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

[NPU] Convert cu_window_seqlens to CPU for npu_flush_attention_unpad operator (#20328)

Browse Source
This commit is contained in:
fy
2026-03-12 09:08:43 +08:00
committed by GitHub
parent 67f02681c9
commit 677e446e51
1 changed files with 1 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/qwen2_5_vl.py
+1
View File
@@ -463,6 +463,7 @@ class Qwen2_5_VisionTransformer(nn.Module, RotaryPosMixin):
# cu_seqlens must be on cpu because of npu_flash_attention_unpad operator restriction # cu_seqlens must be on cpu because of npu_flash_attention_unpad operator restriction
if is_npu(): if is_npu():
cu_seqlens = cu_seqlens.to("cpu") cu_seqlens = cu_seqlens.to("cpu")
cu_window_seqlens = cu_window_seqlens.to("cpu")
# transformers # transformers
x = x.unsqueeze(1) x = x.unsqueeze(1)
for layer_num, blk in enumerate(self.blocks): for layer_num, blk in enumerate(self.blocks):