@@ -305,10 +305,9 @@ def supports_trtllm_attention() -> bool:
|
||||
if envs.VLLM_BATCH_INVARIANT:
|
||||
return False
|
||||
|
||||
# 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()
|
||||
return (
|
||||
current_platform.is_device_capability_family(100) and has_nvidia_artifactory()
|
||||
)
|
||||
|
||||
|
||||
def force_use_trtllm_attention() -> bool | None:
|
||||
|
||||
Reference in New Issue
Block a user