From 24a27d532084dc47e37334a6bb131d3d9b681ffd Mon Sep 17 00:00:00 2001
From: Yuhao Yang <47235274+yhyang201@users.noreply.github.com>
Date: Wed, 18 Mar 2026 00:32:07 +0800
Subject: [PATCH] vlm: support piecewise cuda graph for Kimi-K2.5 (#20747)

---
 python/sglang/srt/models/kimi_k25.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/python/sglang/srt/models/kimi_k25.py b/python/sglang/srt/models/kimi_k25.py
index 9e7688719..490ef9334 100644
--- a/python/sglang/srt/models/kimi_k25.py
+++ b/python/sglang/srt/models/kimi_k25.py
@@ -716,6 +716,8 @@ class KimiK25ForConditionalGeneration(nn.Module):
             ),
         )
 
+        self.model = self.language_model.model
+
         # Ensure that the dtype of the vision_tower and mm_projector matches that of the language_model.
         # This solves the dtype mismatch issue when using device_map="auto" and torch_dtype.
         if hasattr(self.language_model, "dtype"):