From 3dda9aeb54cc15c9ecc6d1498a42ceb372d4472b Mon Sep 17 00:00:00 2001
From: Thomas Parnell <tpa@zurich.ibm.com>
Date: Sat, 9 May 2026 17:30:55 +0200
Subject: [PATCH] [Bugfix] Remove nested torch.compile in GDN
 rearrange_mixed_qkv causing CUDA graph capture failure (#42070)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
Co-authored-by: Jiangyun Zhu <riverclouds.zhu@qq.com>
---
 vllm/model_executor/layers/mamba/gdn_linear_attn.py | 1 -
 1 file changed, 1 deletion(-)

diff --git a/vllm/model_executor/layers/mamba/gdn_linear_attn.py b/vllm/model_executor/layers/mamba/gdn_linear_attn.py
index 518e9d4f0c..10760d6116 100644
--- a/vllm/model_executor/layers/mamba/gdn_linear_attn.py
+++ b/vllm/model_executor/layers/mamba/gdn_linear_attn.py
@@ -637,7 +637,6 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
 
         return mixed_qkv_out, z_out, b_out, a_out
 
-    @torch.compile(fullgraph=True)
     def rearrange_mixed_qkv(self, mixed_qkv):
         """Split packed qkv into contiguous (1, seq, heads, dim) tensors.