[Bugfix] Fallback ViT attn backend to SDPA for blackwell (vllm-project#25851)

Signed-off-by: Roger Wang <[email protected]>

Author: ywang96

vllm/model_executor/models/qwen3_vl.py

@@ -66,7 +66,7 @@
                                         PromptReplacement, PromptUpdate,
                                         PromptUpdateDetails)
 from vllm.multimodal.profiling import BaseDummyInputsBuilder
-from vllm.platforms import _Backend, current_platform
+from vllm.platforms import _Backend
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.config import uses_mrope
 from vllm.utils import is_list_of
@@ -336,14 +336,6 @@ def __init__(
         }:
             raise RuntimeError(
                 f"Qwen3-VL does not support {self.attn_backend} backend now.")
-        if current_platform.is_device_capability(
-                100) and self.attn_backend != _Backend.TORCH_SDPA:
-            # TODO(Roger/Wentao): remove this after FA
-            # or XFORMERS's issue fixed on Blackwell
-            logger.info_once("Qwen3-VL vision attention does not support "
-                             f"{self.attn_backend} backend on Blackwell now. "
-                             "Vision attention backend is set to TORCH_SDPA.")
-            self.attn_backend = _Backend.TORCH_SDPA
 
         self.blocks = nn.ModuleList([
             Qwen3_VisionBlock(

vllm/platforms/cuda.py

@@ -205,6 +205,12 @@ def get_current_memory_usage(cls,
     @classmethod
     def get_vit_attn_backend(cls, head_size: int,
                              dtype: torch.dtype) -> _Backend:
+
+        # For Blackwell GPUs, force TORCH_SDPA for now.
+        # See https://github.com/facebookresearch/xformers/issues/1317#issuecomment-3199392579 # noqa: E501
+        if cls.has_device_capability(100):
+            return _Backend.TORCH_SDPA
+
         if dtype not in (torch.float16, torch.bfloat16):
             return _Backend.XFORMERS