[Bugfix] Re-enable FP8 MoE on NVIDIA Thor (#46339)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Jun 24, 2026 at 22:35 UTC
24d5186138448a4b484bbbb387fe2541eed89b72
2 files changed
+3
-3
CMakeLists.txt
+2
-2
@@ -869,9 +869,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
869
endif()
870
871
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
872
- cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f" "${CUDA_ARCHS}")
872
+ cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
873
else()
874
- cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.3a" "${CUDA_ARCHS}")
874
+ cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
875
endif()
876
if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
877
set(CUTLASS_MOE_SM100_SRCS "csrc/libtorch_stable/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm100.cu")
csrc/libtorch_stable/quantization/w8a8/cutlass/scaled_mm_entry.cu
+1
-1
@@ -180,7 +180,7 @@ bool cutlass_group_gemm_supported(int64_t cuda_device_capability) {
180
181
#if defined CUDA_VERSION
182
#if defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100
183
- if (cuda_device_capability >= 100 && cuda_device_capability < 110) {
183
+ if (cuda_device_capability >= 100 && cuda_device_capability < 120) {
184
return CUDA_VERSION >= 12080;
185
}
186
#endif