LM/llama.cpp for SYCL: Difference between revisions

From Fundamental Ramen
< LM
Jump to navigation Jump to search
 
(8 intermediate revisions by the same user not shown)
Line 6: Line 6:
cmake -B build/sycl -G Ninja \
cmake -B build/sycl -G Ninja \
   -DGGML_SYCL=ON \
   -DGGML_SYCL=ON \
  -DGGML_SYCL_F16=ON \
  -DGGML_SYCL_FORCE_MMQ=ON \
  -DGGML_SYCL_GRAPH=ON \
  -DGGML_SYCL_DNN=ON \
   -DGGML_SYCL_TARGET=INTEL \
   -DGGML_SYCL_TARGET=INTEL \
  -DGGML_SYCL_DNN=OFF \
  -DGGML_SYCL_F16=ON \
   -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
   -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
   -DGGML_NATIVE=ON \
   -DGGML_NATIVE=ON \
Line 24: Line 26:
</syntaxhighlight>
</syntaxhighlight>


=== F16 with DNN ===
=== F16 without DNN ===


<syntaxhighlight lang="bash">
<syntaxhighlight lang="bash">
cmake -B build/sycl-dnn -G Ninja \
cmake -B build/sycl-nodnn -G Ninja \
   -DGGML_SYCL=ON \
   -DGGML_SYCL=ON \
  -DGGML_SYCL_F16=ON \
  -DGGML_SYCL_DNN=OFF \
   -DGGML_SYCL_TARGET=INTEL \
   -DGGML_SYCL_TARGET=INTEL \
  -DGGML_SYCL_DNN=ON \
  -DGGML_SYCL_F16=ON \
   -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
   -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
   -DGGML_NATIVE=ON \
   -DGGML_NATIVE=ON \
Line 43: Line 45:
   -DCMAKE_C_FLAGS="-O3 -fp-model=fast=2 -qopenmp"
   -DCMAKE_C_FLAGS="-O3 -fp-model=fast=2 -qopenmp"


cmake --build build/sycl-dnn
cmake --build build/sycl-nodnn
cmake --install build/sycl-dnn --prefix ~/llama.cpp/sycl-dnn
cmake --install build/sycl-nodnn --prefix ~/llama.cpp/sycl-nodnn
</syntaxhighlight>
</syntaxhighlight>


Line 52: Line 54:
cmake -B build/sycl-f32 -G Ninja \
cmake -B build/sycl-f32 -G Ninja \
   -DGGML_SYCL=ON \
   -DGGML_SYCL=ON \
  -DGGML_SYCL_F16=OFF \
  -DGGML_SYCL_DNN=ON \
   -DGGML_SYCL_TARGET=INTEL \
   -DGGML_SYCL_TARGET=INTEL \
  -DGGML_SYCL_DNN=OFF \
  -DGGML_SYCL_F16=OFF \
   -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
   -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
   -DGGML_NATIVE=ON \
   -DGGML_NATIVE=ON \
Line 105: Line 107:
cmake -B build/sycl -G Ninja \
cmake -B build/sycl -G Ninja \
   -DGGML_SYCL=ON \
   -DGGML_SYCL=ON \
  -DGGML_SYCL_F16=ON \
  -DGGML_SYCL_DNN=ON \
   -DGGML_SYCL_TARGET=INTEL \
   -DGGML_SYCL_TARGET=INTEL \
  -DGGML_SYCL_DNN=OFF \
  -DGGML_SYCL_F16=ON \
   -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
   -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
   -DGGML_NATIVE=ON \
   -DGGML_NATIVE=ON \
Line 148: Line 150:


<syntaxhighlight lang="bash">
<syntaxhighlight lang="bash">
cmake -B build/sycl -G Ninja \
  -DGGML_SYCL=ON \
  -DGGML_SYCL_TARGET=INTEL \
  -DGGML_SYCL_DNN=ON \
  -DGGML_SYCL_F16=ON \
  -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
  -DGGML_NATIVE=ON \
  -DGGML_LTO=ON \
  -DGGML_BLAS=ON \
  -DGGML_BLAS_VENDOR=Intel10_64lp \
  -DMKL_DIR=/opt/intel/oneapi/mkl/latest/lib/cmake/mkl \
  -DCMAKE_C_COMPILER=icx \
  -DCMAKE_CXX_COMPILER=icpx \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_RPATH="\$ORIGIN/../lib" \
  -DCMAKE_CXX_FLAGS="-O3 -fp-model=fast=2 -qopenmp" \
  -DCMAKE_C_FLAGS="-O3 -fp-model=fast=2 -qopenmp"
cmake --build build/sycl 2>&1 | tee build.log
# turbo quant fork
# turbo quant fork
cmake -B build/sycl -G Ninja \
cmake -B build/sycl -G Ninja \

Latest revision as of 04:03, 14 September 2026

Rebuild

F16

cmake -B build/sycl -G Ninja \
  -DGGML_SYCL=ON \
  -DGGML_SYCL_F16=ON \
  -DGGML_SYCL_FORCE_MMQ=ON \
  -DGGML_SYCL_GRAPH=ON \
  -DGGML_SYCL_DNN=ON \
  -DGGML_SYCL_TARGET=INTEL \
  -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
  -DGGML_NATIVE=ON \
  -DGGML_LTO=ON \
  -DMKL_DIR=/opt/intel/oneapi/mkl/latest/lib/cmake/mkl \
  -DCMAKE_C_COMPILER=icx \
  -DCMAKE_CXX_COMPILER=icpx \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_RPATH="\$ORIGIN/../lib" \
  -DCMAKE_CXX_FLAGS="-O3 -fp-model=fast=2 -qopenmp" \
  -DCMAKE_C_FLAGS="-O3 -fp-model=fast=2 -qopenmp"

cmake --build build/sycl
cmake --install build/sycl --prefix ~/llama.cpp/sycl

F16 without DNN

cmake -B build/sycl-nodnn -G Ninja \
  -DGGML_SYCL=ON \
  -DGGML_SYCL_F16=ON \
  -DGGML_SYCL_DNN=OFF \
  -DGGML_SYCL_TARGET=INTEL \
  -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
  -DGGML_NATIVE=ON \
  -DGGML_LTO=ON \
  -DMKL_DIR=/opt/intel/oneapi/mkl/latest/lib/cmake/mkl \
  -DCMAKE_C_COMPILER=icx \
  -DCMAKE_CXX_COMPILER=icpx \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_RPATH="\$ORIGIN/../lib" \
  -DCMAKE_CXX_FLAGS="-O3 -fp-model=fast=2 -qopenmp" \
  -DCMAKE_C_FLAGS="-O3 -fp-model=fast=2 -qopenmp"

cmake --build build/sycl-nodnn
cmake --install build/sycl-nodnn --prefix ~/llama.cpp/sycl-nodnn

F32

cmake -B build/sycl-f32 -G Ninja \
  -DGGML_SYCL=ON \
  -DGGML_SYCL_F16=OFF \
  -DGGML_SYCL_DNN=ON \
  -DGGML_SYCL_TARGET=INTEL \
  -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
  -DGGML_NATIVE=ON \
  -DGGML_LTO=ON \
  -DMKL_DIR=/opt/intel/oneapi/mkl/latest/lib/cmake/mkl \
  -DCMAKE_C_COMPILER=icx \
  -DCMAKE_CXX_COMPILER=icpx \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_RPATH="\$ORIGIN/../lib" \
  -DCMAKE_CXX_FLAGS="-O3 -fp-model=fast=2 -qopenmp" \
  -DCMAKE_C_FLAGS="-O3 -fp-model=fast=2 -qopenmp"

cmake --build build/sycl-f32
cmake --install build/sycl-f32 --prefix ~/llama.cpp/sycl-f32

Download source

sudo apt install -y git cmake libssl-dev ninja-build
sudo apt install git cmake
sudo apt install libssl-dev openssl
mkdir ~/llama.cpp

git clone https://github.com/ggerganov/llama.cpp ~/llama.cpp/src
git clone https://github.com/TheTom/llama-cpp-turboquant ~/llama.cpp/src-tbq

List available options

grep -rn "option(GGML_SYCL" CMakeLists.txt ggml/CMakeLists.txt
grep -rn "SYCL" CMakeLists.txt ggml/CMakeLists.txt

Build

# install required headers
sudo apt install -y \
  intel-oneapi-mkl-devel \
  intel-oneapi-dnnl-devel \
  intel-oneapi-mkl-sycl-distributed-dft-devel \
  intel-oneapi-compiler-dpcpp-cpp \
  intel-ocloc-dev \
  libigc-dev \
  libze-dev

# set options
cmake -B build/sycl -G Ninja \
  -DGGML_SYCL=ON \
  -DGGML_SYCL_F16=ON \
  -DGGML_SYCL_DNN=ON \
  -DGGML_SYCL_TARGET=INTEL \
  -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
  -DGGML_NATIVE=ON \
  -DGGML_LTO=ON \
  -DMKL_DIR=/opt/intel/oneapi/mkl/latest/lib/cmake/mkl \
  -DCMAKE_C_COMPILER=icx \
  -DCMAKE_CXX_COMPILER=icpx \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_RPATH="\$ORIGIN/../lib" \
  -DCMAKE_CXX_FLAGS="-O3 -fp-model=fast=2 -qopenmp" \
  -DCMAKE_C_FLAGS="-O3 -fp-model=fast=2 -qopenmp"

# build
cmake --build build/sycl

# debug compilation error
cmake --build build/sycl 2>&1 | tee build.log
grep -n -i "error" build.log | head -n 1
tail -n +8270 build.log

# install 
cmake --install build/sycl --prefix ~/llama.cpp/sycl

# clean
rm -rf build/sycl

# detect GPU
cd ~/llama.cpp/sycl/bin
./llama-cli --list-devices

Got It! 

Available devices:
  SYCL0: Intel(R) Arc(TM) Pro B70 Graphics (32656 MiB, 32128 MiB free)

TODO: optimized 

# turbo quant fork
cmake -B build/sycl -G Ninja \
  -DGGML_SYCL=ON \
  -DGGML_SYCL_TARGET=INTEL \
  -DGGML_SYCL_DNN=ON \
  -DGGML_SYCL_F16=ON \
  -DGGML_SYCL_DEVICE_ARCH=bmg-g31 \
  -DGGML_NATIVE=ON \
  -DMKL_DIR=/opt/intel/oneapi/mkl/latest/lib/cmake/mkl \
  -DCMAKE_C_COMPILER=icx \
  -DCMAKE_CXX_COMPILER=icpx \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_RPATH="\$ORIGIN/../lib" \
  -DCMAKE_CXX_FLAGS="-ffast-math -O3" \
  -DCMAKE_C_FLAGS="-ffast-math -O3" \
  -DGGML_SYCL_TURBO_QUANT=OFF

Run

.bashrc 

export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
export ZES_ENABLE_SYSMAN=1

List all oneAPI header packages

# list name only
apt list "intel-oneapi-*-devel"

# list name & description
apt search --names-only "^intel-oneapi-.+-devel$"

正在列出... 完成
intel-oneapi-ccl-devel/all 2022.1.1-3 amd64
intel-oneapi-daal-devel/all 2021.5.3-832 amd64
intel-oneapi-dal-devel/all 2026.1.0-7 amd64
intel-oneapi-dnnl-devel/all,now 2026.0.2-40 amd64 [已安裝]
intel-oneapi-ipp-devel/all 2026.0.1-51 amd64
intel-oneapi-ippcp-devel/all 2026.0.1-9 amd64
intel-oneapi-ishmem-devel/all 1.6.0-223 amd64
intel-oneapi-libdpstd-devel/all 2022.13.0-107 amd64
intel-oneapi-mkl-classic-devel/all 2026.1.0-236 amd64
intel-oneapi-mkl-cluster-devel/all 2026.1.0-236 amd64
intel-oneapi-mkl-core-devel/all 2026.1.0-236 amd64
intel-oneapi-mkl-devel/all,now 2026.1.0-236 amd64 [已安裝]
intel-oneapi-mkl-sycl-devel/all 2026.1.0-236 amd64
intel-oneapi-mkl-sycl-distributed-dft-devel/all,now 2026.1.0-236 amd64 [已安裝]
intel-oneapi-mpi-devel/all 2021.18.1-9 amd64
intel-oneapi-onevpl-devel/all 2023.1.0-43488 amd64
intel-oneapi-tbb-devel/all 2023.1.0-151 amd64

References

https://sergiiob.dev/posts/intel-arc-pro-b70-full-recipe-llama-cpp-sycl/

Retrieved from "https://wiki.fundamental-ramen.com/index.php?title=LM/llama.cpp_for_SYCL&oldid=3344"