From 18403f6bfe65aedc767b8e064e807fc8dab39161 Mon Sep 17 00:00:00 2001
From: yinghui <32845984+cicirori@users.noreply.github.com>
Date: Sun, 23 Nov 2025 22:35:40 +0100
Subject: [PATCH] make trtllm attn backend's init_forward_metadat non blocking
 (#13802)

---
 python/sglang/srt/layers/attention/trtllm_mla_backend.py | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/python/sglang/srt/layers/attention/trtllm_mla_backend.py b/python/sglang/srt/layers/attention/trtllm_mla_backend.py
index 7e28341dc..c3fbaa422 100755
--- a/python/sglang/srt/layers/attention/trtllm_mla_backend.py
+++ b/python/sglang/srt/layers/attention/trtllm_mla_backend.py
@@ -564,7 +564,9 @@ class TRTLLMMLABackend(FlashInferMLAAttnBackend):
             seq_lens = forward_batch.seq_lens - forward_batch.extend_prefix_lens
             cum_seq_lens_q = torch.cat(
                 (
-                    torch.tensor([0], device=forward_batch.seq_lens.device),
+                    torch.zeros(
+                        1, dtype=torch.int32, device=forward_batch.seq_lens.device
+                    ),
                     torch.cumsum(seq_lens, dim=0),
                 )
             ).int()