From 9a697ceabb2dce6885f6a1ab8993d9980ade86c1 Mon Sep 17 00:00:00 2001
From: akhilg-nv <165961486+akhilg-nv@users.noreply.github.com>
Date: Tue, 17 Mar 2026 00:42:11 -0700
Subject: [PATCH] [Fix #20389] Illegal memory access in triton attention for
 large token counts (#20390)

---
 python/sglang/srt/layers/attention/triton_backend.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/python/sglang/srt/layers/attention/triton_backend.py b/python/sglang/srt/layers/attention/triton_backend.py
index d5c47d2fa..18ed55572 100644
--- 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(