From 509bdb4661d4c45e518698bbc85abd984a5ada23 Mon Sep 17 00:00:00 2001
From: binbin Deng <108676127+plusbang@users.noreply.github.com>
Date: Wed, 11 Dec 2024 16:49:32 +0800
Subject: [PATCH] [NPU] Fix minicpm-2B error (#12527)

---
 .../transformers/npu_pipeline_model/convert_pipeline.py       | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/python/llm/src/ipex_llm/transformers/npu_pipeline_model/convert_pipeline.py b/python/llm/src/ipex_llm/transformers/npu_pipeline_model/convert_pipeline.py
index 3b223017dce..f429dd4ebe4 100644
--- a/python/llm/src/ipex_llm/transformers/npu_pipeline_model/convert_pipeline.py
+++ b/python/llm/src/ipex_llm/transformers/npu_pipeline_model/convert_pipeline.py
@@ -435,8 +435,10 @@ def convert_llm_for_deploy(model: torch.nn.Module,
     layernorm_const = os.environ.get("IPEX_LLM_NPU_LAYERNORM_CONST", "1") == "1"
 
     lm_head_low_bit = getattr(model.config, "bigdl_transformers_low_bit", "sym_int4_rtn")
-    if not isinstance(model.lm_head, SlicedLMHead):
+    if hasattr(model, "lm_head") and not isinstance(model.lm_head, SlicedLMHead):
         lm_head_low_bit = model.lm_head.qtype
+    elif hasattr(model, "lm_head_0") and not isinstance(model.lm_head_0, SlicedLMHead):
+        lm_head_low_bit = model.lm_head_0.qtype
     else:
         lm_head_low_bit = model.lm_head.lm_heads[0].qtype