@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 511e81e7c9
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 11, 10:00 UTC

[BUILD] use sm_100f when compiling flashmla to fix support on sm103 (#30705)

Signed-off-by: Shengqi Chen <harry-chen@outlook.com>

Shengqi Chen committed Dec 16, 2025 at 06:48 UTC 511e81e7c9a8a6c7ff5e8ce075c75c88513ad29f
1 file changed +11 -5
cmake/external_projects/flashmla.cmake
+11 -5
@@ -35,16 +35,21 @@ message(STATUS "FlashMLA is available at ${flashmla_SOURCE_DIR}")
35 # sm90a
36
37 set(SUPPORT_ARCHS)
38 -if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.3)
39 - list(APPEND SUPPORT_ARCHS 9.0a)
38 +if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3)
39 + list(APPEND SUPPORT_ARCHS "9.0a")
40 endif()
41 -if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8)
42 - list(APPEND SUPPORT_ARCHS 10.0a)
41 +if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.9)
42 + # CUDA 12.9 has introduced "Family-Specific Architecture Features"
43 + # this supports all compute_10x family
44 + list(APPEND SUPPORT_ARCHS "10.0f")
45 +elseif(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
46 + list(APPEND SUPPORT_ARCHS "10.0a")
47 endif()
48
49
50 cuda_archs_loose_intersection(FLASH_MLA_ARCHS "${SUPPORT_ARCHS}" "${CUDA_ARCHS}")
51 if(FLASH_MLA_ARCHS)
52 + message(STATUS "FlashMLA CUDA architectures: ${FLASH_MLA_ARCHS}")
53 set(VLLM_FLASHMLA_GPU_FLAGS ${VLLM_GPU_FLAGS})
54 list(APPEND VLLM_FLASHMLA_GPU_FLAGS "--expt-relaxed-constexpr" "--expt-extended-lambda" "--use_fast_math")
55
@@ -126,7 +131,8 @@ if(FLASH_MLA_ARCHS)
131 $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
132 $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>)
133 else()
129 - # Create empty targets for setup.py when not targeting sm90a systems
134 + message(STATUS "FlashMLA will not compile: unsupported CUDA architecture ${CUDA_ARCHS}")
135 + # Create empty targets for setup.py on unsupported systems
136 add_custom_target(_flashmla_C)
137 add_custom_target(_flashmla_extension_C)
138 endif()