[LLM] update llm server dockerfiles (#9940)

* [LLM] update llm server dockerfiles

* update llm dockerfile

llm/server/dockerfiles/Dockerfile_serving_cuda118_cudnn8

@@ -1,18 +1,19 @@
 FROM registry.baidubce.com/paddlepaddle/fastdeploy:llm-base-gcc12.3-cuda11.8-cudnn8-nccl2.15.5
 
 WORKDIR /opt/output/
-ENV LD_LIBRARY_PATH="/usr/local/cuda-11.8/compat/:$LD_LIBRARY_PATH"
+ENV LD_LIBRARY_PATH="/usr/local/cuda-11.8/lib64:/usr/lib64:/usr/local/cuda-11.8/targets/x86_64-linux/lib/:/opt/nccl-2.15.5-1/build/lib/:/usr/local/nvidia/lib:/usr/local/nvidia/lib64"
 
-RUN python3 -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu123/ \
-    && python3 -m pip install paddlenlp==3.0.0b0 \
-    && python3 -m pip install --no-cache-dir sentencepiece pycryptodome tritonclient[all]==2.41.1 \
+# 安装 paddlepaddle & paddlenlp & paddlenlp_ops
+RUN python3 -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu118/ \
+    && python3 -m pip install --no-cache-dir --force-reinstall https://paddle-qa.bj.bcebos.com/paddlenlp/wheel/2f85a64edd4aa9911c94ccb5ce53e83ac41ce22b/paddlenlp-3.0.0b3.post20250123-py3-none-any.whl \
     && python3 -m pip install --no-cache-dir --force-reinstall https://paddlepaddle-inference-banchmark.bj.bcebos.com/paddlenlp_ops-0.0.0-py3-none-any.whl \
+    && python3 -m pip install --no-cache-dir sentencepiece pycryptodome tritonclient[all]==2.41.1 \
     && apt-get clean && rm -rf /var/lib/apt/lists/*
 
+# clone paddle & paddlenlp 源码（代码版本应与上述安装版本对齐）
 RUN mkdir -p /opt/source/ && cd /opt/source/ \
     && git clone https://github.com/PaddlePaddle/Paddle.git \
-    && git clone https://github.com/PaddlePaddle/PaddleNLP.git \
-    && cp -r /opt/source/PaddleNLP/paddlenlp /usr/local/lib/python3.10/dist-packages/ \
+    && git clone -b release/3.0-beta4 https://github.com/PaddlePaddle/PaddleNLP.git \
     && python3 -m pip install --no-cache-dir -r PaddleNLP/requirements.txt \
     && python3 -m pip install --no-cache-dir -r PaddleNLP/llm/server/server/requirements.txt
 
@@ -21,7 +22,7 @@ RUN mkdir -p /opt/output/Serving/llm_model/model/1 \
     && cp /opt/source/PaddleNLP/llm/server/server/scripts/start_server.sh /opt/output/Serving/ \
     && cp /opt/source/PaddleNLP/llm/server/server/scripts/stop_server.sh /opt/output/Serving/
 
-ENV PYTHONPATH="/opt/source/PaddleNLP/llm/server/server"
+ENV PYTHONPATH="/opt/source/PaddleNLP/llm/server/server:/opt/source/PaddleNLP"
 RUN echo "from server.triton_server import TritonPythonModel" >>/opt/output/Serving/llm_model/model/1/model.py
 
 ENV http_proxy=""

llm/server/dockerfiles/Dockerfile_serving_cuda123_cudnn9 → llm/server/dockerfiles/Dockerfile_serving_cuda124_cudnn9

@@ -1,18 +1,19 @@
-FROM registry.baidubce.com/paddlepaddle/fastdeploy:llm-base-gcc12.3-cuda12.3-cudnn9-nccl2.15.5
+FROM registry.baidubce.com/paddlepaddle/fastdeploy:llm-base-gcc12.3-cuda12.4-cudnn9-nccl2.15.5
 
 WORKDIR /opt/output/
-ENV LD_LIBRARY_PATH="/usr/local/cuda-12.3/compat/:$LD_LIBRARY_PATH"
+ENV LD_LIBRARY_PATH="/usr/local/cuda-12.4/lib64:/usr/lib64:/usr/local/cuda-12.4/targets/x86_64-linux/lib/:/opt/nccl-2.15.5-1/build/lib/:/usr/local/nvidia/lib:/usr/local/nvidia/lib64"
 
+# 安装 paddlepaddle & paddlenlp & paddlenlp_ops
 RUN python3 -m pip install --pre paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/nightly/cu123/ \
-    && python3 -m pip install paddlenlp==3.0.0b0 \
-    && python3 -m pip install --no-cache-dir sentencepiece pycryptodome tritonclient[all]==2.41.1 \
+    && python3 -m pip install --no-cache-dir --force-reinstall https://paddle-qa.bj.bcebos.com/paddlenlp/wheel/2f85a64edd4aa9911c94ccb5ce53e83ac41ce22b/paddlenlp-3.0.0b3.post20250123-py3-none-any.whl \
     && python3 -m pip install --no-cache-dir --force-reinstall https://paddlepaddle-inference-banchmark.bj.bcebos.com/paddlenlp_ops-0.0.0-py3-none-any.whl \
+    && python3 -m pip install --no-cache-dir sentencepiece pycryptodome tritonclient[all]==2.41.1 \
     && apt-get clean && rm -rf /var/lib/apt/lists/*
 
+# clone paddle & paddlenlp 源码（代码版本应与上述安装版本对齐）
 RUN mkdir -p /opt/source/ && cd /opt/source/ \
     && git clone https://github.com/PaddlePaddle/Paddle.git \
-    && git clone https://github.com/PaddlePaddle/PaddleNLP.git \
-    && cp -r /opt/source/PaddleNLP/paddlenlp /usr/local/lib/python3.10/dist-packages/ \
+    && git clone -b release/3.0-beta4 https://github.com/PaddlePaddle/PaddleNLP.git \
     && python3 -m pip install --no-cache-dir -r PaddleNLP/requirements.txt \
     && python3 -m pip install --no-cache-dir -r PaddleNLP/llm/server/server/requirements.txt
 
@@ -21,7 +22,7 @@ RUN mkdir -p /opt/output/Serving/llm_model/model/1 \
     && cp /opt/source/PaddleNLP/llm/server/server/scripts/start_server.sh /opt/output/Serving/ \
     && cp /opt/source/PaddleNLP/llm/server/server/scripts/stop_server.sh /opt/output/Serving/
 
-ENV PYTHONPATH="/opt/source/PaddleNLP/llm/server/server"
+ENV PYTHONPATH="/opt/source/PaddleNLP/llm/server/server:/opt/source/PaddleNLP"
 RUN echo "from server.triton_server import TritonPythonModel" >>/opt/output/Serving/llm_model/model/1/model.py
 
 ENV http_proxy=""