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 14, 05:30 UTC
[Bugfix] use flash attn on sm90 (#22933)
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com> Co-authored-by: Michael Goin <mgoin64@gmail.com>
Yongye Zhu committed
Aug 14, 2025 at 19:37 UTC
39cd09dc86cadd80f3588e8d67ee9288012bcb9e
1 file changed
+1
-1
vllm/platforms/cuda.py
+1
-1
@@ -316,7 +316,7 @@ class CudaPlatformBase(Platform):
316
317
# FlashAttention is the default for SM 8.0+ GPUs
318
if cls.has_device_capability(80):
319
- if has_sink:
319
+ if has_sink and not cls.is_device_capability(90):
320
logger.info_once("Using Triton backend on V1 engine.")
321
return TRITON_ATTN_VLLM_V1
322
if is_default_backend_supported := is_attn_backend_supported(