matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-11 16:26:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

Refine build to support CPU only

Browse Source 
If someone checks out the ollama repo and doesn't install the CUDA
library, this will ensure they can build a CPU only version
This commit is contained in:
Daniel Hiltgen
2023-12-13 17:26:47 -08:00
parent 51082535e1
commit 1b991d0ba9
9 changed files with 152 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
llm/ext_server.go
View File

@@ -21,17 +21,7 @@ package llm
#cgo linux CFLAGS: -D_GNU_SOURCE
#cgo linux windows CFLAGS: -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_MMV_Y=1 -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128 -DGGML_USE_CUBLAS
#cgo linux LDFLAGS: -L/usr/local/cuda/targets/x86_64-linux/lib -L/usr/local/cuda/lib64 -L/usr/local/cuda/targets/x86_64-linux/lib/stubs
#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/gguf/build/cuda/examples/server/libext_server.a
#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/gguf/build/cuda/common/libcommon.a
#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/gguf/build/cuda/libllama.a
#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/gguf/build/cuda/libggml_static.a
// Note: the following requires cuda library presence on linux to build, even if you only have rocm or CPU only
#cgo linux LDFLAGS: /usr/local/cuda/lib64/libcudart_static.a
#cgo linux LDFLAGS: /usr/local/cuda/lib64/libcublas_static.a
#cgo linux LDFLAGS: /usr/local/cuda/lib64/libcublasLt_static.a
#cgo linux LDFLAGS: /usr/local/cuda/lib64/libcudadevrt.a
#cgo linux LDFLAGS: /usr/local/cuda/lib64/libculibos.a
#cgo linux LDFLAGS: ${SRCDIR}/llama.cpp/gguf/build/cuda/libollama.a
#cgo linux LDFLAGS: -lrt -lpthread -ldl -lstdc++ -lm
#cgo windows LDFLAGS: -L${SRCDIR}/llama.cpp/gguf/build/wincuda/dist/bin
#cgo windows LDFLAGS: -lext_server_shared -lpthread

51
llm/llama.cpp/gen_linux.sh
View File

@@ -13,28 +13,43 @@ source $(dirname $0)/gen_common.sh
init_vars
git_module_setup
apply_patches
CMAKE_DEFS="-DLLAMA_CUBLAS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
if [ -d /usr/local/cuda/lib64/ ] ; then
CMAKE_DEFS="-DLLAMA_CUBLAS=on -DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
else
CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
fi
BUILD_DIR="gguf/build/cuda"
LIB_DIR="${BUILD_DIR}/lib"
mkdir -p ../../dist/
build
# TODO - explore mechanism to soften the hard cuda dependency on linux
# by conditionally building some archive here that aggregates the cuda libs if present
# so that the cgo flags link this intermediate archive instead of the underlying cuda libs
#
# gcc -fPIC -g -shared -o ${LIB_DIR}/libcuda_server.so \
# -Wl,--whole-archive \
# ${BUILD_DIR}/examples/server/CMakeFiles/ext_server.dir/server.cpp.o \
# ${BUILD_DIR}/common/libcommon.a \
# ${BUILD_DIR}/libllama.a \
# ${BUILD_DIR}/examples/llava/libllava_static.a \
# -Wl,--no-whole-archive \
# -lrt -lpthread -ldl -lstdc++ -lm \
# /usr/local/cuda/lib64/libcudart_static.a \
# /usr/local/cuda/lib64/libcublas_static.a \
# /usr/local/cuda/lib64/libcublasLt_static.a \
# /usr/local/cuda/lib64/libcudadevrt.a \
# /usr/local/cuda/lib64/libculibos.a
if [ -d /usr/local/cuda/lib64/ ] ; then
pwd
ar -M <<EOF
create ${BUILD_DIR}/libollama.a
addlib ${BUILD_DIR}/examples/server/libext_server.a
addlib ${BUILD_DIR}/common/libcommon.a
addlib ${BUILD_DIR}/libllama.a
addlib ${BUILD_DIR}/libggml_static.a
addlib /usr/local/cuda/lib64/libcudart_static.a
addlib /usr/local/cuda/lib64/libcublas_static.a
addlib /usr/local/cuda/lib64/libcublasLt_static.a
addlib /usr/local/cuda/lib64/libcudadevrt.a
addlib /usr/local/cuda/lib64/libculibos.a
save
end
EOF
else
ar -M <<EOF
create ${BUILD_DIR}/libollama.a
addlib ${BUILD_DIR}/examples/server/libext_server.a
addlib ${BUILD_DIR}/common/libcommon.a
addlib ${BUILD_DIR}/libllama.a
addlib ${BUILD_DIR}/libggml_static.a
save
end
EOF
fi
if [ -z "${ROCM_PATH}" ] ; then
# Try the default location in case it exists