EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

CUDA: replace GGML_CUDA_F16 with CUDA arch checks (#15433)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-08-20 16:58:49 +02:00
committed by GitHub
parent fec9519802
commit 7a6e91ad26
12 changed files with 32 additions and 86 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/src/ggml-cuda/CMakeLists.txt
View File

@@ -24,12 +24,6 @@ if (CUDAToolkit_FOUND)
# for best performance and to also build real architectures for the most commonly used GPUs.
if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
set(CMAKE_CUDA_ARCHITECTURES "native")
elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
else()
set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real")
endif()
else()
if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
@@ -91,10 +85,6 @@ if (CUDAToolkit_FOUND)
add_compile_definitions(GGML_CUDA_NO_FA)
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()