attention fix

Author: AmpereKowk

xllm/python/kernels_npu/attention.py

@@ -53,37 +53,28 @@ def vision_fusion_attention(
         Attention output with same shape as ``q``.
     """
     import torch_npu
-    from torchair.ops import npu_fused_infer_attention_score
-
-    if torch.compiler.is_compiling():
-        actual_seq_lengths = torch.tensor(
-            actual_seq_qlen, dtype=torch.int64, device=q.device
-        )
-        actual_seq_lengths_kv = torch.tensor(
-            actual_seq_kvlen, dtype=torch.int64, device=q.device
-        )
-        return npu_fused_infer_attention_score(
-            q,
-            k,
-            v,
-            actual_seq_lengths=actual_seq_lengths,
-            actual_seq_lengths_kv=actual_seq_lengths_kv,
-            num_heads=num_heads,
-            scale=scale,
-            input_layout=input_layout,
-        )[0]
-    else:
-        return torch_npu.npu_fusion_attention(
-            q,
-            k,
-            v,
-            actual_seq_qlen=actual_seq_qlen,
-            actual_seq_kvlen=actual_seq_kvlen,
-            head_num=num_heads,
-            scale=scale,
-            input_layout=input_layout,
-        )[0]
-
+    return torch_npu.npu_fused_infer_attention_score(
+        q,
+        k,
+        v,
+        actual_seq_lengths=actual_seq_qlen,
+        actual_seq_lengths_kv=actual_seq_kvlen,
+        num_heads=num_heads,
+        scale=scale,
+        input_layout=input_layout,
+    )[0]
+    '''
+    return torch_npu.npu_fusion_attention(
+        q,
+        k,
+        v,
+        actual_seq_qlen=actual_seq_qlen,
+        actual_seq_kvlen=actual_seq_kvlen,
+        head_num=num_heads,
+        scale=scale,
+        input_layout=input_layout,
+    )[0]
+   '''
 
 __all__ = [
     "reshape_paged_cache",

xllm/python/models/qwen3_vl.py

@@ -679,7 +679,6 @@ def forward(
             grid_thw.to(device=self.device), self.device
         )
         actual_seq: List[int] = cu_seqlens[1:].cpu().tolist()
-
         # Transformer blocks with deepstack feature collection
         deepstack_features: List[torch.Tensor] = []
         for layer_num, blk in enumerate(self.blocks):