From 39cd09dc86cadd80f3588e8d67ee9288012bcb9e Mon Sep 17 00:00:00 2001 From: Yongye Zhu Date: Thu, 14 Aug 2025 19:37:22 -0400 Subject: [PATCH] [Bugfix] use flash attn on sm90 (#22933) Signed-off-by: Yongye Zhu Co-authored-by: Michael Goin --- vllm/platforms/cuda.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/platforms/cuda.py b/vllm/platforms/cuda.py index 63f6b373c322f..483d5e1531a92 100644 --- a/vllm/platforms/cuda.py +++ b/vllm/platforms/cuda.py @@ -316,7 +316,7 @@ class CudaPlatformBase(Platform): # FlashAttention is the default for SM 8.0+ GPUs if cls.has_device_capability(80): - if has_sink: + if has_sink and not cls.is_device_capability(90): logger.info_once("Using Triton backend on V1 engine.") return TRITON_ATTN_VLLM_V1 if is_default_backend_supported := is_attn_backend_supported(