diff --git a/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py b/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
index 6626f89aa..d4049b5db 100644
--- a/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
+++ b/python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
@@ -802,7 +802,8 @@ class AscendAttnBackend(AttentionBackend):
                 and _is_dflash_verify(spec_info)
                 and seq_lens_cpu is not None
             ):
-                seq_lens_int = seq_lens_cpu[:bs].int()
+                # seq_lens_cpu may be a CPU tensor; swa_indices is on-device.
+                seq_lens_int = seq_lens_cpu[:bs].int().to(self.device)
             else:
                 seq_lens_int = seq_lens[:bs].int()
             starts = torch.clamp(seq_lens_int - self.sliding_window_size, min=0)