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 10, 12:45 UTC
fix(rocm): add early return in get_flash_attn_version for ROCm (#31286)
Signed-off-by: rabi <ramishra@redhat.com>
Rabi Mishra committed
Jan 8, 2026 at 07:58 UTC
39d82005f7a58fed901526a82bd83ff91a60ea0c
1 file changed
+3
vllm/attention/utils/fa_utils.py
+3
@@ -40,6 +40,9 @@ def get_flash_attn_version(requires_alibi: bool = False) -> int | None:
40
41
if current_platform.is_xpu():
42
return 2
43
+ if current_platform.is_rocm():
44
+ # ROCm doesn't use vllm_flash_attn; return None to skip fa_version arg
45
+ return None
46
try:
47
from vllm.vllm_flash_attn.flash_attn_interface import (
48
fa_version_unsupported_reason,