[Build] Fix CUDA arch build coverage gaps (#45277)

Signed-off-by: Shengqi Chen <harry-chen@outlook.com>
Co-authored-by: Xin Li <xinli-sw@users.noreply.github.com>
Co-authored-by: ShawRong <ShawRong@users.noreply.github.com>
Co-authored-by: Change72 <Change72@users.noreply.github.com>
This commit is contained in:
Shengqi Chen
2026-06-13 22:09:20 -07:00
committed by GitHub
co-authored by Xin Li ShawRong Change72
parent 54bbf51668
commit 78e7293bb1
14 changed files with 240 additions and 132 deletions
@@ -1,3 +1,4 @@
#include <cuda.h>
#include <cudaTypedefs.h>
#include <torch/csrc/stable/tensor.h>
@@ -174,15 +175,20 @@ bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability) {
bool cutlass_group_gemm_supported(int64_t cuda_device_capability) {
// CUTLASS grouped FP8 kernels need at least CUDA 12.3 and SM90 (Hopper)
// or CUDA 12.8 and SM100 (Blackwell)
// or CUDA 12.8 and SM100 (Blackwell). Only report archs that have an
// actual cutlass_moe_mm dispatch compiled into this file.
#if defined CUDA_VERSION
if (cuda_device_capability >= 100) {
#if defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100
if (cuda_device_capability >= 100 && cuda_device_capability < 110) {
return CUDA_VERSION >= 12080;
}
if (cuda_device_capability >= 90) {
#endif
#if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90
if (cuda_device_capability >= 90 && cuda_device_capability < 100) {
return CUDA_VERSION >= 12030;
}
#endif
#endif
return false;