[Fix #20389] Illegal memory access in triton attention for large token counts (#20390)

2026-03-17 00:42:11 -07:00
parent e3277b3be2
commit 9a697ceabb
1 changed files with 1 additions and 1 deletions
--- a/python/sglang/srt/layers/attention/triton_backend.py
+++ b/python/sglang/srt/layers/attention/triton_backend.py
@@ -169,7 +169,7 @@ class TritonAttnBackend(AttentionBackend):

        if not self.skip_prefill:
            self.qo_indptr = torch.zeros(
-                (max_bs + 1,), dtype=torch.int32, device=model_runner.device
+                (max_bs + 1,), dtype=torch.int64, device=model_runner.device
            )

            self.mask_indptr = torch.zeros(