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 9, 10:00 UTC
[Bugfix] Register fp8 cutlass_group_gemm as supported for only SM90+SM100 (#33285)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Jan 28, 2026 at 21:40 UTC
1bd47d6e5a57000e404af60838eadb86b6f457c0
1 file changed
+2
vllm/_custom_ops.py
+2
@@ -789,6 +789,8 @@ def cutlass_sparse_scaled_mm_supported(cuda_device_capability: int) -> bool:
789
790
791
def cutlass_group_gemm_supported(cuda_device_capability: int) -> bool:
792
+ if cuda_device_capability < 90 or cuda_device_capability >= 110:
793
+ return False
794
try:
795
return torch.ops._C.cutlass_group_gemm_supported(cuda_device_capability)
796
except AttributeError: