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

fix(lora): avoid CUDA graph-breaking scalar assignment in seg_indptr (#23738)

Browse Source
This commit is contained in:
Ethan (Yusheng) Su
2026-04-30 01:11:45 -07:00
committed by GitHub
parent da07b22949
commit 125f75db72
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/lora/backend/triton_backend.py
+1 -1
View File
@@ -203,7 +203,7 @@ class TritonLoRABackend(BaseLoRABackend):
return return
sgemm.permutation[:bs] = perm sgemm.permutation[:bs] = perm
sgemm.seg_lens[:] = seg_lens sgemm.seg_lens[:] = seg_lens
sgemm.seg_indptr[0] = 0 sgemm.seg_indptr[0:1].zero_()
torch.cumsum(sgemm.seg_lens, dim=0, out=sgemm.seg_indptr[1:]) torch.cumsum(sgemm.seg_lens, dim=0, out=sgemm.seg_indptr[1:])
sgemm.max_len = bs sgemm.max_len = bs
sgemm.lora_ranks[:mlpb] = bi.lora_ranks[:mlpb] sgemm.lora_ranks[:mlpb] = bi.lora_ranks[:mlpb]