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 15, 06:00 UTC
Only build CUTLASS MoE kernels on Hopper (#19648)
Huy Do committed
Jun 14, 2025 at 11:44 UTC
bc956b38d0cbe51b97be2127573490810e542d69
1 file changed
+2
-2
CMakeLists.txt
+2
-2
@@ -542,10 +542,10 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
542
543
# CUTLASS MoE kernels
544
545
- # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and only works
545
+ # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and ONLY works
546
# on Hopper). get_cutlass_(pplx_)moe_mm_data should only be compiled
547
# if it's possible to compile MoE kernels that use its output.
548
- cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
548
+ cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
549
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
550
set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu"
551
"csrc/quantization/cutlass_w8a8/moe/moe_data.cu")