baalajimaestro/ollama
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

llm: Remove GGML_CUDA_NO_PEER_COPY for ROCm (#7174)

Browse source 
This workaround logic in llama.cpp is causing crashes for users with less system memory than VRAM.
This commit is contained in:
Daniel Hiltgen 2024-10-12 09:56:49 -07:00 committed by GitHub
parent 7fe3902552
commit c3d321d405
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 1 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
llm/generate/gen_linux.sh
View file

@ -251,7 +251,7 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
fi
init_vars
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DGGML_CUDA_NO_PEER_COPY=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
# Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""

1
llm/generate/gen_windows.ps1
View file

@ -340,7 +340,6 @@ function build_rocm() {
"-DCMAKE_C_COMPILER=clang.exe",
"-DCMAKE_CXX_COMPILER=clang++.exe",
"-DGGML_HIPBLAS=on",
"-DGGML_CUDA_NO_PEER_COPY=on",
"-DHIP_PLATFORM=amd",
"-DGGML_AVX=on",
"-DGGML_AVX2=off",