refactor: register CUDA INT8 GEMM and epilogue extensions

This commit is contained in:
leejet committed 2026-10-31 22:29:12 +08:00
1 parent d8964d0302
commit 7ad74e6490
7 files changed
+593 -60

No files matched your search

+5 -4
View File
@@ -348,17 +348,18 @@ if (SD_CUDA)
target_sources(${SD_LIB} PRIVATE src/core/cuda/rope.cu src/core/cuda/int8.cu src/core/cuda/convrot.cu)
set_source_files_properties(src/core/cuda/rope.cu src/core/cuda/int8.cu src/core/cuda/convrot.cu PROPERTIES COMPILE_OPTIONS "--use_fast_math")
if (GGML_CUDA_CUTLASS_PATH)
target_sources(${SD_LIB} PRIVATE src/core/cuda/cute_attention.cu src/core/cuda/i8_projection.cu src/core/cuda/i8_gated_residual.cu)
set_source_files_properties(src/core/cuda/cute_attention.cu src/core/cuda/i8_projection.cu src/core/cuda/i8_gated_residual.cu PROPERTIES
target_sources(${SD_LIB} PRIVATE src/core/cuda/cute_attention.cu src/core/cuda/i8_projection.cu src/core/cuda/i8_gated_residual.cu src/core/cuda/i8_gemm.cu)
set_source_files_properties(src/core/cuda/cute_attention.cu src/core/cuda/i8_projection.cu src/core/cuda/i8_gated_residual.cu src/core/cuda/i8_gemm.cu PROPERTIES
INCLUDE_DIRECTORIES "${GGML_CUDA_CUTLASS_PATH}/include"
COMPILE_OPTIONS "--use_fast_math;--extended-lambda;--expt-relaxed-constexpr")
if (MSVC)
set_property(SOURCE src/core/cuda/cute_attention.cu src/core/cuda/i8_projection.cu src/core/cuda/i8_gated_residual.cu APPEND PROPERTY COMPILE_OPTIONS -Xcompiler=/Zc:preprocessor,/Zc:__cplusplus)
set_property(SOURCE src/core/cuda/cute_attention.cu src/core/cuda/i8_projection.cu src/core/cuda/i8_gated_residual.cu src/core/cuda/i8_gemm.cu APPEND PROPERTY COMPILE_OPTIONS -Xcompiler=/Zc:preprocessor,/Zc:__cplusplus)
endif()
if (GGML_CUDA_FORCE_CUBLAS)
set_property(SOURCE src/core/cuda/i8_projection.cu src/core/cuda/i8_gated_residual.cu APPEND PROPERTY COMPILE_DEFINITIONS SD_CUDA_FORCE_CUBLAS)
set_property(SOURCE src/core/cuda/i8_projection.cu src/core/cuda/i8_gated_residual.cu src/core/cuda/i8_gemm.cu APPEND PROPERTY COMPILE_DEFINITIONS SD_CUDA_FORCE_CUBLAS)
endif()
set_property(SOURCE src/core/ggml_extensions.cpp APPEND PROPERTY COMPILE_DEFINITIONS SD_CUDA_CUTE_ATTENTION)
set_property(SOURCE src/core/cuda/int8.cu APPEND PROPERTY COMPILE_DEFINITIONS SD_CUDA_I8_CUTLASS)
endif()
set_target_properties(${SD_LIB} PROPERTIES CUDA_STANDARD 17 CUDA_STANDARD_REQUIRED ON)
target_link_libraries(${SD_LIB} PRIVATE CUDA::cudart CUDA::cublas)