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

[NPU] Update non-vit vision part for cumulative seqlen (#31867)

Browse Source
This commit is contained in:
Jinyan Yi
2026-07-22 10:41:48 +08:00
committed by GitHub
parent 93cb9a5482
commit 394b0dd13a
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/vision.py
+4 -4
View File
@@ -803,11 +803,11 @@ class VisionAscendAttention(nn.Module):
[b * s, h, head_size] [b * s, h, head_size]
""" """
if forward_metadata is not None: if forward_metadata is not None:
seq_lens = forward_metadata.seq_lens # TND fused attention expects cumulative seqlens (cu_seqlens[1:]),
if seq_lens.is_npu: # not per-sequence lengths in forward_metadata.seq_lens.
seq_lens = seq_lens.to("cpu") cu = forward_metadata.cu_seqlens.to("cpu")
output = torch.empty_like(q) output = torch.empty_like(q)
seq_len_arg = seq_lens.to(torch.int32) seq_len_arg = cu[1:].to(torch.int32)
elif envs.SGLANG_VIT_ENABLE_CUDA_GRAPH.get(): elif envs.SGLANG_VIT_ENABLE_CUDA_GRAPH.get():
if "output_ws" not in kwargs: if "output_ws" not in kwargs:
raise RuntimeError("output_ws should be prepared for npu-graph mode") raise RuntimeError("output_ws should be prepared for npu-graph mode")