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 15, 04:45 UTC
Fix FA2 fallback for Blackwell V1 (#19781)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Jun 19, 2025 at 10:53 UTC
36239f79dd356a8284ed4cf4d261312eac2495e1
1 file changed
+1
-1
vllm/platforms/cuda.py
+1
-1
@@ -255,7 +255,7 @@ class CudaPlatformBase(Platform):
255
"install FlashInfer for better performance.")
256
pass
257
# FlashAttention is the default for SM 8.0+ GPUs
258
- elif cls.has_device_capability(80):
258
+ if cls.has_device_capability(80):
259
logger.info_once("Using Flash Attention backend on V1 engine.")
260
return ("vllm.v1.attention.backends."
261
"flash_attn.FlashAttentionBackend")