From 2b993ad4797d395303a084aa6b72bf2ef20e7593 Mon Sep 17 00:00:00 2001
From: "Wang, Jian4" <61138589+hzjane@users.noreply.github.com>
Date: Wed, 4 Sep 2024 13:50:32 +0800
Subject: [PATCH] vllm update for glm-4 model automatic not_convert (#12003)

---
 python/llm/src/ipex_llm/vllm/xpu/model_convert.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/python/llm/src/ipex_llm/vllm/xpu/model_convert.py b/python/llm/src/ipex_llm/vllm/xpu/model_convert.py
index 7979bfbc62a..065652e7162 100644
--- a/python/llm/src/ipex_llm/vllm/xpu/model_convert.py
+++ b/python/llm/src/ipex_llm/vllm/xpu/model_convert.py
@@ -250,7 +250,8 @@ def _ipex_llm_load_model(self) -> None:
             from ipex_llm import optimize_model
             import os
             not_convert_last_mlp = os.getenv("IPEX_LLM_NOT_CONVERT_LAST_MLP", None)
-            if not_convert_last_mlp is not None:
+            is_glm4_model = "glm-4" in self.model_config.model.lower()
+            if not_convert_last_mlp is not None or is_glm4_model:
                 # only use to avoid nan value in last mlp forward running glm4-9b-chat
                 modules = ["35.mlp", "36.mlp", "37.mlp", "38.mlp", "39.mlp"]
             else: