From 899453ac508e0a5b239ba995a2a92773ca276e1a Mon Sep 17 00:00:00 2001
From: jianan-gu <jianan.gu@intel.com>
Date: Tue, 28 Oct 2025 10:05:57 +0800
Subject: [PATCH] Use explicit uint64 dtype for Tensor data_ptr() to avoid
 overflow (#11994)

---
 python/sglang/srt/mem_cache/common.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/python/sglang/srt/mem_cache/common.py b/python/sglang/srt/mem_cache/common.py
index 8f96ba0fe..d50e7949f 100644
--- a/python/sglang/srt/mem_cache/common.py
+++ b/python/sglang/srt/mem_cache/common.py
@@ -89,6 +89,7 @@ def write_cache_indices(
         prefix_pointers = torch.tensor(
             [t.data_ptr() for t in prefix_tensors],
             device=req_to_token_pool.device,
+            dtype=torch.uint64,
         )
         # TODO: some tensors can be reused for ForwardBatchInfo (e.g., extend_lens, cumsum_start)
         write_req_to_token_pool_triton[(req_pool_indices_tensor.shape[0],)](