Add pp_serving example to serving image (#11433)

* init pp

* update

* update

* no clone ipex-llm again

docker/llm/inference/xpu/docker/Dockerfile

@@ -58,6 +58,9 @@ RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRO
     pip install --upgrade "uvicorn[standard]" && \
     # Download vLLM-Serving
     cp -r ./ipex-llm/python/llm/example/GPU/vLLM-Serving/ ./vLLM-Serving && \
+    # Download pp_serving
+    mkdir -p /llm/pp_serving && \
+    cp ./ipex-llm/python/llm/example/GPU/Pipeline-Parallel-FastAPI/*.py /llm/pp_serving/ && \
     # Install related library of benchmarking
     pip install pandas omegaconf && \
     chmod +x /llm/benchmark.sh && \

docker/llm/serving/xpu/docker/Dockerfile

@@ -21,7 +21,10 @@ RUN apt-get update && \
     pip install outlines==0.0.34 --no-deps && \
     pip install interegular cloudpickle diskcache joblib lark nest-asyncio numba scipy && \
     # For Qwen series models support
-    pip install transformers_stream_generator einops tiktoken
+    pip install transformers_stream_generator einops tiktoken && \
+    # For pipeline serving support
+    pip install mpi4py fastapi uvicorn openai && \
+    pip install gradio # for gradio web UI
 
 COPY ./vllm_online_benchmark.py        /llm/
 COPY ./vllm_offline_inference.py       /llm/
@@ -29,5 +32,7 @@ COPY ./payload-1024.lua                /llm/
 COPY ./start-vllm-service.sh           /llm/
 COPY ./benchmark_vllm_throughput.py   /llm/
 COPY ./start-fastchat-service.sh       /llm/
+COPY ./start-pp_serving-service.sh       /llm/
+
 
 WORKDIR /llm/

docker/llm/serving/xpu/docker/README.md

@@ -45,6 +45,13 @@ After the container is booted, you could get into the container through `docker
 
 Currently, we provide two different serving engines in the image, which are FastChat serving engine and vLLM serving engine.
 
+#### Pipeline parallel serving engine
+
+To run Pipeline parallel serving using `IPEX-LLM` as backend, you can refer to this [readme](https://github.com/intel-analytics/ipex-llm/tree/main/python/llm/example/GPU/Pipeline-Parallel-FastAPI).
+
+For convenience, we have included a file `/llm/start-pp_serving-service.sh` in the image.
+
+
 #### FastChat serving engine
 
 To run model-serving using `IPEX-LLM` as backend using FastChat, you can refer to this [quickstart](https://ipex-llm.readthedocs.io/en/latest/doc/LLM/Quickstart/fastchat_quickstart.html#).

docker/llm/serving/xpu/docker/start-pp_serving-service.sh

@@ -0,0 +1,24 @@
+source /opt/intel/oneapi/setvars.sh --force
+export no_proxy=localhost
+export FI_PROVIDER=tcp
+export OMP_NUM_THREADS=32
+
+export LD_PRELOAD=${LD_PRELOAD}:${CONDA_PREFIX}/lib/libtcmalloc.so
+basekit_root=/opt/intel/oneapi
+source $basekit_root/setvars.sh --force
+source $basekit_root/ccl/latest/env/vars.sh --force
+
+export USE_XETLA=OFF
+if [[ $KERNEL_VERSION != *"6.5"* ]]; then
+    export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
+fi
+export TORCH_LLM_ALLREDUCE=0
+
+export IPEX_LLM_QUANTIZE_KV_CACHE=1
+export num_gpus=2
+export model_path="/llm/models/Llama-2-7b-chat-hf"
+export low_bit="fp8"
+# max requests = max_num_reqs * rank_num
+export max_num_seqs="4"
+cd /llm/pp_serving
+CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $num_gpus pipeline_serving.py --repo-id-or-model-path $model_path --low-bit $low_bit --max-num-seqs $max_num_seqs