[NPU]Replace ascend vision attn operator (#25768)

Signed-off-by: zhaozx-cn <zhaozx2116@163.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

python/sglang/srt/layers/attention/vision.py

@@ -733,27 +733,26 @@ class VisionAscendAttention(nn.Module):
             seq_len_arg = cu_seqlens
         else:
             cu_seqlens = resolve_seqlens(cu_seqlens, bsz, seq_len, device="cpu")
-            seq_lens = cu_seqlens[1:] - cu_seqlens[:-1]
+            seq_len_arg = cu_seqlens[1:].to(torch.int32)
-            if seq_lens.is_npu:
-                seq_lens = seq_lens.to("cpu")
-            output = torch.empty_like(q)
-            seq_len_arg = seq_lens.to(torch.int32)
 
         _, num_heads, head_size = q.shape
         num_kv_heads = k.shape[1]
 
         scale_value = softmax_scale if softmax_scale is not None else head_size**-0.5
 
-        torch_npu._npu_flash_attention_unpad(
+        seq_len_arg = seq_len_arg.tolist()
+        output = torch_npu.npu_fused_infer_attention_score(
             query=q,
             key=k,
             value=v,
-            seq_len=seq_len_arg,
+            actual_seq_lengths=seq_len_arg,
-            scale_value=scale_value,
+            actual_seq_lengths_kv=seq_len_arg,
+            scale=scale_value,
             num_heads=num_heads,
-            num_kv_heads=num_kv_heads,
+            num_key_value_heads=num_kv_heads,
-            out=output,
+            sparse_mode=0,
-        )
+            input_layout="TND",
+        )[0]
         return output