vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into romain/set_arch

Browse source
This commit is contained in:
romain.biessy 2024-11-15 10:22:59 +00:00
commit 1333a47641
267 changed files with 51368 additions and 20898 deletions
Download patch file Download diff file Expand all files Collapse all files

4
.devops/llama-cli-cann.Dockerfile
View file

@ -1,6 +1,6 @@
ARG ASCEND_VERSION=8.0.rc2.alpha003-910b-openeuler22.03-py3.8
FROM cosdt/cann:$ASCEND_VERSION AS build
FROM ascendai/cann:$ASCEND_VERSION AS build
WORKDIR /app
@ -26,7 +26,7 @@ RUN echo "Building with static libs" && \
cmake --build build --config Release --target llama-cli
# TODO: use image with NNRT
FROM cosdt/cann:$ASCEND_VERSION AS runtime
FROM ascendai/cann:$ASCEND_VERSION AS runtime
COPY --from=build /app/build/bin/llama-cli /llama-cli
ENV LC_ALL=C.utf8

9
.devops/llama-cli-cuda.Dockerfile
View file

@ -23,15 +23,16 @@ RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-cli -j$(nproc)
cmake --build build --config Release --target llama-cli -j$(nproc) && \
mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libgomp1
COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
COPY --from=build /app/build/src/libllama.so /libllama.so
COPY --from=build /app/build/bin/llama-cli /llama-cli
COPY --from=build /app/lib/ /
COPY --from=build /app/build/bin/llama-cli /
ENTRYPOINT [ "/llama-cli" ]

7
.devops/llama-cli-musa.Dockerfile
View file

@ -16,15 +16,16 @@ WORKDIR /app
COPY . .
RUN cmake -B build -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-cli -j$(nproc)
cmake --build build --config Release --target llama-cli -j$(nproc) && \
mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libgomp1
COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
COPY --from=build /app/build/src/libllama.so /libllama.so
COPY --from=build /app/lib/ /
COPY --from=build /app/build/bin/llama-cli /llama-cli
ENTRYPOINT [ "/llama-cli" ]

7
.devops/llama-server-cuda.Dockerfile
View file

@ -23,15 +23,16 @@ RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
fi && \
cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-server -j$(nproc)
cmake --build build --config Release --target llama-server -j$(nproc) && \
mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev libgomp1 curl
COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
COPY --from=build /app/build/src/libllama.so /libllama.so
COPY --from=build /app/lib/ /
COPY --from=build /app/build/bin/llama-server /llama-server
# Must be set to 0.0.0.0 so it can listen to requests from host machine

7
.devops/llama-server-musa.Dockerfile
View file

@ -16,15 +16,16 @@ WORKDIR /app
COPY . .
RUN cmake -B build -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
cmake --build build --config Release --target llama-server -j$(nproc)
cmake --build build --config Release --target llama-server -j$(nproc) && \
mkdir -p /app/lib && \
find build -name "*.so" -exec cp {} /app/lib \;
FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
RUN apt-get update && \
apt-get install -y libcurl4-openssl-dev libgomp1 curl
COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
COPY --from=build /app/build/src/libllama.so /libllama.so
COPY --from=build /app/lib/ /
COPY --from=build /app/build/bin/llama-server /llama-server
# Must be set to 0.0.0.0 so it can listen to requests from host machine

6
.devops/nix/package.nix
View file

@ -126,9 +126,9 @@ effectiveStdenv.mkDerivation (finalAttrs: {
};
postPatch = ''
substituteInPlace ./ggml/src/ggml-metal.m \
substituteInPlace ./ggml/src/ggml-metal/ggml-metal.m \
--replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
substituteInPlace ./ggml/src/ggml-metal.m \
substituteInPlace ./ggml/src/ggml-metal/ggml-metal.m \
--replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
'';
@ -173,7 +173,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
(cmakeBool "GGML_NATIVE" false)
(cmakeBool "GGML_BLAS" useBlas)
(cmakeBool "GGML_CUDA" useCuda)
(cmakeBool "GGML_HIPBLAS" useRocm)
(cmakeBool "GGML_HIP" useRocm)
(cmakeBool "GGML_METAL" useMetalKit)
(cmakeBool "GGML_VULKAN" useVulkan)
(cmakeBool "GGML_STATIC" enableStatic)

10
.editorconfig
View file

@ -24,6 +24,16 @@ insert_final_newline = unset
[examples/server/public/*]
indent_size = 2
[examples/server/public/deps_*]
trim_trailing_whitespace = unset
indent_style = unset
indent_size = unset
[examples/server/deps_*]
trim_trailing_whitespace = unset
indent_style = unset
indent_size = unset
[examples/llama.swiftui/llama.swiftui.xcodeproj/*]
indent_style = tab

29
.github/workflows/build.yml vendored
View file

@ -55,7 +55,13 @@ jobs:
sysctl -a
mkdir build
cd build
cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
cmake .. \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_CURL=ON \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DGGML_RPC=ON \
-DBUILD_SHARED_LIBS=OFF
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -113,7 +119,12 @@ jobs:
sysctl -a
# Metal is disabled due to intermittent failures with Github runners not having a GPU:
# https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
cmake -B build \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_CURL=ON \
-DGGML_METAL=OFF \
-DGGML_RPC=ON \
-DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -394,13 +405,13 @@ jobs:
- name: Build with native CMake HIP support
id: cmake_build
run: |
cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIPBLAS=ON
cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIP=ON
cmake --build build --config Release -j $(nproc)
- name: Build with legacy HIP support
id: cmake_build_legacy_hip
run: |
cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DGGML_HIPBLAS=ON
cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DGGML_HIP=ON
cmake --build build2 --config Release -j $(nproc)
ubuntu-22-cmake-sycl:
@ -569,6 +580,7 @@ jobs:
mkdir build
cd build
cmake -G Xcode .. \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TESTS=OFF \
@ -599,6 +611,7 @@ jobs:
mkdir build
cd build
cmake -G Xcode .. \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TESTS=OFF \
@ -734,7 +747,7 @@ jobs:
id: clone_kompute
if: ${{ matrix.build == 'kompute-x64' }}
run: |
git submodule update --init ggml/src/kompute
git submodule update --init ggml/src/ggml-kompute/kompute
- name: Download OpenBLAS
id: get_openblas
@ -917,7 +930,7 @@ jobs:
shell: bash
env:
WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7dff44ba-e3af-4448-841c-0d616c8da6e7/w_BaseKit_p_2024.1.0.595_offline.exe
WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b380d914-366b-4b77-a74a-05e3c38b3514/intel-oneapi-base-toolkit-2025.0.0.882_offline.exe
WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
steps:
@ -1001,7 +1014,7 @@ jobs:
run: |
$env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
$env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIP=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
windows-latest-cmake-hip-release:
@ -1037,7 +1050,7 @@ jobs:
run: |
$env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
$env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIP=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
md "build\bin\rocblas\library\"
cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"

2
.gitmodules vendored
View file

@ -1,3 +1,3 @@
[submodule "kompute"]
path = ggml/src/kompute
path = ggml/src/ggml-kompute/kompute
url = https://github.com/nomic-ai/kompute.git

1
CMakeLists.txt
View file

@ -140,7 +140,6 @@ set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location o
set(LLAMA_LIB_INSTALL_DIR ${CMAKE_INSTALL_LIBDIR} CACHE PATH "Location of library files")
set(LLAMA_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR} CACHE PATH "Location of binary files")
# At the moment some compile definitions are placed within the ggml/src
# directory but not exported on the `ggml` target. This could be improved by
# determining _precisely_ which defines are necessary for the llama-config

292
Makefile
View file

@ -523,65 +523,54 @@ ifndef GGML_NO_ACCELERATE
# Mac OS - include Accelerate framework.
# `-framework Accelerate` works both with Apple Silicon and Mac Intel
ifeq ($(UNAME_S),Darwin)
MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS
MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS -DGGML_BLAS_USE_ACCELERATE
MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
MK_LDFLAGS += -framework Accelerate
OBJ_GGML += ggml/src/ggml-blas.o
OBJ_GGML += ggml/src/ggml-blas/ggml-blas.o
endif
endif # GGML_NO_ACCELERATE
ifdef GGML_MUSA
CC := clang
CXX := clang++
GGML_CUDA := 1
MK_CPPFLAGS += -DGGML_USE_MUSA
endif
ifndef GGML_NO_OPENMP
MK_CPPFLAGS += -DGGML_USE_OPENMP
MK_CFLAGS += -fopenmp
MK_CXXFLAGS += -fopenmp
ifdef GGML_MUSA
MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp
MK_LDFLAGS += -L/usr/lib/llvm-10/lib
endif # GGML_MUSA
endif # GGML_NO_OPENMP
ifdef GGML_OPENBLAS
MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
MK_CFLAGS += $(shell pkg-config --cflags-only-other openblas)
MK_LDFLAGS += $(shell pkg-config --libs openblas)
OBJ_GGML += ggml/src/ggml-blas.o
OBJ_GGML += ggml/src/ggml-blas/ggml-blas.o
endif # GGML_OPENBLAS
ifdef GGML_OPENBLAS64
MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
MK_CFLAGS += $(shell pkg-config --cflags-only-other openblas64)
MK_LDFLAGS += $(shell pkg-config --libs openblas64)
OBJ_GGML += ggml/src/ggml-blas.o
OBJ_GGML += ggml/src/ggml-blas/ggml-blas.o
endif # GGML_OPENBLAS64
ifdef GGML_BLIS
MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_BLIS -I/usr/local/include/blis -I/usr/include/blis
MK_LDFLAGS += -lblis -L/usr/local/lib
OBJ_GGML += ggml/src/ggml-blas.o
OBJ_GGML += ggml/src/ggml-blas/ggml-blas.o
endif # GGML_BLIS
ifdef GGML_NVPL
MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_NVPL -DNVPL_ILP64 -I/usr/local/include/nvpl_blas -I/usr/include/nvpl_blas
MK_LDFLAGS += -L/usr/local/lib -lnvpl_blas_core -lnvpl_blas_ilp64_gomp
OBJ_GGML += ggml/src/ggml-blas.o
OBJ_GGML += ggml/src/ggml-blas/ggml-blas.o
endif # GGML_NVPL
ifndef GGML_NO_LLAMAFILE
MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
OBJ_GGML += ggml/src/llamafile/sgemm.o
OBJ_GGML += ggml/src/ggml-cpu/llamafile/sgemm.o
endif
ifndef GGML_NO_AMX
MK_CPPFLAGS += -DGGML_USE_AMX
OBJ_GGML += ggml/src/ggml-amx.o ggml/src/ggml-amx/mmq.o
OBJ_GGML += ggml/src/ggml-amx/ggml-amx.o ggml/src/ggml-amx/mmq.o
endif
ifdef GGML_RPC
@ -601,29 +590,17 @@ else
endif # GGML_CUDA_FA_ALL_QUANTS
ifdef GGML_CUDA
ifdef GGML_MUSA
ifneq ('', '$(wildcard /opt/musa)')
CUDA_PATH ?= /opt/musa
else
CUDA_PATH ?= /usr/local/musa
endif
MK_CPPFLAGS += -DGGML_USE_CUDA -I$(CUDA_PATH)/include
MK_LDFLAGS += -lmusa -lmublas -lmusart -lpthread -ldl -lrt -L$(CUDA_PATH)/lib -L/usr/lib64
MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22
ifneq ('', '$(wildcard /opt/cuda)')
CUDA_PATH ?= /opt/cuda
else
ifneq ('', '$(wildcard /opt/cuda)')
CUDA_PATH ?= /opt/cuda
else
CUDA_PATH ?= /usr/local/cuda
endif
CUDA_PATH ?= /usr/local/cuda
endif
MK_CPPFLAGS += -DGGML_USE_CUDA -DGGML_CUDA_USE_GRAPHS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
MK_NVCCFLAGS += -use_fast_math
endif # GGML_MUSA
MK_CPPFLAGS += -DGGML_USE_CUDA -DGGML_CUDA_USE_GRAPHS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
MK_NVCCFLAGS += -use_fast_math
OBJ_GGML += ggml/src/ggml-cuda.o
OBJ_GGML += ggml/src/ggml-cuda/ggml-cuda.o
OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
OBJ_GGML += $(OBJ_CUDA_TMPL)
@ -631,11 +608,9 @@ ifdef LLAMA_FATAL_WARNINGS
MK_NVCCFLAGS += -Werror all-warnings
endif # LLAMA_FATAL_WARNINGS
ifndef GGML_MUSA
ifndef JETSON_EOL_MODULE_DETECT
MK_NVCCFLAGS += --forward-unknown-to-host-compiler
endif # JETSON_EOL_MODULE_DETECT
endif # GGML_MUSA
ifdef LLAMA_DEBUG
MK_NVCCFLAGS += -lineinfo
@ -648,11 +623,7 @@ endif # GGML_CUDA_DEBUG
ifdef GGML_CUDA_NVCC
NVCC = $(CCACHE) $(GGML_CUDA_NVCC)
else
ifdef GGML_MUSA
NVCC = $(CCACHE) mcc
else
NVCC = $(CCACHE) nvcc
endif # GGML_MUSA
NVCC = $(CCACHE) nvcc
endif # GGML_CUDA_NVCC
ifdef CUDA_DOCKER_ARCH
@ -724,15 +695,9 @@ define NVCC_COMPILE
$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
endef # NVCC_COMPILE
else
ifdef GGML_MUSA
define NVCC_COMPILE
$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -c $< -o $@
endef # NVCC_COMPILE
else
define NVCC_COMPILE
$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
endef # NVCC_COMPILE
endif # GGML_MUSA
endif # JETSON_EOL_MODULE_DETECT
ggml/src/ggml-cuda/%.o: \
@ -742,8 +707,8 @@ ggml/src/ggml-cuda/%.o: \
ggml/src/ggml-cuda/common.cuh
$(NVCC_COMPILE)
ggml/src/ggml-cuda.o: \
ggml/src/ggml-cuda.cu \
ggml/src/ggml-cuda/ggml-cuda.o: \
ggml/src/ggml-cuda/ggml-cuda.cu \
ggml/include/ggml-cuda.h \
ggml/include/ggml.h \
ggml/include/ggml-backend.h \
@ -819,7 +784,7 @@ ifdef GGML_HIPBLAS
GGML_CUDA_MMV_Y ?= 1
GGML_CUDA_KQUANTS_ITER ?= 2
MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
MK_CPPFLAGS += -DGGML_USE_HIP -DGGML_USE_CUDA
ifdef GGML_HIP_UMA
MK_CPPFLAGS += -DGGML_HIP_UMA
@ -852,12 +817,12 @@ ifdef GGML_CUDA_NO_PEER_COPY
HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY
endif # GGML_CUDA_NO_PEER_COPY
OBJ_GGML += ggml/src/ggml-cuda.o
OBJ_GGML += ggml/src/ggml-cuda/ggml-cuda.o
OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
OBJ_GGML += $(OBJ_CUDA_TMPL)
ggml/src/ggml-cuda.o: \
ggml/src/ggml-cuda.cu \
ggml/src/ggml-cuda/ggml-cuda.o: \
ggml/src/ggml-cuda/ggml-cuda.cu \
ggml/include/ggml-cuda.h \
ggml/include/ggml.h \
ggml/include/ggml-backend.h \
@ -874,39 +839,148 @@ ggml/src/ggml-cuda/%.o: \
$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
endif # GGML_HIPBLAS
ifdef GGML_MUSA
ifeq ($(wildcard /opt/musa),)
MUSA_PATH ?= /usr/local/musa
else
MUSA_PATH ?= /opt/musa
endif
MTGPU_TARGETS ?= mp_21 mp_22
MK_CPPFLAGS += -DGGML_USE_MUSA -DGGML_USE_CUDA
MK_LDFLAGS += -L$(MUSA_PATH)/lib -Wl,-rpath=$(MUSA_PATH)/lib
MK_LDFLAGS += -lmusa -lmusart -lmublas
ifndef GGML_NO_OPENMP
# For Ubuntu Focal
MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp
MK_LDFLAGS += -L/usr/lib/llvm-10/lib
# For Ubuntu Jammy
MK_CPPFLAGS += -I/usr/lib/llvm-14/lib/clang/14.0.0/include
MK_LDFLAGS += -L/usr/lib/llvm-14/lib
endif # GGML_NO_OPENMP
CC := $(MUSA_PATH)/bin/clang
CXX := $(MUSA_PATH)/bin/clang++
MCC := $(CCACHE) $(MUSA_PATH)/bin/mcc
MUSAFLAGS += $(addprefix --cuda-gpu-arch=, $(MTGPU_TARGETS))
ifdef GGML_CUDA_FORCE_DMMV
MUSAFLAGS += -DGGML_CUDA_FORCE_DMMV
endif # GGML_CUDA_FORCE_DMMV
ifdef GGML_CUDA_FORCE_MMQ
MUSAFLAGS += -DGGML_CUDA_FORCE_MMQ
endif # GGML_CUDA_FORCE_MMQ
ifdef GGML_CUDA_FORCE_CUBLAS
MUSAFLAGS += -DGGML_CUDA_FORCE_CUBLAS
endif # GGML_CUDA_FORCE_CUBLAS
ifdef GGML_CUDA_DMMV_X
MUSAFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
else
MUSAFLAGS += -DGGML_CUDA_DMMV_X=32
endif # GGML_CUDA_DMMV_X
ifdef GGML_CUDA_MMV_Y
MUSAFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
else
MUSAFLAGS += -DGGML_CUDA_MMV_Y=1
endif # GGML_CUDA_MMV_Y
ifdef GGML_CUDA_F16
MUSAFLAGS += -DGGML_CUDA_F16
endif # GGML_CUDA_F16
ifdef GGML_CUDA_DMMV_F16
MUSAFLAGS += -DGGML_CUDA_F16
endif # GGML_CUDA_DMMV_F16
ifdef GGML_CUDA_KQUANTS_ITER
MUSAFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
else
MUSAFLAGS += -DK_QUANTS_PER_ITERATION=2
endif
ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
MUSAFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
else
MUSAFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
endif # GGML_CUDA_PEER_MAX_BATCH_SIZE
ifdef GGML_CUDA_NO_PEER_COPY
MUSAFLAGS += -DGGML_CUDA_NO_PEER_COPY
endif # GGML_CUDA_NO_PEER_COPY
ifdef GGML_CUDA_FA_ALL_QUANTS
MUSAFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
endif # GGML_CUDA_FA_ALL_QUANTS
OBJ_GGML += ggml/src/ggml-cuda/ggml-cuda.o
OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
OBJ_GGML += $(OBJ_CUDA_TMPL)
ggml/src/ggml-cuda/ggml-cuda.o: \
ggml/src/ggml-cuda/ggml-cuda.cu \
ggml/include/ggml-cuda.h \
ggml/include/ggml.h \
ggml/include/ggml-backend.h \
ggml/src/ggml-backend-impl.h \
ggml/src/ggml-common.h \
$(wildcard ggml/src/ggml-cuda/*.cuh)
$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -x musa -mtgpu -c -o $@ $<
ggml/src/ggml-cuda/%.o: \
ggml/src/ggml-cuda/%.cu \
ggml/include/ggml.h \
ggml/src/ggml-common.h \
ggml/src/ggml-cuda/common.cuh
$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -x musa -mtgpu -c -o $@ $<
endif # GGML_MUSA
ifndef GGML_NO_CPU_AARCH64
MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
endif
ifdef GGML_METAL
MK_CPPFLAGS += -DGGML_USE_METAL
MK_LDFLAGS += -framework Foundation -framework Metal -framework MetalKit
OBJ_GGML += ggml/src/ggml-metal.o
OBJ_GGML += ggml/src/ggml-metal/ggml-metal.o
ifdef GGML_METAL_USE_BF16
MK_CPPFLAGS += -DGGML_METAL_USE_BF16
endif # GGML_METAL_USE_BF16
ifdef GGML_METAL_NDEBUG
MK_CPPFLAGS += -DGGML_METAL_NDEBUG
endif
ifdef GGML_METAL_EMBED_LIBRARY
MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
OBJ_GGML += ggml/src/ggml-metal-embed.o
OBJ_GGML += ggml/src/ggml-metal-embed.o
endif
endif # GGML_METAL
ifdef GGML_METAL
ggml/src/ggml-metal.o: \
ggml/src/ggml-metal.m \
ggml/src/ggml-metal/ggml-metal.o: \
ggml/src/ggml-metal/ggml-metal.m \
ggml/include/ggml-metal.h \
ggml/include/ggml.h
$(CC) $(CFLAGS) -c $< -o $@
ifdef GGML_METAL_EMBED_LIBRARY
ggml/src/ggml-metal-embed.o: \
ggml/src/ggml-metal.metal \
ggml/src/ggml-metal/ggml-metal.metal \
ggml/src/ggml-common.h
@echo "Embedding Metal library"
@sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal
@sed -e '/__embed_ggml-common.h__/r ggml/src/ggml-common.h' -e '/__embed_ggml-common.h__/d' < ggml/src/ggml-metal/ggml-metal.metal > ggml/src/ggml-metal/ggml-metal-embed.metal
$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".incbin \"ggml/src/ggml-metal/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
@rmdir ${TEMP_ASSEMBLY}
@ -915,11 +989,16 @@ endif # GGML_METAL
OBJ_GGML += \
ggml/src/ggml.o \
ggml/src/ggml-cpu.o \
ggml/src/ggml-aarch64.o \
ggml/src/ggml-alloc.o \
ggml/src/ggml-backend.o \
ggml/src/ggml-backend-reg.o \
ggml/src/ggml-quants.o \
ggml/src/ggml-aarch64.o
ggml/src/ggml-threading.o \
ggml/src/ggml-cpu/ggml-cpu.o \
ggml/src/ggml-cpu/ggml-cpu-cpp.o \
ggml/src/ggml-cpu/ggml-cpu-aarch64.o \
ggml/src/ggml-cpu/ggml-cpu-quants.o
OBJ_LLAMA = \
src/llama.o \
@ -993,7 +1072,6 @@ $(info I CXX: $(shell $(CXX) --version | head -n 1))
ifdef GGML_CUDA
$(info I NVCC: $(shell $(NVCC) --version | tail -n 1))
CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
ifndef GGML_MUSA
ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
ifndef CUDA_DOCKER_ARCH
@ -1003,7 +1081,6 @@ endif # CUDA_POWER_ARCH
endif # CUDA_DOCKER_ARCH
endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
endif # GGML_MUSA
endif # GGML_CUDA
$(info )
@ -1047,12 +1124,23 @@ ggml/src/ggml.o: \
ggml/include/ggml.h
$(CC) $(CFLAGS) -c $< -o $@
ggml/src/ggml-cpu.o: \
ggml/src/ggml-cpu.c \
ggml/src/ggml-threading.o: \
ggml/src/ggml-threading.cpp \
ggml/include/ggml.h
$(CXX) $(XXCFLAGS) -c $< -o $@
ggml/src/ggml-cpu/ggml-cpu.o: \
ggml/src/ggml-cpu/ggml-cpu.c \
ggml/include/ggml.h \
ggml/src/ggml-common.h
$(CC) $(CFLAGS) -c $< -o $@
ggml/src/ggml-cpu/ggml-cpu-cpp.o: \
ggml/src/ggml-cpu/ggml-cpu.cpp \
ggml/include/ggml.h \
ggml/src/ggml-common.h
$(CXX) $(CXXFLAGS) -c $< -o $@
ggml/src/ggml-alloc.o: \
ggml/src/ggml-alloc.c \
ggml/include/ggml.h \
@ -1080,22 +1168,22 @@ ggml/src/ggml-aarch64.o: \
ggml/src/ggml-common.h
$(CC) $(CFLAGS) -c $< -o $@
ggml/src/ggml-blas.o: \
ggml/src/ggml-blas.cpp \
ggml/src/ggml-blas/ggml-blas.o: \
ggml/src/ggml-blas/ggml-blas.cpp \
ggml/include/ggml-blas.h
$(CXX) $(CXXFLAGS) -c $< -o $@
ifndef GGML_NO_LLAMAFILE
ggml/src/llamafile/sgemm.o: \
ggml/src/llamafile/sgemm.cpp \
ggml/src/llamafile/sgemm.h \
ggml/src/ggml-cpu/llamafile/sgemm.o: \
ggml/src/ggml-cpu/llamafile/sgemm.cpp \
ggml/src/ggml-cpu/llamafile/sgemm.h \
ggml/include/ggml.h
$(CXX) $(CXXFLAGS) -c $< -o $@
$(CXX) $(CXXFLAGS) -c $< -o $@ -I ggml/src -I ggml/src/ggml-cpu
endif # GGML_NO_LLAMAFILE
ifndef GGML_NO_AMX
ggml/src/ggml-amx.o: \
ggml/src/ggml-amx.cpp \
ggml/src/ggml-amx/ggml-amx.o: \
ggml/src/ggml-amx/ggml-amx.cpp \
ggml/include/ggml-amx.h
$(CXX) $(CXXFLAGS) -c $< -o $@
@ -1246,13 +1334,24 @@ clean:
rm -rvf ggml/*.a
rm -rvf ggml/*.dll
rm -rvf ggml/*.so
rm -vrf ggml/src/*.o
rm -rvf ggml/src/llamafile/*.o
rm -rvf ggml/src/*.o
rm -rvf common/build-info.cpp
rm -vrf ggml/src/ggml-metal-embed.metal
rm -rvf ggml/src/ggml-cpu/*.o
rm -rvf ggml/src/ggml-cpu/llamafile/*.o
rm -vrf ggml/src/ggml-amx/*.o
rm -vrf ggml/src/ggml-blas/*.o
rm -vrf ggml/src/ggml-cann/*.o
rm -vrf ggml/src/ggml-cpu/*.o
rm -vrf ggml/src/ggml-cuda/*.o
rm -vrf ggml/src/ggml-cuda/template-instances/*.o
rm -vrf ggml/src/ggml-amx/*.o
rm -vrf ggml/src/ggml-hip/*.o
rm -vrf ggml/src/ggml-kompute/*.o
rm -vrf ggml/src/ggml-metal/*.o
rm -vrf ggml/src/ggml-metal/ggml-metal-embed.metal
rm -vrf ggml/src/ggml-rpc/*.o
rm -vrf ggml/src/ggml-sycl/*.o
rm -vrf ggml/src/ggml-vulkan/*.o
rm -vrf ggml/src/ggml-musa/*.o
rm -rvf $(BUILD_TARGETS)
rm -rvf $(TEST_TARGETS)
rm -f vulkan-shaders-gen ggml/src/ggml-vulkan-shaders.hpp ggml/src/ggml-vulkan-shaders.cpp
@ -1455,22 +1554,13 @@ llama-server: \
examples/server/server.cpp \
examples/server/utils.hpp \
examples/server/httplib.h \
examples/server/colorthemes.css.hpp \
examples/server/style.css.hpp \
examples/server/theme-beeninorder.css.hpp \
examples/server/theme-ketivah.css.hpp \
examples/server/theme-mangotango.css.hpp \
examples/server/theme-playground.css.hpp \
examples/server/theme-polarnight.css.hpp \
examples/server/theme-snowstorm.css.hpp \
examples/server/index.html.hpp \
examples/server/index-new.html.hpp \
examples/server/index.js.hpp \
examples/server/completion.js.hpp \
examples/server/system-prompts.js.hpp \
examples/server/prompt-formats.js.hpp \
examples/server/json-schema-to-grammar.mjs.hpp \
examples/server/loading.html.hpp \
examples/server/deps_daisyui.min.css.hpp \
examples/server/deps_markdown-it.js.hpp \
examples/server/deps_tailwindcss.js.hpp \
examples/server/deps_vue.esm-browser.js.hpp \
common/json.hpp \
common/stb_image.h \
$(OBJ_ALL)

19
Package.swift
View file

@ -10,11 +10,16 @@ var sources = [
"src/unicode.cpp",
"src/unicode-data.cpp",
"ggml/src/ggml.c",
"ggml/src/ggml-cpu.c",
"ggml/src/ggml-aarch64.c",
"ggml/src/ggml-alloc.c",
"ggml/src/ggml-backend.cpp",
"ggml/src/ggml-backend-reg.cpp",
"ggml/src/ggml-cpu/ggml-cpu.c",
"ggml/src/ggml-cpu/ggml-cpu.cpp",
"ggml/src/ggml-cpu/ggml-cpu-aarch64.c",
"ggml/src/ggml-cpu/ggml-cpu-quants.c",
"ggml/src/ggml-threading.cpp",
"ggml/src/ggml-quants.c",
"ggml/src/ggml-aarch64.c",
]
var resources: [Resource] = []
@ -22,6 +27,7 @@ var linkerSettings: [LinkerSetting] = []
var cSettings: [CSetting] = [
.unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
.unsafeFlags(["-fno-objc-arc"]),
.headerSearchPath("ggml/src"),
// NOTE: NEW_LAPACK will required iOS version 16.4+
// We should consider add this in the future when we drop support for iOS 14
// (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
@ -30,8 +36,9 @@ var cSettings: [CSetting] = [
]
#if canImport(Darwin)
sources.append("ggml/src/ggml-metal.m")
resources.append(.process("ggml/src/ggml-metal.metal"))
sources.append("ggml/src/ggml-common.h")
sources.append("ggml/src/ggml-metal/ggml-metal.m")
resources.append(.process("ggml/src/ggml-metal/ggml-metal.metal"))
linkerSettings.append(.linkedFramework("Accelerate"))
cSettings.append(
contentsOf: [
@ -61,13 +68,15 @@ let package = Package(
name: "llama",
path: ".",
exclude: [
"build",
"cmake",
"examples",
"scripts",
"models",
"tests",
"CMakeLists.txt",
"Makefile"
"Makefile",
"ggml/src/ggml-metal-embed.metal"
],
sources: sources,
resources: resources,

1
README.md
View file

@ -131,6 +131,7 @@ Typically finetunes of the base models below are supported as well.
- Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
- Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
- Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
- Flutter: [xuegao-tzx/Fllama](https://github.com/xuegao-tzx/Fllama)
- PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326)
- Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
- Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)

2
ci/run.sh
View file

@ -39,7 +39,7 @@ SRC=`pwd`
CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
if [ ! -z ${GG_BUILD_METAL} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
fi
if [ ! -z ${GG_BUILD_CUDA} ]; then

2
cmake/llama-config.cmake.in
View file

@ -6,7 +6,7 @@ set(LLAMA_SHARED_LIB @BUILD_SHARED_LIBS@)
set(GGML_BLAS @GGML_BLAS@)
set(GGML_CUDA @GGML_CUDA@)
set(GGML_METAL @GGML_METAL@)
set(GGML_HIPBLAS @GGML_HIPBLAS@)
set(GGML_HIP @GGML_HIP@)
set(GGML_ACCELERATE @GGML_ACCELERATE@)
set(GGML_VULKAN @GGML_VULKAN@)
set(GGML_VULKAN_CHECK_RESULTS @GGML_VULKAN_CHECK_RESULTS@)

8
common/common.cpp
View file

@ -1003,6 +1003,9 @@ static ggml_type kv_cache_type_from_str(const std::string & s) {
if (s == "f16") {
return GGML_TYPE_F16;
}
if (s == "bf16") {
return GGML_TYPE_BF16;
}
if (s == "q8_0") {
return GGML_TYPE_Q8_0;
}
@ -1964,18 +1967,13 @@ void yaml_dump_non_result_info(FILE * stream, const common_params & params, cons
fprintf(stream, "cpu_has_avx512: %s\n", ggml_cpu_has_avx512() ? "true" : "false");
fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
fprintf(stream, "cpu_has_cuda: %s\n", ggml_cpu_has_cuda() ? "true" : "false");
fprintf(stream, "cpu_has_vulkan: %s\n", ggml_cpu_has_vulkan() ? "true" : "false");
fprintf(stream, "cpu_has_kompute: %s\n", ggml_cpu_has_kompute() ? "true" : "false");
fprintf(stream, "cpu_has_fma: %s\n", ggml_cpu_has_fma() ? "true" : "false");
fprintf(stream, "cpu_has_gpublas: %s\n", ggml_cpu_has_gpublas() ? "true" : "false");
fprintf(stream, "cpu_has_neon: %s\n", ggml_cpu_has_neon() ? "true" : "false");
fprintf(stream, "cpu_has_sve: %s\n", ggml_cpu_has_sve() ? "true" : "false");
fprintf(stream, "cpu_has_f16c: %s\n", ggml_cpu_has_f16c() ? "true" : "false");
fprintf(stream, "cpu_has_fp16_va: %s\n", ggml_cpu_has_fp16_va() ? "true" : "false");
fprintf(stream, "cpu_has_riscv_v: %s\n", ggml_cpu_has_riscv_v() ? "true" : "false");
fprintf(stream, "cpu_has_wasm_simd: %s\n", ggml_cpu_has_wasm_simd() ? "true" : "false");
fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
fprintf(stream, "cpu_has_sse3: %s\n", ggml_cpu_has_sse3() ? "true" : "false");
fprintf(stream, "cpu_has_vsx: %s\n", ggml_cpu_has_vsx() ? "true" : "false");
fprintf(stream, "cpu_has_matmul_int8: %s\n", ggml_cpu_has_matmul_int8() ? "true" : "false");

2
common/common.h
View file

@ -178,7 +178,7 @@ struct common_params {
float yarn_beta_fast = 32.0f; // YaRN low correction dim
float yarn_beta_slow = 1.0f; // YaRN high correction dim
int32_t yarn_orig_ctx = 0; // YaRN original context length
float defrag_thold = -1.0f; // KV cache defragmentation threshold
float defrag_thold = 0.1f; // KV cache defragmentation threshold
struct cpu_params cpuparams;
struct cpu_params cpuparams_batch;

6
convert_hf_to_gguf.py
View file

@ -3748,10 +3748,7 @@ class JaisModel(Model):
# Embeddings scale
self.embeddings_scale = 1.0
# note: For some JAIS flavors, output is tied to (same as) wte in original model
self.output_is_wte = False
if 'mup_embeddings_scale' in self.hparams:
self.output_is_wte = True # Hack (?)
self.embeddings_scale = self.hparams['mup_embeddings_scale']
elif 'embeddings_scale' in self.hparams:
self.embeddings_scale = self.hparams['embeddings_scale']
@ -3808,10 +3805,7 @@ class JaisModel(Model):
if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
tensors.append((new_name, data_torch * self.embeddings_scale))
if self.output_is_wte:
tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch * self.width_scale))
elif new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
assert not self.output_is_wte
tensors.append((new_name, data_torch * self.width_scale))
else:
tensors.append((new_name, data_torch))

4
docs/backend/SYCL.md
View file

@ -41,6 +41,8 @@ The following release is verified with good quality:
## News
- 2024.11
- Use syclcompat to improve the performance on some platforms. This requires to use oneAPI 2025.0 or newer.
- 2024.8
- Use oneDNN as the default GEMM library, improve the compatibility for new Intel GPUs.
@ -380,7 +382,7 @@ found 2 SYCL devices:
|Chosen Device ID|Setting|
|-|-|
|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:0"` or no action|
|1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
|0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|

8
docs/build.md
View file

@ -230,7 +230,7 @@ You can download it from your Linux distro's package manager or from here: [ROCm
- Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
```bash
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
&& cmake --build build --config Release -- -j 16
```
On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
@ -247,7 +247,7 @@ You can download it from your Linux distro's package manager or from here: [ROCm
```bash
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
&& cmake --build build -- -j 16
```
@ -259,7 +259,7 @@ You can download it from your Linux distro's package manager or from here: [ROCm
- Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
```bash
set PATH=%HIP_PATH%\bin;%PATH%
cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIP=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
cmake --build build
```
Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
@ -375,7 +375,7 @@ cmake --build build --config release
You can test with:
`./build/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32`
`./build/bin/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32`
If the fllowing info is output on screen, you are using `llama.cpp by CANN backend`:
```bash

12
examples/chat-persistent.sh
View file

@ -23,8 +23,9 @@ CUR_PROMPT_CACHE="${CHAT_SAVE_DIR}/current-cache.bin"
NEXT_PROMPT_FILE="${CHAT_SAVE_DIR}/next-prompt.txt"
NEXT_PROMPT_CACHE="${CHAT_SAVE_DIR}/next-cache.bin"
SESSION_SIZE_MSG_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'
SAMPLE_TIME_MSG_PATTERN='sample time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
SESSION_AND_SAMPLE_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'\
'|'\
'sampling time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d"
CTX_SIZE=2048
@ -129,15 +130,12 @@ while read -e line; do
printf ' '
# HACK get num tokens from debug message
# TODO get both messages in one go
if ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
if ! session_and_sample_msg=$(tail -n30 "$LOG" | grep -oE "$SESSION_AND_SAMPLE_PATTERN"); then
echo >&2 "Couldn't get number of tokens from ./llama-cli output!"
exit 1
fi
n_tokens=$(($(cut -d/ -f2 <<<"$session_size_msg") + $(cut -d/ -f2 <<<"$sample_time_msg")))
n_tokens=$(awk '{sum+=$1} END {print sum}' <<< "$(cut -d/ -f2 <<< "$session_and_sample_msg")")
if ((n_tokens > CTX_ROTATE_POINT)); then
tail -c+$((n_prompt_len_pre + 1)) "$CUR_PROMPT_FILE" >>"$NEXT_PROMPT_FILE"

26
examples/convert_legacy_llama.py
View file

@ -840,6 +840,8 @@ class OutputFile:
self.gguf.add_base_model_version(key, base_model_entry["version"])
if "organization" in base_model_entry:
self.gguf.add_base_model_organization(key, base_model_entry["organization"])
if "description" in base_model_entry:
self.gguf.add_base_model_description(key, base_model_entry["description"])
if "url" in base_model_entry:
self.gguf.add_base_model_url(key, base_model_entry["url"])
if "doi" in base_model_entry:
@ -849,12 +851,32 @@ class OutputFile:
if "repo_url" in base_model_entry:
self.gguf.add_base_model_repo_url(key, base_model_entry["repo_url"])
if metadata.datasets is not None:
self.gguf.add_dataset_count(len(metadata.datasets))
for key, dataset_entry in enumerate(metadata.datasets):
if "name" in dataset_entry:
self.gguf.add_dataset_name(key, dataset_entry["name"])
if "author" in dataset_entry:
self.gguf.add_dataset_author(key, dataset_entry["author"])
if "version" in dataset_entry:
self.gguf.add_dataset_version(key, dataset_entry["version"])
if "organization" in dataset_entry:
self.gguf.add_dataset_organization(key, dataset_entry["organization"])
if "description" in dataset_entry:
self.gguf.add_dataset_description(key, dataset_entry["description"])
if "url" in dataset_entry:
self.gguf.add_dataset_url(key, dataset_entry["url"])
if "doi" in dataset_entry:
self.gguf.add_dataset_doi(key, dataset_entry["doi"])
if "uuid" in dataset_entry:
self.gguf.add_dataset_uuid(key, dataset_entry["uuid"])
if "repo_url" in dataset_entry:
self.gguf.add_dataset_repo_url(key, dataset_entry["repo_url"])
if metadata.tags is not None:
self.gguf.add_tags(metadata.tags)
if metadata.languages is not None:
self.gguf.add_languages(metadata.languages)
if metadata.datasets is not None:
self.gguf.add_datasets(metadata.datasets)
def add_meta_arch(self, params: Params) -> None:
# Metadata About The Neural Architecture Itself

35
examples/llama-bench/llama-bench.cpp
View file

@ -256,6 +256,9 @@ static ggml_type ggml_type_from_name(const std::string & s) {
if (s == "f16") {
return GGML_TYPE_F16;
}
if (s == "bf16") {
return GGML_TYPE_BF16;
}
if (s == "q8_0") {
return GGML_TYPE_Q8_0;
}
@ -771,13 +774,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
struct test {
static const std::string build_commit;
static const int build_number;
static const bool cuda;
static const bool vulkan;
static const bool kompute;
static const bool metal;
static const bool sycl;
static const bool gpu_blas;
static const bool blas;
static const std::string cpu_info;
static const std::string gpu_info;
std::string model_filename;
@ -790,7 +786,6 @@ struct test {
std::string cpu_mask;
bool cpu_strict;
int poll;
bool has_rpc;
ggml_type type_k;
ggml_type type_v;
int n_gpu_layers;
@ -819,7 +814,6 @@ struct test {
cpu_mask = inst.cpu_mask;
cpu_strict = inst.cpu_strict;
poll = inst.poll;
has_rpc = !inst.rpc_servers.empty();
type_k = inst.type_k;
type_v = inst.type_v;
n_gpu_layers = inst.n_gpu_layers;
@ -878,8 +872,7 @@ struct test {
static const std::vector<std::string> & get_fields() {
static const std::vector<std::string> fields = {
"build_commit", "build_number",
"cuda", "vulkan", "kompute", "metal", "sycl", "rpc", "gpu_blas", "blas",
"cpu_info", "gpu_info",
"cpu_info", "gpu_info", "backends",
"model_filename", "model_type", "model_size", "model_n_params",
"n_batch", "n_ubatch",
"n_threads", "cpu_mask", "cpu_strict", "poll",
@ -905,8 +898,7 @@ struct test {
field == "avg_ns" || field == "stddev_ns") {
return INT;
}
if (field == "cuda" || field == "vulkan" || field == "kompute" || field == "metal" ||
field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" ||
if (field == "f16_kv" || field == "no_kv_offload" ||
field == "cpu_strict" ||
field == "flash_attn" || field == "use_mmap" || field == "embeddings") {
return BOOL;
@ -935,9 +927,7 @@ struct test {
}
std::vector<std::string> values = {
build_commit, std::to_string(build_number),
std::to_string(cuda), std::to_string(vulkan), std::to_string(vulkan),
std::to_string(metal), std::to_string(sycl), std::to_string(has_rpc), std::to_string(gpu_blas), std::to_string(blas),
cpu_info, gpu_info,
cpu_info, gpu_info, get_backend(),
model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
std::to_string(n_batch), std::to_string(n_ubatch),
std::to_string(n_threads), cpu_mask, std::to_string(cpu_strict), std::to_string(poll),
@ -964,13 +954,6 @@ struct test {
const std::string test::build_commit = LLAMA_COMMIT;
const int test::build_number = LLAMA_BUILD_NUMBER;
const bool test::cuda = !!ggml_cpu_has_cuda();
const bool test::vulkan = !!ggml_cpu_has_vulkan();
const bool test::kompute = !!ggml_cpu_has_kompute();
const bool test::metal = !!ggml_cpu_has_metal();
const bool test::gpu_blas = !!ggml_cpu_has_gpublas();
const bool test::blas = !!ggml_cpu_has_blas();
const bool test::sycl = !!ggml_cpu_has_sycl();
const std::string test::cpu_info = get_cpu_info();
const std::string test::gpu_info = get_gpu_info();
@ -1175,7 +1158,8 @@ struct markdown_printer : public printer {
fields.emplace_back("size");
fields.emplace_back("params");
fields.emplace_back("backend");
bool is_cpu_backend = test::get_backend() == "CPU" || test::get_backend() == "BLAS";
bool is_cpu_backend = test::get_backend().find("CPU") != std::string::npos ||
test::get_backend().find("BLAS") != std::string::npos;
if (!is_cpu_backend) {
fields.emplace_back("n_gpu_layers");
}
@ -1265,9 +1249,6 @@ struct markdown_printer : public printer {
value = buf;
} else if (field == "backend") {
value = test::get_backend();
if (t.has_rpc) {
value += "+RPC";
}
} else if (field == "test") {
if (t.n_prompt > 0 && t.n_gen == 0) {
snprintf(buf, sizeof(buf), "pp%d", t.n_prompt);

19
examples/quantize-stats/quantize-stats.cpp
View file

@ -142,7 +142,7 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
}
static void test_roundtrip_on_chunk(
const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits & qfns, bool use_reference,
const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits & qfns, const ggml_type_traits_cpu & qfns_cpu, bool use_reference,
float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats
) {
if (layer->type == GGML_TYPE_F16) {
@ -156,7 +156,7 @@ static void test_roundtrip_on_chunk(
if (use_reference) {
qfns.from_float_ref(input_scratch, quantized_scratch, chunk_size);
} else {
qfns.from_float(input_scratch, quantized_scratch, chunk_size);
qfns_cpu.from_float(input_scratch, quantized_scratch, chunk_size);
}
qfns.to_float(quantized_scratch, output_scratch, chunk_size);
@ -166,7 +166,7 @@ static void test_roundtrip_on_chunk(
// Run quantization function for a single layer and update error stats
static void test_roundtrip_on_layer(
std::string & name, bool print_layer_stats, const ggml_type_traits & qfns, bool use_reference,
std::string & name, bool print_layer_stats, const ggml_type_traits & qfns, const ggml_type_traits_cpu & qfns_cpu, bool use_reference,
const ggml_tensor * layer, std::vector<float> & input_scratch, std::vector<char> & quantized_scratch,
std::vector<float> & output_scratch, error_stats & total_error, int max_thread = 0
) {
@ -187,13 +187,13 @@ static void test_roundtrip_on_layer(
int num_chunks = (nelements + chunk_size - 1)/chunk_size;
if (num_chunks < 2 || max_thread < 2) {
test_roundtrip_on_chunk(layer, 0, nelements, qfns, use_reference, input_scratch_ptr, quantized_scratch.data(),
test_roundtrip_on_chunk(layer, 0, nelements, qfns, qfns_cpu, use_reference, input_scratch_ptr, quantized_scratch.data(),
output_scratch.data(), print_layer_stats ? layer_error : total_error);
} else {
auto & stats = print_layer_stats ? layer_error : total_error;
std::mutex mutex;
uint64_t counter = 0;
auto compute = [&mutex, &counter, &stats, &qfns, nelements, layer, use_reference, input_scratch_ptr,
auto compute = [&mutex, &counter, &stats, &qfns, &qfns_cpu, nelements, layer, use_reference, input_scratch_ptr,
&quantized_scratch, &output_scratch, chunk_size] () {
error_stats local_stats {};
while (true) {
@ -205,7 +205,7 @@ static void test_roundtrip_on_layer(
}
lock.unlock();
uint64_t chunk = offset + chunk_size < nelements ? chunk_size : nelements - offset;
test_roundtrip_on_chunk(layer, offset, chunk, qfns, use_reference, input_scratch_ptr + offset,
test_roundtrip_on_chunk(layer, offset, chunk, qfns, qfns_cpu, use_reference, input_scratch_ptr + offset,
quantized_scratch.data() + 4*offset, output_scratch.data() + offset, local_stats);
}
};
@ -371,8 +371,9 @@ int main(int argc, char ** argv) {
if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
continue;
}
const auto * qfns = ggml_get_type_traits(type);
if (qfns->from_float && qfns->to_float) {
const auto * qfns = ggml_get_type_traits(type);
const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
if (qfns_cpu->from_float && qfns->to_float) {
if (params.verbose) {
printf("testing %s ...\n", ggml_type_name(type));
}
@ -393,7 +394,7 @@ int main(int argc, char ** argv) {
test_roundtrip_on_layer(
layer_name,
params.per_layer_stats,
*qfns,
*qfns, *qfns_cpu,
params.reference,
kv_tensor.second,
input_scratch,

17
examples/server/CMakeLists.txt
View file

@ -15,22 +15,13 @@ set(TARGET_SRCS
httplib.h
)
set(PUBLIC_ASSETS
colorthemes.css
style.css
theme-beeninorder.css
theme-ketivah.css
theme-mangotango.css
theme-playground.css
theme-polarnight.css
theme-snowstorm.css
index.html
index-new.html
index.js
completion.js
system-prompts.js
prompt-formats.js
json-schema-to-grammar.mjs
loading.html
deps_daisyui.min.css
deps_markdown-it.js
deps_tailwindcss.js
deps_vue.esm-browser.js
)
foreach(asset ${PUBLIC_ASSETS})

36
examples/server/README.md
View file

@ -39,7 +39,7 @@ The project is under active development, and we are [looking for feedback and co
| `--cpu-strict-batch <0\|1>` | use strict CPU placement (default: same as --cpu-strict) |
| `--prio-batch N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)<br/> |
| `--poll-batch <0\|1>` | use polling to wait for work (default: same as --poll) |
| `-c, --ctx-size N` | size of the prompt context (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE) |
| `-c, --ctx-size N` | size of the prompt context (default: 4096, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE) |
| `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)<br/>(env: LLAMA_ARG_N_PREDICT) |
| `-b, --batch-size N` | logical maximum batch size (default: 2048)<br/>(env: LLAMA_ARG_BATCH) |
| `-ub, --ubatch-size N` | physical maximum batch size (default: 512)<br/>(env: LLAMA_ARG_UBATCH) |
@ -64,7 +64,7 @@ The project is under active development, and we are [looking for feedback and co
| `-nkvo, --no-kv-offload` | disable KV offload<br/>(env: LLAMA_ARG_NO_KV_OFFLOAD) |
| `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
| `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V) |
| `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: -1.0, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
| `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: 0.1, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
| `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
| `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
| `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_NO_MMAP) |
@ -99,25 +99,27 @@ The project is under active development, and we are [looking for feedback and co
| Argument | Explanation |
| -------- | ----------- |
| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;typ_p;top_p;min_p;temperature) |
| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: dry;top_k;typ_p;top_p;min_p;xtc;temperature) |
| `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
| `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) |
| `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: dkypmxt) |
| `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
| `--penalize-nl` | penalize newline tokens (default: false) |
| `--temp N` | temperature (default: 0.8) |
| `--top-k N` | top-k sampling (default: 40, 0 = disabled) |
| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) |
| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) |
| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
| `--dry-multiplier N` | DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
| `--dry-base N` | DRY sampling base value (default: 1.75) |
| `--dry-allowed-length N` | allowed length for DRY sampling (default: 2) |
| `--dry-penalty-last-n N` | DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
| `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers (`['\n', ':', '"', '*']`) in the process; use `"none"` to not use any sequence breakers
| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
| `--dry-base N` | set DRY sampling base value (default: 1.75) |
| `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
| `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
| `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers<br/> |
| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
| `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
@ -381,6 +383,10 @@ node index.js
`dry_sequence_breakers`: Specify an array of sequence breakers for DRY sampling. Only a JSON array of strings is accepted. Default: `['\n', ':', '"', '*']`
`xtc_probability`: Set the chance for token removal via XTC sampler. Default: `0.0`, which is disabled.
`xtc_threshold`: Set a minimum probability threshold for tokens to be removed via XTC sampler. Default: `0.1` (> `0.5` disables XTC)
`mirostat`: Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.
`mirostat_tau`: Set the Mirostat target entropy, parameter tau. Default: `5.0`
@ -409,7 +415,7 @@ node index.js
`cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `false`
`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", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
`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: `["dry", "top_k", "typ_p", "top_p", "min_p", "xtc", "temperature"]` - these are all the available values.
**Response format**
@ -928,6 +934,16 @@ Apart from error types supported by OAI, we also have custom types that are spec
}
```
### Legacy completion web UI
A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggerganov/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./examples/server/public_legacy`
For example:
```sh
./llama-server -m my_model.gguf -c 8192 --path ./examples/server/public_legacy
```
### Extending or building alternative Web Front End
You can extend the front end by running the server binary with `--path` set to `./your-directory` and importing `/completion.js` to get access to the llamaComplete() method.

2
examples/server/chat.mjs
View file

@ -1,7 +1,7 @@
import * as readline from 'node:readline'
import { stdin, stdout } from 'node:process'
import { readFileSync } from 'node:fs'
import { SchemaConverter } from './public/json-schema-to-grammar.mjs'
import { SchemaConverter } from './public_legacy/json-schema-to-grammar.mjs'
const args = process.argv.slice(2);
const grammarJsonSchemaFile = args.find(

19
examples/server/deps.sh
View file

@ -6,5 +6,20 @@ DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
PUBLIC=$DIR/public
echo "download js bundle files"
curl https://npm.reversehttp.com/@preact/signals-core,@preact/signals,htm/preact,preact,preact/hooks > $PUBLIC/index.js
echo >> $PUBLIC/index.js # add newline
# Note for contributors: Always pin to a specific version "maj.min.patch" to avoid breaking the CI
curl -L https://cdn.tailwindcss.com/3.4.14 > $PUBLIC/deps_tailwindcss.js
echo >> $PUBLIC/deps_tailwindcss.js # add newline
curl -L https://cdnjs.cloudflare.com/ajax/libs/daisyui/4.12.14/styled.min.css > $PUBLIC/deps_daisyui.min.css
curl -L https://cdnjs.cloudflare.com/ajax/libs/daisyui/4.12.14/themes.min.css >> $PUBLIC/deps_daisyui.min.css
echo >> $PUBLIC/deps_daisyui.min.css # add newline
curl -L https://unpkg.com/vue@3.5.12/dist/vue.esm-browser.js > $PUBLIC/deps_vue.esm-browser.js
echo >> $PUBLIC/deps_vue.esm-browser.js # add newline
curl -L https://cdnjs.cloudflare.com/ajax/libs/markdown-it/13.0.2/markdown-it.js > $PUBLIC/deps_markdown-it.js
echo >> $PUBLIC/deps_markdown-it.js # add newline
ls -lah $PUBLIC

29
examples/server/public/completion.js
View file

@ -1,12 +1,16 @@
const paramDefaults = {
stream: true,
n_predict: 500,
temperature: 0.2,
stop: ["</s>"]
};
let generation_settings = null;
export class CompletionError extends Error {
constructor(message, name, data) {
super(message);
this.name = name;
}
};
// Completes the prompt as a generator. Recommended for most use cases.
//
@ -29,7 +33,7 @@ export async function* llama(prompt, params = {}, config = {}) {
const completionParams = { ...paramDefaults, ...params, prompt };
const response = await fetch(`${api_url}/completion`, {
const response = await fetch(`${api_url}${config.endpoint || '/completion'}`, {
method: 'POST',
body: JSON.stringify(completionParams),
headers: {
@ -41,6 +45,18 @@ export async function* llama(prompt, params = {}, config = {}) {
signal: controller.signal,
});
const status = response.status;
if (status !== 200) {
try {
const body = await response.json();
if (body && body.error && body.error.message) {
throw new CompletionError(body.error.message, 'ServerError');
}
} catch (err) {
throw new CompletionError(err.message, 'ServerError');
}
}
const reader = response.body.getReader();
const decoder = new TextDecoder();
@ -78,7 +94,12 @@ export async function* llama(prompt, params = {}, config = {}) {
for (const line of lines) {
const match = regex.exec(line);
if (match) {
result[match[1]] = match[2]
result[match[1]] = match[2];
if (result.data === '[DONE]') {
cont = false;
break;
}
// since we know this is llama.cpp, let's just decode the json in data
if (result.data) {
result.data = JSON.parse(result.data);

13
examples/server/public/deps_daisyui.min.css vendored Normal file
View file

File diff suppressed because one or more lines are too long

8442
examples/server/public/deps_markdown-it.js Normal file
View file

File diff suppressed because it is too large Load diff

82
examples/server/public/deps_tailwindcss.js Normal file
View file

File diff suppressed because one or more lines are too long

18160
examples/server/public/deps_vue.esm-browser.js Normal file
View file

File diff suppressed because it is too large Load diff

1939
examples/server/public/index.html
View file

File diff suppressed because it is too large Load diff

0
examples/server/public/colorthemes.css → examples/server/public_legacy/colorthemes.css
View file

209
examples/server/public_legacy/completion.js Normal file
View file

@ -0,0 +1,209 @@
const paramDefaults = {
stream: true,
n_predict: 500,
temperature: 0.2,
stop: ["</s>"]
};
let generation_settings = null;
// Completes the prompt as a generator. Recommended for most use cases.
//
// Example:
//
// import { llama } from '/completion.js'
//
// const request = llama("Tell me a joke", {n_predict: 800})
// for await (const chunk of request) {
// document.write(chunk.data.content)
// }
//
export async function* llama(prompt, params = {}, config = {}) {
let controller = config.controller;
const api_url = config.api_url?.replace(/\/+$/, '') || "";
if (!controller) {
controller = new AbortController();
}
const completionParams = { ...paramDefaults, ...params, prompt };
const response = await fetch(`${api_url}${config.endpoint || '/completion'}`, {
method: 'POST',
body: JSON.stringify(completionParams),
headers: {
'Connection': 'keep-alive',
'Content-Type': 'application/json',
'Accept': 'text/event-stream',
...(params.api_key ? {'Authorization': `Bearer ${params.api_key}`} : {})
},
signal: controller.signal,
});
const reader = response.body.getReader();
const decoder = new TextDecoder();
let content = "";
let leftover = ""; // Buffer for partially read lines
try {
let cont = true;
while (cont) {
const result = await reader.read();
if (result.done) {
break;
}
// Add any leftover data to the current chunk of data
const text = leftover + decoder.decode(result.value);
// Check if the last character is a line break
const endsWithLineBreak = text.endsWith('\n');
// Split the text into lines
let lines = text.split('\n');
// If the text doesn't end with a line break, then the last line is incomplete
// Store it in leftover to be added to the next chunk of data
if (!endsWithLineBreak) {
leftover = lines.pop();
} else {
leftover = ""; // Reset leftover if we have a line break at the end
}
// Parse all sse events and add them to result
const regex = /^(\S+):\s(.*)$/gm;
for (const line of lines) {
const match = regex.exec(line);
if (match) {
result[match[1]] = match[2];
if (result.data === '[DONE]') {
cont = false;
break;
}
// since we know this is llama.cpp, let's just decode the json in data
if (result.data) {
result.data = JSON.parse(result.data);
content += result.data.content;
// yield
yield result;
// if we got a stop token from server, we will break here
if (result.data.stop) {
if (result.data.generation_settings) {
generation_settings = result.data.generation_settings;
}
cont = false;
break;
}
}
if (result.error) {
try {
result.error = JSON.parse(result.error);
if (result.error.message.includes('slot unavailable')) {
// Throw an error to be caught by upstream callers
throw new Error('slot unavailable');
} else {
console.error(`llama.cpp error [${result.error.code} - ${result.error.type}]: ${result.error.message}`);
}
} catch(e) {
console.error(`llama.cpp error ${result.error}`)
}
}
}
}
}
} catch (e) {
if (e.name !== 'AbortError') {
console.error("llama error: ", e);
}
throw e;
}
finally {
controller.abort();
}
return content;
}
// Call llama, return an event target that you can subscribe to
//
// Example:
//
// import { llamaEventTarget } from '/completion.js'
//
// const conn = llamaEventTarget(prompt)
// conn.addEventListener("message", (chunk) => {
// document.write(chunk.detail.content)
// })
//
export const llamaEventTarget = (prompt, params = {}, config = {}) => {
const eventTarget = new EventTarget();
(async () => {
let content = "";
for await (const chunk of llama(prompt, params, config)) {
if (chunk.data) {
content += chunk.data.content;
eventTarget.dispatchEvent(new CustomEvent("message", { detail: chunk.data }));
}
if (chunk.data.generation_settings) {
eventTarget.dispatchEvent(new CustomEvent("generation_settings", { detail: chunk.data.generation_settings }));
}
if (chunk.data.timings) {
eventTarget.dispatchEvent(new CustomEvent("timings", { detail: chunk.data.timings }));
}
}
eventTarget.dispatchEvent(new CustomEvent("done", { detail: { content } }));
})();
return eventTarget;
}
// Call llama, return a promise that resolves to the completed text. This does not support streaming
//
// Example:
//
// llamaPromise(prompt).then((content) => {
// document.write(content)
// })
//
// or
//
// const content = await llamaPromise(prompt)
// document.write(content)
//
export const llamaPromise = (prompt, params = {}, config = {}) => {
return new Promise(async (resolve, reject) => {
let content = "";
try {
for await (const chunk of llama(prompt, params, config)) {
content += chunk.data.content;
}
resolve(content);
} catch (error) {
reject(error);
}
});
};
/**
* (deprecated)
*/
export const llamaComplete = async (params, controller, callback) => {
for await (const chunk of llama(params.prompt, params, { controller })) {
callback(chunk);
}
}
// Get the model info from the server. This is useful for getting the context window and so on.
export const llamaModelInfo = async (config = {}) => {
if (!generation_settings) {
const api_url = config.api_url?.replace(/\/+$/, '') || "";
const props = await fetch(`${api_url}/props`).then(r => r.json());
generation_settings = props.default_generation_settings;
}
return generation_settings;
}

0
examples/server/public/favicon.ico → examples/server/public_legacy/favicon.ico
View file

Side by side Swipe Overlay

Before

Width:  |  Height:  |  Size: 4 KiB

After

Width:  |  Height:  |  Size: 4 KiB

Before After
Before After

0
examples/server/public/index-new.html → examples/server/public_legacy/index-new.html
View file

1303
examples/server/public_legacy/index.html Normal file
View file

File diff suppressed because it is too large Load diff

0
examples/server/public/index.js → examples/server/public_legacy/index.js
View file

0
examples/server/public/json-schema-to-grammar.mjs → examples/server/public_legacy/json-schema-to-grammar.mjs
View file

12
examples/server/public_legacy/loading.html Normal file
View file

@ -0,0 +1,12 @@
<!DOCTYPE html>
<html>
<head>
<meta http-equiv="refresh" content="5">
</head>
<body>
<div id="loading">
The model is loading. Please wait.<br/>
The user interface will appear soon.
</div>
</body>
</html>

0
examples/server/public/prompt-formats.js → examples/server/public_legacy/prompt-formats.js
View file

0
examples/server/public/style.css → examples/server/public_legacy/style.css
View file

0
examples/server/public/system-prompts.js → examples/server/public_legacy/system-prompts.js
View file

0
examples/server/public/theme-beeninorder.css → examples/server/public_legacy/theme-beeninorder.css
View file

0
examples/server/public/theme-ketivah.css → examples/server/public_legacy/theme-ketivah.css
View file

0
examples/server/public/theme-mangotango.css → examples/server/public_legacy/theme-mangotango.css
View file

0
examples/server/public/theme-playground.css → examples/server/public_legacy/theme-playground.css
View file

0
examples/server/public/theme-polarnight.css → examples/server/public_legacy/theme-polarnight.css
View file

0
examples/server/public/theme-snowstorm.css → examples/server/public_legacy/theme-snowstorm.css
View file

139
examples/server/server.cpp
View file

@ -14,22 +14,13 @@
#define MIMETYPE_JSON "application/json; charset=utf-8"
// auto generated files (update with ./deps.sh)
#include "colorthemes.css.hpp"
#include "style.css.hpp"
#include "theme-beeninorder.css.hpp"
#include "theme-ketivah.css.hpp"
#include "theme-mangotango.css.hpp"
#include "theme-playground.css.hpp"
#include "theme-polarnight.css.hpp"
#include "theme-snowstorm.css.hpp"
#include "index.html.hpp"
#include "index-new.html.hpp"
#include "index.js.hpp"
#include "completion.js.hpp"
#include "system-prompts.js.hpp"
#include "prompt-formats.js.hpp"
#include "json-schema-to-grammar.mjs.hpp"
#include "loading.html.hpp"
#include "deps_daisyui.min.css.hpp"
#include "deps_markdown-it.js.hpp"
#include "deps_tailwindcss.js.hpp"
#include "deps_vue.esm-browser.js.hpp"
#include <atomic>
#include <condition_variable>
@ -378,8 +369,8 @@ struct server_queue {
std::condition_variable condition_tasks;
// callback functions
std::function<void(server_task&)> callback_new_task;
std::function<void(void)> callback_update_slots;
std::function<void(server_task)> callback_new_task;
std::function<void(void)> callback_update_slots;
// Add a new task to the end of the queue
int post(server_task task, bool front = false) {
@ -431,7 +422,7 @@ struct server_queue {
}
// Register function to process a new task
void on_new_task(std::function<void(server_task &)> callback) {
void on_new_task(std::function<void(server_task)> callback) {
callback_new_task = std::move(callback);
}
@ -481,7 +472,7 @@ struct server_queue {
lock.unlock();
QUE_DBG("processing task, id = %d\n", task.id);
callback_new_task(task);
callback_new_task(std::move(task));
}
// all tasks in the current loop is processed, slots data is now ready
@ -644,17 +635,12 @@ struct server_context {
bool load_model(const common_params & params_) {
params = params_;
// reserve one extra sequence (seq_id == 0) for extra features
params.n_parallel += 1;
common_init_result llama_init = common_init_from_params(params);
model = llama_init.model;
ctx = llama_init.context;
loras = llama_init.lora_adapters;
params.n_parallel -= 1; // but be sneaky about it
if (model == nullptr) {
SRV_ERR("failed to load model, '%s'\n", params.model.c_str());
return false;
@ -669,11 +655,16 @@ struct server_context {
}
bool validate_model_chat_template() const {
llama_chat_message chat[] = {{"user", "test"}};
const int res = llama_chat_apply_template(model, nullptr, chat, 1, true, nullptr, 0);
return res > 0;
std::vector<char> model_template(2048, 0); // longest known template is about 1200 bytes
std::string template_key = "tokenizer.chat_template";
int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
if (res >= 0) {
llama_chat_message chat[] = {{"user", "test"}};
std::string tmpl = std::string(model_template.data(), model_template.size());
int32_t chat_res = llama_chat_apply_template(model, tmpl.c_str(), chat, 1, true, nullptr, 0);
return chat_res > 0;
}
return false;
}
void init() {
@ -1288,16 +1279,16 @@ struct server_context {
void send_embedding(const server_slot & slot, const llama_batch & batch) {
server_task_result res;
res.id = slot.id_task;
res.error = false;
res.stop = true;
res.id = slot.id_task;
res.error = false;
res.stop = true;
const int n_embd = llama_n_embd(model);
std::vector<float> embd_res(n_embd, 0.0f);
for (int i = 0; i < batch.n_tokens; ++i) {
if (!batch.logits[i] || batch.seq_id[i][0] != slot.id + 1) {
if (!batch.logits[i] || batch.seq_id[i][0] != slot.id) {
continue;
}
@ -1332,12 +1323,12 @@ struct server_context {
void send_rerank(const server_slot & slot, const llama_batch & batch) {
server_task_result res;
res.id = slot.id_task;
res.error = false;
res.stop = true;
res.id = slot.id_task;
res.error = false;
res.stop = true;
for (int i = 0; i < batch.n_tokens; ++i) {
if (!batch.logits[i] || batch.seq_id[i][0] != slot.id + 1) {
if (!batch.logits[i] || batch.seq_id[i][0] != slot.id) {
continue;
}
@ -1510,7 +1501,7 @@ struct server_context {
// Functions to process the task
//
void process_single_task(const server_task & task) {
void process_single_task(server_task task) {
switch (task.type) {
case SERVER_TASK_TYPE_INFERENCE:
{
@ -1646,7 +1637,7 @@ struct server_context {
std::string filename = task.data.at("filename");
std::string filepath = task.data.at("filepath");
const size_t nwrite = llama_state_seq_save_file(ctx, filepath.c_str(), slot->id + 1, slot->cache_tokens.data(), token_count);
const size_t nwrite = llama_state_seq_save_file(ctx, filepath.c_str(), slot->id, slot->cache_tokens.data(), token_count);
const int64_t t_end = ggml_time_us();
const double t_save_ms = (t_end - t_start) / 1000.0;
@ -1688,7 +1679,7 @@ struct server_context {
slot->cache_tokens.resize(slot->n_ctx);
size_t token_count = 0;
size_t nread = llama_state_seq_load_file(ctx, filepath.c_str(), slot->id + 1, slot->cache_tokens.data(), slot->cache_tokens.size(), &token_count);
size_t nread = llama_state_seq_load_file(ctx, filepath.c_str(), slot->id, slot->cache_tokens.data(), slot->cache_tokens.size(), &token_count);
if (nread == 0) {
slot->cache_tokens.resize(0);
send_error(task, "Unable to restore slot, no available space in KV cache or invalid slot save file", ERROR_TYPE_INVALID_REQUEST);
@ -1731,7 +1722,7 @@ struct server_context {
// Erase token cache
const size_t n_erased = slot->cache_tokens.size();
llama_kv_cache_seq_rm(ctx, slot->id + 1, -1, -1);
llama_kv_cache_seq_rm(ctx, slot->id, -1, -1);
slot->cache_tokens.clear();
server_task_result result;
@ -1808,8 +1799,8 @@ struct server_context {
SLT_WRN(slot, "slot context shift, n_keep = %d, n_left = %d, n_discard = %d\n", n_keep, n_left, n_discard);
llama_kv_cache_seq_rm (ctx, slot.id + 1, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, slot.id + 1, n_keep + n_discard, slot.n_past, -n_discard);
llama_kv_cache_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past, -n_discard);
if (slot.params.cache_prompt) {
for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) {
@ -1836,7 +1827,7 @@ struct server_context {
slot.i_batch = batch.n_tokens;
common_batch_add(batch, slot.sampled, slot.n_past, { slot.id + 1 }, true);
common_batch_add(batch, slot.sampled, slot.n_past, { slot.id }, true);
slot.n_past += 1;
@ -1983,8 +1974,8 @@ struct server_context {
const int64_t kv_shift = (int64_t) head_p - (int64_t) head_c;
llama_kv_cache_seq_rm (ctx, slot.id + 1, head_p, head_c);
llama_kv_cache_seq_add(ctx, slot.id + 1, head_c, -1, kv_shift);
llama_kv_cache_seq_rm (ctx, slot.id, head_p, head_c);
llama_kv_cache_seq_add(ctx, slot.id, head_c, -1, kv_shift);
for (size_t i = 0; i < n_match; i++) {
slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i];
@ -2033,9 +2024,9 @@ struct server_context {
}
// keep only the common part
if (!llama_kv_cache_seq_rm(ctx, slot.id + 1, slot.n_past, -1)) {
if (!llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1)) {
// could not partially delete (likely using a non-Transformer model)
llama_kv_cache_seq_rm(ctx, slot.id + 1, -1, -1);
llama_kv_cache_seq_rm(ctx, slot.id, -1, -1);
// there is no common part left
slot.n_past = 0;
@ -2048,7 +2039,7 @@ struct server_context {
// add prompt tokens for processing in the current batch
while (slot.n_past < slot.n_prompt_tokens && batch.n_tokens < n_batch) {
common_batch_add(batch, prompt_tokens[slot.n_past], slot.n_past, { slot.id + 1 }, false);
common_batch_add(batch, prompt_tokens[slot.n_past], slot.n_past, { slot.id }, false);
if (slot.params.cache_prompt) {
slot.cache_tokens.push_back(prompt_tokens[slot.n_past]);
@ -2290,16 +2281,6 @@ int main(int argc, char ** argv) {
std::atomic<server_state> state{SERVER_STATE_LOADING_MODEL};
svr->set_default_headers({{"Server", "llama.cpp"}});
// CORS preflight
svr->Options(R"(.*)", [](const httplib::Request &, httplib::Response & res) {
// Access-Control-Allow-Origin is already set by middleware
res.set_header("Access-Control-Allow-Credentials", "true");
res.set_header("Access-Control-Allow-Methods", "POST");
res.set_header("Access-Control-Allow-Headers", "*");
return res.set_content("", "text/html"); // blank response, no data
});
svr->set_logger(log_server_request);
auto res_error = [](httplib::Response & res, const json & error_data) {
@ -2412,6 +2393,14 @@ int main(int argc, char ** argv) {
// register server middlewares
svr->set_pre_routing_handler([&middleware_validate_api_key, &middleware_server_state](const httplib::Request & req, httplib::Response & res) {
res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
// If this is OPTIONS request, skip validation because browsers don't include Authorization header
if (req.method == "OPTIONS") {
res.set_header("Access-Control-Allow-Credentials", "true");
res.set_header("Access-Control-Allow-Methods", "GET, POST");
res.set_header("Access-Control-Allow-Headers", "*");
res.set_content("", "text/html"); // blank response, no data
return httplib::Server::HandlerResponse::Handled; // skip further processing
}
if (!middleware_server_state(req, res)) {
return httplib::Server::HandlerResponse::Handled;
}
@ -3121,33 +3110,19 @@ int main(int argc, char ** argv) {
// register static assets routes
if (!params.public_path.empty()) {
// Set the base directory for serving static files
svr->set_base_dir(params.public_path);
}
if (!params.api_keys.empty()) {
// for now, if API key is set, web UI is unusable
svr->Get("/", [&](const httplib::Request &, httplib::Response & res) {
return res.set_content("Web UI is disabled because API key is set.", "text/html; charset=utf-8");
});
bool is_found = svr->set_mount_point("/", params.public_path);
if (!is_found) {
LOG_ERR("%s: static assets path not found: %s\n", __func__, params.public_path.c_str());
return 1;
}
} else {
// using embedded static files
svr->Get("/", handle_static_file(index_html, index_html_len, "text/html; charset=utf-8"));
svr->Get("/index.js", handle_static_file(index_js, index_js_len, "text/javascript; charset=utf-8"));
svr->Get("/completion.js", handle_static_file(completion_js, completion_js_len, "text/javascript; charset=utf-8"));
svr->Get("/json-schema-to-grammar.mjs", handle_static_file(json_schema_to_grammar_mjs, json_schema_to_grammar_mjs_len, "text/javascript; charset=utf-8"));
// add new-ui files
svr->Get("/colorthemes.css", handle_static_file(colorthemes_css, colorthemes_css_len, "text/css; charset=utf-8"));
svr->Get("/style.css", handle_static_file(style_css, style_css_len, "text/css; charset=utf-8"));
svr->Get("/theme-beeninorder.css", handle_static_file(theme_beeninorder_css, theme_beeninorder_css_len, "text/css; charset=utf-8"));
svr->Get("/theme-ketivah.css", handle_static_file(theme_ketivah_css, theme_ketivah_css_len, "text/css; charset=utf-8"));
svr->Get("/theme-mangotango.css", handle_static_file(theme_mangotango_css, theme_mangotango_css_len, "text/css; charset=utf-8"));
svr->Get("/theme-playground.css", handle_static_file(theme_playground_css, theme_playground_css_len, "text/css; charset=utf-8"));
svr->Get("/theme-polarnight.css", handle_static_file(theme_polarnight_css, theme_polarnight_css_len, "text/css; charset=utf-8"));
svr->Get("/theme-snowstorm.css", handle_static_file(theme_snowstorm_css, theme_snowstorm_css_len, "text/css; charset=utf-8"));
svr->Get("/index-new.html", handle_static_file(index_new_html, index_new_html_len, "text/html; charset=utf-8"));
svr->Get("/system-prompts.js", handle_static_file(system_prompts_js, system_prompts_js_len, "text/javascript; charset=utf-8"));
svr->Get("/prompt-formats.js", handle_static_file(prompt_formats_js, prompt_formats_js_len, "text/javascript; charset=utf-8"));
svr->Get("/", handle_static_file(index_html, index_html_len, "text/html; charset=utf-8"));
svr->Get("/completion.js", handle_static_file(completion_js, completion_js_len, "text/javascript; charset=utf-8"));
svr->Get("/deps_daisyui.min.css", handle_static_file(deps_daisyui_min_css, deps_daisyui_min_css_len, "text/css; charset=utf-8"));
svr->Get("/deps_markdown-it.js", handle_static_file(deps_markdown_it_js, deps_markdown_it_js_len, "text/javascript; charset=utf-8"));
svr->Get("/deps_tailwindcss.js", handle_static_file(deps_tailwindcss_js, deps_tailwindcss_js_len, "text/javascript; charset=utf-8"));
svr->Get("/deps_vue.esm-browser.js", handle_static_file(deps_vue_esm_browser_js, deps_vue_esm_browser_js_len, "text/javascript; charset=utf-8"));
}
// register API routes

2
examples/server/tests/features/security.feature
View file

@ -64,5 +64,5 @@ Feature: Security
| localhost | Access-Control-Allow-Origin | localhost |
| web.mydomain.fr | Access-Control-Allow-Origin | web.mydomain.fr |
| origin | Access-Control-Allow-Credentials | true |
| web.mydomain.fr | Access-Control-Allow-Methods | POST |
| web.mydomain.fr | Access-Control-Allow-Methods | GET, POST |
| web.mydomain.fr | Access-Control-Allow-Headers | * |

5
examples/speculative/speculative.cpp
View file

@ -267,11 +267,12 @@ int main(int argc, char ** argv) {
for (size_t i = 0; i < dist_tgt.size; i++) {
if (dist_tgt.data[i].id == drafts[s].tokens[i_dft]) {
p_tgt = dist_tgt.data[i].p;
break;
}
}
for (size_t i = 0; i < dist_dft.size; i++) {
if (dist_dft.data[i].id == drafts[s].tokens[i_dft]) {
p_dft = dist_dft.data[i].p;
}
if (p_tgt && p_dft) {
break;
}
}

6
flake.lock generated
View file

@ -20,11 +20,11 @@
},
"nixpkgs": {
"locked": {
"lastModified": 1730200266,
"narHash": "sha256-l253w0XMT8nWHGXuXqyiIC/bMvh1VRszGXgdpQlfhvU=",
"lastModified": 1730785428,
"narHash": "sha256-Zwl8YgTVJTEum+L+0zVAWvXAGbWAuXHax3KzuejaDyo=",
"owner": "NixOS",
"repo": "nixpkgs",
"rev": "807e9154dcb16384b1b765ebe9cd2bba2ac287fd",
"rev": "4aa36568d413aca0ea84a1684d2d46f55dbabad7",
"type": "github"
},
"original": {

14
ggml/CMakeLists.txt
View file

@ -92,6 +92,7 @@ else()
endif()
option(GGML_CPU_HBM "ggml: use memkind for CPU HBM" OFF)
option(GGML_CPU_AARCH64 "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
option(GGML_AVX "ggml: enable AVX" ${INS_ENB})
option(GGML_AVX2 "ggml: enable AVX2" ${INS_ENB})
@ -116,6 +117,7 @@ endif()
# ggml core
set(GGML_SCHED_MAX_COPIES "4" CACHE STRING "ggml: max input copies for pipeline parallelism")
option(GGML_CPU "ggml: enable CPU backend" ON)
# 3rd party libs / backends
option(GGML_ACCELERATE "ggml: enable Accelerate framework" ON)
@ -141,7 +143,7 @@ option(GGML_CUDA_NO_VMM "ggml: do not try to use CUDA VMM"
option(GGML_CUDA_FA_ALL_QUANTS "ggml: compile all quants for FlashAttention" OFF)
option(GGML_CUDA_GRAPHS "ggml: use CUDA graphs (llama.cpp only)" ${GGML_CUDA_GRAPHS_DEFAULT})
option(GGML_HIPBLAS "ggml: use hipBLAS" OFF)
option(GGML_HIP "ggml: use HIP" OFF)
option(GGML_HIP_UMA "ggml: use HIP unified memory architecture" OFF)
option(GGML_VULKAN "ggml: use Vulkan" OFF)
option(GGML_VULKAN_CHECK_RESULTS "ggml: run Vulkan op checks" OFF)
@ -153,6 +155,7 @@ option(GGML_VULKAN_VALIDATE "ggml: enable Vulkan validation"
option(GGML_VULKAN_RUN_TESTS "ggml: run Vulkan tests" OFF)
option(GGML_KOMPUTE "ggml: use Kompute" OFF)
option(GGML_METAL "ggml: use Metal" ${GGML_METAL_DEFAULT})
option(GGML_METAL_USE_BF16 "ggml: use bfloat if available" OFF)
option(GGML_METAL_NDEBUG "ggml: disable Metal debugging" OFF)
option(GGML_METAL_SHADER_DEBUG "ggml: compile Metal with -fno-fast-math" OFF)
option(GGML_METAL_EMBED_LIBRARY "ggml: embed Metal library" ${GGML_METAL})
@ -220,12 +223,12 @@ include(CMakePackageConfigHelpers)
# all public headers
set(GGML_PUBLIC_HEADERS
include/ggml.h
include/ggml-cpu.h
include/ggml-alloc.h
include/ggml-backend.h
include/ggml-blas.h
include/ggml-cann.h
include/ggml-cuda.h
include/ggml.h
include/ggml-kompute.h
include/ggml-metal.h
include/ggml-rpc.h
@ -239,12 +242,15 @@ set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
install(TARGETS ggml PUBLIC_HEADER)
if (BUILD_SHARED_LIBS)
install(TARGETS ggml LIBRARY)
install(TARGETS ggml LIBRARY)
install(TARGETS ggml-base LIBRARY)
endif()
# FIXME: this should be done in the backend cmake files
if (GGML_METAL)
# FIXME: does this need to be installed with GGML_METAL_EMBED_LIBRARY?
install(
FILES src/ggml-metal.metal
FILES src/ggml-metal/ggml-metal.metal
PERMISSIONS
OWNER_READ
OWNER_WRITE

10
ggml/include/ggml-amx.h
View file

@ -9,16 +9,16 @@ extern "C" {
#endif
// buffer_type API
GGML_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
GGML_API bool ggml_backend_is_amx(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_amx(ggml_backend_t backend);
// backend API
GGML_API ggml_backend_t ggml_backend_amx_init(void);
GGML_BACKEND_API ggml_backend_t ggml_backend_amx_init(void);
GGML_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
GGML_BACKEND_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
GGML_API ggml_backend_reg_t ggml_backend_amx_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_amx_reg(void);
#ifdef __cplusplus
}

14
ggml/include/ggml-backend.h
View file

@ -3,6 +3,20 @@
#include "ggml.h"
#include "ggml-alloc.h"
#ifdef GGML_BACKEND_SHARED
# if defined(_WIN32) && !defined(__MINGW32__)
# ifdef GGML_BACKEND_BUILD
# define GGML_BACKEND_API __declspec(dllexport) extern
# else
# define GGML_BACKEND_API __declspec(dllimport) extern
# endif
# else
# define GGML_BACKEND_API __attribute__ ((visibility ("default"))) extern
# endif
#else
# define GGML_BACKEND_API extern
#endif
#ifdef __cplusplus
extern "C" {
#endif

8
ggml/include/ggml-blas.h
View file

@ -9,15 +9,15 @@ extern "C" {
#endif
// backend API
GGML_API ggml_backend_t ggml_backend_blas_init(void);
GGML_BACKEND_API ggml_backend_t ggml_backend_blas_init(void);
GGML_API bool ggml_backend_is_blas(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_blas(ggml_backend_t backend);
// number of threads used for conversion to float
// for openblas and blis, this will also set the number of threads used for blas operations
GGML_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
GGML_BACKEND_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
GGML_API ggml_backend_reg_t ggml_backend_blas_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_blas_reg(void);
#ifdef __cplusplus

16
ggml/include/ggml-cann.h
View file

@ -34,7 +34,7 @@ extern "C" {
*/
#define GGML_CANN_MAX_DEVICES 16
GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cann_reg(void);
/**
* @brief Initializes the CANN backend for a specified device.
@ -46,7 +46,7 @@ GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
* @param device The index of the device to initialize.
* @return A pointer to the initialized backend instance, or nullptr on failure.
*/
GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
GGML_BACKEND_API ggml_backend_t ggml_backend_cann_init(int32_t device);
/**
* @brief Checks if a given backend is a CANN backend.
@ -57,7 +57,7 @@ GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
* @param backend The backend instance to check.
* @return True if the backend is a CANN backend, false otherwise.
*/
GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_cann(ggml_backend_t backend);
/**
* @brief Retrieves the CANN buffer type for a specified device.
@ -69,7 +69,7 @@ GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
* @return A pointer to the buffer type interface for the specified device, or
* nullptr if the device index is out of range.
*/
GGML_API ggml_backend_buffer_type_t
GGML_BACKEND_API ggml_backend_buffer_type_t
ggml_backend_cann_buffer_type(int32_t device);
/**
@ -80,14 +80,14 @@ ggml_backend_cann_buffer_type(int32_t device);
*
* @return The number of CANN devices available.
*/
GGML_API int32_t ggml_backend_cann_get_device_count(void);
GGML_BACKEND_API int32_t ggml_backend_cann_get_device_count(void);
/**
* @brief pinned host buffer for use with the CPU backend for faster copies between CPU and NPU.
*
* @return A pointer to the host buffer type interface.
*/
GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
/**
* @brief Retrieves the description of a specific CANN device.
@ -99,7 +99,7 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
* @param description Pointer to a buffer where the description will be written.
* @param description_size Size of the description buffer.
*/
GGML_API void ggml_backend_cann_get_device_description(
GGML_BACKEND_API void ggml_backend_cann_get_device_description(
int32_t device, char* description, size_t description_size);
/**
@ -114,7 +114,7 @@ GGML_API void ggml_backend_cann_get_device_description(
* @param total Pointer to a variable where the total memory size will be
* stored.
*/
GGML_API void ggml_backend_cann_get_device_memory(int32_t device,
GGML_BACKEND_API void ggml_backend_cann_get_device_memory(int32_t device,
size_t* free,
size_t* total);

107
ggml/include/ggml-cpu.h
View file

@ -54,54 +54,77 @@ extern "C" {
GGML_NUMA_STRATEGY_COUNT
};
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_BACKEND_API void ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
GGML_BACKEND_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node
GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
GGML_BACKEND_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
GGML_BACKEND_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
GGML_BACKEND_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
GGML_BACKEND_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
GGML_API void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
GGML_BACKEND_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
GGML_BACKEND_API void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_API void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
GGML_BACKEND_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_BACKEND_API void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
GGML_API float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
GGML_API void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
GGML_BACKEND_API float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
GGML_BACKEND_API void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
GGML_API float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_API void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
GGML_BACKEND_API float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
GGML_BACKEND_API void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
GGML_API void ggml_threadpool_params_init (struct ggml_threadpool_params * p, int n_threads);
GGML_API bool ggml_threadpool_params_match (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
GGML_API struct ggml_threadpool * ggml_threadpool_new (struct ggml_threadpool_params * params);
GGML_API void ggml_threadpool_free (struct ggml_threadpool * threadpool);
GGML_API int ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
GGML_API void ggml_threadpool_pause (struct ggml_threadpool * threadpool);
GGML_API void ggml_threadpool_resume (struct ggml_threadpool * threadpool);
GGML_BACKEND_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
GGML_BACKEND_API void ggml_threadpool_params_init (struct ggml_threadpool_params * p, int n_threads);
GGML_BACKEND_API bool ggml_threadpool_params_match (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
GGML_BACKEND_API struct ggml_threadpool * ggml_threadpool_new (struct ggml_threadpool_params * params);
GGML_BACKEND_API void ggml_threadpool_free (struct ggml_threadpool * threadpool);
GGML_BACKEND_API int ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
GGML_BACKEND_API void ggml_threadpool_pause (struct ggml_threadpool * threadpool);
GGML_BACKEND_API void ggml_threadpool_resume (struct ggml_threadpool * threadpool);
// ggml_graph_plan() has to be called before ggml_graph_compute()
// when plan.work_size > 0, caller must allocate memory for plan.work_data
GGML_API struct ggml_cplan ggml_graph_plan(
GGML_BACKEND_API struct ggml_cplan ggml_graph_plan(
const struct ggml_cgraph * cgraph,
int n_threads, /* = GGML_DEFAULT_N_THREADS */
struct ggml_threadpool * threadpool /* = NULL */ );
GGML_API enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
GGML_BACKEND_API enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
// same as ggml_graph_compute() but the work data is allocated as a part of the context
// note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
GGML_API enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
GGML_BACKEND_API enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
// TODO: move to backend interface
GGML_API int ggml_cpu_has_neon (void);
GGML_API int ggml_cpu_has_sve (void);
GGML_API int ggml_cpu_has_matmul_int8(void);
// get the sve vector length in bytes
GGML_API int ggml_cpu_get_sve_cnt(void);
//
// system info
//
// x86
GGML_BACKEND_API int ggml_cpu_has_sse3 (void);
GGML_BACKEND_API int ggml_cpu_has_ssse3 (void);
GGML_BACKEND_API int ggml_cpu_has_avx (void);
GGML_BACKEND_API int ggml_cpu_has_avx2 (void);
GGML_BACKEND_API int ggml_cpu_has_f16c (void);
GGML_BACKEND_API int ggml_cpu_has_fma (void);
GGML_BACKEND_API int ggml_cpu_has_avx_vnni (void);
GGML_BACKEND_API int ggml_cpu_has_avx512 (void);
GGML_BACKEND_API int ggml_cpu_has_avx512_vbmi(void);
GGML_BACKEND_API int ggml_cpu_has_avx512_vnni(void);
GGML_BACKEND_API int ggml_cpu_has_avx512_bf16(void);
GGML_BACKEND_API int ggml_cpu_has_amx_int8 (void);
// ARM
GGML_BACKEND_API int ggml_cpu_has_neon (void);
GGML_BACKEND_API int ggml_cpu_has_arm_fma (void);
GGML_BACKEND_API int ggml_cpu_has_fp16_va (void);
GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void);
GGML_BACKEND_API int ggml_cpu_has_sve (void);
GGML_BACKEND_API int ggml_cpu_get_sve_cnt (void); // sve vector length in bytes
// other
GGML_BACKEND_API int ggml_cpu_has_riscv_v (void);
GGML_BACKEND_API int ggml_cpu_has_vsx (void);
GGML_BACKEND_API int ggml_cpu_has_wasm_simd (void);
GGML_BACKEND_API int ggml_cpu_has_llamafile (void);
// Internal types and functions exposed for tests and benchmarks
@ -115,6 +138,7 @@ extern "C" {
const void * GGML_RESTRICT y, int nr, int nc);
struct ggml_type_traits_cpu {
ggml_from_float_t from_float;
ggml_from_float_to_mat_t from_float_to_mat;
ggml_vec_dot_t vec_dot;
enum ggml_type vec_dot_type;
@ -124,27 +148,30 @@ extern "C" {
ggml_gemm_t gemm;
};
GGML_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
GGML_API void ggml_cpu_init(void);
GGML_BACKEND_API void ggml_cpu_init(void);
//
// CPU backend
//
GGML_API ggml_backend_t ggml_backend_cpu_init(void);
GGML_BACKEND_API ggml_backend_t ggml_backend_cpu_init(void);
GGML_API bool ggml_backend_is_cpu (ggml_backend_t backend);
GGML_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads);
GGML_API void ggml_backend_cpu_set_threadpool (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
GGML_BACKEND_API bool ggml_backend_is_cpu (ggml_backend_t backend);
GGML_BACKEND_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads);
GGML_BACKEND_API void ggml_backend_cpu_set_threadpool (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
GGML_BACKEND_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
#ifdef GGML_USE_CPU_HBM
GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
#endif
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);
GGML_BACKEND_API bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft);
#ifdef __cplusplus
}
#endif

24
ggml/include/ggml-cuda.h
View file

@ -7,7 +7,7 @@
extern "C" {
#endif
#ifdef GGML_USE_HIPBLAS
#ifdef GGML_USE_HIP
#define GGML_CUDA_NAME "ROCm"
#define GGML_CUBLAS_NAME "hipBLAS"
#elif defined(GGML_USE_MUSA)
@ -20,27 +20,27 @@ extern "C" {
#define GGML_CUDA_MAX_DEVICES 16
// backend API
GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
GGML_BACKEND_API ggml_backend_t ggml_backend_cuda_init(int device);
GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_cuda(ggml_backend_t backend);
// device buffer
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
// split tensor buffer that splits matrices by rows across multiple devices
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
GGML_API int ggml_backend_cuda_get_device_count(void);
GGML_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
GGML_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
GGML_BACKEND_API int ggml_backend_cuda_get_device_count(void);
GGML_BACKEND_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
GGML_BACKEND_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
GGML_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
GGML_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
GGML_BACKEND_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
GGML_BACKEND_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
GGML_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
#ifdef __cplusplus
}

8
ggml/include/ggml-kompute.h
View file

@ -37,13 +37,13 @@ struct ggml_vk_device ggml_vk_current_device(void);
// forward declaration
typedef struct ggml_backend * ggml_backend_t;
GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
GGML_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
#ifdef __cplusplus
}

16
ggml/include/ggml-metal.h
View file

@ -39,27 +39,27 @@ extern "C" {
// user-code should use only these functions
//
GGML_API ggml_backend_t ggml_backend_metal_init(void);
GGML_BACKEND_API ggml_backend_t ggml_backend_metal_init(void);
GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
GGML_DEPRECATED(
GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
"obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713");
GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
// helper to check if the device supports a specific family
// ideally, the user code should be doing these checks
// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
GGML_BACKEND_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
// capture all command buffers committed the next time `ggml_backend_graph_compute` is called
GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
GGML_BACKEND_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
GGML_API ggml_backend_reg_t ggml_backend_metal_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_metal_reg(void);
#ifdef __cplusplus
}

14
ggml/include/ggml-rpc.h
View file

@ -10,18 +10,18 @@ extern "C" {
#define GGML_RPC_MAX_SERVERS 16
// backend API
GGML_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend);
GGML_BACKEND_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
GGML_BACKEND_API bool ggml_backend_is_rpc(ggml_backend_t backend);
GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
GGML_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
GGML_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
GGML_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
GGML_BACKEND_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
#ifdef __cplusplus
}

26
ggml/include/ggml-sycl.h
View file

@ -17,32 +17,32 @@ extern "C" {
#endif
// backend API
GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
GGML_BACKEND_API ggml_backend_t ggml_backend_sycl_init(int device);
GGML_API bool ggml_backend_is_sycl(ggml_backend_t backend);
GGML_BACKEND_API bool ggml_backend_is_sycl(ggml_backend_t backend);
// devide buffer
GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
// split tensor buffer that splits matrices by rows across multiple devices
GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
GGML_API void ggml_backend_sycl_print_sycl_devices(void);
GGML_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len);
GGML_API void ggml_backend_sycl_get_device_description(int device,
GGML_BACKEND_API void ggml_backend_sycl_print_sycl_devices(void);
GGML_BACKEND_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len);
GGML_BACKEND_API void ggml_backend_sycl_get_device_description(int device,
char *description,
size_t description_size);
GGML_API int ggml_backend_sycl_get_device_count();
GGML_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
GGML_BACKEND_API int ggml_backend_sycl_get_device_count();
GGML_BACKEND_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
// SYCL doesn't support registering host memory, keep here for reference
// GGML_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
// GGML_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
// GGML_BACKEND_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
// GGML_BACKEND_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
GGML_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
#ifdef __cplusplus
}

18
ggml/include/ggml-vulkan.h
View file

@ -10,21 +10,21 @@ extern "C" {
#define GGML_VK_NAME "Vulkan"
#define GGML_VK_MAX_DEVICES 16
GGML_API void ggml_vk_instance_init(void);
GGML_BACKEND_API void ggml_vk_instance_init(void);
// backend API
GGML_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
GGML_BACKEND_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
GGML_API bool ggml_backend_is_vk(ggml_backend_t backend);
GGML_API int ggml_backend_vk_get_device_count(void);
GGML_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
GGML_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
GGML_BACKEND_API bool ggml_backend_is_vk(ggml_backend_t backend);
GGML_BACKEND_API int ggml_backend_vk_get_device_count(void);
GGML_BACKEND_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
GGML_BACKEND_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
GGML_API ggml_backend_reg_t ggml_backend_vk_reg(void);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_vk_reg(void);
#ifdef __cplusplus
}

50
ggml/include/ggml.h
View file

@ -176,15 +176,15 @@
#ifdef GGML_SHARED
# if defined(_WIN32) && !defined(__MINGW32__)
# ifdef GGML_BUILD
# define GGML_API __declspec(dllexport)
# define GGML_API __declspec(dllexport) extern
# else
# define GGML_API __declspec(dllimport)
# define GGML_API __declspec(dllimport) extern
# endif
# else
# define GGML_API __attribute__ ((visibility ("default")))
# define GGML_API __attribute__ ((visibility ("default"))) extern
# endif
#else
# define GGML_API
# define GGML_API extern
#endif
// TODO: support for clang
@ -509,7 +509,7 @@ extern "C" {
GGML_OP_WIN_UNPART,
GGML_OP_GET_REL_POS,
GGML_OP_ADD_REL_POS,
GGML_OP_RWKV_WKV,
GGML_OP_RWKV_WKV6,
GGML_OP_UNARY,
@ -1490,7 +1490,7 @@ extern "C" {
"use ggml_rope_ext_inplace instead");
// compute correction dims for YaRN RoPE scaling
void ggml_rope_yarn_corr_dims(
GGML_API void ggml_rope_yarn_corr_dims(
int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
// rotary position embedding backward, i.e compute dx from dy
@ -1746,6 +1746,9 @@ extern "C" {
struct ggml_tensor * a,
enum ggml_prec prec);
GGML_API enum ggml_prec ggml_flash_attn_ext_get_prec(
const struct ggml_tensor * a);
// TODO: needs to be adapted to ggml_flash_attn_ext
GGML_API struct ggml_tensor * ggml_flash_attn_back(
struct ggml_context * ctx,
@ -1819,7 +1822,7 @@ extern "C" {
struct ggml_tensor * pw,
struct ggml_tensor * ph);
GGML_API struct ggml_tensor * ggml_rwkv_wkv(
GGML_API struct ggml_tensor * ggml_rwkv_wkv6(
struct ggml_context * ctx,
struct ggml_tensor * k,
struct ggml_tensor * v,
@ -2381,38 +2384,6 @@ extern "C" {
GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
GGML_API void gguf_get_meta_data(const struct gguf_context * ctx, void * data);
//
// system info
//
GGML_API int ggml_cpu_has_avx (void);
GGML_API int ggml_cpu_has_avx_vnni (void);
GGML_API int ggml_cpu_has_avx2 (void);
GGML_API int ggml_cpu_has_avx512 (void);
GGML_API int ggml_cpu_has_avx512_vbmi(void);
GGML_API int ggml_cpu_has_avx512_vnni(void);
GGML_API int ggml_cpu_has_avx512_bf16(void);
GGML_API int ggml_cpu_has_amx_int8 (void);
GGML_API int ggml_cpu_has_fma (void);
GGML_API int ggml_cpu_has_arm_fma (void);
GGML_API int ggml_cpu_has_metal (void);
GGML_API int ggml_cpu_has_f16c (void);
GGML_API int ggml_cpu_has_fp16_va (void);
GGML_API int ggml_cpu_has_wasm_simd (void);
GGML_API int ggml_cpu_has_blas (void);
GGML_API int ggml_cpu_has_cuda (void);
GGML_API int ggml_cpu_has_vulkan (void);
GGML_API int ggml_cpu_has_kompute (void);
GGML_API int ggml_cpu_has_gpublas (void);
GGML_API int ggml_cpu_has_sse3 (void);
GGML_API int ggml_cpu_has_ssse3 (void);
GGML_API int ggml_cpu_has_riscv_v (void);
GGML_API int ggml_cpu_has_sycl (void);
GGML_API int ggml_cpu_has_rpc (void);
GGML_API int ggml_cpu_has_vsx (void);
GGML_API int ggml_cpu_has_cann (void);
GGML_API int ggml_cpu_has_llamafile (void);
#ifdef __cplusplus
// restrict not standard in C++
#define GGML_RESTRICT
@ -2429,7 +2400,6 @@ extern "C" {
size_t type_size;
bool is_quantized;
ggml_to_float_t to_float;
ggml_from_float_t from_float;
ggml_from_float_t from_float_ref;
};

1262
ggml/src/CMakeLists.txt
View file

File diff suppressed because it is too large Load diff

3386
ggml/src/ggml-aarch64.c
View file

File diff suppressed because it is too large Load diff

20
ggml/src/ggml-aarch64.h
View file

@ -1,9 +1,5 @@
// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
#pragma once
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "ggml.h"
// GGML internal header
@ -12,27 +8,11 @@
extern "C" {
#endif
// Quantization
void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
// GEMV
void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
// GEMM
void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
#ifdef __cplusplus
}
#endif

107
ggml/src/ggml-amx/CMakeLists.txt Normal file
View file

@ -0,0 +1,107 @@
if (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
(NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$") AND
CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 11.0)
message(STATUS "Using AMX")
file(GLOB GGML_HEADERS_AMX "*.h")
list(APPEND GGML_HEADERS_AMX "../../include/ggml-amx.h")
file(GLOB GGML_SOURCES_AMX "*.cpp")
add_library(ggml-amx
${GGML_HEADERS_AMX}
${GGML_SOURCES_AMX})
target_link_libraries(ggml-amx PRIVATE ggml-base)
target_include_directories(ggml-amx PRIVATE . ..)
# this is duplicated from the CPU backend, since the AMX backend also depends on the architecture flags
# TODO: integrate AMX backend into the CPU backend
if (MSVC)
# instruction set detection for MSVC only
if (GGML_NATIVE)
# TODO: improve, should not reference files from the parent folder
include(../ggml-cpu/cmake/FindSIMD.cmake)
endif ()
if (GGML_AVX512)
list(APPEND ARCH_FLAGS /arch:AVX512)
# MSVC has no compile-time flags enabling specific
# AVX512 extensions, neither it defines the
# macros corresponding to the extensions.
# Do it manually.
if (GGML_AVX512_VBMI)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
endif()
if (GGML_AVX512_VNNI)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
endif()
if (GGML_AVX512_BF16)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
endif()
if (GGML_AMX_TILE)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
endif()
if (GGML_AMX_INT8)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
endif()
if (GGML_AMX_BF16)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
endif()
elseif (GGML_AVX2)
list(APPEND ARCH_FLAGS /arch:AVX2)
elseif (GGML_AVX)
list(APPEND ARCH_FLAGS /arch:AVX)
endif()
else()
if (GGML_NATIVE)
list(APPEND ARCH_FLAGS -march=native)
endif()
if (GGML_F16C)
list(APPEND ARCH_FLAGS -mf16c)
endif()
if (GGML_FMA)
list(APPEND ARCH_FLAGS -mfma)
endif()
if (GGML_AVX)
list(APPEND ARCH_FLAGS -mavx)
endif()
if (GGML_AVX2)
list(APPEND ARCH_FLAGS -mavx2)
endif()
if (GGML_AVX512)
list(APPEND ARCH_FLAGS -mavx512f)
list(APPEND ARCH_FLAGS -mavx512dq)
list(APPEND ARCH_FLAGS -mavx512bw)
endif()
if (GGML_AVX512_VBMI)
list(APPEND ARCH_FLAGS -mavx512vbmi)
endif()
if (GGML_AVX512_VNNI)
list(APPEND ARCH_FLAGS -mavx512vnni)
endif()
if (GGML_AVX512_BF16)
list(APPEND ARCH_FLAGS -mavx512bf16)
endif()
if (GGML_AMX_TILE)
list(APPEND ARCH_FLAGS -mamx-tile)
endif()
if (GGML_AMX_INT8)
list(APPEND ARCH_FLAGS -mamx-int8)
endif()
if (GGML_AMX_BF16)
list(APPEND ARCH_FLAGS -mamx-bf16)
endif()
endif()
target_compile_options(ggml-amx PRIVATE ${ARCH_FLAGS})
else()
set(GGML_AMX OFF PARENT_SCOPE)
message(WARNING "AMX requires x86 and gcc version > 11.0. Turning off GGML_AMX.")
endif()

3
ggml/src/ggml-amx/common.h
View file

@ -1,7 +1,8 @@
#pragma once
#include "ggml.h"
#include "ggml-cpu-impl.h" // <immintrin.h>
// hack until AMX is moved into the CPU backend
#include "../ggml-cpu/ggml-cpu-impl.h" // <immintrin.h>
#include <algorithm>
#include <memory>

15
ggml/src/ggml-amx.cpp → ggml/src/ggml-amx/ggml-amx.cpp
View file

@ -421,9 +421,18 @@ ggml_backend_reg_t ggml_backend_amx_reg(void) {
#else // if defined(__AMX_INT8__)
ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void) {
return nullptr;
}
bool ggml_backend_is_amx(ggml_backend_t backend) {
GGML_UNUSED(backend);
return false;
}
ggml_backend_t ggml_backend_amx_init(void) {
fprintf(stderr, "GGML is not compiled with AMX support!\n");
return ggml_backend_t{};
return nullptr;
}
void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
@ -433,4 +442,8 @@ void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
GGML_UNUSED(n_threads);
}
ggml_backend_reg_t ggml_backend_amx_reg(void) {
return nullptr;
}
#endif

7
ggml/src/ggml-amx/mmq.cpp
View file

@ -496,19 +496,20 @@ inline void from_float(const float * x, char * vy, int64_t k);
template <>
inline void from_float<block_q8_0>(const float * x, char * vy, int64_t k) {
quantize_row_q8_0(x, vy, k);
// FIXME: using unoptimized reference impl until moved to CPU backend
quantize_row_q8_0_ref(x, (block_q8_0 *)vy, k);
}
template <>
inline void from_float<block_q8_1>(const float * x, char * vy, int64_t k) {
quantize_row_q8_1(x, vy, k);
quantize_row_q8_1_ref(x, (block_q8_1 *)vy, k);
}
template <>
inline void from_float<block_q8_K>(const float * x, char * vy, int64_t k) {
#if 1
// TODO: this is reference impl!
quantize_row_q8_K(x, vy, k);
quantize_row_q8_K_ref(x, (block_q8_K *)vy, k);
#else
quantize_row_q8_K_vnni(x, vy, k);
#endif

195
ggml/src/ggml-backend-reg.cpp Normal file
View file

@ -0,0 +1,195 @@
#include "ggml-backend-impl.h"
#include "ggml-backend.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include <cstring>
#include <vector>
// Backend registry
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif
#ifdef GGML_USE_SYCL
#include "ggml-sycl.h"
#endif
#ifdef GGML_USE_VULKAN
#include "ggml-vulkan.h"
#endif
#ifdef GGML_USE_BLAS
#include "ggml-blas.h"
#endif
#ifdef GGML_USE_RPC
#include "ggml-rpc.h"
#endif
#ifdef GGML_USE_AMX
# include "ggml-amx.h"
#endif
#ifdef GGML_USE_CANN
#include "ggml-cann.h"
#endif
#ifdef GGML_USE_KOMPUTE
#include "ggml-kompute.h"
#endif
struct ggml_backend_registry {
std::vector<ggml_backend_reg_t> backends;
std::vector<ggml_backend_dev_t> devices;
ggml_backend_registry() {
#ifdef GGML_USE_CUDA
register_backend(ggml_backend_cuda_reg());
#endif
#ifdef GGML_USE_METAL
register_backend(ggml_backend_metal_reg());
#endif
#ifdef GGML_USE_SYCL
register_backend(ggml_backend_sycl_reg());
#endif
#ifdef GGML_USE_VULKAN
register_backend(ggml_backend_vk_reg());
#endif
#ifdef GGML_USE_CANN
register_backend(ggml_backend_cann_reg());
#endif
#ifdef GGML_USE_BLAS
register_backend(ggml_backend_blas_reg());
#endif
#ifdef GGML_USE_RPC
register_backend(ggml_backend_rpc_reg());
#endif
#ifdef GGML_USE_AMX
register_backend(ggml_backend_amx_reg());
#endif
#ifdef GGML_USE_KOMPUTE
register_backend(ggml_backend_kompute_reg());
#endif
register_backend(ggml_backend_cpu_reg());
}
void register_backend(ggml_backend_reg_t reg) {
if (!reg) {
return;
}
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
__func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
#endif
backends.push_back(reg);
for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
register_device(ggml_backend_reg_dev_get(reg, i));
}
}
void register_device(ggml_backend_dev_t device) {
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
#endif
devices.push_back(device);
}
};
static ggml_backend_registry & get_reg() {
static ggml_backend_registry reg;
return reg;
}
// Internal API
void ggml_backend_register(ggml_backend_reg_t reg) {
get_reg().register_backend(reg);
}
void ggml_backend_device_register(ggml_backend_dev_t device) {
get_reg().register_device(device);
}
// Backend (reg) enumeration
size_t ggml_backend_reg_count() {
return get_reg().backends.size();
}
ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
GGML_ASSERT(index < ggml_backend_reg_count());
return get_reg().backends[index];
}
ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
ggml_backend_reg_t reg = ggml_backend_reg_get(i);
if (std::strcmp(ggml_backend_reg_name(reg), name) == 0) {
return reg;
}
}
return NULL;
}
// Device enumeration
size_t ggml_backend_dev_count() {
return get_reg().devices.size();
}
ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
GGML_ASSERT(index < ggml_backend_dev_count());
return get_reg().devices[index];
}
ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
return dev;
}
}
return NULL;
}
ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
if (ggml_backend_dev_type(dev) == type) {
return dev;
}
}
return NULL;
}
// Convenience functions
ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
if (!dev) {
return NULL;
}
return ggml_backend_dev_init(dev, params);
}
ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
if (!dev) {
return NULL;
}
return ggml_backend_dev_init(dev, params);
}
ggml_backend_t ggml_backend_init_best(void) {
ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
if (!dev) {
dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
}
if (!dev) {
return NULL;
}
return ggml_backend_dev_init(dev, NULL);
}

676
ggml/src/ggml-backend.cpp
View file

@ -525,197 +525,6 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
return reg->iface.get_proc_address(reg, name);
}
// Backend registry
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
#ifdef GGML_USE_METAL
#include "ggml-metal.h"
#endif
#ifdef GGML_USE_SYCL
#include "ggml-sycl.h"
#endif
#ifdef GGML_USE_VULKAN
#include "ggml-vulkan.h"
#endif
#ifdef GGML_USE_BLAS
#include "ggml-blas.h"
#endif
#ifdef GGML_USE_RPC
#include "ggml-rpc.h"
#endif
#ifndef __AMX_INT8__
#undef GGML_USE_AMX
#endif
#ifdef GGML_USE_AMX
# include "ggml-amx.h"
#endif
#ifdef GGML_USE_CANN
#include "ggml-cann.h"
#endif
#ifdef GGML_USE_KOMPUTE
#include "ggml-kompute.h"
#endif
#include "ggml-cpu.h"
struct ggml_backend_registry {
std::vector<ggml_backend_reg_t> backends;
std::vector<ggml_backend_dev_t> devices;
ggml_backend_registry() {
#ifdef GGML_USE_CUDA
register_backend(ggml_backend_cuda_reg());
#endif
#ifdef GGML_USE_METAL
register_backend(ggml_backend_metal_reg());
#endif
#ifdef GGML_USE_SYCL
register_backend(ggml_backend_sycl_reg());
#endif
#ifdef GGML_USE_VULKAN
register_backend(ggml_backend_vk_reg());
#endif
#ifdef GGML_USE_CANN
register_backend(ggml_backend_cann_reg());
#endif
#ifdef GGML_USE_BLAS
register_backend(ggml_backend_blas_reg());
#endif
#ifdef GGML_USE_RPC
register_backend(ggml_backend_rpc_reg());
#endif
#ifdef GGML_USE_AMX
register_backend(ggml_backend_amx_reg());
#endif
#ifdef GGML_USE_KOMPUTE
register_backend(ggml_backend_kompute_reg());
#endif
register_backend(ggml_backend_cpu_reg());
}
void register_backend(ggml_backend_reg_t reg) {
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
__func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
#endif
backends.push_back(reg);
for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
register_device(ggml_backend_reg_dev_get(reg, i));
}
}
void register_device(ggml_backend_dev_t device) {
#ifndef NDEBUG
GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
#endif
devices.push_back(device);
}
};
static ggml_backend_registry & get_reg() {
static ggml_backend_registry reg;
return reg;
}
// Internal API
void ggml_backend_register(ggml_backend_reg_t reg) {
get_reg().register_backend(reg);
}
void ggml_backend_device_register(ggml_backend_dev_t device) {
get_reg().register_device(device);
}
// Backend (reg) enumeration
size_t ggml_backend_reg_count() {
return get_reg().backends.size();
}
ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
GGML_ASSERT(index < ggml_backend_reg_count());
return get_reg().backends[index];
}
ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
ggml_backend_reg_t reg = ggml_backend_reg_get(i);
if (strcmp(ggml_backend_reg_name(reg), name) == 0) {
return reg;
}
}
return NULL;
}
// Device enumeration
size_t ggml_backend_dev_count() {
return get_reg().devices.size();
}
ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
GGML_ASSERT(index < ggml_backend_dev_count());
return get_reg().devices[index];
}
ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
return dev;
}
}
return NULL;
}
ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
ggml_backend_dev_t dev = ggml_backend_dev_get(i);
if (ggml_backend_dev_type(dev) == type) {
return dev;
}
}
return NULL;
}
// Convenience functions
ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
if (!dev) {
return NULL;
}
return ggml_backend_dev_init(dev, params);
}
ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
if (!dev) {
return NULL;
}
return ggml_backend_dev_init(dev, params);
}
ggml_backend_t ggml_backend_init_best(void) {
ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
if (!dev) {
dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
}
if (!dev) {
return NULL;
}
return ggml_backend_dev_init(dev, NULL);
}
// multi-buffer buffer
struct ggml_backend_multi_buffer_context {
@ -1640,7 +1449,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
bool parallel) {
GGML_ASSERT(n_backends > 0);
GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
struct ggml_backend_sched * sched = (ggml_backend_sched *) calloc(1, sizeof(struct ggml_backend_sched));
@ -2036,17 +1845,6 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
return true;
}
#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include <cctype>
#include <string>
// ggml-backend interface
// CPU backend - buffer
static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
@ -2120,7 +1918,9 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
/* .reset = */ NULL,
};
// CPU backend - buffer type
// CPU backend buffer type
// this buffer type is defined here to make it available to all backends
static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU";
@ -2161,7 +1961,7 @@ ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .device = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ NULL,
};
@ -2184,478 +1984,14 @@ static ggml_backend_buffer_type_t ggml_backend_cpu_buffer_from_ptr_type(void) {
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .device = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type;
}
#ifdef GGML_USE_CPU_HBM
// buffer type HBM
#include <hbwmalloc.h>
static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_HBM";
GGML_UNUSED(buft);
}
static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
hbw_free(buffer->context);
}
static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
void * ptr;
int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
if (result != 0) {
GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
return NULL;
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type_hbm;
}
#endif
static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
static ggml_backend_buffer_type_t bufts[] = {
#ifdef GGML_USE_CPU_HBM
ggml_backend_cpu_hbm_buffer_type(),
#endif
NULL
};
return bufts;
GGML_UNUSED(device);
}
// CPU backend - backend (stream)
struct ggml_backend_cpu_context {
int n_threads;
ggml_threadpool_t threadpool;
uint8_t * work_data;
size_t work_size;
ggml_abort_callback abort_callback;
void * abort_callback_data;
};
static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
return "CPU";
GGML_UNUSED(backend);
}
static void ggml_backend_cpu_free(ggml_backend_t backend) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
delete[] cpu_ctx->work_data;
delete cpu_ctx;
delete backend;
}
struct ggml_backend_plan_cpu {
struct ggml_cplan cplan;
struct ggml_cgraph cgraph;
};
static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
cpu_plan->cgraph = *cgraph; // FIXME: deep copy
if (cpu_plan->cplan.work_size > 0) {
cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
if (cpu_plan->cplan.work_data == NULL) {
delete cpu_plan;
return NULL;
}
}
cpu_plan->cplan.abort_callback = cpu_ctx->abort_callback;
cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return cpu_plan;
}
static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
delete[] cpu_plan->cplan.work_data;
delete cpu_plan;
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
if (cpu_ctx->work_size < cplan.work_size) {
delete[] cpu_ctx->work_data;
cpu_ctx->work_data = new uint8_t[cplan.work_size];
if (cpu_ctx->work_data == NULL) {
cpu_ctx->work_size = 0;
return GGML_STATUS_ALLOC_FAILED;
}
cpu_ctx->work_size = cplan.work_size;
}
cplan.work_data = (uint8_t *)cpu_ctx->work_data;
cplan.abort_callback = cpu_ctx->abort_callback;
cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return ggml_graph_compute(cgraph, &cplan);
}
static const struct ggml_backend_i ggml_backend_cpu_i = {
/* .get_name = */ ggml_backend_cpu_get_name,
/* .free = */ ggml_backend_cpu_free,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create,
/* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute,
/* .graph_compute = */ ggml_backend_cpu_graph_compute,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
};
static ggml_guid_t ggml_backend_cpu_guid(void) {
static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
return &guid;
}
ggml_backend_t ggml_backend_cpu_init(void) {
// initialize CPU backend now to avoid slowing the first graph computation
ggml_cpu_init();
struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
if (ctx == NULL) {
return NULL;
}
ctx->n_threads = GGML_DEFAULT_N_THREADS;
ctx->threadpool = NULL;
ctx->work_data = NULL;
ctx->work_size = 0;
ctx->abort_callback = NULL;
ctx->abort_callback_data = NULL;
ggml_backend_t cpu_backend = new ggml_backend {
/* .guid = */ ggml_backend_cpu_guid(),
/* .interface = */ ggml_backend_cpu_i,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ ctx,
};
if (cpu_backend == NULL) {
delete ctx;
return NULL;
}
return cpu_backend;
}
bool ggml_backend_is_cpu(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
}
void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->n_threads = n_threads;
}
void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
if (ctx->threadpool && ctx->threadpool != threadpool) {
// already had a different threadpool, pause/suspend it before switching
ggml_threadpool_pause(ctx->threadpool);
}
ctx->threadpool = threadpool;
}
void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->abort_callback = abort_callback;
ctx->abort_callback_data = abort_callback_data;
}
ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
return ggml_backend_buffer_init(ggml_backend_cpu_buffer_from_ptr_type(), ggml_backend_cpu_buffer_from_ptr_i, ptr, size);
}
// CPU backend - device
struct ggml_backend_cpu_device_context {
std::string description = "CPU";
ggml_backend_cpu_device_context() {
#ifdef __APPLE__
size_t len = 0;
if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
description.resize(len);
sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
}
#elif defined(__linux__)
FILE * f = fopen("/proc/cpuinfo", "r");
if (f) {
char buf[1024];
while (fgets(buf, sizeof(buf), f)) {
if (strncmp(buf, "model name", 10) == 0) {
char * p = strchr(buf, ':');
if (p) {
p++;
while (std::isspace(*p)) {
p++;
}
while (std::isspace(p[strlen(p) - 1])) {
p[strlen(p) - 1] = '\0';
}
description = p;
break;
}
}
}
fclose(f);
}
#elif defined(_WIN32)
HKEY hKey;
if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
0,
KEY_READ,
&hKey) == ERROR_SUCCESS) {
DWORD cpu_brand_size = 0;
if (RegQueryValueExA(hKey,
TEXT("ProcessorNameString"),
NULL,
NULL,
NULL,
&cpu_brand_size) == ERROR_SUCCESS) {
description.resize(cpu_brand_size);
if (RegQueryValueExA(hKey,
TEXT("ProcessorNameString"),
NULL,
NULL,
(LPBYTE)&description[0], // NOLINT
&cpu_brand_size) == ERROR_SUCCESS) {
if (description.find('\0') != std::string::npos) {
description.resize(description.find('\0'));
}
}
}
RegCloseKey(hKey);
}
#endif
}
};
static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
return "CPU";
GGML_UNUSED(dev);
}
static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
return ctx->description.c_str();
}
static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
// TODO
*free = 0;
*total = 0;
GGML_UNUSED(dev);
}
static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
return GGML_BACKEND_DEVICE_TYPE_CPU;
GGML_UNUSED(dev);
}
static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
props->name = ggml_backend_cpu_device_get_name(dev);
props->description = ggml_backend_cpu_device_get_description(dev);
props->type = ggml_backend_cpu_device_get_type(dev);
ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
props->caps = {
/* .async = */ false,
/* .host_buffer = */ false,
/* .buffer_from_host_ptr = */ true,
/* .events = */ false,
};
}
static ggml_backend_t ggml_backend_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
return ggml_backend_cpu_init();
GGML_UNUSED(dev);
GGML_UNUSED(params);
}
static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
return ggml_backend_cpu_buffer_type();
GGML_UNUSED(dev);
}
static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
return ggml_backend_cpu_buffer_from_ptr(ptr, size);
GGML_UNUSED(dev);
GGML_UNUSED(max_tensor_size);
}
static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, 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 &&
op->type != GGML_TYPE_IQ1_S &&
op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
case GGML_OP_MUL_MAT:
return op->src[1]->type == GGML_TYPE_F32;// FIXME || op->src[1]->type == ggml_get_type_traits(op->src[0]->type)->vec_dot_type;
case GGML_OP_ROPE_BACK:
return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
case GGML_OP_IM2COL_BACK:
return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
case GGML_OP_OUT_PROD:
return (op->src[0]->type == GGML_TYPE_F32 || ggml_is_quantized(op->src[0]->type)) && op->src[1]->type == GGML_TYPE_F32;
default:
return true;
}
GGML_UNUSED(dev);
}
static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
return ggml_backend_buft_is_host(buft);
GGML_UNUSED(dev);
}
static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
/* .get_name = */ ggml_backend_cpu_device_get_name,
/* .get_description = */ ggml_backend_cpu_device_get_description,
/* .get_memory = */ ggml_backend_cpu_device_get_memory,
/* .get_type = */ ggml_backend_cpu_device_get_type,
/* .get_props = */ ggml_backend_cpu_device_get_props,
/* .init_backend = */ ggml_backend_cpu_device_init_backend,
/* .get_buffer_type = */ ggml_backend_cpu_device_get_buffer_type,
/* .get_host_buffer_type = */ NULL,
/* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
/* .supports_op = */ ggml_backend_cpu_device_supports_op,
/* .supports_buft = */ ggml_backend_cpu_device_supports_buft,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_synchronize = */ NULL,
};
// CPU backend - backend (reg)
static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
return "CPU";
GGML_UNUSED(reg);
}
static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
return 1;
GGML_UNUSED(reg);
}
static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
GGML_ASSERT(index == 0);
static ggml_backend_cpu_device_context ctx;
static ggml_backend_device ggml_backend_cpu_device = {
/* .iface = */ ggml_backend_cpu_device_i,
/* .reg = */ reg,
/* .context = */ &ctx,
};
return &ggml_backend_cpu_device;
}
static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
return (void *)ggml_backend_cpu_set_n_threads;
}
if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
return (void *)ggml_backend_cpu_get_extra_bufts;
}
return NULL;
GGML_UNUSED(reg);
}
static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
/* .get_name = */ ggml_backend_cpu_reg_get_name,
/* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
/* .get_device = */ ggml_backend_cpu_reg_get_device,
/* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
};
ggml_backend_reg_t ggml_backend_cpu_reg(void) {
static struct ggml_backend_reg ggml_backend_cpu_reg = {
/* .iface = */ ggml_backend_cpu_reg_i,
/* .context = */ NULL,
};
return &ggml_backend_cpu_reg;
}

91
ggml/src/ggml-blas/CMakeLists.txt Normal file
View file

@ -0,0 +1,91 @@
if (GGML_STATIC)
set(BLA_STATIC ON)
endif()
#if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
# set(BLA_SIZEOF_INTEGER 8)
#endif()
set(BLA_VENDOR ${GGML_BLAS_VENDOR})
find_package(BLAS)
if (BLAS_FOUND)
message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
add_library(ggml-blas
ggml-blas.cpp
)
target_link_libraries(ggml-blas PRIVATE ggml-base)
target_include_directories(ggml-blas PRIVATE . ..)
if (${GGML_BLAS_VENDOR} MATCHES "Apple")
add_compile_definitions(ACCELERATE_NEW_LAPACK)
add_compile_definitions(ACCELERATE_LAPACK_ILP64)
add_compile_definitions(GGML_BLAS_USE_ACCELERATE)
elseif ("${BLAS_INCLUDE_DIRS}" STREQUAL "")
# BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
# see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
find_package(PkgConfig REQUIRED)
if (${GGML_BLAS_VENDOR} MATCHES "Generic")
pkg_check_modules(DepBLAS blas)
elseif (${GGML_BLAS_VENDOR} MATCHES "OpenBLAS")
# As of openblas v0.3.22, the 64-bit is named openblas64.pc
pkg_check_modules(DepBLAS openblas64)
if (NOT DepBLAS_FOUND)
pkg_check_modules(DepBLAS openblas)
endif()
elseif (${GGML_BLAS_VENDOR} MATCHES "FLAME")
add_compile_definitions(GGML_BLAS_USE_BLIS)
pkg_check_modules(DepBLAS blis)
elseif (${GGML_BLAS_VENDOR} MATCHES "ATLAS")
pkg_check_modules(DepBLAS blas-atlas)
elseif (${GGML_BLAS_VENDOR} MATCHES "FlexiBLAS")
pkg_check_modules(DepBLAS flexiblas_api)
elseif (${GGML_BLAS_VENDOR} MATCHES "Intel")
add_compile_definitions(GGML_BLAS_USE_MKL)
# all Intel* libraries share the same include path
pkg_check_modules(DepBLAS mkl-sdl)
elseif (${GGML_BLAS_VENDOR} MATCHES "NVHPC")
# this doesn't provide pkg-config
# suggest to assign BLAS_INCLUDE_DIRS on your own
if ("${NVHPC_VERSION}" STREQUAL "")
message(WARNING "Better to set NVHPC_VERSION")
else()
set(DepBLAS_FOUND ON)
set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
endif()
endif()
if (DepBLAS_FOUND)
set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
else()
message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
" detected by pkgconfig, trying to find cblas.h from possible paths...")
find_path(BLAS_INCLUDE_DIRS
NAMES cblas.h
HINTS
/usr/include
/usr/local/include
/usr/include/openblas
/opt/homebrew/opt/openblas/include
/usr/local/opt/openblas/include
/usr/include/x86_64-linux-gnu/openblas/include
)
endif()
endif()
message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
#add_compile_options(${BLAS_LINKER_FLAGS})
target_compile_options(ggml-blas PRIVATE ${BLAS_LINKER_FLAGS})
if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
add_compile_definitions(GGML_BLAS_USE_MKL)
endif()
target_link_libraries (ggml-blas PRIVATE ${BLAS_LIBRARIES})
target_include_directories(ggml-blas PRIVATE ${BLAS_INCLUDE_DIRS})
else()
message(ERROR "BLAS not found, please refer to "
"https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
" to set correct GGML_BLAS_VENDOR")
endif()

4
ggml/src/ggml-blas.cpp → ggml/src/ggml-blas/ggml-blas.cpp
View file

@ -6,7 +6,7 @@
#include <vector>
#include <cstring>
#if defined(GGML_USE_ACCELERATE)
#if defined(GGML_BLAS_USE_ACCELERATE)
# include <Accelerate/Accelerate.h>
#elif defined(GGML_BLAS_USE_MKL)
# include <mkl.h>
@ -320,7 +320,7 @@ static const char * ggml_backend_blas_device_get_name(ggml_backend_dev_t dev) {
}
static const char * ggml_backend_blas_device_get_description(ggml_backend_dev_t dev) {
#if defined(GGML_USE_ACCELERATE)
#if defined(GGML_BLAS_USE_ACCELERATE)
return "Accelerate";
#elif defined(GGML_BLAS_USE_MKL)
return "MKL";

46
ggml/src/ggml-cann/CMakeLists.txt Normal file
View file

@ -0,0 +1,46 @@
if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOME})
set(CANN_INSTALL_DIR $ENV{ASCEND_TOOLKIT_HOME})
message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$ENV{ASCEND_TOOLKIT_HOME}")
endif()
if (CANN_INSTALL_DIR)
# Only Support Linux.
if (NOT UNIX)
message(FATAL_ERROR "CANN: CANN toolkit supports unix but not ${CMAKE_SYSTEM_NAME}")
endif()
# Supported platforms: x86-64, arm64
if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64")
else()
message(FATAL_ERROR "CANN: CANN toolkit supports x86-64 and arm64 but not ${CMAKE_SYSTEM_PROCESSOR}")
endif()
# Set header and libs
set(CANN_INCLUDE_DIRS
${CANN_INSTALL_DIR}/include
${CANN_INSTALL_DIR}/include/aclnn
${CANN_INSTALL_DIR}/acllib/include
)
add_subdirectory(kernels)
list(APPEND CANN_LIBRARIES
ascendcl
nnopbase
opapi
acl_op_compiler
ascendc_kernels
)
file(GLOB GGML_SOURCES_CANN "*.cpp")
add_library(ggml-cann ${GGML_SOURCES_CANN})
target_link_libraries(ggml-cann PRIVATE ggml-base ${CANN_LIBRARIES})
target_include_directories(ggml-cann PRIVATE . .. ${CANN_INCLUDE_DIRS})
target_link_directories(ggml-cann PRIVATE ${CANN_INSTALL_DIR}/lib64)
message(STATUS "CANN: CANN_INCLUDE_DIRS = ${CANN_INCLUDE_DIRS}")
message(STATUS "CANN: CANN_LIBRARIES = ${CANN_LIBRARIES}")
else()
message(FATAL_ERROR "CANN: Can't find CANN_INSTALL_DIR, did you forget to source set_var.sh?")
endif()

0
ggml/src/ggml-cann.cpp → ggml/src/ggml-cann/ggml-cann.cpp
View file

249
ggml/src/ggml-cpu/CMakeLists.txt Normal file
View file

@ -0,0 +1,249 @@
add_library(ggml-cpu
ggml-cpu.c
ggml-cpu.cpp
ggml-cpu-aarch64.c
ggml-cpu-aarch64.h
ggml-cpu-quants.c
ggml-cpu-quants.h
)
target_link_libraries(ggml-cpu PRIVATE ggml-base)
target_include_directories(ggml-cpu PRIVATE . ..)
if (APPLE AND GGML_ACCELERATE)
find_library(ACCELERATE_FRAMEWORK Accelerate)
if (ACCELERATE_FRAMEWORK)
message(STATUS "Accelerate framework found")
add_compile_definitions(GGML_USE_ACCELERATE)
add_compile_definitions(ACCELERATE_NEW_LAPACK)
add_compile_definitions(ACCELERATE_LAPACK_ILP64)
target_link_libraries(ggml-cpu PRIVATE ${ACCELERATE_FRAMEWORK})
else()
message(WARNING "Accelerate framework not found")
endif()
endif()
if (GGML_OPENMP)
find_package(OpenMP)
if (OpenMP_FOUND)
message(STATUS "OpenMP found")
add_compile_definitions(GGML_USE_OPENMP)
target_link_libraries(ggml-cpu PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
# FIXME: should be replaced with a compiler id check
#if (GGML_MUSA)
# list(APPEND GGML_CPU_EXTRA_INCLUDES "/usr/lib/llvm-14/lib/clang/14.0.0/include")
# list(APPEND GGML_CPU_EXTRA_LIBS_PRIVATE "/usr/lib/llvm-14/lib/libomp.so")
#endif()
else()
message(WARNING "OpenMP not found")
endif()
endif()
if (GGML_LLAMAFILE)
message(STATUS "Using llamafile")
add_compile_definitions(GGML_USE_LLAMAFILE)
target_sources(ggml-cpu PRIVATE
llamafile/sgemm.cpp
llamafile/sgemm.h)
endif()
if (GGML_CPU_HBM)
find_library(memkind memkind REQUIRED)
message(STATUS "Using memkind for CPU HBM")
add_compile_definitions(GGML_USE_CPU_HBM)
target_link_libraries(ggml-cpu PUBLIC memkind)
endif()
if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR
CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
(NOT CMAKE_OSX_ARCHITECTURES AND
NOT CMAKE_GENERATOR_PLATFORM_LWR AND
CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
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)
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() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
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 "")
list(APPEND ARCH_FLAGS -mfp16-format=ieee)
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
# Raspberry Pi 1, Zero
list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
# Android armeabi-v7a
list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
else()
# Raspberry Pi 2
list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
endif()
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
# Android arm64-v8a
# Raspberry Pi 3, 4, Zero 2 (32-bit)
list(APPEND ARCH_FLAGS -mno-unaligned-access)
endif()
if (GGML_SVE)
list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
endif()
endif()
elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
(NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
message(STATUS "x86 detected")
if (MSVC)
# instruction set detection for MSVC only
if (GGML_NATIVE)
# TODO: improve, should not reference files from the parent folder
include(cmake/FindSIMD.cmake)
endif ()
if (GGML_AVX512)
list(APPEND ARCH_FLAGS /arch:AVX512)
# MSVC has no compile-time flags enabling specific
# AVX512 extensions, neither it defines the
# macros corresponding to the extensions.
# Do it manually.
if (GGML_AVX512_VBMI)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
endif()
if (GGML_AVX512_VNNI)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
endif()
if (GGML_AVX512_BF16)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
endif()
if (GGML_AMX_TILE)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
endif()
if (GGML_AMX_INT8)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
endif()
if (GGML_AMX_BF16)
add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
endif()
elseif (GGML_AVX2)
list(APPEND ARCH_FLAGS /arch:AVX2)
elseif (GGML_AVX)
list(APPEND ARCH_FLAGS /arch:AVX)
endif()
else()
if (GGML_NATIVE)
list(APPEND ARCH_FLAGS -march=native)
endif()
if (GGML_F16C)
list(APPEND ARCH_FLAGS -mf16c)
endif()
if (GGML_FMA)
list(APPEND ARCH_FLAGS -mfma)
endif()
if (GGML_AVX)
list(APPEND ARCH_FLAGS -mavx)
endif()
if (GGML_AVX2)
list(APPEND ARCH_FLAGS -mavx2)
endif()
if (GGML_AVX512)
list(APPEND ARCH_FLAGS -mavx512f)
list(APPEND ARCH_FLAGS -mavx512dq)
list(APPEND ARCH_FLAGS -mavx512bw)
endif()
if (GGML_AVX512_VBMI)
list(APPEND ARCH_FLAGS -mavx512vbmi)
endif()
if (GGML_AVX512_VNNI)
list(APPEND ARCH_FLAGS -mavx512vnni)
endif()
if (GGML_AVX512_BF16)
list(APPEND ARCH_FLAGS -mavx512bf16)
endif()
if (GGML_AMX_TILE)
list(APPEND ARCH_FLAGS -mamx-tile)
endif()
if (GGML_AMX_INT8)
list(APPEND ARCH_FLAGS -mamx-int8)
endif()
if (GGML_AMX_BF16)
list(APPEND ARCH_FLAGS -mamx-bf16)
endif()
endif()
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
message(STATUS "PowerPC detected")
execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1"
OUTPUT_VARIABLE POWER10_M)
string(FIND ${POWER10_M} "POWER10" substring_index)
if(${substring_index} GREATER_EQUAL 0)
list(APPEND ARCH_FLAGS -mcpu=power10)
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
else()
list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
#TODO: Add targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
endif()
elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
message(STATUS "loongarch64 detected")
list(APPEND ARCH_FLAGS -march=loongarch64)
if (GGML_LASX)
list(APPEND ARCH_FLAGS -mlasx)
endif()
if (GGML_LSX)
list(APPEND ARCH_FLAGS -mlsx)
endif()
else()
message(STATUS "Unknown architecture")
endif()
if (GGML_CPU_AARCH64)
message(STATUS "Using runtime weight conversion of Q4_0 to Q4_0_x_x to enable optimized GEMM/GEMV kernels")
add_compile_definitions(GGML_USE_CPU_AARCH64)
endif()
target_compile_options(ggml-cpu PRIVATE "$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
target_compile_options(ggml-cpu PRIVATE "$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
if (EMSCRIPTEN)
set_target_properties(ggml-cpu PROPERTIES COMPILE_FLAGS "-msimd128")
endif()

0
ggml/cmake/FindSIMD.cmake → ggml/src/ggml-cpu/cmake/FindSIMD.cmake
View file

3531
ggml/src/ggml-cpu/ggml-cpu-aarch64.c Normal file
View file

File diff suppressed because it is too large Load diff

30
ggml/src/ggml-cpu/ggml-cpu-aarch64.h Normal file
View file

@ -0,0 +1,30 @@
#pragma once
#include "ggml.h"
// GGML internal header
#ifdef __cplusplus
extern "C" {
#endif
// Quantization
void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
// GEMV
void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
// GEMM
void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * data, size_t data_size);
enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur);
#ifdef __cplusplus
}
#endif

243
ggml/src/ggml-cpu-impl.h → ggml/src/ggml-cpu/ggml-cpu-impl.h
View file

@ -27,80 +27,6 @@ extern "C" {
#endif
/**
* Converts brain16 to float32.
*
* The bfloat16 floating point format has the following structure:
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌──┴───┐┌─┴───┐
* 0b0000000000000000 brain16
*
* Since bf16 has the same number of exponent bits as a 32bit float,
* encoding and decoding numbers becomes relatively straightforward.
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌──┴───┐┌─┴───────────────────┐
* 0b00000000000000000000000000000000 IEEE binary32
*
* For comparison, the standard fp16 format has fewer exponent bits.
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌─┴─┐┌─┴──────┐
* 0b0000000000000000 IEEE binary16
*
* @see IEEE 754-2008
*/
static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
union {
float f;
uint32_t i;
} u;
u.i = (uint32_t)h.bits << 16;
return u.f;
}
/**
* Converts float32 to brain16.
*
* This is binary identical with Google Brain float conversion.
* Floats shall round to nearest even, and NANs shall be quiet.
* Subnormals aren't flushed to zero, except perhaps when used.
* This code should vectorize nicely if using modern compilers.
*/
static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
ggml_bf16_t h;
union {
float f;
uint32_t i;
} u;
u.f = s;
if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
h.bits = (u.i >> 16) | 64; /* force to quiet */
return h;
}
h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
return h;
}
#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
#ifndef __FMA__
@ -388,28 +314,6 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
#endif // defined(__ARM_NEON)
#if defined(__ARM_NEON) && !defined(_MSC_VER)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
ggml_fp16_internal_t tmp;
memcpy(&tmp, &h, sizeof(ggml_fp16_t));
return (float)tmp;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
ggml_fp16_t res;
ggml_fp16_internal_t tmp = f;
memcpy(&res, &tmp, sizeof(ggml_fp16_t));
return res;
}
#else
#ifdef __wasm_simd128__
#include <wasm_simd128.h>
#else
@ -462,153 +366,6 @@ static __m256 __lasx_xvreplfr2vr_s(float val) {
}
#endif
#ifdef __F16C__
#ifdef _MSC_VER
#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
#else
#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
#endif
#elif defined(__POWER9_VECTOR__)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
/* the inline asm below is about 12% faster than the lookup method */
#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
register float f;
register double d;
__asm__(
"mtfprd %0,%2\n"
"xscvhpdp %0,%0\n"
"frsp %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=f"(f):
/* in */ "r"(h));
return f;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
register double d;
register ggml_fp16_t r;
__asm__( /* xscvdphp can work on double or single precision */
"xscvdphp %0,%2\n"
"mffprd %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=r"(r):
/* in */ "f"(f));
return r;
}
#else
// FP16 <-> FP32
// ref: https://github.com/Maratyszcza/FP16
static inline float fp32_from_bits(uint32_t w) {
union {
uint32_t as_bits;
float as_value;
} fp32;
fp32.as_bits = w;
return fp32.as_value;
}
static inline uint32_t fp32_to_bits(float f) {
union {
float as_value;
uint32_t as_bits;
} fp32;
fp32.as_value = f;
return fp32.as_bits;
}
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
const uint32_t w = (uint32_t) h << 16;
const uint32_t sign = w & UINT32_C(0x80000000);
const uint32_t two_w = w + w;
const uint32_t exp_offset = UINT32_C(0xE0) << 23;
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float exp_scale = 0x1.0p-112f;
#else
const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
#endif
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float scale_to_inf = 0x1.0p+112f;
const float scale_to_zero = 0x1.0p-110f;
#else
const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
#endif
float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
const uint32_t w = fp32_to_bits(f);
const uint32_t shl1_w = w + w;
const uint32_t sign = w & UINT32_C(0x80000000);
uint32_t bias = shl1_w & UINT32_C(0xFF000000);
if (bias < UINT32_C(0x71000000)) {
bias = UINT32_C(0x71000000);
}
base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
const uint32_t bits = fp32_to_bits(base);
const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
const uint32_t nonsign = exp_bits + mantissa_bits;
return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
}
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#endif // __F16C__
#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
#ifdef __ARM_FEATURE_SVE
#include <arm_sve.h>
#endif // __ARM_FEATURE_SVE
// precomputed f32 table for f16 (256 KB)
// defined in ggml.c, initialized in ggml_init()
extern float ggml_table_f32_f16[1 << 16];
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
// This is also true for POWER9.
#if !defined(GGML_FP16_TO_FP32)
inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
uint16_t s;
memcpy(&s, &f, sizeof(uint16_t));
return ggml_table_f32_f16[s];
}
#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
#endif
#if !defined(GGML_FP32_TO_FP16)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#endif
#ifdef __cplusplus
}
#endif

10796
ggml/src/ggml-cpu/ggml-cpu-quants.c Normal file
View file

File diff suppressed because it is too large Load diff

63
ggml/src/ggml-cpu/ggml-cpu-quants.h Normal file
View file

@ -0,0 +1,63 @@
#pragma once
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "ggml.h"
// GGML CPU internal header
#ifdef __cplusplus
extern "C" {
#endif
// Quantization
void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t 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);
void ggml_vec_dot_q4_1_q8_1(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_q5_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);
void ggml_vec_dot_q5_1_q8_1(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_q8_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);
void ggml_vec_dot_q2_K_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_q3_K_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_q4_K_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_q5_K_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_q6_K_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_tq1_0_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_tq2_0_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_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_iq2_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);
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);
void ggml_vec_dot_iq1_m_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_iq4_nl_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);
void ggml_vec_dot_iq4_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_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);
#ifdef __cplusplus
}
#endif

481
ggml/src/ggml-cpu.c → ggml/src/ggml-cpu/ggml-cpu.c
View file

@ -1,13 +1,15 @@
#define _CRT_SECURE_NO_DEPRECATE // Disables "unsafe" warnings on Windows
#define _USE_MATH_DEFINES // For M_PI on MSVC
#include "ggml-aarch64.h"
#include "ggml-backend-impl.h"
#include "ggml-backend.h"
#include "ggml-cpu-aarch64.h"
#include "ggml-cpu-impl.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include "ggml-quants.h"
#include "ggml-cpu-quants.h"
#include "ggml-threading.h"
#include "ggml.h"
#if defined(_MSC_VER) || defined(__MINGW32__)
@ -42,7 +44,7 @@
#endif
#ifdef GGML_USE_LLAMAFILE
#include <llamafile/sgemm.h>
#include "llamafile/sgemm.h"
#endif
#if defined(_MSC_VER)
@ -104,9 +106,6 @@ static ggml_fp16_t ggml_table_gelu_f16[1 << 16];
// precomputed quick gelu table for f16 (128 KB)
static ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
// precomputed f32 table for f16 (256 KB) (ggml-impl.h)
float ggml_table_f32_f16[1 << 16];
#if defined(__ARM_ARCH)
struct ggml_arm_arch_features_type {
int has_neon;
@ -261,11 +260,13 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.nrows = 1,
},
[GGML_TYPE_F16] = {
.from_float = (ggml_from_float_t) ggml_fp32_to_fp16_row,
.vec_dot = (ggml_vec_dot_t) ggml_vec_dot_f16,
.vec_dot_type = GGML_TYPE_F16,
.nrows = 1,
},
[GGML_TYPE_Q4_0] = {
.from_float = quantize_row_q4_0,
.vec_dot = ggml_vec_dot_q4_0_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0,
#if defined (__ARM_FEATURE_MATMUL_INT8)
@ -275,6 +276,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
#endif
},
[GGML_TYPE_Q4_1] = {
.from_float = quantize_row_q4_1,
.vec_dot = ggml_vec_dot_q4_1_q8_1,
.vec_dot_type = GGML_TYPE_Q8_1,
#if defined (__ARM_FEATURE_MATMUL_INT8)
@ -283,27 +285,20 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.nrows = 1,
#endif
},
[4] = { // GGML_TYPE_Q4_2
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_COUNT,
.nrows = 1,
},
[5] = { // GGML_TYPE_Q4_3
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_COUNT,
.nrows = 1,
},
[GGML_TYPE_Q5_0] = {
.from_float = quantize_row_q5_0,
.vec_dot = ggml_vec_dot_q5_0_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
},
[GGML_TYPE_Q5_1] = {
.from_float = quantize_row_q5_1,
.vec_dot = ggml_vec_dot_q5_1_q8_1,
.vec_dot_type = GGML_TYPE_Q8_1,
.nrows = 1,
},
[GGML_TYPE_Q8_0] = {
.from_float = quantize_row_q8_0,
.from_float_to_mat = quantize_mat_q8_0,
.vec_dot = ggml_vec_dot_q8_0_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0,
@ -314,85 +309,106 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
#endif
},
[GGML_TYPE_Q8_1] = {
.from_float = quantize_row_q8_1,
.vec_dot_type = GGML_TYPE_Q8_1,
.nrows = 1,
},
[GGML_TYPE_Q2_K] = {
.from_float = quantize_row_q2_K,
.vec_dot = ggml_vec_dot_q2_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_Q3_K] = {
.from_float = quantize_row_q3_K,
.vec_dot = ggml_vec_dot_q3_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_Q4_K] = {
.from_float = quantize_row_q4_K,
.vec_dot = ggml_vec_dot_q4_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_Q5_K] = {
.from_float = quantize_row_q5_K,
.vec_dot = ggml_vec_dot_q5_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_Q6_K] = {
.from_float = quantize_row_q6_K,
.vec_dot = ggml_vec_dot_q6_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ2_XXS] = {
.from_float = NULL,
.vec_dot = ggml_vec_dot_iq2_xxs_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ2_XS] = {
.from_float = NULL,
.vec_dot = ggml_vec_dot_iq2_xs_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ3_XXS] = {
// NOTE: from_float for iq3 and iq2_s was removed because these quants require initialization in ggml_quantize_init
//.from_float = quantize_row_iq3_xxs,
.vec_dot = ggml_vec_dot_iq3_xxs_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ3_S] = {
//.from_float = quantize_row_iq3_s,
.vec_dot = ggml_vec_dot_iq3_s_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ2_S] = {
//.from_float = quantize_row_iq2_s,
.vec_dot = ggml_vec_dot_iq2_s_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ1_S] = {
.from_float = NULL,
.vec_dot = ggml_vec_dot_iq1_s_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ1_M] = {
.from_float = NULL,
.vec_dot = ggml_vec_dot_iq1_m_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_IQ4_NL] = {
.from_float = quantize_row_iq4_nl,
.vec_dot = ggml_vec_dot_iq4_nl_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
},
[GGML_TYPE_IQ4_XS] = {
.from_float = quantize_row_iq4_xs,
.vec_dot = ggml_vec_dot_iq4_xs_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_Q8_K] = {
.from_float = quantize_row_q8_K,
},
[GGML_TYPE_BF16] = {
.from_float = (ggml_from_float_t) ggml_fp32_to_bf16_row,
.vec_dot = (ggml_vec_dot_t) ggml_vec_dot_bf16,
.vec_dot_type = GGML_TYPE_BF16,
.nrows = 1,
},
[GGML_TYPE_Q4_0_4_4] = {
.from_float = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
@ -401,6 +417,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.gemm = ggml_gemm_q4_0_4x4_q8_0,
},
[GGML_TYPE_Q4_0_4_8] = {
.from_float = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
@ -409,17 +426,22 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
.gemm = ggml_gemm_q4_0_4x8_q8_0,
},
[GGML_TYPE_Q4_0_8_8] = {
.from_float = NULL,
.vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1,
.ncols = 8,
.gemv = ggml_gemv_q4_0_8x8_q8_0,
.gemm = ggml_gemm_q4_0_8x8_q8_0,
},
[GGML_TYPE_TQ1_0] = {
.from_float = quantize_row_tq1_0,
.vec_dot = ggml_vec_dot_tq1_0_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
},
[GGML_TYPE_TQ2_0] = {
.from_float = quantize_row_tq2_0,
.vec_dot = ggml_vec_dot_tq2_0_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1,
@ -2248,22 +2270,7 @@ struct ggml_state {
struct ggml_numa_nodes numa;
};
// global state
static struct ggml_state g_state = {0};
static atomic_flag g_state_critical = ATOMIC_FLAG_INIT;
// TODO: move to threading file
// critical section via spin lock
void ggml_critical_section_start(void) {
while (atomic_flag_test_and_set(&g_state_critical)) {
// spin
sched_yield();
}
}
void ggml_critical_section_end(void) {
atomic_flag_clear(&g_state_critical);
}
static void ggml_barrier(struct ggml_threadpool * tp) {
int n_threads = atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed);
@ -2995,8 +3002,8 @@ static void ggml_compute_forward_dup_f16(
id += ne00 * (ne01 - ir1);
}
}
} else if (ggml_get_type_traits(dst->type)->from_float) {
ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
} else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
size_t id = 0;
@ -3276,8 +3283,8 @@ static void ggml_compute_forward_dup_bf16(
id += ne00 * (ne01 - ir1);
}
}
} else if (ggml_get_type_traits(dst->type)->from_float) {
ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
} else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
size_t id = 0;
@ -3592,8 +3599,8 @@ static void ggml_compute_forward_dup_f32(
id += rs * (ne01 - ir1);
}
}
} else if (ggml_get_type_traits(dst->type)->from_float) {
ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
} else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
size_t id = 0;
size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type));
@ -4375,7 +4382,7 @@ static void ggml_compute_forward_add_q_f32(
const enum ggml_type type = src0->type;
const enum ggml_type dtype = dst->type;
ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float;
ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dtype)->from_float;
ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dtype)->from_float;
// we don't support permuted src0 or src1
GGML_ASSERT(nb00 == ggml_type_size(type));
@ -4677,7 +4684,7 @@ static void ggml_compute_forward_add1_q_f32(
const enum ggml_type type = src0->type;
ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float;
ggml_from_float_t const quantize_row_q = ggml_get_type_traits(type)->from_float;
ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(type)->from_float;
// we don't support permuted src0
GGML_ASSERT(nb00 == ggml_type_size(type));
@ -7323,6 +7330,7 @@ static void ggml_compute_forward_group_norm(
static void ggml_compute_forward_mul_mat_one_chunk(
const struct ggml_compute_params * params,
struct ggml_tensor * dst,
const enum ggml_type type,
const int64_t num_rows_per_vec_dot,
const int64_t ir0_start,
const int64_t ir0_end,
@ -7334,8 +7342,6 @@ static void ggml_compute_forward_mul_mat_one_chunk(
GGML_TENSOR_BINARY_OP_LOCALS
const enum ggml_type type = src0->type;
const bool src1_cont = ggml_is_contiguous(src1);
ggml_vec_dot_t const vec_dot = type_traits_cpu[type].vec_dot;
@ -7423,10 +7429,14 @@ static void ggml_compute_forward_mul_mat(
const int ith = params->ith;
const int nth = params->nth;
const enum ggml_type type = src0->type;
enum ggml_type type = src0->type;
if (src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) {
type = (enum ggml_type)(intptr_t)src0->extra;
}
enum ggml_type const vec_dot_type = type_traits_cpu[type].vec_dot_type;
ggml_from_float_t const from_float = ggml_get_type_traits(vec_dot_type)->from_float;
ggml_from_float_t const from_float = type_traits_cpu[vec_dot_type].from_float;
ggml_from_float_to_mat_t const from_float_to_mat = type_traits_cpu[vec_dot_type].from_float_to_mat;
int64_t const vec_dot_num_rows = type_traits_cpu[type].nrows;
int64_t const matmul_num_cols = type_traits_cpu[type].ncols;
@ -7462,15 +7472,15 @@ static void ggml_compute_forward_mul_mat(
if (src1_cont) {
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(src0->type),
nb01/ggml_type_size(type),
(const char *)src1->data + i12*nb12 + i13*nb13,
nb11/ggml_type_size(src1->type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),
ith, nth,
src0->type,
type,
src1->type,
dst->type))
goto UseGgmlGemm1;
@ -7523,15 +7533,15 @@ UseGgmlGemm1:;
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(src0->type),
nb01/ggml_type_size(type),
(const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
row_size/ggml_type_size(vec_dot_type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),
ith, nth,
src0->type,
type,
vec_dot_type,
dst->type))
goto UseGgmlGemm2;
@ -7616,7 +7626,7 @@ UseGgmlGemm2:;
const int64_t ir1_start = dr1 * ith1;
const int64_t ir1_end = MIN(ir1_start + dr1, nr1);
ggml_compute_forward_mul_mat_one_chunk(params, dst, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
ggml_compute_forward_mul_mat_one_chunk(params, dst, type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
if (nth >= nchunk0 * nchunk1) {
break;
@ -7647,7 +7657,7 @@ static void ggml_compute_forward_mul_mat_id(
ggml_vec_dot_t const vec_dot = type_traits_cpu[type].vec_dot;
enum ggml_type const vec_dot_type = type_traits_cpu[type].vec_dot_type;
ggml_from_float_t const from_float = ggml_get_type_traits(vec_dot_type)->from_float;
ggml_from_float_t const from_float = type_traits_cpu[vec_dot_type].from_float;
int64_t const matmul_num_cols = type_traits_cpu[type].ncols;
ggml_gemv_t const gemv = type_traits_cpu[type].gemv;
@ -9157,12 +9167,6 @@ static void rope_yarn(
*sin_theta = sinf(theta) * mscale;
}
// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
// `corr_dim(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
static float ggml_rope_yarn_corr_dim(int n_dims, int n_ctx_orig, float n_rot, float base) {
return n_dims * logf(n_ctx_orig / (n_rot * 2 * (float)M_PI)) / (2 * logf(base));
}
static void ggml_rope_cache_init(
float theta_base, float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
float * cache, float sin_sign, float theta_scale) {
@ -9179,16 +9183,6 @@ static void ggml_rope_cache_init(
}
}
void ggml_rope_yarn_corr_dims(
int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
) {
// start and end correction dims
float start = floorf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_fast, freq_base));
float end = ceilf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_slow, freq_base));
dims[0] = MAX(0, start);
dims[1] = MIN(n_dims - 1, end);
}
static void ggml_compute_forward_rope_f32(
const struct ggml_compute_params * params,
struct ggml_tensor * dst,
@ -10666,7 +10660,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
enum ggml_type const k_vec_dot_type = type_traits_cpu[k->type].vec_dot_type;
ggml_from_float_t const q_to_vec_dot = ggml_get_type_traits(k_vec_dot_type)->from_float;
ggml_from_float_t const q_to_vec_dot = type_traits_cpu[k_vec_dot_type].from_float;
ggml_vec_dot_t const kq_vec_dot = type_traits_cpu[k->type].vec_dot;
ggml_to_float_t const v_to_float = ggml_get_type_traits(v->type)->to_float;
@ -11642,24 +11636,30 @@ static void ggml_compute_forward_add_rel_pos(
}
}
// ggml_compute_forward_rwkv_wkv
// ggml_compute_forward_rwkv_wkv6
static void ggml_compute_forward_rwkv_wkv_f32(
static void ggml_compute_forward_rwkv_wkv6_f32(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
const size_t T = dst->src[1]->ne[3];
const size_t C = dst->ne[0];
const size_t H = dst->src[1]->ne[2];
const size_t n_seqs = dst->src[5]->ne[1];
const int64_t T = dst->src[1]->ne[3];
const int64_t C = dst->ne[0];
const int64_t HEADS = dst->src[1]->ne[2];
const int64_t n_seqs = dst->src[5]->ne[1];
const int64_t head_size = C / HEADS;
float * dst_data = (float *) dst->data;
float * state = ((float *) dst->data) + C * T;
if (params->ith != 0) {
const int ith = params->ith;
const int nth = params->nth;
if (ith >= HEADS) {
return;
}
memset(dst_data, 0, T * C * sizeof(float));
const int h_start = (HEADS * ith) / nth;
const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ?
(HEADS * (ith + 1)) / nth : HEADS;
float * k = (float *) dst->src[0]->data;
float * v = (float *) dst->src[1]->data;
@ -11667,54 +11667,160 @@ static void ggml_compute_forward_rwkv_wkv_f32(
float * time_faaaa = (float *) dst->src[3]->data;
float * time_decay = (float *) dst->src[4]->data;
size_t t_stride = H * (C / H);
size_t t_stride = HEADS * head_size; // Same to C
size_t h_stride = C / H;
size_t h_stride_2d = (C / H) * (C / H);
size_t h_stride = C / HEADS;
GGML_ASSERT(C % HEADS == 0); // C must be divisible by HEADS
size_t h_stride_2d = head_size * head_size;
// basically fused operations:
// dst = r @ (time_faaaa * (k @ v) + state),
// state = time_decay * state + (k @ v),
// recursive through each token
for (size_t t = 0; t < T; t++) {
size_t t_offset = t * t_stride;
size_t state_offset = (C / H) * C * (t / (T / n_seqs));
float * state_cur = state + state_offset;
float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[5]->data + state_offset;
if (ith == 0) {
memset(dst_data, 0, T * C * sizeof(float));
}
ggml_barrier(params->threadpool);
for (size_t h = 0; h < H; h++) {
size_t h_offset = h * h_stride;
size_t t_h_offset = t_offset + h_offset;
size_t h_2d_offset = h * h_stride_2d;
for (size_t i = 0; i < C / H; i++) {
size_t t_h_i_offset = t_h_offset + i;
size_t h_i_offset = h_offset + i;
size_t h_2d_i_offset = h_2d_offset + i * h_stride;
#if defined(__AVX__) && !defined(__AVX512F__)
#define GGML_F32X GGML_F32x8
#define GGML_F32X_SET1 GGML_F32x8_SET1
#define GGML_F32X_LOAD GGML_F32x8_LOAD
#define GGML_F32X_STORE GGML_F32x8_STORE
#define GGML_F32X_MUL GGML_F32x8_MUL
#define GGML_F32X_FMA GGML_F32x8_FMA
#define WKV_VECTOR_SIZE 8
#elif defined(__AVX512F__)
#define GGML_F32X GGML_F32x16
#define GGML_F32X_SET1 GGML_F32x16_SET1
#define GGML_F32X_LOAD GGML_F32x16_LOAD
#define GGML_F32X_STORE GGML_F32x16_STORE
#define GGML_F32X_MUL GGML_F32x16_MUL
#define GGML_F32X_FMA GGML_F32x16_FMA
#define WKV_VECTOR_SIZE 16
#elif defined(__ARM_NEON) && defined(__aarch64__)
#define GGML_F32X GGML_F32x4
#define GGML_F32X_SET1 GGML_F32x4_SET1
#define GGML_F32X_LOAD GGML_F32x4_LOAD
#define GGML_F32X_STORE GGML_F32x4_STORE
#define GGML_F32X_MUL GGML_F32x4_MUL
#define GGML_F32X_FMA GGML_F32x4_FMA
#define WKV_VECTOR_SIZE 4
#endif
float k_val = k[t_h_i_offset];
float r_val = r[t_h_i_offset];
float time_faaaa_val = time_faaaa[h_i_offset];
// RWKV v6: different time_decay for each token.
float time_decay_val = time_decay[t_h_i_offset];
#ifdef WKV_VECTOR_SIZE
const int64_t vec_count = head_size / WKV_VECTOR_SIZE;
for (size_t j = 0; j < C / H; j ++) {
size_t t_h_j_offset = t_h_offset + j;
size_t h_2d_i_j_offset = h_2d_i_offset + j;
for (int64_t t = 0; t < T; t++) {
size_t t_offset = t * t_stride;
size_t state_offset = head_size * C * (t / (T / n_seqs));
float * state_cur = state + state_offset;
float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[5]->data + state_offset;
float v_val = v[t_h_j_offset];
float kv_val = v_val * k_val;
float prev_state_val = state_prev[h_2d_i_j_offset];
float temp_val = kv_val * time_faaaa_val + prev_state_val;
dst_data[t_h_j_offset] += temp_val * r_val;
state_cur[h_2d_i_j_offset] = prev_state_val * time_decay_val + kv_val;
for (int64_t h = h_start; h < h_end; h++) {
size_t h_offset = h * h_stride;
size_t t_h_offset = t_offset + h_offset;
size_t h_2d_offset = h * h_stride_2d;
for (int64_t i = 0; i < head_size; i++) {
size_t t_h_i_offset = t_h_offset + i;
size_t h_i_offset = h_offset + i;
size_t h_2d_i_offset = h_2d_offset + i * h_stride;
float k_val = k[t_h_i_offset];
float r_val = r[t_h_i_offset];
float time_faaaa_val = time_faaaa[h_i_offset];
float time_decay_val = time_decay[t_h_i_offset];
// Broadcast scalar values to vectors
GGML_F32X k_vec = GGML_F32X_SET1(k_val);
GGML_F32X r_vec = GGML_F32X_SET1(r_val);
GGML_F32X time_faaaa_vec = GGML_F32X_SET1(time_faaaa_val);
GGML_F32X time_decay_vec = GGML_F32X_SET1(time_decay_val);
for (int64_t j = 0; j < vec_count; j++) {
size_t base_j = j * WKV_VECTOR_SIZE;
size_t t_h_j_offset = t_h_offset + base_j;
size_t h_2d_i_j_offset = h_2d_i_offset + base_j;
// Load x elements at once
GGML_F32X v_vec = GGML_F32X_LOAD(&v[t_h_j_offset]);
GGML_F32X prev_state_vec = GGML_F32X_LOAD(&state_prev[h_2d_i_j_offset]);
GGML_F32X dst_vec = GGML_F32X_LOAD(&dst_data[t_h_j_offset]);
// Compute kv = v * k
GGML_F32X kv_vec = GGML_F32X_MUL(v_vec, k_vec);
// Compute temp = kv * time_faaaa + prev_state
GGML_F32X temp_vec = GGML_F32X_FMA(prev_state_vec, kv_vec, time_faaaa_vec);
// Update dst: dst += temp * r
dst_vec = GGML_F32X_FMA(dst_vec, temp_vec, r_vec);
GGML_F32X_STORE(&dst_data[t_h_j_offset], dst_vec);
// Update state: state = prev_state * time_decay + kv
GGML_F32X new_state_vec = GGML_F32X_FMA(kv_vec, prev_state_vec, time_decay_vec);
GGML_F32X_STORE(&state_cur[h_2d_i_j_offset], new_state_vec);
}
// Handle remaining elements, this will not be used.
for (int64_t j = vec_count * WKV_VECTOR_SIZE; j < head_size; j++) {
size_t t_h_j_offset = t_h_offset + j;
size_t h_2d_i_j_offset = h_2d_i_offset + j;
float v_val = v[t_h_j_offset];
float kv_val = v_val * k_val;
float prev_state_val = state_prev[h_2d_i_j_offset];
float temp_val = kv_val * time_faaaa_val + prev_state_val;
dst_data[t_h_j_offset] += temp_val * r_val;
state_cur[h_2d_i_j_offset] = prev_state_val * time_decay_val + kv_val;
}
}
}
}
}
#else
// basically fused operations:
// dst = r @ (time_faaaa * (k @ v) + state),
// state = time_decay * state + (k @ v),
// recursive through each token
for (int64_t t = 0; t < T; t++) {
size_t t_offset = t * t_stride;
size_t state_offset = head_size * C * (t / (T / n_seqs));
float * state_cur = state + state_offset;
float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[5]->data + state_offset;
for (int64_t h = h_start; h < h_end; h++) {
size_t h_offset = h * h_stride;
size_t t_h_offset = t_offset + h_offset;
size_t h_2d_offset = h * h_stride_2d;
for (int64_t i = 0; i < head_size; i++) {
size_t t_h_i_offset = t_h_offset + i;
size_t h_i_offset = h_offset + i;
size_t h_2d_i_offset = h_2d_offset + i * h_stride;
float k_val = k[t_h_i_offset];
float r_val = r[t_h_i_offset];
float time_faaaa_val = time_faaaa[h_i_offset];
// RWKV v6: different time_decay for each token.
float time_decay_val = time_decay[t_h_i_offset];
for (int64_t j = 0; j < head_size; j++) {
size_t t_h_j_offset = t_h_offset + j;
size_t h_2d_i_j_offset = h_2d_i_offset + j;
float v_val = v[t_h_j_offset];
float kv_val = v_val * k_val;
float prev_state_val = state_prev[h_2d_i_j_offset];
float temp_val = kv_val * time_faaaa_val + prev_state_val;
dst_data[t_h_j_offset] += temp_val * r_val;
state_cur[h_2d_i_j_offset] = prev_state_val * time_decay_val + kv_val;
}
}
}
}
#endif
}
static void ggml_compute_forward_rwkv_wkv(
static void ggml_compute_forward_rwkv_wkv6(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
@ -11723,7 +11829,7 @@ static void ggml_compute_forward_rwkv_wkv(
switch (src0->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_rwkv_wkv_f32(params, dst);
ggml_compute_forward_rwkv_wkv6_f32(params, dst);
} break;
default:
{
@ -12475,9 +12581,9 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_add_rel_pos(params, tensor);
} break;
case GGML_OP_RWKV_WKV:
case GGML_OP_RWKV_WKV6:
{
ggml_compute_forward_rwkv_wkv(params, tensor);
ggml_compute_forward_rwkv_wkv6(params, tensor);
} break;
case GGML_OP_MAP_UNARY:
{
@ -12775,7 +12881,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_OP_WIN_PART:
case GGML_OP_WIN_UNPART:
case GGML_OP_GET_REL_POS:
case GGML_OP_RWKV_WKV:
case GGML_OP_RWKV_WKV6:
case GGML_OP_MAP_UNARY:
case GGML_OP_MAP_BINARY:
case GGML_OP_MAP_CUSTOM1_F32:
@ -13645,6 +13751,151 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
return ggml_graph_compute(cgraph, &cplan);
}
int ggml_cpu_has_avx(void) {
#if defined(__AVX__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_avx_vnni(void) {
#if defined(__AVXVNNI__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_avx2(void) {
#if defined(__AVX2__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_avx512(void) {
#if defined(__AVX512F__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_avx512_vbmi(void) {
#if defined(__AVX512VBMI__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_avx512_vnni(void) {
#if defined(__AVX512VNNI__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_avx512_bf16(void) {
#if defined(__AVX512BF16__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_amx_int8(void) {
#if defined(__AMX_INT8__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_fma(void) {
#if defined(__FMA__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_arm_fma(void) {
#if defined(__ARM_FEATURE_FMA)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_riscv_v(void) {
#if defined(__riscv_v_intrinsic)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_f16c(void) {
#if defined(__F16C__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_fp16_va(void) {
#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_wasm_simd(void) {
#if defined(__wasm_simd128__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_llamafile(void) {
#if defined(GGML_USE_LLAMAFILE)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_sse3(void) {
#if defined(__SSE3__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_ssse3(void) {
#if defined(__SSSE3__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_vsx(void) {
#if defined(__POWER9_VECTOR__)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_neon(void) {
#if defined(__ARM_ARCH)
return ggml_arm_arch_features.has_neon;

663
ggml/src/ggml-cpu/ggml-cpu.cpp Normal file
View file

@ -0,0 +1,663 @@
#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-cpu.h"
#include "ggml-cpu-aarch64.h"
#include "ggml-impl.h"
#include <cctype>
#include <string>
#include <vector>
#if defined(__APPLE__)
#include <sys/types.h>
#include <sys/sysctl.h>
#endif
#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
#define NOMINMAX
#endif
#include <windows.h>
#endif
// ggml-backend interface
#ifdef GGML_USE_CPU_HBM
// buffer type HBM
#include <hbwmalloc.h>
static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_HBM";
GGML_UNUSED(buft);
}
static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
hbw_free(buffer->context);
}
static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
void * ptr;
int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
if (result != 0) {
GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
return NULL;
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_hbm_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type_get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ ggml_backend_cpu_buffer_type_is_host,
},
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type_hbm;
}
#endif
// buffer type AARCH64
static void ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
tensor->extra = (void *)ggml_aarch64_get_optimal_repack_type(tensor); // NOLINT
GGML_UNUSED(buffer);
}
static void ggml_backend_cpu_aarch64_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
GGML_ASSERT(offset == 0);
GGML_ASSERT(size == ggml_nbytes(tensor));
enum ggml_type repack_type = (enum ggml_type)(intptr_t)tensor->extra;
ggml_aarch64_repack_tensor(tensor, repack_type, data, size);
GGML_UNUSED(buffer);
}
static const char * ggml_backend_cpu_aarch64_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "CPU_AARCH64";
GGML_UNUSED(buft);
}
static ggml_backend_buffer_t ggml_backend_cpu_aarch64_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
auto * buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
if (buffer == NULL) {
return NULL;
}
buffer->buft = buft;
buffer->iface.init_tensor = ggml_backend_cpu_aarch64_buffer_init_tensor;
buffer->iface.set_tensor = ggml_backend_cpu_aarch64_buffer_set_tensor;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_aarch64 = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cpu_aarch64_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_cpu_aarch64_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .is_host = */ NULL,
},
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ NULL,
};
return &ggml_backend_cpu_buffer_type_aarch64;
}
bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft) {
return buft == ggml_backend_cpu_aarch64_buffer_type();
}
static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
static std::vector<ggml_backend_buffer_type_t> bufts = []() {
std::vector<ggml_backend_buffer_type_t> bufts;
#ifdef GGML_USE_CPU_HBM
bufts.push_back(ggml_backend_cpu_hbm_buffer_type());
#endif
#ifdef GGML_USE_CPU_AARCH64
bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
#endif
bufts.push_back(NULL);
return bufts;
}();
return bufts.data();
GGML_UNUSED(device);
}
// CPU backend - backend (stream)
struct ggml_backend_cpu_context {
int n_threads;
ggml_threadpool_t threadpool;
uint8_t * work_data;
size_t work_size;
ggml_abort_callback abort_callback;
void * abort_callback_data;
};
static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
return "CPU";
GGML_UNUSED(backend);
}
static void ggml_backend_cpu_free(ggml_backend_t backend) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
delete[] cpu_ctx->work_data;
delete cpu_ctx;
delete backend;
}
struct ggml_backend_plan_cpu {
struct ggml_cplan cplan;
struct ggml_cgraph cgraph;
};
static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
cpu_plan->cgraph = *cgraph; // FIXME: deep copy
if (cpu_plan->cplan.work_size > 0) {
cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
if (cpu_plan->cplan.work_data == NULL) {
delete cpu_plan;
return NULL;
}
}
cpu_plan->cplan.abort_callback = cpu_ctx->abort_callback;
cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return cpu_plan;
}
static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
delete[] cpu_plan->cplan.work_data;
delete cpu_plan;
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
GGML_UNUSED(backend);
}
static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
if (cpu_ctx->work_size < cplan.work_size) {
delete[] cpu_ctx->work_data;
cpu_ctx->work_data = new uint8_t[cplan.work_size];
if (cpu_ctx->work_data == NULL) {
cpu_ctx->work_size = 0;
return GGML_STATUS_ALLOC_FAILED;
}
cpu_ctx->work_size = cplan.work_size;
}
cplan.work_data = (uint8_t *)cpu_ctx->work_data;
cplan.abort_callback = cpu_ctx->abort_callback;
cplan.abort_callback_data = cpu_ctx->abort_callback_data;
return ggml_graph_compute(cgraph, &cplan);
}
static const struct ggml_backend_i ggml_backend_cpu_i = {
/* .get_name = */ ggml_backend_cpu_get_name,
/* .free = */ ggml_backend_cpu_free,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ ggml_backend_cpu_graph_plan_create,
/* .graph_plan_free = */ ggml_backend_cpu_graph_plan_free,
/* .graph_plan_update = */ NULL,
/* .graph_plan_compute = */ ggml_backend_cpu_graph_plan_compute,
/* .graph_compute = */ ggml_backend_cpu_graph_compute,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
};
static ggml_guid_t ggml_backend_cpu_guid(void) {
static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
return &guid;
}
ggml_backend_t ggml_backend_cpu_init(void) {
// initialize CPU backend now to avoid slowing the first graph computation
ggml_cpu_init();
struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
if (ctx == NULL) {
return NULL;
}
ctx->n_threads = GGML_DEFAULT_N_THREADS;
ctx->threadpool = NULL;
ctx->work_data = NULL;
ctx->work_size = 0;
ctx->abort_callback = NULL;
ctx->abort_callback_data = NULL;
ggml_backend_t cpu_backend = new ggml_backend {
/* .guid = */ ggml_backend_cpu_guid(),
/* .interface = */ ggml_backend_cpu_i,
/* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
/* .context = */ ctx,
};
if (cpu_backend == NULL) {
delete ctx;
return NULL;
}
return cpu_backend;
}
bool ggml_backend_is_cpu(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
}
void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->n_threads = n_threads;
}
void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
if (ctx->threadpool && ctx->threadpool != threadpool) {
// already had a different threadpool, pause/suspend it before switching
ggml_threadpool_pause(ctx->threadpool);
}
ctx->threadpool = threadpool;
}
void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
ctx->abort_callback = abort_callback;
ctx->abort_callback_data = abort_callback_data;
}
// CPU backend - device
struct ggml_backend_cpu_device_context {
std::string description = "CPU";
ggml_backend_cpu_device_context() {
#ifdef __APPLE__
size_t len = 0;
if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
description.resize(len);
sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
}
#elif defined(__linux__)
FILE * f = fopen("/proc/cpuinfo", "r");
if (f) {
char buf[1024];
while (fgets(buf, sizeof(buf), f)) {
if (strncmp(buf, "model name", 10) == 0) {
char * p = strchr(buf, ':');
if (p) {
p++;
while (std::isspace(*p)) {
p++;
}
while (std::isspace(p[strlen(p) - 1])) {
p[strlen(p) - 1] = '\0';
}
description = p;
break;
}
}
}
fclose(f);
}
#elif defined(_WIN32)
HKEY hKey;
if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
0,
KEY_READ,
&hKey) == ERROR_SUCCESS) {
DWORD cpu_brand_size = 0;
if (RegQueryValueExA(hKey,
TEXT("ProcessorNameString"),
NULL,
NULL,
NULL,
&cpu_brand_size) == ERROR_SUCCESS) {
description.resize(cpu_brand_size);
if (RegQueryValueExA(hKey,
TEXT("ProcessorNameString"),
NULL,
NULL,
(LPBYTE)&description[0], // NOLINT
&cpu_brand_size) == ERROR_SUCCESS) {
if (description.find('\0') != std::string::npos) {
description.resize(description.find('\0'));
}
}
}
RegCloseKey(hKey);
}
#endif
}
};
static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
return "CPU";
GGML_UNUSED(dev);
}
static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
return ctx->description.c_str();
}
static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
// TODO
*free = 0;
*total = 0;
GGML_UNUSED(dev);
}
static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
return GGML_BACKEND_DEVICE_TYPE_CPU;
GGML_UNUSED(dev);
}
static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
props->name = ggml_backend_cpu_device_get_name(dev);
props->description = ggml_backend_cpu_device_get_description(dev);
props->type = ggml_backend_cpu_device_get_type(dev);
ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
props->caps = {
/* .async = */ false,
/* .host_buffer = */ false,
/* .buffer_from_host_ptr = */ true,
/* .events = */ false,
};
}
static ggml_backend_t ggml_backend_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
return ggml_backend_cpu_init();
GGML_UNUSED(dev);
GGML_UNUSED(params);
}
static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
return ggml_backend_cpu_buffer_type();
GGML_UNUSED(dev);
}
static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
return ggml_backend_cpu_buffer_from_ptr(ptr, size);
GGML_UNUSED(dev);
GGML_UNUSED(max_tensor_size);
}
static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
const struct ggml_tensor * src0 = op->src[0];
const struct ggml_tensor * src1 = op->src[1];
if (src0 && src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) {
if (op->op != GGML_OP_MUL_MAT || src0->type != GGML_TYPE_Q4_0 || ggml_aarch64_get_optimal_repack_type(src0) == GGML_TYPE_Q4_0) {
return false;
}
}
for (int i = 1; i < GGML_MAX_SRC; i++) {
if (op->src[i] && op->src[i]->buffer && ggml_backend_cpu_buft_is_aarch64(op->src[i]->buffer->buft)) {
return false;
}
}
switch (op->op) {
case GGML_OP_CPY:
return
op->type != GGML_TYPE_IQ2_XXS &&
op->type != GGML_TYPE_IQ2_XS &&
op->type != GGML_TYPE_IQ1_S &&
op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
case GGML_OP_MUL_MAT:
return src1->type == GGML_TYPE_F32 || src1->type == ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
case GGML_OP_ROPE_BACK:
return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
case GGML_OP_IM2COL_BACK:
return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
case GGML_OP_OUT_PROD:
return (src0->type == GGML_TYPE_F32 || ggml_is_quantized(src0->type)) && src1->type == GGML_TYPE_F32;
default:
return true;
}
GGML_UNUSED(dev);
}
static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
return ggml_backend_buft_is_host(buft) || ggml_backend_cpu_buft_is_aarch64(buft);
GGML_UNUSED(dev);
}
static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
/* .get_name = */ ggml_backend_cpu_device_get_name,
/* .get_description = */ ggml_backend_cpu_device_get_description,
/* .get_memory = */ ggml_backend_cpu_device_get_memory,
/* .get_type = */ ggml_backend_cpu_device_get_type,
/* .get_props = */ ggml_backend_cpu_device_get_props,
/* .init_backend = */ ggml_backend_cpu_device_init_backend,
/* .get_buffer_type = */ ggml_backend_cpu_device_get_buffer_type,
/* .get_host_buffer_type = */ NULL,
/* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
/* .supports_op = */ ggml_backend_cpu_device_supports_op,
/* .supports_buft = */ ggml_backend_cpu_device_supports_buft,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_synchronize = */ NULL,
};
// CPU backend - backend (reg)
static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
return "CPU";
GGML_UNUSED(reg);
}
static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
return 1;
GGML_UNUSED(reg);
}
static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
GGML_ASSERT(index == 0);
static ggml_backend_cpu_device_context ctx;
static ggml_backend_device ggml_backend_cpu_device = {
/* .iface = */ ggml_backend_cpu_device_i,
/* .reg = */ reg,
/* .context = */ &ctx,
};
return &ggml_backend_cpu_device;
}
struct ggml_backend_feature {
const char * name;
const char * value;
};
// Not used yet
// This is intended to replace the the ggml_cpu_has_* functions when loading the CPU backend dynamically,
// and additionally to allow other backends to expose their own list of features that applications can query using the same API.
static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t reg) {
static std::vector<ggml_backend_feature> features = []() {
std::vector<ggml_backend_feature> features;
if (ggml_cpu_has_sse3()) {
features.push_back({ "SSE3", "1" });
}
if (ggml_cpu_has_ssse3()) {
features.push_back({ "SSSE3", "1" });
}
if (ggml_cpu_has_avx()) {
features.push_back({ "AVX", "1" });
}
if (ggml_cpu_has_avx2()) {
features.push_back({ "AVX2", "1" });
}
if (ggml_cpu_has_f16c()) {
features.push_back({ "F16C", "1" });
}
if (ggml_cpu_has_fma()) {
features.push_back({ "FMA", "1" });
}
if (ggml_cpu_has_avx_vnni()) {
features.push_back({ "AVX_VNNI", "1" });
}
if (ggml_cpu_has_avx512()) {
features.push_back({ "AVX512", "1" });
}
if (ggml_cpu_has_avx512_vbmi()) {
features.push_back({ "AVX512_VBMI", "1" });
}
if (ggml_cpu_has_avx512_vnni()) {
features.push_back({ "AVX512_VNNI", "1" });
}
if (ggml_cpu_has_avx512_bf16()) {
features.push_back({ "AVX512_BF16", "1" });
}
if (ggml_cpu_has_amx_int8()) {
features.push_back({ "AMX_INT8", "1" });
}
if (ggml_cpu_has_neon()) {
features.push_back({ "NEON", "1" });
}
if (ggml_cpu_has_arm_fma()) {
features.push_back({ "ARM_FMA", "1" });
}
if (ggml_cpu_has_fp16_va()) {
features.push_back({ "FP16_VA", "1" });
}
if (ggml_cpu_has_matmul_int8()) {
features.push_back({ "MATMUL_INT8", "1" });
}
if (ggml_cpu_has_sve()) {
features.push_back({ "SVE", "1" });
}
if (ggml_cpu_get_sve_cnt() > 0) {
static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
features.push_back({ "SVE_CNT", sve_cnt.c_str() });
}
if (ggml_cpu_has_riscv_v()) {
features.push_back({ "RISCV_V", "1" });
}
if (ggml_cpu_has_vsx()) {
features.push_back({ "VSX", "1" });
}
if (ggml_cpu_has_wasm_simd()) {
features.push_back({ "WASM_SIMD", "1" });
}
if (ggml_cpu_has_llamafile()) {
features.push_back({ "LLAMAFILE", "1" });
}
features.push_back({ nullptr, nullptr });
return features;
}();
return features.data();
GGML_UNUSED(reg);
}
static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
return (void *)ggml_backend_cpu_set_n_threads;
}
if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
return (void *)ggml_backend_cpu_get_extra_bufts;
}
return NULL;
GGML_UNUSED(reg);
}
static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
/* .get_name = */ ggml_backend_cpu_reg_get_name,
/* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
/* .get_device = */ ggml_backend_cpu_reg_get_device,
/* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
};
ggml_backend_reg_t ggml_backend_cpu_reg(void) {
// init CPU feature detection
ggml_cpu_init();
static struct ggml_backend_reg ggml_backend_cpu_reg = {
/* .iface = */ ggml_backend_cpu_reg_i,
/* .context = */ NULL,
};
return &ggml_backend_cpu_reg;
}

608
ggml/src/llamafile/sgemm.cpp → ggml/src/ggml-cpu/llamafile/sgemm.cpp
View file

@ -106,6 +106,10 @@ inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
#if defined(__MMA__)
typedef vector unsigned char vec_t;
typedef __vector_quad acc_t;
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// VECTORIZED FUSED MULTIPLY ADD
@ -1026,6 +1030,600 @@ class tinyBLAS_Q0_AVX {
};
#endif // __AVX__
//PPC Implementation
#if defined(__MMA__)
#define SAVE_ACC(ACC, ii, jj) \
__builtin_mma_disassemble_acc(vec_C, ACC); \
for (int I = 0; I < 4; I++) { \
for (int J = 0; J < 4; J++) { \
*((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J); \
} \
} \
template <typename TA, typename TB, typename TC>
class tinyBLAS_PPC {
public:
tinyBLAS_PPC(int64_t k,
const TA *A, int64_t lda,
const TB *B, int64_t ldb,
TC *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int64_t m, int64_t n) {
mnpack(0, m, 0, n);
}
private:
void (tinyBLAS_PPC::*kernel)(int64_t, int64_t);
void READ_BLOCK(const float* a, int64_t lda, int rows, int cols, float* vec) {
int64_t i, j;
float *aoffset = NULL, *boffset = NULL;
float *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
float *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
aoffset = const_cast<float*>(a);
boffset = vec;
j = (rows >> 3);
if (j > 0) {
do {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
aoffset4 = aoffset3 + lda;
aoffset5 = aoffset4 + lda;
aoffset6 = aoffset5 + lda;
aoffset7 = aoffset6 + lda;
aoffset8 = aoffset7 + lda;
aoffset += 8 * lda;
i = (cols >> 3);
if (i > 0) {
__vector_pair C1, C2, C3, C4, C5, C6, C7, C8;
vector float c1[2], c2[2], c3[2], c4[2], c5[2], c6[2], c7[2], c8[2];
vector float t1, t2, t3, t4, t5, t6, t7, t8;
do {
C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1);
C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2);
C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3);
C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4);
C5 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset5);
C6 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset6);
C7 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset7);
C8 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset8);
__builtin_vsx_disassemble_pair(c1, &C1);
__builtin_vsx_disassemble_pair(c2, &C2);
__builtin_vsx_disassemble_pair(c3, &C3);
__builtin_vsx_disassemble_pair(c4, &C4);
__builtin_vsx_disassemble_pair(c5, &C5);
__builtin_vsx_disassemble_pair(c6, &C6);
__builtin_vsx_disassemble_pair(c7, &C7);
__builtin_vsx_disassemble_pair(c8, &C8);
t1 = vec_mergeh(c1[0], c2[0]);
t2 = vec_mergeh(c3[0], c4[0]);
t3 = vec_mergeh(c5[0], c6[0]);
t4 = vec_mergeh(c7[0], c8[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset);
vec_xst(t6, 0, boffset+4);
vec_xst(t7, 0, boffset+8);
vec_xst(t8, 0, boffset+12);
t1 = vec_mergel(c1[0], c2[0]);
t2 = vec_mergel(c3[0], c4[0]);
t3 = vec_mergel(c5[0], c6[0]);
t4 = vec_mergel(c7[0], c8[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+16);
vec_xst(t6, 0, boffset+20);
vec_xst(t7, 0, boffset+24);
vec_xst(t8, 0, boffset+28);
t1 = vec_mergeh(c1[1], c2[1]);
t2 = vec_mergeh(c3[1], c4[1]);
t3 = vec_mergeh(c5[1], c6[1]);
t4 = vec_mergeh(c7[1], c8[1]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+32);
vec_xst(t6, 0, boffset+36);
vec_xst(t7, 0, boffset+40);
vec_xst(t8, 0, boffset+44);
t1 = vec_mergel(c1[1], c2[1]);
t2 = vec_mergel(c3[1], c4[1]);
t3 = vec_mergel(c5[1], c6[1]);
t4 = vec_mergel(c7[1], c8[1]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+48);
vec_xst(t6, 0, boffset+52);
vec_xst(t7, 0, boffset+56);
vec_xst(t8, 0, boffset+60);
aoffset1 += 8*lda;
aoffset2 += 8*lda;
aoffset3 += 8*lda;
aoffset4 += 8*lda;
boffset += 64;
i--;
} while(i > 0);
}
if (cols & 4) {
vector float c1, c2, c3, c4, c5, c6, c7, c8;
vector float t1, t2, t3, t4, t5, t6, t7, t8;
c1 = vec_xl(0, aoffset1);
c2 = vec_xl(0, aoffset2);
c3 = vec_xl(0, aoffset3);
c4 = vec_xl(0, aoffset4);
c5 = vec_xl(0, aoffset5);
c6 = vec_xl(0, aoffset6);
c7 = vec_xl(0, aoffset7);
c8 = vec_xl(0, aoffset8);
t1 = vec_mergeh(c1, c2);
t2 = vec_mergeh(c3, c4);
t3 = vec_mergeh(c5, c6);
t4 = vec_mergeh(c7, c8);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset);
vec_xst(t6, 0, boffset+4);
vec_xst(t7, 0, boffset+8);
vec_xst(t8, 0, boffset+12);
t1 = vec_mergel(c1, c2);
t2 = vec_mergel(c3, c4);
t3 = vec_mergel(c5, c6);
t4 = vec_mergel(c7, c8);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+16);
vec_xst(t6, 0, boffset+20);
vec_xst(t7, 0, boffset+24);
vec_xst(t8, 0, boffset+28);
}
j--;
} while(j > 0);
}
if (rows & 4) {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
aoffset4 = aoffset3 + lda;
aoffset += 4 * lda;
i = (cols >> 3);
if (i > 0) {
__vector_pair C1, C2, C3, C4;
vector float c1[2], c2[2], c3[2], c4[2];
vector float t1, t2, t3, t4, t5, t6, t7, t8;
do {
C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1);
C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2);
C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3);
C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4);
__builtin_vsx_disassemble_pair(c1, &C1);
__builtin_vsx_disassemble_pair(c2, &C2);
__builtin_vsx_disassemble_pair(c3, &C3);
__builtin_vsx_disassemble_pair(c4, &C4);
t1 = vec_mergeh(c1[0], c2[0]);
t2 = vec_mergeh(c3[0], c4[0]);
t3 = vec_mergel(c1[0], c2[0]);
t4 = vec_mergel(c3[0], c4[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t1, t2, 3);
t7 = vec_xxpermdi(t3, t4, 0);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset);
vec_xst(t6, 0, boffset+4);
vec_xst(t7, 0, boffset+8);
vec_xst(t8, 0, boffset+12);
t1 = vec_mergeh(c1[1], c2[1]);
t2 = vec_mergeh(c3[1], c4[1]);
t3 = vec_mergel(c1[1], c2[1]);
t4 = vec_mergel(c3[1], c4[1]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t1, t2, 3);
t7 = vec_xxpermdi(t3, t4, 0);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+16);
vec_xst(t6, 0, boffset+20);
vec_xst(t7, 0, boffset+24);
vec_xst(t8, 0, boffset+28);
aoffset1 += 8*lda;
aoffset2 += 8*lda;
aoffset3 += 8*lda;
aoffset4 += 8*lda;
boffset += 32;
i--;
} while(i > 0);
}
if (cols & 4) {
vector float c1, c2, c3, c4;
vector float t1, t2, t3, t4;
c1 = vec_xl(0, aoffset1);
c2 = vec_xl(0, aoffset2);
c3 = vec_xl(0, aoffset3);
c4 = vec_xl(0, aoffset4);
t1 = vec_mergeh(c1, c2);
t2 = vec_mergeh(c3, c4);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset);
vec_xst(t4, 0, boffset+4);
t1 = vec_mergel(c1, c2);
t2 = vec_mergel(c3, c4);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset+8);
vec_xst(t4, 0, boffset+12);
}
}
if (rows & 3) {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
if (cols & 4) {
vector float c1, c2, c3, c4 = {0};
vector float t1, t2, t3, t4;
c1 = vec_xl(0, aoffset1);
c2 = vec_xl(0, aoffset2);
c3 = vec_xl(0, aoffset3);
t1 = vec_mergeh(c1, c2);
t2 = vec_mergeh(c3, c4);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset);
vec_xst(t4, 0, boffset+4);
t1 = vec_mergel(c1, c2);
t2 = vec_mergel(c3, c4);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset+8);
vec_xst(t4, 0, boffset+12);
}
}
}
void KERNEL_4x4(int64_t ii, int64_t jj) {
vec_t vec_A[4], vec_B[4], vec_C[4];
acc_t acc_0;
__builtin_mma_xxsetaccz(&acc_0);
for (int l = 0; l < k; l+=4) {
READ_BLOCK(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[3], vec_B[3]);
}
SAVE_ACC(&acc_0, ii, jj);
}
void KERNEL_4x8(int64_t ii, int64_t jj) {
vec_t vec_A[4], vec_B[8], vec_C[4];
acc_t acc_0, acc_1;
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
for (int64_t l = 0; l < k; l+=4) {
READ_BLOCK(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, 8, 4, (float*)vec_B);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[0], (vec_t)vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[0], (vec_t)vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[1], (vec_t)vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[1], (vec_t)vec_B[3]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[2], (vec_t)vec_B[4]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[2], (vec_t)vec_B[5]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[3], (vec_t)vec_B[6]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[3], (vec_t)vec_B[7]);
}
SAVE_ACC(&acc_0, ii, jj);
SAVE_ACC(&acc_1, ii, jj+4);
}
void KERNEL_8x4(int64_t ii, int64_t jj) {
vec_t vec_A[8], vec_B[4], vec_C[4];
acc_t acc_0, acc_1;
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
for (int64_t l = 0; l < k; l+=4) {
READ_BLOCK(A+(ii*lda)+l, lda, 8, 4, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[0], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[1], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[2], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[3], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[4], vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[5], vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[6], vec_B[3]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[7], vec_B[3]);
}
SAVE_ACC(&acc_0, ii, jj);
SAVE_ACC(&acc_1, ii+4, jj);
}
void KERNEL_8x8(int64_t ii, int64_t jj) {
vec_t vec_A[16], vec_B[16], vec_C[4];
acc_t acc_0, acc_1, acc_2, acc_3;
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
__builtin_mma_xxsetaccz(&acc_2);
__builtin_mma_xxsetaccz(&acc_3);
for (int l = 0; l < k; l+=8) {
READ_BLOCK(A+(ii*lda)+l, lda, 8, 8, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, 8, 8, (float*)vec_B);
for(int x = 0; x < 16; x+=2) {
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[x], vec_B[x]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[x], vec_B[x+1]);
__builtin_mma_xvf32gerpp(&acc_2, (vec_t)vec_A[x+1], vec_B[x]);
__builtin_mma_xvf32gerpp(&acc_3, (vec_t)vec_A[x+1], vec_B[x+1]);
}
}
SAVE_ACC(&acc_0, ii, jj);
SAVE_ACC(&acc_1, ii, jj+4);
SAVE_ACC(&acc_2, ii+4, jj);
SAVE_ACC(&acc_3, ii+4, jj+4);
}
void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t mc, nc, mp, np;
int m_rem = MIN(m - m0, 16);
int n_rem = MIN(n - n0, 16);
if (m_rem >= 16 && n_rem >= 8) {
mc = 8;
nc = 8;
gemm<8,8>(m0, m, n0, n);
} else if(m_rem >= 8 && n_rem >= 16) {
mc = 8;
nc = 8;
gemm<8,8>(m0, m, n0, n);
} else if (m_rem >= 8 && n_rem >= 8) {
mc = 8;
nc = 8;
gemm<8,8>(m0, m, n0, n);
} else if (m_rem >= 4 && n_rem >= 8) {
mc = 4;
nc = 8;
gemm<4,8>(m0, m, n0, n);
} else if (m_rem >= 8 && n_rem >= 4) {
mc = 8;
nc = 4;
gemm<8,4>(m0, m, n0, n);
} else if (m_rem >= 4 && n_rem >= 4) {
mc = 4;
nc = 4;
gemm<4,4>(m0, m, n0, n);
} else if ((m_rem < 4) && (n_rem > 4)) {
nc = 4;
switch(m_rem) {
case 1:
mc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 2:
mc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 3:
mc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
default:
return;
}
} else if ((m_rem > 4) && (n_rem < 4)) {
mc = 4;
switch(n_rem) {
case 1:
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 2:
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 3:
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
default:
return;
}
} else {
switch((m_rem << 4) | n_rem) {
case 0x43:
mc = 4;
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x42:
mc = 4;
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x41:
mc = 4;
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x34:
mc = 3;
nc = 4;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x33:
mc = 3;
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x32:
mc = 3;
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x31:
mc = 3;
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x24:
mc = 2;
nc = 4;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x23:
mc = 2;
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x22:
mc = 2;
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x21:
mc = 2;
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x14:
mc = 1;
nc = 4;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x13:
mc = 1;
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x12:
mc = 1;
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x11:
mc = 1;
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
default:
return;
}
}
mp = m0 + (m - m0) / mc * mc;
np = n0 + (n - n0) / nc * nc;
mnpack(mp, m, n0, np);
mnpack(m0, m, np, n);
}
void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
vec_t vec_C[4];
acc_t acc_0;
__builtin_mma_xxsetaccz(&acc_0);
vec_t vec_A[4], vec_B[4];
for (int l=0; l<k; l+=4) {
if (RN >= 4 && RM == 1) {
float* a = const_cast<float*>(A+(ii)*lda+l);
READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
vec_A[0] = (vec_t)vec_xl(0,a);
vec_A[1] = (vec_t)vec_splats(*((float*)&vec_A+1));
vec_A[2] = (vec_t)vec_splats(*((float*)&vec_A+2));
vec_A[3] = (vec_t)vec_splats(*((float*)&vec_A+3));
} else {
READ_BLOCK(A+(ii*lda)+l, lda, RM, 4, (float*)vec_A);
READ_BLOCK(B+(jj*ldb)+l, ldb, RN, 4, (float*)vec_B);
}
__builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[3], vec_B[3]);
}
__builtin_mma_disassemble_acc(vec_C, &acc_0);
for (int I = 0; I < RM; I++) {
for (int J = 0; J < RN; J++) {
*((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J);
}
}
}
}
template <int RM, int RN>
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (RM == 4 && RN == 4) {
kernel = &tinyBLAS_PPC::KERNEL_4x4;
} else if (RM == 4 && RN == 8) {
kernel = &tinyBLAS_PPC::KERNEL_4x8;
} else if (RM == 8 && RN == 4) {
kernel = &tinyBLAS_PPC::KERNEL_8x4;
} else if (RM == 8 && RN == 8) {
kernel = &tinyBLAS_PPC::KERNEL_8x8;
}
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
(this->*kernel)(ii, jj);
}
}
const TA *const A;
const TB *const B;
TC *C;
TA *At;
TB *Bt;
const int64_t k;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
const int ith;
const int nth;
};
#endif
} // namespace
/**
@ -1114,6 +1712,16 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
ith, nth};
tb.matmul(m, n);
return true;
#elif defined(__MMA__)
if (k % 8)
return false;
tinyBLAS_PPC<float, float, float> tb{
k, (const float *)A, lda,
(const float *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n);
return true;
#else
return false;
#endif

0
ggml/src/llamafile/sgemm.h → ggml/src/ggml-cpu/llamafile/sgemm.h
View file

165
ggml/src/ggml-cuda/CMakeLists.txt Normal file
View file

@ -0,0 +1,165 @@
cmake_minimum_required(VERSION 3.18) # for CMAKE_CUDA_ARCHITECTURES
find_package(CUDAToolkit)
if (CUDAToolkit_FOUND)
message(STATUS "CUDA Toolkit found")
if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
# 52 == lowest CUDA 12 standard
# 60 == FP16 CUDA intrinsics
# 61 == integer CUDA intrinsics
# 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
else()
set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
#set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
endif()
endif()
message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
enable_language(CUDA)
file(GLOB GGML_HEADERS_CUDA "*.cuh")
list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
file(GLOB GGML_SOURCES_CUDA "*.cu")
file(GLOB SRCS "template-instances/fattn-wmma*.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
file(GLOB SRCS "template-instances/mmq*.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
if (GGML_CUDA_FA_ALL_QUANTS)
file(GLOB SRCS "template-instances/fattn-vec*.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
else()
file(GLOB SRCS "template-instances/fattn-vec*q4_0-q4_0.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
file(GLOB SRCS "template-instances/fattn-vec*q8_0-q8_0.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
file(GLOB SRCS "template-instances/fattn-vec*f16-f16.cu")
list(APPEND GGML_SOURCES_CUDA ${SRCS})
endif()
add_library(ggml-cuda
${GGML_HEADERS_CUDA}
${GGML_SOURCES_CUDA}
)
target_link_libraries(ggml-cuda PRIVATE ggml-base)
target_include_directories(ggml-cuda PRIVATE . ..)
# TODO: change the definitions to this target only
add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
if (GGML_CUDA_GRAPHS)
add_compile_definitions(GGML_CUDA_USE_GRAPHS)
endif()
if (GGML_CUDA_FORCE_DMMV)
add_compile_definitions(GGML_CUDA_FORCE_DMMV)
endif()
if (GGML_CUDA_FORCE_MMQ)
add_compile_definitions(GGML_CUDA_FORCE_MMQ)
endif()
if (GGML_CUDA_FORCE_CUBLAS)
add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
endif()
if (GGML_CUDA_NO_VMM)
add_compile_definitions(GGML_CUDA_NO_VMM)
endif()
if (DEFINED GGML_CUDA_DMMV_Y)
add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_DMMV_Y}) # for backwards compatibility
endif()
if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
add_compile_definitions(GGML_CUDA_F16)
endif()
if (GGML_CUDA_NO_PEER_COPY)
add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
endif()
if (GGML_STATIC)
if (WIN32)
# As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
else ()
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
endif()
else()
target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
endif()
if (GGML_CUDA_NO_VMM)
# No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
else()
target_link_libraries(ggml-cuda PRIVATE CUDA::cuda_driver)
endif()
set(CUDA_CXX_FLAGS "")
set(CUDA_FLAGS -use_fast_math)
if (GGML_FATAL_WARNINGS)
list(APPEND CUDA_FLAGS -Werror all-warnings)
endif()
if (GGML_ALL_WARNINGS AND NOT MSVC)
set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
endif()
execute_process(
COMMAND ${NVCC_CMD} -Xcompiler --version
OUTPUT_VARIABLE CUDA_CCFULLVER
ERROR_QUIET
)
if (NOT CUDA_CCFULLVER MATCHES clang)
set(CUDA_CCID "GNU")
execute_process(
COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
OUTPUT_VARIABLE CUDA_CCVER
ERROR_QUIET
)
else()
if (CUDA_CCFULLVER MATCHES Apple)
set(CUDA_CCID "AppleClang")
else()
set(CUDA_CCID "Clang")
endif()
string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
endif()
message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
get_flags(${CUDA_CCID} ${CUDA_CCVER})
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()
list(JOIN CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED) # pass host compiler flags as a single argument
if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
endif()
add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
else()
message(FATAL_ERROR "CUDA Toolkit not found")
endif()

50
ggml/src/ggml-cuda/common.cuh
View file

@ -6,7 +6,7 @@
#include <cstdint>
#include <memory>
#if defined(GGML_USE_HIPBLAS)
#if defined(GGML_USE_HIP)
#define GGML_COMMON_DECL_HIP
#define GGML_COMMON_IMPL_HIP
#else
@ -26,13 +26,13 @@
#include <string>
#include <vector>
#if defined(GGML_USE_HIPBLAS)
#if defined(GGML_USE_HIP)
#include "vendors/hip.h"
#elif defined(GGML_USE_MUSA)
#include "vendors/musa.h"
#else
#include "vendors/cuda.h"
#endif // defined(GGML_USE_HIPBLAS)
#endif // defined(GGML_USE_HIP)
#define STRINGIZE_IMPL(...) #__VA_ARGS__
#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
@ -97,7 +97,7 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
#define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str)
#if !defined(GGML_USE_HIPBLAS)
#if !defined(GGML_USE_HIP)
static const char * cu_get_error_str(CUresult err) {
const char * err_str;
cuGetErrorString(err, &err_str);
@ -120,21 +120,21 @@ typedef float dfloat; // dequantize float
typedef float2 dfloat2;
#endif // GGML_CUDA_F16
#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
#define FP16_AVAILABLE
#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
#if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
#define FAST_FP16_AVAILABLE
#endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
#define FP16_MMA_AVAILABLE
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
#define INT8_MMA_AVAILABLE
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
#define FLASH_ATTN_AVAILABLE
@ -156,14 +156,14 @@ static constexpr bool int8_mma_available(const int cc) {
static __device__ void no_device_code(
const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
file_name, line, function_name, arch);
GGML_UNUSED(arch_list);
#else
printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
file_name, line, function_name, arch, arch_list);
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
__trap();
GGML_UNUSED(no_device_code); // suppress unused function warning
@ -176,7 +176,7 @@ static __device__ void no_device_code(
#endif // __CUDA_ARCH__
static __device__ __forceinline__ int warp_reduce_sum(int x) {
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
return __reduce_add_sync(0xffffffff, x);
#else
#pragma unroll
@ -184,7 +184,7 @@ static __device__ __forceinline__ int warp_reduce_sum(int x) {
x += __shfl_xor_sync(0xffffffff, x, mask, 32);
}
return x;
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
}
static __device__ __forceinline__ float warp_reduce_sum(float x) {
@ -207,7 +207,7 @@ static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
#ifdef FP16_AVAILABLE
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
@ -221,7 +221,7 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
}
return a;
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#else
NO_DEVICE_CODE;
@ -240,11 +240,11 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
#ifdef FP16_AVAILABLE
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
return __float2half(fmaxf(__half2float(a), __half2float(b)));
#else
return __hmax(a, b);
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
#else
NO_DEVICE_CODE;
@ -254,7 +254,7 @@ static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b
}
static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
#if CUDART_VERSION >= CUDART_HMAX
return __hmax2(a, b);
@ -269,11 +269,11 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal
GGML_UNUSED(a);
GGML_UNUSED(b);
NO_DEVICE_CODE;
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
}
static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
@ -282,7 +282,7 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
#else
GGML_UNUSED(x);
NO_DEVICE_CODE;
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
}
#if CUDART_VERSION < CUDART_HMASK
@ -294,7 +294,7 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half
#endif // CUDART_VERSION < CUDART_HMASK
static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
c = __builtin_amdgcn_sdot4(a, b, c, false);
#elif defined(RDNA3)
@ -320,7 +320,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
#endif
return c;
#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
#if __CUDA_ARCH__ >= MIN_CC_DP4A
return __dp4a(a, b, c);
@ -330,7 +330,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
}
// TODO: move to ggml-common.h

2
ggml/src/ggml-cuda/count-equal.cu
View file

@ -44,7 +44,7 @@ void ggml_cuda_count_equal(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const int64_t ne = ggml_nelements(src0);
GGML_ASSERT(ne < (1 << 30) && "atomicAdd implementation only supports int");
const int64_t dne = GGML_PAD(ne / (4*nsm), CUDA_COUNT_EQUAL_CHUNK_SIZE);
const int64_t dne = GGML_PAD((ne + 4*nsm - 1) / (4*nsm), CUDA_COUNT_EQUAL_CHUNK_SIZE);
CUDA_CHECK(cudaMemsetAsync(dst_d, 0, ggml_nbytes(dst), stream));

4
ggml/src/ggml-cuda/fattn-common.cuh
View file

@ -517,9 +517,9 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
}
template<int D, int parallel_blocks> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(D, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_combine_results(
const float * __restrict__ VKQ_parts,
const float2 * __restrict__ VKQ_meta,

4
ggml/src/ggml-cuda/fattn-tile-f16.cu
View file

@ -5,9 +5,9 @@
#define FATTN_KQ_STRIDE_TILE_F16 64
template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(nwarps*WARP_SIZE, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_tile_ext_f16(
const char * __restrict__ Q,
const char * __restrict__ K,

4
ggml/src/ggml-cuda/fattn-tile-f32.cu
View file

@ -5,9 +5,9 @@
#define FATTN_KQ_STRIDE_TILE_F32 32
template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(nwarps*WARP_SIZE, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_tile_ext_f32(
const char * __restrict__ Q,
const char * __restrict__ K,

4
ggml/src/ggml-cuda/fattn-vec-f16.cuh
View file

@ -2,9 +2,9 @@
#include "fattn-common.cuh"
template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(D, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_vec_ext_f16(
const char * __restrict__ Q,
const char * __restrict__ K,

4
ggml/src/ggml-cuda/fattn-vec-f32.cuh
View file

@ -2,9 +2,9 @@
#include "fattn-common.cuh"
template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
__launch_bounds__(D, 1)
#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
static __global__ void flash_attn_vec_ext_f32(
const char * __restrict__ Q,
const char * __restrict__ K,

Some files were not shown because too many files have changed in this diff Show more