[Bugfix] Restrict TRTLLM attention to SM100, fixing GB300 (SM103) hang (#38730)

Signed-off-by: Stefano Castagnetta <scastagnetta@nvidia.com>
This commit is contained in:
Stefano Castagnetta
2026-04-01 12:08:40 -07:00
committed by GitHub
parent c09ad767cd
commit 6183cae1bd
3 changed files with 14 additions and 10 deletions
+4 -4
View File
@@ -290,10 +290,10 @@ def supports_trtllm_attention() -> bool:
if envs.VLLM_BATCH_INVARIANT:
return False
# Requires SM100 and NVIDIA artifactory to be accessible to download cubins
return (
current_platform.is_device_capability_family(100) and has_nvidia_artifactory()
)
# TRTLLM attention is currently only validated on SM100 (CC 10.0).
# SM103 (GB300) hangs with FlashInfer >= 0.6.7.
# See: https://github.com/flashinfer-ai/flashinfer/issues/2939
return current_platform.is_device_capability(100) and has_nvidia_artifactory()
def force_use_trtllm_attention() -> bool | None: