Fix qwen model adapter in docker (#9969)

* fix qwen in docker

* add patch for model_adapter.py in fastchat

* add patch for model_adapter.py in fastchat

docker/llm/serving/cpu/docker/Dockerfile

@@ -8,10 +8,13 @@ ARG TINI_VERSION=v0.18.0
 ARG PIP_NO_CACHE_DIR=false
 
 COPY ./entrypoint.sh /opt/entrypoint.sh
+COPY ./model_adapter.py/patch /llm/model_adapter.py.patch
 ADD  https://github.com/krallin/tini/releases/download/${TINI_VERSION}/tini /sbin/tini
 # Install Serving Dependencies
 RUN cd /llm && \
     pip install --pre --upgrade bigdl-llm[serving] && \
+# Fix Qwen model adpater in fastchat
+    patch /usr/local/lib/python3.9/dist-packages/fastchat/model/model_adapter.py < /llm/model_adapter.py.patch && \
     chmod +x /opt/entrypoint.sh && \
     chmod +x /sbin/tini && \
     cp /sbin/tini /usr/bin/tini

docker/llm/serving/cpu/docker/model_adapter.py.patch

@@ -0,0 +1,21 @@
+--- model_adapter.py.old	2024-01-24 01:56:23.903144335 +0000
++++ model_adapter.py	2024-01-24 01:59:22.605062765 +0000
+@@ -1346,15 +1346,17 @@
+         )
+         # NOTE: if you use the old version of model file, please remove the comments below
+         # config.use_flash_attn = False
+-        config.fp16 = True
++        # config.fp16 = True
+         generation_config = GenerationConfig.from_pretrained(
+             model_path, trust_remote_code=True
+         )
++        from bigdl.llm.transformers import AutoModelForCausalLM
+         model = AutoModelForCausalLM.from_pretrained(
+             model_path,
+             config=config,
+             low_cpu_mem_usage=True,
+             trust_remote_code=True,
++            load_in_4bit=True,
+             **from_pretrained_kwargs,
+         ).eval()
+         if hasattr(model.config, "use_dynamic_ntk") and model.config.use_dynamic_ntk: