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

2026-04-30 01:11:45 -07:00
parent da07b22949
commit 125f75db72
1 changed files with 1 additions and 1 deletions
@@ -203,7 +203,7 @@ class TritonLoRABackend(BaseLoRABackend):
                return
            sgemm.permutation[:bs] = perm
            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:])
            sgemm.max_len = bs
            sgemm.lora_ranks[:mlpb] = bi.lora_ranks[:mlpb]