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000,
0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00,
0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000,
0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff,
0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101,
0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff,
0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00,
0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff,
};
constexpr constant static uint8_t ksigns_iq2xs[128] = {
0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15,
@ -4259,7 +4432,10 @@ void kernel_mul_mv_iq2_xxs_f32_impl(
y4 += 32 * 32;
}
#else
// TODO
(void) x;
(void) y;
(void) yl;
(void) nb32;
#endif
for (int row = 0; row < N_DST; ++row) {
@ -4402,7 +4578,10 @@ void kernel_mul_mv_iq2_xs_f32_impl(
y4 += 32 * 32;
}
#else
// TODO
(void) x;
(void) y;
(void) yl;
(void) nb32;
#endif
for (int row = 0; row < N_DST; ++row) {
@ -4538,7 +4717,10 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
y4 += 32 * 32;
}
#else
// TODO
(void) x;
(void) y;
(void) yl;
(void) nb32;
#endif
for (int row = 0; row < N_DST; ++row) {
@ -4578,6 +4760,126 @@ kernel void kernel_mul_mv_iq3_xxs_f32(
kernel_mul_mv_iq3_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
}
void kernel_mul_mv_iq1_s_f32_impl(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne10,
constant int64_t & ne12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
const int nb = ne00/QK_K;
const int r0 = tgpig.x;
const int r1 = tgpig.y;
const int im = tgpig.z;
const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
const int ib_row = first_row * nb;
const uint i12 = im%ne12;
const uint i13 = im/ne12;
const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
device const block_iq1_s * x = (device const block_iq1_s *) src0 + ib_row + offset0;
device const float * y = (device const float *) src1 + r1*ne10 + im*ne00*ne1;
float yl[16];
float sumf[N_DST]={0.f}, all_sum;
const int nb32 = nb * (QK_K / 32);
#if QK_K == 256
const int ix = tiisg/2;
const int il = tiisg%2;
device const float * y4 = y + 32 * ix + 16 * il;
for (int ib32 = ix; ib32 < nb32; ib32 += 16) {
for (int i = 0; i < 16; ++i) {
yl[i] = y4[i];
}
const int ibl = ib32 / (QK_K / 32);
const int ib = ib32 % (QK_K / 32);
device const block_iq1_s * xr = x + ibl;
device const uint8_t * qs = xr->qs + 4 * ib + 2 * il;
device const uint8_t * sc = xr->scales + 2 * ib + il;
device const half * dh = &xr->d;
for (int row = 0; row < N_DST; row++) {
constant int8_t * grid1 = (constant int8_t *)(iq1s_grid + (qs[0] | ((sc[0] & 0x08) << 5)));
constant int8_t * grid2 = (constant int8_t *)(iq1s_grid + (qs[1] | ((sc[0] & 0x80) << 1)));
float2 sum = {0};
for (int j = 0; j < 8; ++j) {
sum[0] += yl[j+ 0] * grid1[j];
sum[1] += yl[j+ 8] * grid2[j];
}
sumf[row] += (float)dh[0] * (sum[0] * (2*(sc[0] & 7) + 1) + sum[1] * (2*((sc[0] >> 4) & 7) + 1));
dh += nb*sizeof(block_iq1_s)/2;
qs += nb*sizeof(block_iq1_s);
sc += nb*sizeof(block_iq1_s);
}
y4 += 16 * 32;
}
#else
(void) x;
(void) y;
(void) yl;
(void) nb32;
#endif
for (int row = 0; row < N_DST; ++row) {
all_sum = simd_sum(sumf[row]);
if (tiisg == 0) {
dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
}
}
}
[[host_name("kernel_mul_mv_iq1_s_f32")]]
kernel void kernel_mul_mv_iq1_s_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint & r2,
constant uint & r3,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
kernel_mul_mv_iq1_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
}
//============================= templates and their specializations =============================
@ -4774,6 +5076,8 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg
const float dl = d * sc[0];
const float ml = min * sc[1];
#else
(void) get_scale_min_k4_just2;
q = q + 16 * (il&1);
device const uint8_t * s = xb->scales;
device const half2 * dh = (device const half2 *)xb->d;
@ -4923,6 +5227,22 @@ void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x
}
}
template <typename type4x4>
void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
// il is 0...15 for QK_K = 256 => index of block of 32 is il/2
const float d = xb->d;
device const uint8_t * qs = xb->qs + 2*il;
device const uint8_t * sc = xb->scales + il;
const float dl1 = d * (2*(sc[0] & 7) + 1);
const float dl2 = d * (2*((sc[0] >> 4) & 7) + 1);
constant int8_t * grid1 = (constant int8_t *)(iq1s_grid + (qs[0] | ((sc[0] & 0x08) << 5)));
constant int8_t * grid2 = (constant int8_t *)(iq1s_grid + (qs[1] | ((sc[0] & 0x80) << 1)));
for (int i = 0; i < 8; ++i) {
reg[i/4+0][i%4] = dl1 * grid1[i];
reg[i/4+2][i%4] = dl2 * grid2[i];
}
}
template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
kernel void kernel_get_rows(
device const void * src0,
@ -5465,6 +5785,7 @@ template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows
template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_rows<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
template [[host_name("kernel_get_rows_iq2_xs")]] kernel get_rows_t kernel_get_rows<block_iq2_xs, QK_NL, dequantize_iq2_xs>;
template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_t kernel_get_rows<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
template [[host_name("kernel_get_rows_iq1_s")]] kernel get_rows_t kernel_get_rows<block_iq1_s, QK_NL, dequantize_iq1_s>;
//
// matrix-matrix multiplication
@ -5504,6 +5825,7 @@ template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<b
template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
template [[host_name("kernel_mul_mm_iq2_xs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xs, QK_NL, dequantize_iq2_xs>;
template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
template [[host_name("kernel_mul_mm_iq1_s_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq1_s, QK_NL, dequantize_iq1_s>;
//
// indirect matrix-matrix multiplication
@ -5555,6 +5877,7 @@ template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mu
template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs, QK_NL, dequantize_iq2_xs>;
template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
template [[host_name("kernel_mul_mm_id_iq1_s_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s, QK_NL, dequantize_iq1_s>;
//
// matrix-vector multiplication
@ -6522,3 +6845,66 @@ kernel void kernel_mul_mv_id_iq3_xxs_f32(
tiisg,
sgitg);
}
[[host_name("kernel_mul_mv_id_iq1_s_f32")]]
kernel void kernel_mul_mv_id_iq1_s_f32(
device const char * ids,
device const char * src1,
device float * dst,
constant uint64_t & nbi1,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
constant uint64_t & nb1,
constant uint & r2,
constant uint & r3,
constant int & idx,
device const char * src00,
device const char * src01,
device const char * src02,
device const char * src03,
device const char * src04,
device const char * src05,
device const char * src06,
device const char * src07,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiitg[[thread_index_in_threadgroup]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
const int64_t bid = tgpig.z/(ne12*ne13);
tgpig.z = tgpig.z%(ne12*ne13);
const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
kernel_mul_mv_iq1_s_f32_impl(
src0[id],
(device const float *) (src1 + bid*nb11),
dst + bid*ne0,
ne00,
ne01,
ne02,
ne10,
ne12,
ne0,
ne1,
r2,
r3,
tgpig,
tiisg,
sgitg);
}

715
ggml-quants.c
View File

@ -1837,9 +1837,9 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri
float sigma2 = sumx2/QK_K;
for (int j = 0; j < QK_K/16; ++j) {
const float * restrict qw = quant_weights + QK_K * i + 16*j;
for (int l = 0; l < 16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j + l]*x[16*j + l]);
for (int l = 0; l < 16; ++l) sw[j] += weight[l];
scales[j] = make_qkx3_quants(16, 3, x + 16*j, weight, L + 16*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
for (int l = 0; l < QK_K/16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j + l]*x[16*j + l]);
for (int l = 0; l < QK_K/16; ++l) sw[j] += weight[l];
scales[j] = make_qkx3_quants(QK_K/16, 3, x + 16*j, weight, L + 16*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
}
float dm = make_qp_quants(QK_K/16, 15, scales, Ls, sw);
@ -3480,6 +3480,139 @@ static const uint32_t iq3xxs_grid[256] = {
0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
};
#define NGRID_IQ2XXS 512
static const uint64_t iq1s_grid[NGRID_IQ2XXS] = {
0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01,
0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100,
0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00,
0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101,
0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100,
0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00,
0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff,
0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000,
0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000,
0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001,
0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff,
0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01,
0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001,
0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00,
0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001,
0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100,
0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000,
0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000,
0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000,
0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff,
0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff,
0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01,
0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100,
0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff,
0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000,
0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101,
0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff,
0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff,
0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001,
0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01,
0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101,
0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100,
0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00,
0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001,
0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff,
0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000,
0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000,
0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100,
0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100,
0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01,
0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff,
0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101,
0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000,
0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff,
0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000,
0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff,
0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00,
0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101,
0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000,
0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000,
0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000,
0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100,
0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000,
0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff,
0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000,
0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000,
0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000,
0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff,
0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000,
0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001,
0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100,
0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000,
0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00,
0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100,
0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000,
0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001,
0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00,
0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff,
0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100,
0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff,
0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000,
0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff,
0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff,
0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00,
0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001,
0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001,
0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01,
0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000,
0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101,
0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00,
0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100,
0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101,
0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101,
0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000,
0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff,
0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff,
0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101,
0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff,
0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101,
0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001,
0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff,
0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff,
0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01,
0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff,
0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100,
0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001,
0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00,
0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff,
0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff,
0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000,
0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000,
0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101,
0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001,
0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000,
0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101,
0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000,
0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001,
0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000,
0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100,
0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000,
0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000,
0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100,
0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff,
0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff,
0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00,
0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101,
0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000,
0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00,
0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000,
0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff,
0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101,
0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff,
0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00,
0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff,
};
static const uint8_t ksigns_iq2xs[128] = {
0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15,
144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159,
@ -3578,6 +3711,49 @@ void dequantize_row_iq3_xxs(const block_iq3_xxs * restrict x, float * restrict y
}
}
// ====================== 1.5625 bpw (de)-quantization
void dequantize_row_iq1_s(const block_iq1_s * restrict x, float * restrict y, int k) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
float db[4];
uint16_t idx[4];
//const int8_t * grid[4];
for (int i = 0; i < nb; i++) {
const float d = GGML_FP16_TO_FP32(x[i].d);
const uint8_t * sc = x[i].scales;
const uint8_t * qs = x[i].qs;
for (int i8 = 0; i8 < QK_K/8; i8 += 4) {
idx[0] = qs[0] | ((sc[0] & 0x08) << 5);
idx[1] = qs[1] | ((sc[0] & 0x80) << 1);
idx[2] = qs[2] | ((sc[1] & 0x08) << 5);
idx[3] = qs[3] | ((sc[1] & 0x80) << 1);
//grid[0] = (const int8_t *)(iq1s_grid + (qs[0] | ((sc[0] & 0x08) << 5)));
//grid[1] = (const int8_t *)(iq1s_grid + (qs[1] | ((sc[0] & 0x80) << 1)));
//grid[2] = (const int8_t *)(iq1s_grid + (qs[2] | ((sc[1] & 0x08) << 5)));
//grid[3] = (const int8_t *)(iq1s_grid + (qs[3] | ((sc[1] & 0x80) << 1)));
db[0] = d * (2*(sc[0] & 7) + 1);
db[1] = d * (2*((sc[0] >> 4) & 7) + 1);
db[2] = d * (2*(sc[1] & 7) + 1);
db[3] = d * (2*((sc[1] >> 4) & 7) + 1);
for (int l = 0; l < 4; ++l) {
const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]);
for (int j = 0; j < 8; ++j) {
//y[j] = db[l] * grid[l][j];
y[j] = db[l] * grid[j];
}
y += 8;
}
qs += 4;
sc += 2;
}
}
}
//===================================== Q8_K ==============================================
void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k) {
@ -3819,15 +3995,15 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
/* Compute combined scale for the block */
const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
__m256i bx = bytes_from_nibbles_32(x[i].qs);
__m256i qx = bytes_from_nibbles_32(x[i].qs);
// Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
const __m256i off = _mm256_set1_epi8( 8 );
bx = _mm256_sub_epi8( bx, off );
qx = _mm256_sub_epi8( qx, off );
__m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
__m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_i8_pairs_float(bx, by);
const __m256 q = mul_sum_i8_pairs_float(qx, qy);
/* Multiply q with scale and accumulate */
acc = _mm256_fmadd_ps( d, q, acc );
@ -3848,15 +4024,15 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
const __m128i tmp = _mm_loadu_si128((const __m128i *)x[i].qs);
__m128i bx = _mm_and_si128(lowMask, tmp);
__m128i by = _mm_loadu_si128((const __m128i *)y[i].qs);
bx = _mm_sub_epi8(bx, off);
const __m128i i32_0 = mul_sum_i8_pairs(bx, by);
__m128i bx_0 = _mm_and_si128(lowMask, tmp);
__m128i by_0 = _mm_loadu_si128((const __m128i *)y[i].qs);
bx_0 = _mm_sub_epi8(bx_0, off);
const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
bx = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4));
by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16));
bx = _mm_sub_epi8(bx, off);
const __m128i i32_1 = mul_sum_i8_pairs(bx, by);
bx_0 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4));
by_0 = _mm_loadu_si128((const __m128i *)(y[i].qs + 16));
bx_0 = _mm_sub_epi8(bx_0, off);
const __m128i i32_1 = mul_sum_i8_pairs(bx_0, by_0);
// Convert int32_t to float
__m256 p = _mm256_cvtepi32_ps(MM256_SET_M128I(i32_0, i32_1));
@ -4196,10 +4372,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
// Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
const __m256i bx = bytes_from_nibbles_32(x[i].qs);
const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs );
const __m256i qx = bytes_from_nibbles_32(x[i].qs);
const __m256i qy = _mm256_loadu_si256( (const __m256i *)y[i].qs );
const __m256 xy = mul_sum_us8_pairs_float(bx, by);
const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
// Accumulate d0*d1*x*y
#if defined(__AVX2__)
@ -4418,14 +4594,14 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
/* Compute combined scale for the block */
const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
__m256i bx = bytes_from_nibbles_32(x[i].qs);
__m256i qx = bytes_from_nibbles_32(x[i].qs);
__m256i bxhi = bytes_from_bits_32(x[i].qh);
bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0));
bx = _mm256_or_si256(bx, bxhi);
qx = _mm256_or_si256(qx, bxhi);
__m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
__m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_i8_pairs_float(bx, by);
const __m256 q = mul_sum_i8_pairs_float(qx, qy);
/* Multiply q with scale and accumulate */
acc = _mm256_fmadd_ps(d, q, acc);
@ -4442,21 +4618,21 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r
/* Compute combined scale for the block */
const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
__m256i bx = bytes_from_nibbles_32(x[i].qs);
__m256i bx_0 = bytes_from_nibbles_32(x[i].qs);
const __m256i bxhi = bytes_from_bits_32(x[i].qh);
__m128i bxhil = _mm256_castsi256_si128(bxhi);
__m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
bxhil = _mm_andnot_si128(bxhil, mask);
bxhih = _mm_andnot_si128(bxhih, mask);
__m128i bxl = _mm256_castsi256_si128(bx);
__m128i bxh = _mm256_extractf128_si256(bx, 1);
__m128i bxl = _mm256_castsi256_si128(bx_0);
__m128i bxh = _mm256_extractf128_si256(bx_0, 1);
bxl = _mm_or_si128(bxl, bxhil);
bxh = _mm_or_si128(bxh, bxhih);
bx = MM256_SET_M128I(bxh, bxl);
bx_0 = MM256_SET_M128I(bxh, bxl);
const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_i8_pairs_float(bx, by);
const __m256 q = mul_sum_i8_pairs_float(bx_0, by_0);
/* Multiply q with scale and accumulate */
acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc);
@ -4722,15 +4898,15 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
__m256i bx = bytes_from_nibbles_32(x[i].qs);
__m256i qx = bytes_from_nibbles_32(x[i].qs);
__m256i bxhi = bytes_from_bits_32(x[i].qh);
bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
bx = _mm256_or_si256(bx, bxhi);
qx = _mm256_or_si256(qx, bxhi);
const __m256 dy = _mm256_set1_ps(y[i].d);
const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_us8_pairs_float(bx, by);
const __m256 q = mul_sum_us8_pairs_float(qx, qy);
acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
}
@ -4749,22 +4925,22 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
__m256i bx = bytes_from_nibbles_32(x[i].qs);
__m256i bx_0 = bytes_from_nibbles_32(x[i].qs);
const __m256i bxhi = bytes_from_bits_32(x[i].qh);
__m128i bxhil = _mm256_castsi256_si128(bxhi);
__m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
bxhil = _mm_and_si128(bxhil, mask);
bxhih = _mm_and_si128(bxhih, mask);
__m128i bxl = _mm256_castsi256_si128(bx);
__m128i bxh = _mm256_extractf128_si256(bx, 1);
__m128i bxl = _mm256_castsi256_si128(bx_0);
__m128i bxh = _mm256_extractf128_si256(bx_0, 1);
bxl = _mm_or_si128(bxl, bxhil);
bxh = _mm_or_si128(bxh, bxhih);
bx = MM256_SET_M128I(bxh, bxl);
bx_0 = MM256_SET_M128I(bxh, bxl);
const __m256 dy = _mm256_set1_ps(y[i].d);
const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_us8_pairs_float(bx, by);
const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
}
@ -4973,10 +5149,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
for (int i = 0; i < nb; ++i) {
// Compute combined scale for the block
const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
__m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs);
__m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
__m256i qx = _mm256_loadu_si256((const __m256i *)x[i].qs);
__m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_i8_pairs_float(bx, by);
const __m256 q = mul_sum_i8_pairs_float(qx, qy);
// Multiply q with scale and accumulate
#if defined(__AVX2__)
@ -4993,10 +5169,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
for (int i = 0; i < nb; i++) {
// load elements
vint8m1_t bx = __riscv_vle8_v_i8m1(x[i].qs, vl);
vint8m1_t by = __riscv_vle8_v_i8m1(y[i].qs, vl);
vint8m1_t bx_0 = __riscv_vle8_v_i8m1(x[i].qs, vl);
vint8m1_t by_0 = __riscv_vle8_v_i8m1(y[i].qs, vl);
vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx, by, vl);
vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx_0, by_0, vl);
vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl);
vint32m1_t v_sum = __riscv_vwredsum_vs_i16m2_i32m1(vw_mul, v_zero, vl);
@ -9107,6 +9283,178 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void
#endif
}
#ifdef __AVX2__
static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
const __m256i ax = _mm256_sign_epi8(x, x);
const __m256i sy = _mm256_sign_epi8(y, x);
return _mm256_maddubs_epi16(ax, sy);
}
#endif
void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
assert(n % QK_K == 0);
assert(nrc == 1);
UNUSED(nrc);
UNUSED(bx);
UNUSED(by);
UNUSED(bs);
const block_iq1_s * restrict x = vx;
const block_q8_K * restrict y = vy;
const int nb = n / QK_K;
#if defined __ARM_NEON
const uint8x16_t m8 = vdupq_n_u8(0x08);
const uint8x16_t m7 = vdupq_n_u8(0x07);
const uint8x16_t m1 = vdupq_n_u8(0x01);
const int32x4_t vzero = vdupq_n_s32(0);
uint16_t gindex[8];
uint16x8x2_t vindex;
int8x16x4_t q1b;
int8x16x4_t q8b;
uint16x8x4_t scales;
int32x4x2_t sumi;
int32x4x2_t dotq;
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs;
const uint8_t * sc = x[i].scales;
sumi.val[0] = sumi.val[1] = vzero;
for (int i128 = 0; i128 < QK_K/128; ++i128) {
const uint8x16_t ql = vld1q_u8(qs); qs += 16;
const uint8x8_t tm1 = vld1_u8 (sc); sc += 8;
const uint8x8_t tm2 = vshr_n_u8(tm1, 4);
const uint8x16_t qh = vcombine_u8(vzip1_u8(tm1, tm2), vzip2_u8(tm1, tm2));
const uint8x16_t hbit = vandq_u8(qh, m8);
vindex.val[0] = vorrq_u16(vmovl_u8(vget_low_u8 (ql)), vshlq_n_u16(vmovl_u8(vget_low_u8 (hbit)), 5));
vindex.val[1] = vorrq_u16(vmovl_u8(vget_high_u8(ql)), vshlq_n_u16(vmovl_u8(vget_high_u8(hbit)), 5));
const uint8x16_t scales8 = vorrq_u8(vshlq_n_u8(vandq_u8(qh, m7), 1), m1);
scales.val[0] = vmovl_u8(vget_low_u8 (scales8));
scales.val[1] = vmovl_u8(vget_high_u8 (scales8));
for (int l = 0; l < 2; ++l) {
vst1q_u16(gindex+0, vindex.val[l]);
q1b.val[0] = vcombine_s8(vld1_s8((const void *)(iq1s_grid+gindex[0])), vld1_s8((const void *)(iq1s_grid+gindex[1])));
q1b.val[1] = vcombine_s8(vld1_s8((const void *)(iq1s_grid+gindex[2])), vld1_s8((const void *)(iq1s_grid+gindex[3])));
q1b.val[2] = vcombine_s8(vld1_s8((const void *)(iq1s_grid+gindex[4])), vld1_s8((const void *)(iq1s_grid+gindex[5])));
q1b.val[3] = vcombine_s8(vld1_s8((const void *)(iq1s_grid+gindex[6])), vld1_s8((const void *)(iq1s_grid+gindex[7])));
q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
dotq.val[0] = vpaddq_s32(ggml_vdotq_s32(vzero, q1b.val[0], q8b.val[0]), ggml_vdotq_s32(vzero, q1b.val[1], q8b.val[1]));
dotq.val[1] = vpaddq_s32(ggml_vdotq_s32(vzero, q1b.val[2], q8b.val[2]), ggml_vdotq_s32(vzero, q1b.val[3], q8b.val[3]));
sumi.val[0] = vmlaq_s32(sumi.val[0], dotq.val[0], vreinterpretq_s32_u32(vmovl_u16(vget_low_u16 (scales.val[l]))));
sumi.val[1] = vmlaq_s32(sumi.val[1], dotq.val[1], vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales.val[l]))));
}
}
sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * vaddvq_s32(vaddq_s32(sumi.val[0], sumi.val[1]));
}
*s = sumf;
#elif defined __AVX2__
const __m128i m8 = _mm_set1_epi8(0x08);
const __m128i m7 = _mm_set1_epi8(0x07);
const __m128i m1 = _mm_set1_epi8(0x01);
const __m128i shuffle_h = _mm_set_epi8(15, 7, 14, 6, 13, 5, 12, 4, 11, 3, 10, 2, 9, 1, 8, 0);
const __m128i shuffle_s[4] = {
_mm_set_epi32(0x03030303, 0x02020202, 0x01010101, 0x00000000),
_mm_set_epi32(0x07070707, 0x06060606, 0x05050505, 0x04040404),
_mm_set_epi32(0x0b0b0b0b, 0x0a0a0a0a, 0x09090909, 0x08080808),
_mm_set_epi32(0x0f0f0f0f, 0x0e0e0e0e, 0x0d0d0d0d, 0x0c0c0c0c)
};
uint64_t aux64;
__m256i v_gindex;
const uint16_t * gindex = (const uint16_t *)&v_gindex;
__m256 accum = _mm256_setzero_ps();
for (int i = 0; i < nb; ++i) {
const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs;
const uint8_t * sc = x[i].scales;
__m256i sumi = _mm256_setzero_si256();
for (int i128 = 0; i128 < QK_K/128; ++i128) {
const __m128i ql = _mm_loadu_si128((const __m128i*)qs); qs += 16;
memcpy(&aux64, sc, 8); sc += 8;
const __m128i qh = _mm_shuffle_epi8(_mm_set_epi64x(aux64 >> 4, aux64), shuffle_h);
const __m256i hbit = _mm256_cvtepu8_epi16(_mm_and_si128(qh, m8));
v_gindex = _mm256_or_si256(_mm256_cvtepu8_epi16(ql), _mm256_slli_epi16(hbit, 5));
const __m128i scales = _mm_or_si128(_mm_slli_epi16(_mm_and_si128(qh, m7), 1), m1);
for (int i32 = 0; i32 < 4; ++i32) {
const __m256i q8b = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
const __m256i q1b = _mm256_set_epi64x(iq1s_grid[gindex[4*i32+3]], iq1s_grid[gindex[4*i32+2]],
iq1s_grid[gindex[4*i32+1]], iq1s_grid[gindex[4*i32+0]]);
const __m256i dot = mul_add_epi8(q1b, q8b);
const __m256i s16 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, shuffle_s[i32]));
const __m256i p = _mm256_madd_epi16(s16, dot);
sumi = _mm256_add_epi32(sumi, p);
}
}
accum = _mm256_fmadd_ps(_mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d)), _mm256_cvtepi32_ps(sumi), accum);
}
*s = hsum_float_8(accum);
#else
int db[4];
uint16_t idx[4];
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const int8_t * q8 = y[i].qs;
const uint8_t * qs = x[i].qs;
const uint8_t * sc = x[i].scales;
int sumi = 0;
for (int i32 = 0; i32 < QK_K/32; ++i32) {
idx[0] = qs[0] | ((sc[0] & 0x08) << 5);
idx[1] = qs[1] | ((sc[0] & 0x80) << 1);
idx[2] = qs[2] | ((sc[1] & 0x08) << 5);
idx[3] = qs[3] | ((sc[1] & 0x80) << 1);
db[0] = (2*(sc[0] & 7) + 1);
db[1] = (2*((sc[0] >> 4) & 7) + 1);
db[2] = (2*(sc[1] & 7) + 1);
db[3] = (2*((sc[1] >> 4) & 7) + 1);
for (int l = 0; l < 4; ++l) {
const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]);
int suml = 0;
for (int j = 0; j < 8; ++j) suml += q8[j] * grid[j];
sumi += db[l] * suml;
q8 += 8;
}
qs += 4;
sc += 2;
}
sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * sumi;
}
*s = sumf;
#endif
}
// ================================ IQ2 quantization =============================================
typedef struct {
@ -9115,14 +9463,22 @@ typedef struct {
uint16_t * neighbours;
} iq2_entry_t;
static iq2_entry_t iq2_data[2] = {
static iq2_entry_t iq2_data[3] = {
{NULL, NULL, NULL},
{NULL, NULL, NULL},
{NULL, NULL, NULL},
};
static inline int iq2_data_index(int grid_size) {
GGML_ASSERT(grid_size == 256 || grid_size == 512);
return grid_size == 256 ? 0 : 1;
static inline int iq2_data_index(enum ggml_type type) {
GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S);
return type == GGML_TYPE_IQ2_XXS ? 0 :
type == GGML_TYPE_IQ2_XS ? 1 : 2;
}
static inline int iq2_grid_size(enum ggml_type type) {
GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S);
return type == GGML_TYPE_IQ2_XXS ? 256 :
type == GGML_TYPE_IQ2_XS ? 512 : 512;
}
static int iq2_compare_func(const void * left, const void * right) {
@ -9131,12 +9487,13 @@ static int iq2_compare_func(const void * left, const void * right) {
return l[0] < r[0] ? -1 : l[0] > r[0] ? 1 : l[1] < r[1] ? -1 : l[1] > r[1] ? 1 : 0;
}
void iq2xs_init_impl(int grid_size) {
const int gindex = iq2_data_index(grid_size);
void iq2xs_init_impl(enum ggml_type type) {
const int gindex = iq2_data_index(type);
const int grid_size = iq2_grid_size(type);
if (iq2_data[gindex].grid) {
return;
}
static const uint16_t kgrid_256[256] = {
static const uint16_t kgrid_2bit_256[256] = {
0, 2, 5, 8, 10, 17, 20, 32, 34, 40, 42, 65, 68, 80, 88, 97,
100, 128, 130, 138, 162, 257, 260, 272, 277, 320, 388, 408, 512, 514, 546, 642,
1025, 1028, 1040, 1057, 1060, 1088, 1090, 1096, 1120, 1153, 1156, 1168, 1188, 1280, 1282, 1288,
@ -9154,7 +9511,7 @@ void iq2xs_init_impl(int grid_size) {
33888, 34048, 34118, 34196, 34313, 34368, 34400, 34818, 35076, 35345, 36868, 36880, 36900, 36928, 37025, 37142,
37248, 37445, 37888, 37922, 37956, 38225, 39041, 39200, 40962, 41040, 41093, 41225, 41472, 42008, 43088, 43268,
};
static const uint16_t kgrid_512[512] = {
static const uint16_t kgrid_2bit_512[512] = {
0, 2, 5, 8, 10, 17, 20, 22, 25, 32, 34, 37, 40, 65, 68, 70,
73, 80, 82, 85, 88, 97, 100, 128, 130, 133, 136, 145, 148, 153, 160, 257,
260, 262, 265, 272, 274, 277, 280, 282, 289, 292, 320, 322, 325, 328, 337, 340,
@ -9188,9 +9545,45 @@ void iq2xs_init_impl(int grid_size) {
40962, 40968, 40970, 40992, 41002, 41120, 41297, 41305, 41382, 41472, 41474, 41480, 41514, 41600, 41632, 42048,
42133, 42597, 42648, 43018, 43040, 43042, 43048, 43168, 43176, 43268, 43396, 43398, 43560, 43562, 43665, 43690,
};
static const uint16_t kgrid_1bit_512[512] = {
10, 33, 41, 85, 132, 134, 160, 162, 277, 337, 340, 345, 357, 405, 516, 545,
553, 598, 641, 650, 681, 1042, 1044, 1097, 1169, 1176, 1320, 1345, 1365, 1378, 1434, 1444,
1545, 1617, 1642, 1685, 2053, 2080, 2089, 2133, 2176, 2182, 2208, 2214, 2306, 2384, 2393, 2440,
2453, 2581, 2664, 2690, 2721, 4117, 4161, 4182, 4184, 4261, 4357, 4369, 4372, 4377, 4390, 4422,
4432, 4437, 4449, 4457, 4485, 4497, 4505, 4629, 4677, 4696, 4774, 5205, 5217, 5225, 5386, 5397,
5409, 5445, 5457, 5460, 5461, 5462, 5465, 5472, 5477, 5525, 5545, 5650, 5668, 5717, 5729, 5769,
5777, 6212, 6234, 6244, 6293, 6424, 6482, 6485, 6502, 6505, 6529, 6538, 6565, 6656, 6682, 6788,
6806, 6820, 8218, 8224, 8226, 8232, 8277, 8326, 8354, 8469, 8521, 8530, 8549, 8596, 8737, 8794,
9221, 9253, 9348, 9369, 9380, 9474, 9557, 9633, 9732, 9753, 9793, 9830, 9862, 9880, 10240, 10272,
10282, 10321, 10406, 10517, 10530, 10566, 10585, 10645, 10896, 16466, 16468, 16473, 16485, 16646, 16660, 16665,
16725, 16793, 16806, 16914, 16969, 16977, 16996, 17028, 17057, 17408, 17416, 17434, 17493, 17512, 17578, 17685,
17696, 17733, 17745, 17748, 17749, 17750, 17753, 17765, 17794, 17813, 17946, 17984, 18005, 18072, 18453, 18529,
18569, 18722, 18756, 18762, 18773, 18794, 18833, 18853, 18945, 19026, 19033, 19077, 20489, 20497, 20500, 20517,
20565, 20586, 20610, 20633, 20757, 20769, 20776, 20805, 20817, 20820, 20821, 20822, 20825, 20837, 20864, 20872,
20885, 20896, 21002, 21029, 21077, 21146, 21510, 21525, 21573, 21585, 21588, 21589, 21590, 21593, 21605, 21653,
21665, 21765, 21777, 21780, 21781, 21782, 21785, 21797, 21825, 21828, 21829, 21830, 21833, 21840, 21841, 21842,
21844, 21846, 21848, 21849, 21850, 21857, 21860, 21861, 21862, 21865, 21893, 21905, 21908, 21909, 21910, 21913,
21925, 22024, 22037, 22085, 22097, 22100, 22101, 22102, 22105, 22117, 22165, 22545, 22566, 22568, 22594, 22608,
22613, 22676, 22697, 22793, 22805, 22853, 22865, 22868, 22869, 22870, 22873, 22885, 22933, 22946, 23046, 23072,
23125, 23209, 24597, 24640, 24665, 24673, 24725, 24833, 24840, 24869, 24917, 24934, 24965, 25001, 25108, 25110,
25152, 25184, 25192, 25234, 25616, 25618, 25625, 25685, 25704, 25738, 25744, 25770, 25877, 25897, 25925, 25937,
25940, 25941, 25942, 25945, 25957, 25986, 26005, 26186, 26197, 26276, 26632, 26634, 26725, 26757, 26770, 26885,
26965, 26976, 26986, 27032, 27153, 27174, 27200, 27208, 27240, 27269, 27282, 27290, 32778, 32800, 32802, 32808,
32810, 32853, 32904, 32922, 32930, 32932, 33105, 33110, 33112, 33125, 33157, 33280, 33288, 33301, 33312, 33320,
33424, 33797, 33829, 33858, 34068, 34133, 34146, 34176, 34217, 34306, 34342, 34441, 34454, 34468, 34832, 34918,
34965, 34984, 35094, 35137, 35161, 35208, 35232, 35332, 35338, 35368, 35429, 36932, 36934, 36953, 37009, 37125,
37136, 37138, 37145, 37157, 37205, 37220, 37258, 37290, 37444, 37446, 37465, 37478, 37525, 37905, 37968, 37973,
38040, 38054, 38145, 38154, 38165, 38180, 38186, 38213, 38225, 38228, 38229, 38230, 38233, 38245, 38293, 38485,
38504, 38530, 38938, 38985, 38993, 39012, 39040, 39173, 39192, 39253, 39265, 39301, 39316, 39322, 39442, 39497,
39504, 39590, 40970, 40984, 40992, 41002, 41045, 41120, 41128, 41237, 41289, 41297, 41317, 41364, 41366, 41514,
41557, 41633, 41989, 42021, 42056, 42068, 42074, 42113, 42242, 42265, 42274, 42325, 42340, 42402, 42501, 42512,
42533, 42624, 42632, 42666, 43040, 43093, 43106, 43168, 43176, 43264, 43286, 43345, 43429, 43590, 43618, 43680,
};
const int kmap_size = 43692;
const int nwant = 2;
const uint16_t * kgrid = grid_size == 256 ? kgrid_256 : kgrid_512;
const int nwant = type == GGML_TYPE_IQ1_S ? 3 : 2;
const uint16_t * kgrid = type == GGML_TYPE_IQ2_XXS ? kgrid_2bit_256 :
type == GGML_TYPE_IQ2_XS ? kgrid_2bit_512 : kgrid_1bit_512;
uint64_t * kgrid_q2xs;
int * kmap_q2xs;
uint16_t * kneighbors_q2xs;
@ -9286,9 +9679,9 @@ void iq2xs_init_impl(int grid_size) {
free(dist2);
}
void iq2xs_free_impl(int grid_size) {
GGML_ASSERT(grid_size == 256 || grid_size == 512 || grid_size == 1024);
const int gindex = iq2_data_index(grid_size);
void iq2xs_free_impl(enum ggml_type type) {
GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S);
const int gindex = iq2_data_index(type);
if (iq2_data[gindex].grid) {
free(iq2_data[gindex].grid); iq2_data[gindex].grid = NULL;
free(iq2_data[gindex].map); iq2_data[gindex].map = NULL;
@ -9322,7 +9715,7 @@ static int iq2_find_best_neighbour(const uint16_t * restrict neighbours, const u
static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict vy, int n, const float * restrict quant_weights) {
const int gindex = iq2_data_index(256);
const int gindex = iq2_data_index(GGML_TYPE_IQ2_XXS);
const uint64_t * kgrid_q2xs = iq2_data[gindex].grid;
const int * kmap_q2xs = iq2_data[gindex].map;
@ -9495,7 +9888,7 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict
static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict vy, int n, const float * restrict quant_weights) {
const int gindex = iq2_data_index(512);
const int gindex = iq2_data_index(GGML_TYPE_IQ2_XS);
const uint64_t * kgrid_q2xs = iq2_data[gindex].grid;
const int * kmap_q2xs = iq2_data[gindex].map;
@ -10132,3 +10525,207 @@ void quantize_row_iq3_xxs_reference(const float * restrict x, block_iq3_xxs * re
assert(k % QK_K == 0);
quantize_row_iq3_xxs_impl(x, y, k, NULL);
}
// =================================== 1.5 bpw ===================================================
static int iq1_find_best_neighbour(const uint16_t * restrict neighbours, const uint64_t * restrict grid,
const float * restrict xval, const float * restrict weight, float * scale, int8_t * restrict L, int ngrid) {
int num_neighbors = neighbours[0];
GGML_ASSERT(num_neighbors > 0);
float best_score = 0;
int grid_index = -1;
for (int j = 1; j <= num_neighbors; ++j) {
const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
float sumqx = 0, sumq2 = 0;
for (int i = 0; i < 8; ++i) {
float q = (pg[i] - 3)/2;
float w = weight[i];
sumqx += w*q*xval[i];
sumq2 += w*q*q;
}
if (sumqx > 0 && sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
*scale = sumqx/sumq2; best_score = *scale * sumqx;
grid_index = neighbours[j];
}
}
if (grid_index < 0) {
for (int i = 0; i < ngrid; ++i) {
const int8_t * grid_i = (const int8_t *)(grid + i);
float sumqx = 0, sumq2 = 0;
for (int j = 0; j < 8; ++j) {
float w = weight[j];
float q = (grid_i[j] - 3)/2;
sumqx += w*q*xval[j];
sumq2 += w*q*q;
}
if (sumqx > 0 && sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
*scale = sumqx/sumq2; best_score = *scale*sumqx;
grid_index = i;
}
}
}
if (grid_index < 0) {
printf("Oops, did not find grid point\n");
printf("Have %d neighbours\n", num_neighbors);
for (int j = 1; j <= num_neighbors; ++j) {
const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
float sumqx = 0, sumq2 = 0;
for (int i = 0; i < 8; ++i) {
float q = (pg[i] - 3)/2;
float w = weight[i];
sumqx += w*q*xval[i];
sumq2 += w*q*q;
}
printf(" neighbour %d: sumqx = %g sumq2 = %g\n", j, (double)sumqx, (double)sumq2);
}
}
GGML_ASSERT(grid_index >= 0);
//!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
*scale *= 1.05f; // This is a fudge factor. Don't ask me why it improves the result.
//!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
const int8_t * pg = (const int8_t *)(grid + grid_index);
for (int i = 0; i < 8; ++i) L[i] = (pg[i] - 1)/2;
return grid_index;
}
static int iq1_sort_helper(const void * left, const void * right) {
const float * l = left;
const float * r = right;
return *l < *r ? -1 : *l > *r ? 1 : 0;
}
static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy, int n, const float * restrict quant_weights) {
const int gindex = iq2_data_index(GGML_TYPE_IQ1_S);
const uint64_t * kgrid_q2xs = iq2_data[gindex].grid;
const int * kmap_q2xs = iq2_data[gindex].map;
const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
GGML_ASSERT(quant_weights && "missing quantization weights");
GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?");
GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?");
GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
GGML_ASSERT(n%QK_K == 0);
const int nbl = n/256;
block_iq1_s * y = vy;
float scales[QK_K/8];
float weight[8];
int8_t L[8];
float sumx[9];
float sumw[9];
float pairs[16];
int * idx = (int *)(pairs + 1);
uint8_t hbit[QK_K/8];
for (int ibl = 0; ibl < nbl; ++ibl) {
y[ibl].d = GGML_FP32_TO_FP16(0.f);
memset(y[ibl].qs, 0, QK_K/8);
memset(y[ibl].scales, 0, QK_K/16);
float max_scale = 0;
const float * xbl = x + QK_K*ibl;
float sumx2 = 0;
for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
float sigma2 = sumx2/QK_K;
for (int ib = 0; ib < QK_K/8; ++ib) {
const float * xb = xbl + 8*ib;
const float * qw = quant_weights + QK_K*ibl + 8*ib;
for (int i = 0; i < 8; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
float max = fabsf(xb[0]);
for (int i = 1; i < 8; ++i) max = MAX(max, fabsf(xb[i]));
if (!max) {
scales[ib] = 0;
memset(L, 1, 8);
continue;
}
// Here we solve exactly the sum of squared difference (SSD) weighted minimization problem.
// With just 3 allowed quant values (-1, 0, 1), we can search exhaustively for the two
// boundaries that split the weights xb[i] into 3 groups. To do so, we sort the weights
// in ascending order, compute Si = sum[weight[j] xb[j], j = 0...i] and
// Wi = sum[weight[j], j = 0...i], and use these to quckly get get the optimum scale
// for each possible and score for each split.
for (int j = 0; j < 8; ++j) {
pairs[2*j] = xb[j];
idx[2*j] = j;
}
qsort(pairs, 8, 2*sizeof(float), iq1_sort_helper);
{
sumx[0] = sumw[0] = 0;
for (int j = 0; j < 8; ++j) {
int i = idx[2*j];
sumx[j+1] = sumx[j] + weight[i]*xb[i];
sumw[j+1] = sumw[j] + weight[i];
}
}
float best_score = 0, scale = max;
int besti1 = 0, besti2 = 0;
for (int i1 = 0; i1 <= 8; ++i1) {
for (int i2 = i1; i2 <= 8; ++i2) {
float sumqx = -(sumx[i1] - sumx[0]) + (sumx[8] - sumx[i2]);
float sumq2 = (sumw[i1] - sumw[0]) + (sumw[8] - sumw[i2]);
if (sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
scale = sumqx/sumq2; best_score = scale*sumqx;
besti1 = i1; besti2 = i2;
}
}
}
for (int j = 0; j < besti1; ++j) L[idx[2*j]] = 0;
for (int j = besti1; j < besti2; ++j) L[idx[2*j]] = 1;
for (int j = besti2; j < 8; ++j) L[idx[2*j]] = 2;
if (scale < 0) {
for (int j = 0; j < 8; ++j) L[j] = 2 - L[j];
scale = -scale;
}
// Now we check if the solution found above corresponds to a grid point and, if not, use a neighbouring
// grid point that minimizes SSD.
uint16_t u = 0;
for (int j = 0; j < 8; ++j) u |= (L[j] << 2*j);
int grid_index = kmap_q2xs[u];
if (grid_index < 0) {
const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
grid_index = iq1_find_best_neighbour(neighbours, kgrid_q2xs, xb, weight, &scale, L, NGRID_IQ2XXS);
GGML_ASSERT(grid_index >= 0);
}
y[ibl].qs[ib] = grid_index & 255;
hbit[ib] = grid_index >> 8;
GGML_ASSERT(scale >= 0);
scales[ib] = scale;
max_scale = MAX(max_scale, scale);
}
if (!max_scale) {
memset(y[ibl].qs, 0, QK_K/8);
continue;
}
float d = max_scale/15;
y[ibl].d = GGML_FP32_TO_FP16(d*1.085f); // 1.085f is another fudge factor. Don't ask me why it is needed.
float id = 1/d;
for (int ib = 0; ib < QK_K/8; ++ib) {
int l = nearest_int(0.5f*(id*scales[ib]-1));
l = MAX(0, MIN(7, l));
if (hbit[ib]) l |= 8;
y[ibl].scales[ib/2] |= (l << 4*(ib%2));
}
}
}
size_t quantize_iq1_s(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
char * qrow = (char *)dst;
for (int row = 0; row < nrow; ++row) {
quantize_row_iq1_s_impl(src, qrow, n_per_row, quant_weights);
src += n_per_row;
qrow += nblock*sizeof(block_iq1_s);
}
return nrow * nblock * sizeof(block_iq1_s);
}

14
ggml-quants.h
View File

@ -191,6 +191,13 @@ typedef struct {
} block_iq3_xxs;
static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
typedef struct {
ggml_fp16_t d;
uint8_t qs[QK_K/8];
uint8_t scales[QK_K/16];
} block_iq1_s;
static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
#ifdef __cplusplus
extern "C" {
#endif
@ -243,6 +250,7 @@ void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRI
void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq2_xs (const block_iq2_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq1_s (const block_iq1_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
// Dot product
void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@ -259,6 +267,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
//
// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
@ -266,6 +275,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
size_t quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq3_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq1_s (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q2_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q3_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q4_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
@ -276,8 +286,8 @@ size_t quantize_q4_1 (const float * src, void * dst, int nrows, int n_per_row,
size_t quantize_q5_0 (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q5_1 (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
void iq2xs_init_impl(int grid_size);
void iq2xs_free_impl(int grid_size);
void iq2xs_init_impl(enum ggml_type type);
void iq2xs_free_impl(enum ggml_type type);
void iq3xs_init_impl(int grid_size);
void iq3xs_free_impl(int grid_size);

77
ggml-vulkan.cpp
View File

@ -707,9 +707,21 @@ static void ggml_vk_queue_cleanup(ggml_backend_vk_context * ctx, vk_queue& q) {
q.cmd_buffer_idx = 0;
}
static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags) {
static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) {
for (uint32_t i = 0; i < mem_props->memoryTypeCount; ++i) {
vk::MemoryType memory_type = mem_props->memoryTypes[i];
if ((mem_req->memoryTypeBits & ((uint64_t)1 << i)) &&
(flags & memory_type.propertyFlags) == flags &&
mem_props->memoryHeaps[memory_type.heapIndex].size >= mem_req->size) {
return static_cast<int32_t>(i);
}
}
return UINT32_MAX;
}
static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_create_buffer(" << size << ", " << to_string(req_flags) << ")" << std::endl;
std::cerr << "ggml_vk_create_buffer(" << size << ", " << to_string(req_flags) << ", " << to_string(fallback_flags) << ")" << std::endl;
#endif
vk_buffer buf = std::make_shared<vk_buffer_struct>();
@ -736,15 +748,15 @@ static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t siz
uint32_t memory_type_index = UINT32_MAX;
for (uint32_t i = 0; i < mem_props.memoryTypeCount; ++i) {
vk::MemoryType memory_type = mem_props.memoryTypes[i];
if ((mem_req.memoryTypeBits & ((uint64_t)1 << i)) && (req_flags & memory_type.propertyFlags) == req_flags && mem_props.memoryHeaps[memory_type.heapIndex].size >= mem_req.size) {
memory_type_index = i;
break;
}
memory_type_index = find_properties(&mem_props, &mem_req, req_flags);
buf->memory_property_flags = req_flags;
if (memory_type_index == UINT32_MAX && fallback_flags) {
memory_type_index = find_properties(&mem_props, &mem_req, fallback_flags);
buf->memory_property_flags = fallback_flags;
}
if (memory_type_index >= mem_props.memoryTypeCount) {
if (memory_type_index == UINT32_MAX) {
ctx->device.lock()->device.destroyBuffer(buf->buffer);
buf->size = 0;
throw vk::OutOfDeviceMemoryError("No suitable memory type found");
@ -758,10 +770,9 @@ static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t siz
buf->size = 0;
throw e;
}
buf->memory_property_flags = req_flags;
buf->ptr = nullptr;
if (req_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
if (buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
buf->ptr = ctx->device.lock()->device.mapMemory(buf->device_memory, 0, VK_WHOLE_SIZE);
}
@ -778,9 +789,9 @@ static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t siz
return buf;
}
static vk_buffer ggml_vk_create_buffer_check(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags) {
static vk_buffer ggml_vk_create_buffer_check(ggml_backend_vk_context * ctx, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) {
try {
return ggml_vk_create_buffer(ctx, size, req_flags);
return ggml_vk_create_buffer(ctx, size, req_flags, fallback_flags);
} catch (const vk::SystemError& e) {
std::cerr << "ggml_vulkan: Memory allocation of size " << size << " failed." << std::endl;
std::cerr << "ggml_vulkan: " << e.what() << std::endl;
@ -791,16 +802,16 @@ static vk_buffer ggml_vk_create_buffer_check(ggml_backend_vk_context * ctx, size
static vk_buffer ggml_vk_create_buffer_device(ggml_backend_vk_context * ctx, size_t size) {
vk_buffer buf;
try {
buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eDeviceLocal);
} catch (const vk::SystemError& e) {
if (ctx->device.lock()->uma) {
// Fall back to host memory type
buf = ggml_vk_create_buffer_check(ctx, size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
} else {
std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
std::cerr << "ggml_vulkan: " << e.what() << std::endl;
throw e;
buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eDeviceLocal);
}
} catch (const vk::SystemError& e) {
std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
std::cerr << "ggml_vulkan: " << e.what() << std::endl;
throw e;
}
return buf;
@ -1080,7 +1091,7 @@ static void ggml_vk_print_gpu_info(size_t idx) {
}
}
void ggml_vk_instance_init() {
static void ggml_vk_instance_init() {
if (vk_instance_initialized) {
return;
}
@ -1139,7 +1150,7 @@ void ggml_vk_instance_init() {
vk_instance_initialized = true;
}
void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
GGML_ASSERT(idx < vk_instance.device_indices.size());
size_t dev_num = vk_instance.device_indices[idx];
#ifdef GGML_VULKAN_DEBUG
@ -1422,7 +1433,9 @@ static void * ggml_vk_host_malloc(ggml_backend_vk_context * ctx, size_t size) {
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_host_malloc(" << size << ")" << std::endl;
#endif
vk_buffer buf = ggml_vk_create_buffer(ctx, size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
vk_buffer buf = ggml_vk_create_buffer(ctx, size,
vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
if(!(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible)) {
fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory\n",
@ -1568,7 +1581,9 @@ static void deferred_memcpy(void * dst, const void * src, size_t size, std::vect
static void ggml_vk_ensure_sync_staging_buffer(ggml_backend_vk_context * ctx, size_t size) {
if (ctx->sync_staging == nullptr || ctx->sync_staging->size < size) {
ggml_vk_destroy_buffer(ctx->sync_staging);
ctx->sync_staging = ggml_vk_create_buffer_check(ctx, size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
ctx->sync_staging = ggml_vk_create_buffer_check(ctx, size,
vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
}
}
@ -4082,7 +4097,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
std::cerr << "ggml_vk_preallocate_buffers(qx_size: " << ctx->prealloc_size_qx << " qy_size: " << ctx->prealloc_size_qy << " x_size: " << ctx->prealloc_size_x << " y_size: " << ctx->prealloc_size_y << " split_k_size: " << ctx->prealloc_size_split_k << ")" << std::endl;
#endif
#if defined(GGML_VULKAN_RUN_TESTS)
ctx->staging = ggml_vk_create_buffer_check(ctx, 100ul * 1024ul * 1024ul, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
ctx->staging = ggml_vk_create_buffer_check(ctx, 100ul * 1024ul * 1024ul,
vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached
vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
ggml_vk_test_transfer(ctx, 8192 * 1000, false);
ggml_vk_test_transfer(ctx, 8192 * 1000, true);
@ -4174,7 +4191,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
if (ctx->staging != nullptr) {
ggml_vk_destroy_buffer(ctx->staging);
}
ctx->staging = ggml_vk_create_buffer_check(ctx, ctx->staging_size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);
ctx->staging = ggml_vk_create_buffer_check(ctx, ctx->staging_size,
vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
}
}
@ -4537,13 +4556,13 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
}
}
GGML_CALL int ggml_vk_get_device_count() {
GGML_CALL static int ggml_vk_get_device_count() {
ggml_vk_instance_init();
return vk_instance.device_indices.size();
}
GGML_CALL void ggml_vk_get_device_description(int device, char * description, size_t description_size) {
GGML_CALL static void ggml_vk_get_device_description(int device, char * description, size_t description_size) {
ggml_vk_instance_init();
std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
@ -4561,7 +4580,7 @@ void ggml_vk_init_cpu_assist() {
std::cerr << "ggml_vulkan: Found " << ggml_vk_get_device_count() << " Vulkan devices:" << std::endl;
for (size_t i = 0; i < ggml_vk_get_device_count(); i++) {
for (int i = 0; i < ggml_vk_get_device_count(); i++) {
ggml_vk_print_gpu_info(i);
}
// Initialize the first backend to make sure CPU matrix multiplications can be offloaded.
@ -5248,7 +5267,7 @@ GGML_CALL void ggml_backend_vk_get_device_description(int device, char * descrip
}
GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
GGML_ASSERT(device < vk_instance.device_indices.size());
GGML_ASSERT(device < (int) vk_instance.device_indices.size());
vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];

276
ggml.c
View File

@ -23,6 +23,9 @@
#include <limits.h>
#include <stdarg.h>
#include <signal.h>
#if defined(__gnu_linux__)
#include <syscall.h>
#endif
#ifdef GGML_USE_METAL
#include <unistd.h>
@ -673,6 +676,18 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ1_S] = {
.type_name = "iq1_s",
.blck_size = QK_K,
.type_size = sizeof(block_iq1_s),
.is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_iq1_s,
.from_float = NULL,
.from_float_reference = NULL,
.vec_dot = ggml_vec_dot_iq1_s_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_Q8_K] = {
.type_name = "q8_K",
.blck_size = QK_K,
@ -868,7 +883,7 @@ do { \
const __m128 t0 = _mm_add_ps(_mm256_castps256_ps128(x[0]), \
_mm256_extractf128_ps(x[0], 1)); \
const __m128 t1 = _mm_hadd_ps(t0, t0); \
res = _mm_cvtss_f32(_mm_hadd_ps(t1, t1)); \
res = (ggml_float) _mm_cvtss_f32(_mm_hadd_ps(t1, t1)); \
} while (0)
// TODO: is this optimal ?
@ -1149,7 +1164,7 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
x[i] = _mm_add_ps(x[i], x[offset+i]); \
} \
const __m128 t0 = _mm_hadd_ps(x[0], x[0]); \
res = _mm_cvtss_f32(_mm_hadd_ps(t0, t0)); \
res = (ggml_float) _mm_cvtss_f32(_mm_hadd_ps(t0, t0)); \
}
// TODO: is this optimal ?
@ -2016,9 +2031,16 @@ struct ggml_numa_node {
};
struct ggml_numa_nodes {
enum ggml_numa_strategy numa_strategy;
struct ggml_numa_node nodes[GGML_NUMA_MAX_NODES];
uint32_t n_nodes;
uint32_t total_cpus; // hardware threads on system
uint32_t current_node; // node on which main process is execting
#if defined(__gnu_linux__)
cpu_set_t cpuset; // cpuset from numactl
#else
uint32_t cpuset; // no NUMA support outside of Linux at this time. Use a portable datatype
#endif
};
//
@ -2052,18 +2074,40 @@ inline static void ggml_critical_section_end(void) {
atomic_fetch_sub(&g_state_barrier, 1);
}
void ggml_numa_init(void) {
#if defined(__gnu_linux__)
static cpu_set_t ggml_get_numa_affinity(void) {
cpu_set_t cpuset;
pthread_t thread;
thread = pthread_self();
CPU_ZERO(&cpuset);
pthread_getaffinity_np(thread, sizeof(cpu_set_t), &cpuset);
return cpuset;
}
#else
static uint32_t ggml_get_numa_affinity(void) {
return 0; // no NUMA support
}
#endif
void ggml_numa_init(enum ggml_numa_strategy numa_flag) {
if (g_state.numa.n_nodes > 0) {
fprintf(stderr, "ggml_numa_init: NUMA already initialized\n");
return;
}
#ifdef __linux__
#if defined(__gnu_linux__)
struct stat st;
char path[256];
int rv;
// set numa scheme
g_state.numa.numa_strategy = numa_flag;
GGML_PRINT_DEBUG("numa strategy %u\n",g_state.numa.numa_strategy);
g_state.numa.cpuset = ggml_get_numa_affinity();
// enumerate nodes
while (g_state.numa.n_nodes < GGML_NUMA_MAX_NODES) {
rv = snprintf(path, sizeof(path), "/sys/devices/system/node/node%u", g_state.numa.n_nodes);
@ -2082,11 +2126,23 @@ void ggml_numa_init(void) {
GGML_PRINT_DEBUG("found %u numa nodes, %u CPUs\n", g_state.numa.n_nodes, g_state.numa.total_cpus);
if (g_state.numa.n_nodes < 1 || g_state.numa.total_cpus < 1) {
// figure out which node we're on
uint current_cpu;
int getcpu_ret = 0;
#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 28)
getcpu_ret = getcpu(&current_cpu, &g_state.numa.current_node);
#else
// old glibc doesn't have a wrapper for this call. Fall back on direct syscall
getcpu_ret = syscall(SYS_getcpu,&current_cpu,&g_state.numa.current_node);
#endif
if (g_state.numa.n_nodes < 1 || g_state.numa.total_cpus < 1 || getcpu_ret != 0) {
g_state.numa.n_nodes = 0;
return;
}
GGML_PRINT_DEBUG("found our process on numa node %u, CPU %u\n", g_state.numa.current_node, current_cpu);
for (uint32_t n = 0; n < g_state.numa.n_nodes; ++n) {
struct ggml_numa_node * node = &g_state.numa.nodes[n];
GGML_PRINT_DEBUG("CPUs on node %u:", n);
@ -2113,6 +2169,7 @@ void ggml_numa_init(void) {
}
}
#else
GGML_UNUSED(numa_flag);
// TODO
#endif
}
@ -2293,6 +2350,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
case GGML_FTYPE_MOSTLY_IQ2_XXS: wtype = GGML_TYPE_IQ2_XXS; break;
case GGML_FTYPE_MOSTLY_IQ2_XS: wtype = GGML_TYPE_IQ2_XS; break;
case GGML_FTYPE_MOSTLY_IQ3_XXS: wtype = GGML_TYPE_IQ3_XXS; break;
case GGML_FTYPE_MOSTLY_IQ1_S: wtype = GGML_TYPE_IQ1_S; break;
case GGML_FTYPE_UNKNOWN: wtype = GGML_TYPE_COUNT; break;
case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
}
@ -3246,7 +3304,7 @@ const char * ggml_get_name(const struct ggml_tensor * tensor) {
}
struct ggml_tensor * ggml_set_name(struct ggml_tensor * tensor, const char * name) {
strncpy(tensor->name, name, sizeof(tensor->name));
strncpy(tensor->name, name, sizeof(tensor->name) - 1);
tensor->name[sizeof(tensor->name) - 1] = '\0';
return tensor;
}
@ -5124,15 +5182,27 @@ static struct ggml_tensor * ggml_soft_max_impl(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * mask,
struct ggml_tensor * pos,
float scale,
float max_bias,
bool inplace) {
GGML_ASSERT(ggml_is_contiguous(a));
if (mask) {
GGML_ASSERT(mask->type == GGML_TYPE_F16);
GGML_ASSERT(ggml_is_contiguous(mask));
GGML_ASSERT(mask->ne[2] == 1);
GGML_ASSERT(mask->ne[3] == 1);
GGML_ASSERT(mask->ne[1] >= a->ne[1]);
GGML_ASSERT(ggml_is_matrix(mask));
GGML_ASSERT(ggml_can_repeat_rows(mask, a));
}
if (pos) {
GGML_ASSERT(ggml_is_vector(pos));
GGML_ASSERT(pos->type == GGML_TYPE_F16);
GGML_ASSERT(pos->ne[0] == a->ne[0]);
}
if (max_bias > 0.0f) {
GGML_ASSERT(pos);
}
bool is_node = false;
@ -5143,13 +5213,14 @@ static struct ggml_tensor * ggml_soft_max_impl(
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
float params[] = { scale };
float params[] = { scale, max_bias };
ggml_set_op_params(result, params, sizeof(params));
result->op = GGML_OP_SOFT_MAX;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src[0] = a;
result->src[1] = mask;
result->src[2] = pos;
return result;
}
@ -5157,21 +5228,23 @@ static struct ggml_tensor * ggml_soft_max_impl(
struct ggml_tensor * ggml_soft_max(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_soft_max_impl(ctx, a, NULL, 1.0f, false);
return ggml_soft_max_impl(ctx, a, NULL, NULL, 1.0f, 0.0f, false);
}
struct ggml_tensor * ggml_soft_max_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_soft_max_impl(ctx, a, NULL, 1.0f, true);
return ggml_soft_max_impl(ctx, a, NULL, NULL, 1.0f, 0.0f, true);
}
struct ggml_tensor * ggml_soft_max_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * mask,
float scale) {
return ggml_soft_max_impl(ctx, a, mask, scale, false);
struct ggml_tensor * pos,
float scale,
float max_bias) {
return ggml_soft_max_impl(ctx, a, mask, pos, scale, max_bias, false);
}
// ggml_soft_max_back
@ -7744,6 +7817,7 @@ static void ggml_compute_forward_add(
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
{
ggml_compute_forward_add_q_f32(params, src0, src1, dst);
} break;
@ -8011,6 +8085,7 @@ static void ggml_compute_forward_add1(
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
{
ggml_compute_forward_add1_q_f32(params, src0, src1, dst);
} break;
@ -8131,6 +8206,7 @@ static void ggml_compute_forward_acc(
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
default:
{
GGML_ASSERT(false);
@ -10897,6 +10973,7 @@ static void ggml_compute_forward_out_prod(
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
{
ggml_compute_forward_out_prod_q_f32(params, src0, src1, dst);
} break;
@ -11077,6 +11154,7 @@ static void ggml_compute_forward_set(
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
default:
{
GGML_ASSERT(false);
@ -11274,6 +11352,7 @@ static void ggml_compute_forward_get_rows(
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
{
ggml_compute_forward_get_rows_q(params, src0, src1, dst);
} break;
@ -11577,6 +11656,7 @@ static void ggml_compute_forward_soft_max_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
const struct ggml_tensor * src2,
struct ggml_tensor * dst) {
assert(ggml_is_contiguous(dst));
assert(ggml_are_same_shape(src0, dst));
@ -11585,16 +11665,29 @@ static void ggml_compute_forward_soft_max_f32(
return;
}
float scale = 1.0f;
memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
float scale = 1.0f;
float max_bias = 0.0f;
memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
// TODO: handle transposed/permuted matrices
const int ith = params->ith;
const int nth = params->nth;
GGML_TENSOR_UNARY_OP_LOCALS
const int64_t ne11 = src1 ? src1->ne[1] : 1;
// TODO: is this supposed to be ceil instead of floor?
// https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370
const uint32_t n_head_kv = ne02;
const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head_kv));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
const int nc = src0->ne[0];
const int nr = ggml_nrows(src0);
@ -11607,6 +11700,9 @@ static void ggml_compute_forward_soft_max_f32(
float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith;
// when max_bias <= 0.0f, src2 is not used and we default it to src0 to avoid branching
float * pos = src2 ? (float *) src2->data : src0->data;
for (int i1 = ir0; i1 < ir1; i1++) {
float * sp = (float *)((char *) src0->data + i1*src0->nb[1]);
float * dp = (float *)((char *) dst->data + i1*dst->nb[1]);
@ -11622,6 +11718,16 @@ static void ggml_compute_forward_soft_max_f32(
}
}
// ALiBi bias
if (max_bias > 0.0f) {
const uint32_t h = (i1/ne01)%ne02; // head
const float slope = h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1);
for (int i = 0; i < nc; i++) {
wp[i] = wp[i] + slope*pos[i];
}
}
#ifndef NDEBUG
for (int i = 0; i < nc; ++i) {
//printf("p[%d] = %f\n", i, p[i]);
@ -11666,11 +11772,12 @@ static void ggml_compute_forward_soft_max(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
const struct ggml_tensor * src2,
struct ggml_tensor * dst) {
switch (src0->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_soft_max_f32(params, src0, src1, dst);
ggml_compute_forward_soft_max_f32(params, src0, src1, src2, dst);
} break;
default:
{
@ -11814,22 +11921,20 @@ static void ggml_compute_forward_alibi_f32(
const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
for (int64_t i = 0; i < ne0; i++) {
for (int64_t j = 0; j < ne1; j++) {
for (int64_t k = 0; k < ne2_ne3; k++) {
for (int64_t k = 0; k < ne2_ne3; k++) {
// TODO: k*nb2 or k*nb3
float m_k;
if (k < n_heads_log2_floor) {
m_k = powf(m0, k + 1);
} else {
m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
}
for (int64_t i = 0; i < ne0; i++) {
for (int64_t j = 0; j < ne1; j++) {
float * const src = (float *)((char *) src0->data + i*nb0 + j*nb1 + k*nb2);
float * pdst = (float *)((char *) dst->data + i*nb0 + j*nb1 + k*nb2);
// TODO: k*nb2 or k*nb3
float m_k;
if (k < n_heads_log2_floor) {
m_k = powf(m0, k + 1);
} else {
m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
}
pdst[0] = i * m_k + src[0];
}
}
@ -11874,21 +11979,20 @@ static void ggml_compute_forward_alibi_f16(
const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
for (int i = 0; i < ne0; i++) {
for (int j = 0; j < ne1; j++) {
for (int k = 0; k < ne2_ne3; k++) {
for (int k = 0; k < ne2_ne3; k++) {
// TODO: k*nb2 or k*nb3
float m_k;
if (k < n_heads_log2_floor) {
m_k = powf(m0, k + 1);
} else {
m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
}
for (int i = 0; i < ne0; i++) {
for (int j = 0; j < ne1; j++) {
ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i*nb0 + j*nb1 + k*nb2);
float * pdst = (float *)((char *) dst->data + i*nb0 + j*nb1 + k*nb2);
// TODO: k*nb2 or k*nb3
float m_k;
if (k < n_heads_log2_floor) {
m_k = powf(m0, k + 1);
} else {
m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
}
float * pdst = (float *)((char *) dst->data + i*nb0 + j*nb1 + k*nb2);
// we return F32
pdst[0] = i * m_k + GGML_FP16_TO_FP32(src[0]);
@ -11924,6 +12028,7 @@ static void ggml_compute_forward_alibi(
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_Q8_K:
case GGML_TYPE_I8:
case GGML_TYPE_I16:
@ -12001,6 +12106,7 @@ static void ggml_compute_forward_clamp(
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ3_XXS:
case GGML_TYPE_IQ1_S:
case GGML_TYPE_Q8_K:
case GGML_TYPE_I8:
case GGML_TYPE_I16:
@ -15391,7 +15497,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
} break;
case GGML_OP_SOFT_MAX:
{
ggml_compute_forward_soft_max(params, tensor->src[0], tensor->src[1], tensor);
ggml_compute_forward_soft_max(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
} break;
case GGML_OP_SOFT_MAX_BACK:
{
@ -16953,27 +17059,47 @@ typedef pthread_t ggml_thread_t;
#endif
// Android's libc implementation "bionic" does not support setting affinity
#if defined(__linux__) && !defined(__BIONIC__)
static void set_numa_thread_affinity(int thread_n, int n_threads) {
#if defined(__gnu_linux__)
static void set_numa_thread_affinity(int thread_n) {
if (!ggml_is_numa()) {
return;
}
// run thread on node_num thread_n / (threads per node)
const int node_num = thread_n / ((n_threads + g_state.numa.n_nodes - 1) / g_state.numa.n_nodes);
struct ggml_numa_node * node = &g_state.numa.nodes[node_num];
int node_num;
int rv;
size_t setsize = CPU_ALLOC_SIZE(g_state.numa.total_cpus);
switch(g_state.numa.numa_strategy) {
case GGML_NUMA_STRATEGY_DISTRIBUTE:
// run thread on node_num thread_n / (threads per node)
node_num = thread_n % g_state.numa.n_nodes;
break;
case GGML_NUMA_STRATEGY_ISOLATE:
// run thread on current_node
node_num = g_state.numa.current_node;
break;
case GGML_NUMA_STRATEGY_NUMACTL:
// use the cpuset that numactl gave us
rv = pthread_setaffinity_np(pthread_self(), setsize, &g_state.numa.cpuset);
if (rv) {
fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n",strerror(rv));
}
return;
default:
return;
}
struct ggml_numa_node * node = &g_state.numa.nodes[node_num];
cpu_set_t * cpus = CPU_ALLOC(g_state.numa.total_cpus);
CPU_ZERO_S(setsize, cpus);
for (size_t i = 0; i < node->n_cpus; ++i) {
CPU_SET_S(node->cpus[i], setsize, cpus);
}
int rv = pthread_setaffinity_np(pthread_self(), setsize, cpus);
rv = pthread_setaffinity_np(pthread_self(), setsize, cpus);
if (rv) {
fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n",
strerror(rv));
fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n", strerror(rv));
}
CPU_FREE(cpus);
@ -16994,8 +17120,7 @@ static void clear_numa_thread_affinity(void) {
int rv = pthread_setaffinity_np(pthread_self(), setsize, cpus);
if (rv) {
fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n",
strerror(rv));
fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n", strerror(rv));
}
CPU_FREE(cpus);
@ -17003,7 +17128,7 @@ static void clear_numa_thread_affinity(void) {
#else
// TODO: Windows etc.
// (the linux implementation may also work on BSD, someone should test)
static void set_numa_thread_affinity(int thread_n, int n_threads) { UNUSED(thread_n); UNUSED(n_threads); }
static void set_numa_thread_affinity(int thread_n) { UNUSED(thread_n); }
static void clear_numa_thread_affinity(void) {}
#endif
@ -17304,7 +17429,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
const int n_threads = state->shared->n_threads;
set_numa_thread_affinity(state->ith, n_threads);
set_numa_thread_affinity(state->ith);
int node_n = -1;
int task_phase = GGML_TASK_FINALIZE;
@ -18116,7 +18241,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
ptr += ggml_nbytes(tensor);
fprintf(stderr, "%s: loaded leaf %d: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
fprintf(stderr, "%s: loaded leaf %u: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
}
}
@ -18219,7 +18344,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
result->nodes[i] = tensor;
fprintf(stderr, "%s: loaded node %d: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
fprintf(stderr, "%s: loaded node %u: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
}
}
}
@ -18844,7 +18969,9 @@ static enum ggml_opt_result linesearch_backtracking(
(*step) *= width;
}
GGML_UNREACHABLE();
GGML_ASSERT(false && "line search failed");
return GGML_LINESEARCH_FAIL;
}
static enum ggml_opt_result ggml_opt_lbfgs(
@ -19112,7 +19239,9 @@ static enum ggml_opt_result ggml_opt_lbfgs(
step[0] = 1.0;
}
GGML_UNREACHABLE();
GGML_ASSERT(false && "lbfgs failed");
return GGML_OPT_DID_NOT_CONVERGE;
}
struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) {
@ -19360,8 +19489,9 @@ void ggml_quantize_init(enum ggml_type type) {
ggml_critical_section_start();
switch (type) {
case GGML_TYPE_IQ2_XXS: iq2xs_init_impl(256); break;
case GGML_TYPE_IQ2_XS: iq2xs_init_impl(512); break;
case GGML_TYPE_IQ2_XXS:
case GGML_TYPE_IQ2_XS:
case GGML_TYPE_IQ1_S: iq2xs_init_impl(type); break;
case GGML_TYPE_IQ3_XXS: iq3xs_init_impl(256); break;
default: // nothing
break;
@ -19373,8 +19503,10 @@ void ggml_quantize_init(enum ggml_type type) {
void ggml_quantize_free(void) {
ggml_critical_section_start();
iq2xs_free_impl(256);
iq2xs_free_impl(512);
iq2xs_free_impl(GGML_TYPE_IQ2_XXS);
iq2xs_free_impl(GGML_TYPE_IQ2_XS);
iq2xs_free_impl(GGML_TYPE_IQ1_S);
iq3xs_free_impl(256);
ggml_critical_section_end();
}
@ -19509,7 +19641,8 @@ size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t *
bool ggml_quantize_requires_imatrix(enum ggml_type type) {
return
type == GGML_TYPE_IQ2_XXS ||
type == GGML_TYPE_IQ2_XS;
type == GGML_TYPE_IQ2_XS ||
type == GGML_TYPE_IQ1_S;
}
size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start,
@ -19634,6 +19767,15 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
result = quantize_iq3_xxs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_IQ1_S:
{
GGML_ASSERT(start % QK_K == 0);
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_iq1_s(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_F16:
{
size_t elemsize = sizeof(ggml_fp16_t);

27
ggml.h
View File

@ -354,6 +354,7 @@ extern "C" {
GGML_TYPE_IQ2_XXS = 16,
GGML_TYPE_IQ2_XS = 17,
GGML_TYPE_IQ3_XXS = 18,
GGML_TYPE_IQ1_S = 19,
GGML_TYPE_I8,
GGML_TYPE_I16,
GGML_TYPE_I32,
@ -391,6 +392,7 @@ extern "C" {
GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ2_XS = 16, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ1_S = 18, // except 1d tensors
};
// available tensor operations:
@ -659,6 +661,16 @@ extern "C" {
void * wdata;
};
// numa strategies
enum ggml_numa_strategy {
GGML_NUMA_STRATEGY_DISABLED = 0,
GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
GGML_NUMA_STRATEGY_ISOLATE = 2,
GGML_NUMA_STRATEGY_NUMACTL = 3,
GGML_NUMA_STRATEGY_MIRROR = 4,
GGML_NUMA_STRATEGY_COUNT
};
// misc
GGML_API void ggml_time_init(void); // call this once at the beginning of the program
@ -669,7 +681,7 @@ extern "C" {
GGML_API void ggml_print_backtrace(void);
GGML_API void ggml_numa_init(void); // call once for better performance on NUMA systems
GGML_API void ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
GGML_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node
GGML_API void ggml_print_object (const struct ggml_object * obj);
@ -1374,13 +1386,17 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
// fused soft_max(a*scale + mask)
// fused soft_max(a*scale + mask + pos[i]*(ALiBi slope))
// mask is optional
// pos is required when max_bias > 0.0f
// max_bias = 0.0f for no ALiBi
GGML_API struct ggml_tensor * ggml_soft_max_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * mask,
float scale);
struct ggml_tensor * pos,
float scale,
float max_bias);
GGML_API struct ggml_tensor * ggml_soft_max_back(
struct ggml_context * ctx,
@ -1482,12 +1498,13 @@ extern "C" {
// alibi position embedding
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_alibi(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_alibi(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past,
int n_head,
float bias_max);
float bias_max),
"use ggml_soft_max_ext instead (will be removed in Mar 2024)");
// clamp
// in-place, returns view(a)

45
gguf-py/examples/reader.py Normal file
View File

@ -0,0 +1,45 @@
#!/usr/bin/env python3
import sys
from pathlib import Path
from gguf.gguf_reader import GGUFReader
sys.path.insert(0, str(Path(__file__).parent.parent))
def read_gguf_file(gguf_file_path):
"""
Reads and prints key-value pairs and tensor information from a GGUF file in an improved format.
Parameters:
- gguf_file_path: Path to the GGUF file.
"""
reader = GGUFReader(gguf_file_path)
# List all key-value pairs in a columnized format
print("Key-Value Pairs:")
max_key_length = max(len(key) for key in reader.fields.keys())
for key, field in reader.fields.items():
value = field.parts[field.data[0]]
print(f"{key:{max_key_length}} : {value}")
print("----")
# List all tensors
print("Tensors:")
tensor_info_format = "{:<30} | Shape: {:<15} | Size: {:<12} | Quantization: {}"
print(tensor_info_format.format("Tensor Name", "Shape", "Size", "Quantization"))
print("-" * 80)
for tensor in reader.tensors:
shape_str = "x".join(map(str, tensor.shape))
size_str = str(tensor.n_elements)
quantization_str = tensor.tensor_type.name
print(tensor_info_format.format(tensor.name, shape_str, size_str, quantization_str))
if __name__ == '__main__':
if len(sys.argv) < 2:
print("Usage: reader.py <path_to_gguf_file>")
sys.exit(1)
gguf_file_path = sys.argv[1]
read_gguf_file(gguf_file_path)

67
gguf-py/gguf/constants.py
View File

@ -40,6 +40,7 @@ class Keys:
TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout"
EXPERT_COUNT = "{arch}.expert_count"
EXPERT_USED_COUNT = "{arch}.expert_used_count"
POOLING_TYPE = "{arch}.pooling_type"
class Attention:
HEAD_COUNT = "{arch}.attention.head_count"
@ -72,6 +73,8 @@ class Keys:
UNK_ID = "tokenizer.ggml.unknown_token_id"
SEP_ID = "tokenizer.ggml.seperator_token_id"
PAD_ID = "tokenizer.ggml.padding_token_id"
CLS_ID = "tokenizer.ggml.cls_token_id"
MASK_ID = "tokenizer.ggml.mask_token_id"
ADD_BOS = "tokenizer.ggml.add_bos_token"
ADD_EOS = "tokenizer.ggml.add_eos_token"
ADD_PREFIX = "tokenizer.ggml.add_space_prefix"
@ -86,27 +89,28 @@ class Keys:
class MODEL_ARCH(IntEnum):
LLAMA = auto()
FALCON = auto()
BAICHUAN = auto()
GPT2 = auto()
GPTJ = auto()
GPTNEOX = auto()
MPT = auto()
STARCODER = auto()
PERSIMMON = auto()
REFACT = auto()
BERT = auto()
BLOOM = auto()
STABLELM = auto()
QWEN = auto()
QWEN2 = auto()
PHI2 = auto()
PLAMO = auto()
CODESHELL = auto()
ORION = auto()
LLAMA = auto()
FALCON = auto()
BAICHUAN = auto()
GPT2 = auto()
GPTJ = auto()
GPTNEOX = auto()
MPT = auto()
STARCODER = auto()
PERSIMMON = auto()
REFACT = auto()
BERT = auto()
NOMIC_BERT = auto()
BLOOM = auto()
STABLELM = auto()
QWEN = auto()
QWEN2 = auto()
PHI2 = auto()
PLAMO = auto()
CODESHELL = auto()
ORION = auto()
INTERNLM2 = auto()
MINICPM = auto()
MINICPM = auto()
class MODEL_TENSOR(IntEnum):
@ -152,6 +156,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.PERSIMMON: "persimmon",
MODEL_ARCH.REFACT: "refact",
MODEL_ARCH.BERT: "bert",
MODEL_ARCH.NOMIC_BERT: "nomic-bert",
MODEL_ARCH.BLOOM: "bloom",
MODEL_ARCH.STABLELM: "stablelm",
MODEL_ARCH.QWEN: "qwen",
@ -281,6 +286,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.LAYER_OUT_NORM,
],
MODEL_ARCH.NOMIC_BERT: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.TOKEN_EMBD_NORM,
MODEL_TENSOR.TOKEN_TYPES,
MODEL_TENSOR.POS_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.ATTN_OUT_NORM,
MODEL_TENSOR.ATTN_QKV,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.LAYER_OUT_NORM,
],
MODEL_ARCH.MPT: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@ -542,6 +561,12 @@ class RopeScalingType(Enum):
YARN = 'yarn'
class PoolingType(IntEnum):
NONE = 0
MEAN = 1
CLS = 2
class GGMLQuantizationType(IntEnum):
F32 = 0
F16 = 1
@ -668,5 +693,7 @@ KEY_TOKENIZER_EOS_ID = Keys.Tokenizer.EOS_ID
KEY_TOKENIZER_UNK_ID = Keys.Tokenizer.UNK_ID
KEY_TOKENIZER_SEP_ID = Keys.Tokenizer.SEP_ID
KEY_TOKENIZER_PAD_ID = Keys.Tokenizer.PAD_ID
KEY_TOKENIZER_CLS_ID = Keys.Tokenizer.CLS_ID
KEY_TOKENIZER_MASK_ID = Keys.Tokenizer.MASK_ID
KEY_TOKENIZER_HF_JSON = Keys.Tokenizer.HF_JSON
KEY_TOKENIZER_RWKV = Keys.Tokenizer.RWKV

10
gguf-py/gguf/gguf_writer.py
View File

@ -19,6 +19,7 @@ from .constants import (
GGUFValueType,
Keys,
RopeScalingType,
PoolingType,
TokenType,
)
@ -360,6 +361,9 @@ class GGUFWriter:
def add_causal_attention(self, value: bool) -> None:
self.add_bool(Keys.Attention.CAUSAL.format(arch=self.arch), value)
def add_pooling_type(self, value: PoolingType) -> None:
self.add_uint32(Keys.LLM.POOLING_TYPE.format(arch=self.arch), value)
def add_rope_dimension_count(self, count: int) -> None:
self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count)
@ -411,6 +415,12 @@ class GGUFWriter:
def add_pad_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.PAD_ID, id)
def add_cls_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.CLS_ID, id)
def add_mask_token_id(self, id: int) -> None:
self.add_uint32(Keys.Tokenizer.MASK_ID, id)
def add_add_bos_token(self, value: bool) -> None:
self.add_bool(Keys.Tokenizer.ADD_BOS, value)

12
gguf-py/gguf/tensor_mapping.py
View File

@ -15,7 +15,7 @@ class TensorNameMap:
"word_embeddings", # bloom
"model.embed_tokens", # llama-hf
"tok_embeddings", # llama-pth
"embeddings.word_embeddings", # bert
"embeddings.word_embeddings", # bert nomic-bert
"language_model.embedding.word_embeddings", # persimmon
"wte", # gpt2
"transformer.embd.wte", # phi2
@ -24,13 +24,14 @@ class TensorNameMap:
# Token type embeddings
MODEL_TENSOR.TOKEN_TYPES: (
"embeddings.token_type_embeddings", # bert
"embeddings.token_type_embeddings", # bert nomic-bert
),
# Normalization of token embeddings
MODEL_TENSOR.TOKEN_EMBD_NORM: (
"word_embeddings_layernorm", # bloom
"embeddings.LayerNorm", # bert
"emb_ln", # nomic-bert
),
# Position embeddings
@ -103,6 +104,7 @@ class TensorNameMap:
"model.layers.{bid}.self_attn.query_key_value", # persimmon
"h.{bid}.attn.c_attn", # gpt2
"transformer.h.{bid}.mixer.Wqkv", # phi2
"encoder.layers.{bid}.attn.Wqkv", # nomic-bert
),
# Attention query
@ -152,11 +154,13 @@ class TensorNameMap:
"transformer.h.{bid}.mixer.out_proj", # phi2
"model.layers.layers.{bid}.self_attn.o_proj", # plamo
"model.layers.{bid}.attention.wo", # internlm2
"encoder.layers.{bid}.attn.out_proj", # nomic-bert
),
# Attention output norm
MODEL_TENSOR.ATTN_OUT_NORM: (
"encoder.layer.{bid}.attention.output.LayerNorm", # bert
"encoder.layers.{bid}.norm1", # nomic-bert
),
# Rotary embeddings
@ -205,6 +209,7 @@ class TensorNameMap:
"model.layers.{bid}.mlp.fc1", # phi2
"model.layers.layers.{bid}.mlp.up_proj", # plamo
"model.layers.{bid}.feed_forward.w3", # internlm2
"encoder.layers.{bid}.mlp.fc11", # nomic-bert
),
MODEL_TENSOR.FFN_UP_EXP: (
@ -224,6 +229,7 @@ class TensorNameMap:
"transformer.h.{bid}.mlp.w2", # qwen
"model.layers.layers.{bid}.mlp.gate_proj", # plamo
"model.layers.{bid}.feed_forward.w1", # internlm2
"encoder.layers.{bid}.mlp.fc12", # nomic-bert
),
MODEL_TENSOR.FFN_GATE_EXP: (
@ -249,6 +255,7 @@ class TensorNameMap:
"model.layers.{bid}.mlp.fc2", # phi2
"model.layers.layers.{bid}.mlp.down_proj", # plamo
"model.layers.{bid}.feed_forward.w2", # internlm2
"encoder.layers.{bid}.mlp.fc2", # nomic-bert
),
MODEL_TENSOR.FFN_DOWN_EXP: (
@ -272,6 +279,7 @@ class TensorNameMap:
MODEL_TENSOR.LAYER_OUT_NORM: (
"encoder.layer.{bid}.output.LayerNorm", # bert
"encoder.layers.{bid}.norm2", # nomic-bert
)
}

6
gguf-py/gguf/vocab.py
View File

@ -29,7 +29,7 @@ class SpecialVocab:
if special_token_types is not None:
self.special_token_types = special_token_types
else:
self.special_token_types = ('bos', 'eos', 'unk', 'sep', 'pad')
self.special_token_types = ('bos', 'eos', 'unk', 'sep', 'pad', 'cls', 'mask')
self._load(Path(path))
def __repr__(self) -> str:
@ -152,10 +152,6 @@ class SpecialVocab:
add_entry = tokenizer_config.get(f'add_{typ}_token')
if isinstance(add_entry, bool):
self.add_special_token[typ] = add_entry
if not added_tokens:
# We will need this to get the content for the token, so if it's empty
# may as well just give up.
continue
entry = tokenizer_config.get(f'{typ}_token')
if isinstance(entry, str):
tc_content = entry

641
llama.cpp
View File

File diff suppressed because it is too large Load Diff

42
llama.h
View File

@ -100,6 +100,7 @@ extern "C" {
LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ1_S = 24, // except 1d tensors
LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
};
@ -112,6 +113,12 @@ extern "C" {
LLAMA_ROPE_SCALING_MAX_VALUE = LLAMA_ROPE_SCALING_YARN,
};
enum llama_pooling_type {
LLAMA_POOLING_NONE = 0,
LLAMA_POOLING_MEAN = 1,
LLAMA_POOLING_CLS = 2,
};
enum llama_split_mode {
LLAMA_SPLIT_NONE = 0, // single GPU
LLAMA_SPLIT_LAYER = 1, // split layers and KV across GPUs
@ -236,6 +243,7 @@ extern "C" {
bool logits_all; // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
bool embedding; // embedding mode only
bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
bool do_pooling; // whether to pool (sum) embedding results by sequence id (ignored if no pooling layer)
};
// model quantization parameters
@ -297,6 +305,12 @@ extern "C" {
int32_t n_eval;
};
// used in chat template
typedef struct llama_chat_message {
const char * role;
const char * content;
} llama_chat_message;
// Helpers for getting default parameters
LLAMA_API struct llama_model_params llama_model_default_params(void);
LLAMA_API struct llama_context_params llama_context_default_params(void);
@ -305,7 +319,10 @@ extern "C" {
// Initialize the llama + ggml backend
// If numa is true, use NUMA optimizations
// Call once at the start of the program
LLAMA_API void llama_backend_init(bool numa);
LLAMA_API void llama_backend_init(void);
//optional:
LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa);
// Call once at the end of the program - currently only used for MPI
LLAMA_API void llama_backend_free(void);
@ -628,6 +645,10 @@ extern "C" {
// shape: [n_embd] (1-dimensional)
LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
// Get the embeddings for the ith sequence
// llama_get_embeddings(ctx) + i*n_embd
LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
//
// Vocab
//
@ -684,6 +705,25 @@ extern "C" {
char * buf,
int32_t length);
/// Apply chat template. Inspired by hf apply_chat_template() on python.
/// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
/// NOTE: This function only support some known jinja templates. It is not a jinja parser.
/// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
/// @param chat Pointer to a list of multiple llama_chat_message
/// @param n_msg Number of llama_chat_message in this chat
/// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message.
/// @param buf A buffer to hold the output formatted prompt. The recommended alloc size is 2 * (total number of characters of all messages)
/// @param length The size of the allocated buffer
/// @return The total number of bytes of the formatted prompt. If is it larger than the size of buffer, you may need to re-alloc it and then re-apply the template.
LLAMA_API int32_t llama_chat_apply_template(
const struct llama_model * model,
const char * tmpl,
const struct llama_chat_message * chat,
size_t n_msg,
bool add_ass,
char * buf,
int32_t length);
//
// Grammar
//

37
scripts/compare-commits.sh Executable file
View File

@ -0,0 +1,37 @@
#!/bin/bash
if [ $# -lt 2 ]; then
echo "usage: ./scripts/compare-commits.sh <commit1> <commit2> [additional llama-bench arguments]"
exit 1
fi
set -e
set -x
bench_args="${@:3}"
rm -f llama-bench.sqlite
backend="cpu"
if [[ "$OSTYPE" == "darwin"* ]]; then
backend="metal"
elif command -v nvcc &> /dev/null; then
backend="cuda"
fi
make_opts=""
if [[ "$backend" == "cuda" ]]; then
make_opts="LLAMA_CUBLAS=1"
fi
git checkout $1
make clean && make -j32 $make_opts llama-bench
./llama-bench -o sql $bench_args | tee /dev/tty | sqlite3 llama-bench.sqlite
git checkout $2
make clean && make -j32 $make_opts llama-bench
./llama-bench -o sql $bench_args | tee /dev/tty | sqlite3 llama-bench.sqlite
./scripts/compare-llama-bench.py -b $1 -c $2

2
scripts/get-flags.mk
View File

@ -1,6 +1,6 @@
ifeq '' '$(findstring clang,$(shell $(GF_CC) --version))'
GF_CC_IS_GCC = 1
GF_CC_VER := $(shell { $(GF_CC) -dumpfullversion 2>/dev/null || $(GF_CC) -dumpversion; } | awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
GF_CC_VER := $(shell { $(GF_CC) -dumpfullversion 2>/dev/null; echo; $(GF_CC) -dumpversion; } | awk -F. '/./ { printf("%02d%02d%02d", $$1, $$2, $$3); exit }')
else
GF_CC_IS_CLANG = 1
ifeq '' '$(findstring Apple,$(shell $(GF_CC) --version))'

2
scripts/get-wikitext-2.sh
View File

@ -1,6 +1,6 @@
#!/bin/bash
wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
wget https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
echo "Usage:"
echo ""

107
scripts/hf.sh Executable file
View File

@ -0,0 +1,107 @@
#!/bin/bash
#
# Shortcut for downloading HF models
#
# Usage:
# ./main -m $(./examples/hf.sh https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
# ./main -m $(./examples/hf.sh --url https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/blob/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
# ./main -m $(./examples/hf.sh --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf)
#
# all logs go to stderr
function log {
echo "$@" 1>&2
}
function usage {
log "Usage: $0 [[--url] <url>] [--repo <repo>] [--file <file>] [-h|--help]"
exit 1
}
# check for curl or wget
function has_cmd {
if ! [ -x "$(command -v $1)" ]; then
return 1
fi
}
if has_cmd wget; then
cmd="wget -q --show-progress -c -O %s %s"
elif has_cmd curl; then
cmd="curl -C - -f -o %s -L %s"
else
log "[E] curl or wget not found"
exit 1
fi
url=""
repo=""
file=""
# parse args
while [[ $# -gt 0 ]]; do
case "$1" in
--url)
url="$2"
shift 2
;;
--repo)
repo="$2"
shift 2
;;
--file)
file="$2"
shift 2
;;
-h|--help)
usage
;;
*)
url="$1"
shift
;;
esac
done
if [ -n "$repo" ] && [ -n "$file" ]; then
url="https://huggingface.co/$repo/resolve/main/$file"
fi
if [ -z "$url" ]; then
log "[E] missing --url"
usage
fi
# check if the URL is a HuggingFace model, and if so, try to download it
is_url=false
if [[ ${#url} -gt 22 ]]; then
if [[ ${url:0:22} == "https://huggingface.co" ]]; then
is_url=true
fi
fi
if [ "$is_url" = false ]; then
log "[E] invalid URL, must start with https://huggingface.co"
exit 0
fi
# replace "blob/main" with "resolve/main"
url=${url/blob\/main/resolve\/main}
basename=$(basename $url)
log "[+] attempting to download $basename"
if [ -n "$cmd" ]; then
cmd=$(printf "$cmd" "$basename" "$url")
log "[+] $cmd"
if $cmd; then
echo $basename
exit 0
fi
fi
log "[-] failed to download"
exit 1

1
tests/CMakeLists.txt
View File

@ -28,6 +28,7 @@ endfunction()
llama_build_and_test_executable(test-quantize-fns.cpp)
llama_build_and_test_executable(test-quantize-perf.cpp)
llama_build_and_test_executable(test-sampling.cpp)
llama_build_and_test_executable(test-chat-template.cpp)
llama_build_executable(test-tokenizer-0-llama.cpp)
llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)

2
tests/test-autorelease.cpp
View File

@ -12,7 +12,7 @@ int main(int argc, char ** argv) {
auto * model_path = get_model_or_exit(argc, argv);
std::thread([&model_path]() {
llama_backend_init(false);
llama_backend_init();
auto * model = llama_load_model_from_file(model_path, llama_model_default_params());
auto * ctx = llama_new_context_with_model(model, llama_context_default_params());
llama_free(ctx);

81
tests/test-backend-ops.cpp
View File

@ -1095,24 +1095,32 @@ struct test_diag_mask_inf : public test_case {
struct test_soft_max : public test_case {
const ggml_type type;
const std::array<int64_t, 4> ne;
const float scale;
const bool mask;
const float scale;
const float max_bias;
std::string vars() override {
return VARS_TO_STR4(type, ne, scale, mask);
return VARS_TO_STR5(type, ne, mask, scale, max_bias);
}
test_soft_max(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {10, 10, 10, 10},
bool mask = false,
float scale = 1.0f,
bool mask = false)
: type(type), ne(ne), scale(scale), mask(mask) {}
float max_bias = 0.0f)
: type(type), ne(ne), mask(mask), scale(scale), max_bias(max_bias) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_tensor * b = nullptr;
if (mask) { b = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, ne[0], ne[1]); }
ggml_tensor * out = ggml_soft_max_ext(ctx, a, b, scale);
ggml_tensor * mask = nullptr;
if (this->mask) {
mask = ggml_new_tensor_2d(ctx, type, ne[0], ne[1]);
}
ggml_tensor * pos = nullptr;
if (max_bias > 0.0f) {
pos = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ne[0]);
}
ggml_tensor * out = ggml_soft_max_ext(ctx, a, mask, pos, scale, max_bias);
return out;
}
};
@ -1157,30 +1165,6 @@ struct test_rope : public test_case {
}
};
// GGML_OP_ALIBI
struct test_alibi : public test_case {
const ggml_type type;
const std::array<int64_t, 4> ne;
int n_past;
int n_head;
float bias_max;
std::string vars() override {
return VARS_TO_STR5(type, ne, n_past, n_head, bias_max);
}
test_alibi(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {10, 10, 10, 10},
int n_past = 512, int n_head = 10, float bias_max = 0.5f)
: type(type), ne(ne), n_past(n_past), n_head(n_head), bias_max(bias_max) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_tensor * out = ggml_alibi(ctx, a, n_past, n_head, bias_max);
return out;
}
};
// GGML_OP_POOL2D
struct test_pool2d : public test_case {
enum ggml_op_pool pool_type;
@ -1568,7 +1552,7 @@ struct test_moe : public test_case {
ggml_tensor * cur = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_tokens);
ggml_tensor * logits = ggml_mul_mat(ctx, ffn_gate_inp, cur);
ggml_tensor * probs = ggml_soft_max_ext(ctx, logits, nullptr, 1.0f/sqrtf(n_embd));
ggml_tensor * probs = ggml_soft_max_ext(ctx, logits, nullptr, nullptr, 1.0f/sqrtf(n_embd), 0.0f);
// select experts
ggml_tensor * selected_experts = ggml_top_k(ctx, probs, n_experts_per_tok);
@ -1697,7 +1681,6 @@ public:
ggml_cpy(ctx, v_cur_t, v_cache_view);
}
// if max_alibi_bias > 0 then apply ALiBi
struct ggml_tensor * llm_build_kqv(
struct ggml_context * ctx,
struct ggml_tensor * k_l,
@ -1716,7 +1699,7 @@ public:
struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale);
kq = ggml_soft_max_ext(ctx, kq, kq_mask, nullptr, kq_scale, 0.0f);
// split cached v into n_head heads
struct ggml_tensor * v =
@ -2014,7 +1997,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
GGML_TYPE_Q4_K, GGML_TYPE_Q5_K,
GGML_TYPE_Q6_K,
GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS,
GGML_TYPE_IQ3_XXS,
GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S,
};
// unary ops
@ -2163,6 +2146,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10, 1}, 5));
test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10, 10}, 5));
#if 0
std::uniform_int_distribution<> dist_ne1(1, 50);
int exponent = 1;
while (exponent < (1 << 17)) {
@ -2171,14 +2155,29 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
for (int n = 0; n < 10; ++n) {
int64_t ne0 = dist_ne0(rng);
int64_t ne1 = dist_ne1(rng);
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, n/2 == 0, 0.1f, ne0 < 1000 ? 4.0f : 0.0f));
}
exponent <<= 1;
}
#endif
for (bool mask : {false, true}) {
for (float max_bias : {0.0f, 8.0f}) {
for (float scale : {1.0f, 0.1f}) {
for (int64_t ne0 : {16, 1024}) {
for (int64_t ne1 : {16, 1024}) {
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, mask, scale, max_bias));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, scale, max_bias));
}
}
}
}
}
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, 0.1f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, 0.1f, true));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, 0.1f, 8.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, 0.1f, 8.0f));
for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
test_cases.emplace_back(new test_rope(type, {128, 32, 10, 1}, 128, 0, 512)); // llama 7B
@ -2193,7 +2192,6 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
test_cases.emplace_back(new test_rope(type, { 80, 32, 10, 1}, 32, 2, 512)); // neox (phi-2)
}
test_cases.emplace_back(new test_alibi());
test_cases.emplace_back(new test_concat(GGML_TYPE_F32));
test_cases.emplace_back(new test_concat(GGML_TYPE_I32));
@ -2233,14 +2231,13 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
}
#endif
// these tests are disabled to save execution time, but they can be handy for debugging
#if 0
#if !defined(__SANITIZE_THREAD__)
// FIXME: these tests use too much memory with thread sanitizer
test_cases.emplace_back(new test_moe(8, 2, 1, 4096, 8*1024));
//test_cases.emplace_back(new test_moe(8, 2, 8, 4096, 14336));
#endif
// these tests are disabled to save execution time, but they can be handy for debugging
#if 0
test_cases.emplace_back(new test_llama(1));
test_cases.emplace_back(new test_llama(2));
test_cases.emplace_back(new test_falcon(1));

64
tests/test-chat-template.cpp Normal file
View File

@ -0,0 +1,64 @@
#include <iostream>
#include <string>
#include <vector>
#include <sstream>
#undef NDEBUG
#include <cassert>
#include "llama.h"
int main(void) {
llama_chat_message conversation[] = {
{"system", "You are a helpful assistant"},
{"user", "Hello"},
{"assistant", "Hi there"},
{"user", "Who are you"},
{"assistant", " I am an assistant "},
{"user", "Another question"},
};
size_t message_count = 6;
std::vector<std::string> templates = {
// teknium/OpenHermes-2.5-Mistral-7B
"{% for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}{% endif %}",
// mistralai/Mistral-7B-Instruct-v0.2
"{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",
// TheBloke/FusionNet_34Bx2_MoE-AWQ
"{%- for idx in range(0, messages|length) -%}\\n{%- if messages[idx]['role'] == 'user' -%}\\n{%- if idx > 1 -%}\\n{{- bos_token + '[INST] ' + messages[idx]['content'] + ' [/INST]' -}}\\n{%- else -%}\\n{{- messages[idx]['content'] + ' [/INST]' -}}\\n{%- endif -%}\\n{% elif messages[idx]['role'] == 'system' %}\\n{{- '[INST] <<SYS>>\\\\n' + messages[idx]['content'] + '\\\\n<</SYS>>\\\\n\\\\n' -}}\\n{%- elif messages[idx]['role'] == 'assistant' -%}\\n{{- ' ' + messages[idx]['content'] + ' ' + eos_token -}}\\n{% endif %}\\n{% endfor %}",
// bofenghuang/vigogne-2-70b-chat
"{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif true == true and not '<<SYS>>' in messages[0]['content'] %}{% set loop_messages = messages %}{% set system_message = 'Vous êtes Vigogne, un assistant IA créé par Zaion Lab. Vous suivez extrêmement bien les instructions. Aidez autant que vous le pouvez.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<<SYS>>\\\\n' + system_message + '\\\\n<</SYS>>\\\\n\\\\n' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'system' %}{{ '<<SYS>>\\\\n' + content.strip() + '\\\\n<</SYS>>\\\\n\\\\n' }}{% elif message['role'] == 'assistant' %}{{ ' ' + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}",
};
std::vector<std::string> expected_substr = {
"<|im_start|>assistant\n I am an assistant <|im_end|>\n<|im_start|>user\nAnother question<|im_end|>\n<|im_start|>assistant",
"[/INST]Hi there</s>[INST] Who are you [/INST] I am an assistant </s>[INST] Another question [/INST]",
"</s><s>[INST] Who are you [/INST] I am an assistant </s><s>[INST] Another question [/INST]",
"[/INST] Hi there </s>[INST] Who are you [/INST] I am an assistant </s>[INST] Another question [/INST]",
};
std::vector<char> formatted_chat(1024);
int32_t res;
// test invalid chat template
res = llama_chat_apply_template(nullptr, "INVALID TEMPLATE", conversation, message_count, true, formatted_chat.data(), formatted_chat.size());
assert(res < 0);
for (size_t i = 0; i < templates.size(); i++) {
std::string custom_template = templates[i];
std::string substr = expected_substr[i];
formatted_chat.resize(1024);
res = llama_chat_apply_template(
nullptr,
custom_template.c_str(),
conversation,
message_count,
true,
formatted_chat.data(),
formatted_chat.size()
);
formatted_chat.resize(res);
std::string output(formatted_chat.data(), formatted_chat.size());
std::cout << output << "\n-------------------------\n";
// expect the "formatted_chat" to contain pre-defined strings
assert(output.find(substr) != std::string::npos);
}
return 0;
}

20
tests/test-grammar-parser.cpp
View File

@ -38,8 +38,8 @@ term ::= [0-9]+)""";
// pretty print error message before asserting
if (expected_pair.first != key || expected_pair.second != value)
{
fprintf(stderr, "expected_pair: %s, %d\n", expected_pair.first.c_str(), expected_pair.second);
fprintf(stderr, "actual_pair: %s, %d\n", key.c_str(), value);
fprintf(stderr, "expected_pair: %s, %u\n", expected_pair.first.c_str(), expected_pair.second);
fprintf(stderr, "actual_pair: %s, %u\n", key.c_str(), value);
fprintf(stderr, "expected_pair != actual_pair\n");
}
@ -96,9 +96,9 @@ term ::= [0-9]+)""";
// pretty print error message before asserting
if (expected_element.type != element.type || expected_element.value != element.value)
{
fprintf(stderr, "index: %d\n", index);
fprintf(stderr, "expected_element: %d, %d\n", expected_element.type, expected_element.value);
fprintf(stderr, "actual_element: %d, %d\n", element.type, element.value);
fprintf(stderr, "index: %u\n", index);
fprintf(stderr, "expected_element: %d, %u\n", expected_element.type, expected_element.value);
fprintf(stderr, "actual_element: %d, %u\n", element.type, element.value);
fprintf(stderr, "expected_element != actual_element\n");
}
@ -144,8 +144,8 @@ term ::= [0-9]+)""";
// pretty print error message before asserting
if (expected_pair.first != key || expected_pair.second != value)
{
fprintf(stderr, "expected_pair: %s, %d\n", expected_pair.first.c_str(), expected_pair.second);
fprintf(stderr, "actual_pair: %s, %d\n", key.c_str(), value);
fprintf(stderr, "expected_pair: %s, %u\n", expected_pair.first.c_str(), expected_pair.second);
fprintf(stderr, "actual_pair: %s, %u\n", key.c_str(), value);
fprintf(stderr, "expected_pair != actual_pair\n");
}
@ -235,9 +235,9 @@ term ::= [0-9]+)""";
// pretty print error message before asserting
if (expected_element.type != element.type || expected_element.value != element.value)
{
fprintf(stderr, "index: %d\n", index);
fprintf(stderr, "expected_element: %d, %d\n", expected_element.type, expected_element.value);
fprintf(stderr, "actual_element: %d, %d\n", element.type, element.value);
fprintf(stderr, "index: %u\n", index);
fprintf(stderr, "expected_element: %d, %u\n", expected_element.type, expected_element.value);
fprintf(stderr, "actual_element: %d, %u\n", element.type, element.value);
fprintf(stderr, "expected_element != actual_element\n");
}

4
tests/test-llama-grammar.cpp
View File

@ -180,8 +180,8 @@ int main()
if (expected_element.type != element->type || expected_element.value != element->value)
{
fprintf(stderr, "index: %d\n", index);
fprintf(stderr, "expected_element: %d, %d\n", expected_element.type, expected_element.value);
fprintf(stderr, "actual_element: %d, %d\n", element->type, element->value);
fprintf(stderr, "expected_element: %d, %u\n", expected_element.type, expected_element.value);
fprintf(stderr, "actual_element: %d, %u\n", element->type, element->value);
fprintf(stderr, "expected_element != actual_element\n");
}

2
tests/test-model-load-cancel.cpp
View File

@ -14,7 +14,7 @@ int main(int argc, char *argv[] ) {
fprintf(stderr, "using '%s'\n", model_path);
fclose(file);
llama_backend_init(false);
llama_backend_init();
auto params = llama_model_params{};
params.use_mmap = false;
params.progress_callback = [](float progress, void * ctx){

2
tests/test-tokenizer-0-falcon.cpp
View File

@ -61,7 +61,7 @@ int main(int argc, char **argv) {
llama_model * model;
llama_context * ctx;
llama_backend_init(false);
llama_backend_init();
// load the vocab
{

2
tests/test-tokenizer-0-llama.cpp
View File

@ -60,7 +60,7 @@ int main(int argc, char **argv) {
llama_model * model;
llama_context * ctx;
llama_backend_init(false);
llama_backend_init();
// load the vocab
{

79
tests/test-tokenizer-1-bpe.cpp
View File

@ -4,13 +4,13 @@
#include "console.h"
#include <cassert>
#include <codecvt>
#include <cstdio>
#include <cstring>
#include <string>
#include <codecvt>
#include <map>
#include <vector>
#include <locale>
#include <string>
#include <thread>
#include <vector>
int main(int argc, char **argv) {
if (argc < 2) {
@ -25,7 +25,7 @@ int main(int argc, char **argv) {
llama_model * model;
llama_context * ctx;
llama_backend_init(false);
llama_backend_init();
// load the vocab
{
@ -74,45 +74,46 @@ int main(int argc, char **argv) {
}
}
catch (const std::invalid_argument &) {
fprintf(stderr, "%s : info: utf8 conversion %d '%s'\n", __func__, i, str.c_str());
//fprintf(stderr, "%s : info: utf8 conversion %d '%s'\n", __func__, i, str.c_str());
}
}
for (uint32_t cp = 0x0000; cp < 0xffff; ++cp) {
// NOTE: these exceptions seem to be necessary, because the GPT2 tokenizer doesn't want to interfere with some ASCII control characters
if ((cp < 0x03 || cp > 0x05) && cp != 0x0b && cp != 0x11 && (cp < 0x13 || cp > 0x17) && cp != 0x19 && (cp < 0x1c || cp > 0x1e) && (cp < 0xd800 || cp > 0xdfff)) {
std::string str = " " + codepoint_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_bpe(ctx, tokens);
if (str != check) {
fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
__func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
return 3;
}
}
}
// Restrict to assigned unicode planes
// for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
for (uint32_t cp = 0x10000; cp < 0x00040000; ++cp) {
std::string str = codepoint_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_bpe(ctx, tokens);
if (str != check) {
fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
__func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
return 4;
}
}
for (uint32_t cp = 0x000e0000; cp < 0x0010ffff; ++cp) {
std::string str = codepoint_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_bpe(ctx, tokens);
if (str != check) {
fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
__func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
return 4;
// unicode
{
const int nthread = std::thread::hardware_concurrency();
std::vector<std::thread> threads(nthread);
for (int i = 0; i < nthread; ++i) {
threads[i] = std::thread([i, nthread, ctx]() {
for (uint32_t cp = i; cp < 0x0010ffff; cp += nthread) {
if (!( // NOLINT
(cp < 0x03 || cp > 0x05) && cp != 0x0b && cp != 0x11 &&
(cp < 0x13 || cp > 0x17) && cp != 0x19 &&
(cp < 0x1c || cp > 0x1e) &&
(cp < 0xd800 || cp > 0xdfff) &&
(cp < 0x00040000 || cp >= 0x000e0000)
)) {
continue;
}
std::string str = codepoint_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_bpe(ctx, tokens);
if (cp != 9601 && str != check) {
fprintf(stderr, "error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
cp, check.c_str(), check.length(), str.c_str(), str.length());
std::exit(3);
}
}
});
}
for (auto & t : threads) {
t.join();
}
}
llama_free_model(model);
llama_free(ctx);

55
tests/test-tokenizer-1-llama.cpp
View File

@ -4,13 +4,13 @@
#include "console.h"
#include <cassert>
#include <codecvt>
#include <cstdio>
#include <cstring>
#include <string>
#include <codecvt>
#include <map>
#include <vector>
#include <locale>
#include <string>
#include <thread>
#include <vector>
int main(int argc, char **argv) {
if (argc < 2) {
@ -25,7 +25,7 @@ int main(int argc, char **argv) {
llama_model * model;
llama_context * ctx;
llama_backend_init(false);
llama_backend_init();
// load the vocab
{
@ -72,26 +72,33 @@ int main(int argc, char **argv) {
}
}
for (uint32_t cp = 0x0000; cp < 0xffff; ++cp) {
if (cp < 0xd800 || cp > 0xdfff) {
std::string str = codepoint_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_spm(ctx, tokens);
if (cp != 9601 && str != check) {
fprintf(stderr, "%s : error: codepoint %d detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
__func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
return 3;
}
// unicode
{
const int nthread = std::thread::hardware_concurrency();
std::vector<std::thread> threads(nthread);
for (int i = 0; i < nthread; ++i) {
threads[i] = std::thread([i, nthread, ctx]() {
for (uint32_t cp = i; cp < 0x0010ffff; cp += nthread) {
if (cp >= 0xd800 && cp <= 0xdfff) {
continue;
}
std::string str = codepoint_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_spm(ctx, tokens);
if (cp != 9601 && str != check) {
fprintf(stderr, "error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
cp, check.c_str(), check.length(), str.c_str(), str.length());
std::exit(3);
}
}
});
}
}
for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
std::string str = codepoint_to_utf8(cp);
std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
std::string check = llama_detokenize_spm(ctx, tokens);
if (str != check) {
fprintf(stderr, "%s : error: codepoint %d detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
__func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
return 4;
for (auto & t : threads) {
t.join();
}
}

72
unicode.h
View File

@ -264,26 +264,29 @@ static uint32_t codepoint_from_utf8(const std::string & utf8, size_t & offset) {
offset += 1;
return result;
}
else if (!(utf8[offset + 0] & 0x40)) {
if (!(utf8[offset + 0] & 0x40)) {
throw std::invalid_argument("invalid character");
}
else if (!(utf8[offset + 0] & 0x20)) {
if (offset + 1 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80))
if (!(utf8[offset + 0] & 0x20)) {
if (offset + 1 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80)) {
throw std::invalid_argument("invalid character");
}
auto result = ((utf8[offset + 0] & 0x1f) << 6) | (utf8[offset + 1] & 0x3f);
offset += 2;
return result;
}
else if (!(utf8[offset + 0] & 0x10)) {
if (offset + 2 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80))
if (!(utf8[offset + 0] & 0x10)) {
if (offset + 2 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80)) {
throw std::invalid_argument("invalid character");
}
auto result = ((utf8[offset + 0] & 0x0f) << 12) | ((utf8[offset + 1] & 0x3f) << 6) | (utf8[offset + 2] & 0x3f);
offset += 3;
return result;
}
else if (!(utf8[offset + 0] & 0x08)) {
if (offset + 3 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80))
if (!(utf8[offset + 0] & 0x08)) {
if (offset + 3 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80)) {
throw std::invalid_argument("invalid character");
}
auto result = ((utf8[offset + 0] & 0x07) << 18) | ((utf8[offset + 1] & 0x3f) << 12) | ((utf8[offset + 2] & 0x3f) << 6) | (utf8[offset + 3] & 0x3f);
offset += 4;
return result;
@ -331,21 +334,22 @@ static uint32_t codepoint_from_utf16(const std::vector<uint16_t> & utf16, size_t
offset += 1;
return result;
}
else {
if (offset + 1 >= utf16.size() || !((utf16[1] & 0xdc00) == 0xdc00))
throw std::invalid_argument("invalid character");
auto result = 0x10000 + (((utf16[0] & 0x03ff) << 10) | (utf16[1] & 0x03ff));
offset += 2;
return result;
if (offset + 1 >= utf16.size() || !((utf16[1] & 0xdc00) == 0xdc00)) {
throw std::invalid_argument("invalid character");
}
throw std::invalid_argument("invalid string");
auto result = 0x10000 + (((utf16[0] & 0x03ff) << 10) | (utf16[1] & 0x03ff));
offset += 2;
return result;
}
static std::vector<uint32_t> codepoints_from_utf16(const std::vector<uint16_t> & utf16) {
std::vector<uint32_t> result;
size_t offset = 0;
while (offset < utf16.size())
while (offset < utf16.size()) {
result.push_back(codepoint_from_utf16(utf16, offset));
}
return result;
}
@ -361,44 +365,52 @@ static std::vector<uint32_t> codepoints_from_utf16(const std::vector<uint16_t> &
static std::unordered_map<uint32_t, int> codepoint_type_map() {
std::unordered_map<uint32_t, int> codepoint_types;
for (auto p : digit_ranges) {
for(auto i = p.first; i <= p.second; ++ i)
for (auto i = p.first; i <= p.second; ++ i) {
codepoint_types[i] = CODEPOINT_TYPE_DIGIT;
}
}
for(auto p : letter_ranges) {
for(auto i = p.first; i <= p.second; ++ i)
for (auto p : letter_ranges) {
for (auto i = p.first; i <= p.second; ++ i) {
codepoint_types[i] = CODEPOINT_TYPE_LETTER;
}
}
for(auto p : whitespace_ranges) {
for(auto i = p.first; i <= p.second; ++ i)
for (auto p : whitespace_ranges) {
for (auto i = p.first; i <= p.second; ++ i) {
codepoint_types[i] = CODEPOINT_TYPE_WHITESPACE;
}
}
for(auto p : accent_mark_ranges) {
for(auto i = p.first; i <= p.second; ++ i)
for (auto p : accent_mark_ranges) {
for (auto i = p.first; i <= p.second; ++ i) {
codepoint_types[i] = CODEPOINT_TYPE_ACCENT_MARK;
}
}
for(auto p : punctuation_ranges) {
for(auto i = p.first; i <= p.second; ++ i)
for (auto p : punctuation_ranges) {
for (auto i = p.first; i <= p.second; ++ i) {
codepoint_types[i] = CODEPOINT_TYPE_PUNCTUATION;
}
}
for (auto p : symbol_ranges) {
for (auto i = p.first; i <= p.second; ++i)
for (auto p : symbol_ranges) {
for (auto i = p.first; i <= p.second; ++i) {
codepoint_types[i] = CODEPOINT_TYPE_SYMBOL;
}
}
for(auto p : control_ranges) {
for(auto i = p.first; i <= p.second; ++ i)
for (auto p : control_ranges) {
for (auto i = p.first; i <= p.second; ++ i) {
codepoint_types[i] = CODEPOINT_TYPE_CONTROL;
}
}
return codepoint_types;
}
static int codepoint_type(uint32_t cp) {
static std::unordered_map<uint32_t, int> codepoint_types = codepoint_type_map();
return codepoint_types[cp];
return codepoint_types.find(cp) == codepoint_types.end() ? CODEPOINT_TYPE_UNIDENTIFIED : codepoint_types.at(cp);
}
static int codepoint_type(const std::string & utf8) {
if (utf8.length() == 0)
if (utf8.length() == 0) {
return CODEPOINT_TYPE_UNIDENTIFIED;
}
size_t offset = 0;
return codepoint_type(codepoint_from_utf8(utf8, offset));
}