From 62d7929b3a9a5b388f52173777eb3254901ba0a1 Mon Sep 17 00:00:00 2001
From: AndyLi429 <68410213+AndyLi429@users.noreply.github.com>
Date: Mon, 29 Jun 2026 14:37:45 +0800
Subject: [PATCH] [NPU][Bugfix] Accept in_capture in Ascend replay metadata
 (#29598)

---
 .../npu/attention/ascend_hybrid_linear_attn_backend.py           | 1 +
 1 file changed, 1 insertion(+)

diff --git a/python/sglang/srt/hardware_backend/npu/attention/ascend_hybrid_linear_attn_backend.py b/python/sglang/srt/hardware_backend/npu/attention/ascend_hybrid_linear_attn_backend.py
index eca7ce9b9..3d1e7af81 100644
--- a/python/sglang/srt/hardware_backend/npu/attention/ascend_hybrid_linear_attn_backend.py
+++ b/python/sglang/srt/hardware_backend/npu/attention/ascend_hybrid_linear_attn_backend.py
@@ -131,6 +131,7 @@ class AscendMambaAttnBackendBase(MambaAttnBackendBase):
         spec_info: Optional[SpecInput],
         seq_lens_cpu: Optional[torch.Tensor],
         num_padding: Optional[int] = None,
+        in_capture: bool = False,
     ):
         # out_graph passes seq_lens_cpu=None at capture; mirror the base guard.
         if seq_lens_cpu is None: