Fix 083 lm_head error (#13132)

* fix no quantize error

* update

* update style

docker/llm/serving/xpu/docker/vllm_offline_inference_vision_language.py

@@ -94,10 +94,9 @@ else:
 
 dtype = "float16"
 if "gemma-3" in model_path:
-    mm_processor_kwarg = {"do_pan_and_scan": True}
     dtype = "float32"
 else:
-    mm_processor_kwarg = None
+    pass
 
 
 llm = LLM(
@@ -106,7 +105,7 @@ llm = LLM(
           dtype=dtype,
           enforce_eager=True,
           hf_overrides=hf_override,
-          mm_processor_kwargs=mm_processor_kwarg,
+          mm_processor_kwargs=None,
           load_in_low_bit="sym_int4",
           tensor_parallel_size=2,
           disable_async_output_proc=True,

python/llm/src/ipex_llm/vllm/xpu/model_convert.py

@@ -41,9 +41,9 @@ def _sample_get_logits(
     # HINT: we do not support other types of quantization for now
     # TODO: we may encounter tie-word-embedding problems
     if isinstance(lm_head, VocabParallelEmbedding):
-        logits = lm_head.linear_method.apply(lm_head,
-                                             hidden_states,
-                                             bias=embedding_bias)
+        logits = lm_head.quant_method.apply(lm_head,
+                                            hidden_states,
+                                            bias=embedding_bias)
     else:
         logits = lm_head(hidden_states)
         if embedding_bias is not None: