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(