[sgl-kernel] misc: update deepgemm version for sgl-kernel (#9340)
Co-authored-by: Yineng Zhang <me@zhyncs.com> Co-authored-by: fzyzcjy <ch271828n@outlook.com>
This commit is contained in:
PGFLMG
@@ -50,25 +50,17 @@ FetchContent_Declare(
|
||||
)
|
||||
FetchContent_Populate(repo-cutlass)
|
||||
|
||||
# DeepGEMM
|
||||
if("${CUDA_VERSION}" VERSION_EQUAL "12.8")
|
||||
set(DeepGEMM_REPO "https://github.com/sgl-project/DeepGEMM")
|
||||
set(DeepGEMM_TAG "blackwell")
|
||||
elseif("${CUDA_VERSION}" VERSION_EQUAL "12.9")
|
||||
set(DeepGEMM_REPO "https://github.com/sgl-project/DeepGEMM")
|
||||
set(DeepGEMM_TAG "blackwell")
|
||||
elseif("${CUDA_VERSION}" VERSION_EQUAL "13.0")
|
||||
set(DeepGEMM_REPO "https://github.com/sgl-project/DeepGEMM")
|
||||
set(DeepGEMM_TAG "blackwell")
|
||||
else()
|
||||
set(DeepGEMM_REPO "https://github.com/deepseek-ai/DeepGEMM")
|
||||
set(DeepGEMM_TAG "391755ada0ffefa9a6a52b6f14dcaf22d1a463e0")
|
||||
endif()
|
||||
FetchContent_Declare(
|
||||
repo-fmt
|
||||
GIT_REPOSITORY https://github.com/fmtlib/fmt
|
||||
GIT_TAG 553ec11ec06fbe0beebfbb45f9dc3c9eabd83d28
|
||||
GIT_SHALLOW OFF
|
||||
)
|
||||
|
||||
FetchContent_Declare(
|
||||
repo-deepgemm
|
||||
GIT_REPOSITORY ${DeepGEMM_REPO}
|
||||
GIT_TAG ${DeepGEMM_TAG}
|
||||
GIT_REPOSITORY https://github.com/sgl-project/DeepGEMM
|
||||
GIT_TAG sgl
|
||||
GIT_SHALLOW OFF
|
||||
)
|
||||
FetchContent_Populate(repo-deepgemm)
|
||||
@@ -86,7 +78,7 @@ FetchContent_Populate(repo-triton)
|
||||
FetchContent_Declare(
|
||||
repo-flashinfer
|
||||
GIT_REPOSITORY https://github.com/flashinfer-ai/flashinfer.git
|
||||
GIT_TAG 018b551825c8e5579206e6eb9d3229fa679202b3
|
||||
GIT_TAG 9220fb3443b5a5d274f00ca5552f798e225239b7
|
||||
GIT_SHALLOW OFF
|
||||
)
|
||||
FetchContent_Populate(repo-flashinfer)
|
||||
@@ -182,28 +174,11 @@ if ("${CUDA_VERSION}" VERSION_GREATER_EQUAL "12.8" OR SGL_KERNEL_ENABLE_SM100A)
|
||||
list(APPEND SGL_KERNEL_CUDA_FLAGS
|
||||
"-gencode=arch=compute_100,code=sm_100"
|
||||
"-gencode=arch=compute_100a,code=sm_100a"
|
||||
"-gencode=arch=compute_103,code=sm_103"
|
||||
"-gencode=arch=compute_103a,code=sm_103a"
|
||||
"-gencode=arch=compute_101,code=sm_101"
|
||||
"-gencode=arch=compute_101a,code=sm_101a"
|
||||
"-gencode=arch=compute_120,code=sm_120"
|
||||
"-gencode=arch=compute_120a,code=sm_120a"
|
||||
)
|
||||
|
||||
# refer sm_121, sm_110 and sm_101 description https://github.com/pytorch/pytorch/pull/156176
|
||||
if ("${CUDA_VERSION}" VERSION_GREATER_EQUAL "13.0")
|
||||
list(APPEND SGL_KERNEL_CUDA_FLAGS
|
||||
"-gencode=arch=compute_110,code=sm_110"
|
||||
"-gencode=arch=compute_110a,code=sm_110a"
|
||||
"-gencode=arch=compute_121,code=sm_121"
|
||||
"-gencode=arch=compute_121a,code=sm_121a"
|
||||
"--compress-mode=size"
|
||||
)
|
||||
else()
|
||||
list(APPEND SGL_KERNEL_CUDA_FLAGS
|
||||
"-gencode=arch=compute_101,code=sm_101"
|
||||
"-gencode=arch=compute_101a,code=sm_101a"
|
||||
)
|
||||
endif()
|
||||
|
||||
else()
|
||||
list(APPEND SGL_KERNEL_CUDA_FLAGS
|
||||
"-use_fast_math"
|
||||
@@ -286,6 +261,12 @@ set(SOURCES
|
||||
"csrc/moe/cutlass_moe/w4a8/w4a8_moe_data.cu"
|
||||
"csrc/moe/cutlass_moe/w4a8/w4a8_grouped_mm_c3x.cu"
|
||||
"csrc/moe/marlin_moe_wna16/ops.cu"
|
||||
"csrc/moe/marlin_moe_wna16/kernel_bf16_ku4.cu"
|
||||
"csrc/moe/marlin_moe_wna16/kernel_bf16_ku4b8.cu"
|
||||
"csrc/moe/marlin_moe_wna16/kernel_bf16_ku8b128.cu"
|
||||
"csrc/moe/marlin_moe_wna16/kernel_fp16_ku4.cu"
|
||||
"csrc/moe/marlin_moe_wna16/kernel_fp16_ku4b8.cu"
|
||||
"csrc/moe/marlin_moe_wna16/kernel_fp16_ku8b128.cu"
|
||||
"csrc/moe/moe_align_kernel.cu"
|
||||
"csrc/moe/moe_fused_gate.cu"
|
||||
"csrc/moe/moe_topk_softmax_kernels.cu"
|
||||
@@ -321,8 +302,6 @@ target_include_directories(common_ops PRIVATE
|
||||
${repo-cutlass_SOURCE_DIR}/examples/common
|
||||
${repo-flash-attention_SOURCE_DIR}/csrc/flash_attn/src
|
||||
)
|
||||
set_source_files_properties("csrc/gemm/per_token_group_quant_8bit" PROPERTIES COMPILE_OPTIONS "--use_fast_math")
|
||||
|
||||
|
||||
find_package(Python3 COMPONENTS Interpreter REQUIRED)
|
||||
execute_process(
|
||||
@@ -464,13 +443,38 @@ install(TARGETS spatial_ops LIBRARY DESTINATION sgl_kernel)
|
||||
set(DEEPGEMM_SOURCES
|
||||
"${repo-deepgemm_SOURCE_DIR}/csrc/python_api.cpp"
|
||||
)
|
||||
# JIT Logic
|
||||
# DeepGEMM
|
||||
|
||||
install(DIRECTORY "${repo-deepgemm_SOURCE_DIR}/deep_gemm/"
|
||||
DESTINATION "deep_gemm"
|
||||
PATTERN ".git*" EXCLUDE
|
||||
PATTERN "__pycache__" EXCLUDE)
|
||||
Python_add_library(deep_gemm_cpp MODULE USE_SABI ${SKBUILD_SABI_VERSION} WITH_SOABI ${DEEPGEMM_SOURCES})
|
||||
|
||||
# Link against necessary libraries, including nvrtc for JIT compilation.
|
||||
target_link_libraries(deep_gemm_cpp PRIVATE ${TORCH_LIBRARIES} c10 cuda nvrtc mscclpp_static)
|
||||
|
||||
# Add include directories needed by DeepGEMM.
|
||||
target_include_directories(deep_gemm_cpp PRIVATE
|
||||
${repo-deepgemm_SOURCE_DIR}/deep_gemm/include
|
||||
${repo-cutlass_SOURCE_DIR}/include
|
||||
${repo-fmt_SOURCE_DIR}/include
|
||||
)
|
||||
|
||||
# Apply the same compile options as common_ops.
|
||||
target_compile_options(deep_gemm_cpp PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:${SGL_KERNEL_CUDA_FLAGS}>)
|
||||
|
||||
# Create an empty __init__.py to make `deepgemm` a Python package.
|
||||
file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/deepgemm_pkg_init.py "")
|
||||
install(
|
||||
FILES ${CMAKE_CURRENT_BINARY_DIR}/deepgemm_pkg_init.py
|
||||
DESTINATION deep_gemm
|
||||
RENAME __init__.py
|
||||
)
|
||||
|
||||
# Install the compiled DeepGEMM API library.
|
||||
install(TARGETS deep_gemm_cpp LIBRARY DESTINATION deep_gemm)
|
||||
|
||||
# Install the source files required by DeepGEMM for runtime JIT compilation.
|
||||
install(
|
||||
DIRECTORY ${repo-deepgemm_SOURCE_DIR}/deep_gemm/
|
||||
DESTINATION deep_gemm
|
||||
)
|
||||
|
||||
install(DIRECTORY "${repo-cutlass_SOURCE_DIR}/include/cute/"
|
||||
DESTINATION "deep_gemm/include/cute")
|
||||
|
||||
Reference in New Issue
Block a user