Make the Qwen 2.5VL run (#17435)

Summary:

There are two things to fix

1) some python files in vllm_flash_attn
2) move the import out of the dynamo scope.

Reviewed By: zhewenl, 22quinn

Differential Revision: D73899797

Author: houseroad

vllm/model_executor/layers/rotary_embedding.py

@@ -32,6 +32,9 @@
 from vllm.model_executor.custom_op import CustomOp
 from vllm.platforms import current_platform
 
+if current_platform.is_cuda_alike():
+    from vllm.vllm_flash_attn.layers.rotary import apply_rotary_emb
+
 
 def _rotate_neox(x: torch.Tensor) -> torch.Tensor:
     x1 = x[..., :x.shape[-1] // 2]
@@ -78,7 +81,6 @@ def _apply_rotary_emb(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor,
             positional embeddings.
     """
     if current_platform.is_cuda_alike():
-        from vllm.vllm_flash_attn.layers.rotary import apply_rotary_emb
         return apply_rotary_emb(x.unsqueeze(0), cos, sin,
                                 not is_neox_style).squeeze(0)
     else: