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 13, 00:45 UTC
[NVIDIA] Guard SM100 CUTLASS MoE macro to SM100 builds v2 (#28938)
Signed-off-by: johnnynunez <johnnynuca14@gmail.com> Signed-off-by: Johnny <johnnynuca14@gmail.com>
Johnny committed
Nov 19, 2025 at 01:44 UTC
49ef847aa82c93615d5d86fac81e4716e9cd27cd
1 file changed
+7
-7
CMakeLists.txt
+7
-7
@@ -512,9 +512,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
512
# The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
513
# require CUDA 12.8 or later
514
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
515
- cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
515
+ cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
516
else()
517
- cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
517
+ cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
518
endif()
519
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
520
set(SRCS
@@ -619,9 +619,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
619
620
# FP4 Archs and flags
621
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
622
- cuda_archs_loose_intersection(FP4_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
622
+ cuda_archs_loose_intersection(FP4_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
623
else()
624
- cuda_archs_loose_intersection(FP4_ARCHS "10.0a;10.1a;12.0a;12.1a" "${CUDA_ARCHS}")
624
+ cuda_archs_loose_intersection(FP4_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
625
endif()
626
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
627
set(SRCS
@@ -695,7 +695,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
695
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
696
cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
697
else()
698
- cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
698
+ cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
699
endif()
700
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
701
set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm100.cu")
@@ -741,9 +741,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
741
endif()
742
743
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
744
- cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
744
+ cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
745
else()
746
- cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
746
+ cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
747
endif()
748
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
749
set(SRCS "csrc/quantization/w8a8/cutlass/moe/blockwise_scaled_group_mm_sm100.cu")