minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[NPU] Fix device mismatch in SWA mask for DSpark verify graph capture (#39353)

Browse Source
This commit is contained in:
iridiumine
2026-09-14 19:36:33 +08:00
committed by GitHub
parent 66c7bc838e
commit 87db743021
1 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/hardware_backend/npu/attention/ascend_backend.py
+2 -1
View File
@@ -802,7 +802,8 @@ class AscendAttnBackend(AttentionBackend):
and _is_dflash_verify(spec_info) and _is_dflash_verify(spec_info)
and seq_lens_cpu is not None 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: else:
seq_lens_int = seq_lens[:bs].int() seq_lens_int = seq_lens[:bs].int()
starts = torch.clamp(seq_lens_int - self.sliding_window_size, min=0) starts = torch.clamp(seq_lens_int - self.sliding_window_size, min=0)