Add vllm Predictor (intel#20)

* add vllm_predictor * add tests skeleton * add tests skeleton * add pytest.ini * wip * complete, debug wip * nit * nit * nit * complete generate supporting str and List[str] * add model * add streaming * remove tests * Add install-vllm-cpu script * nit Signed-off-by: Wu, Xiaochang <[email protected]> * nit Signed-off-by: Wu, Xiaochang <[email protected]> * nit * fix package inference * update install script and add doc * nit * nit * nit * add dtype support * nit * nit * nit * add ci * nit * nit * add libpthread-stubs0-dev * fix install-vllm-cpu * fix * revert inference.inference_config * debug ci * debug ci * debug ci * debug ci * debug ci * debug ci * debug ci * debug ci * update --------- Signed-off-by: Wu, Xiaochang <[email protected]>
yutianchen666 · Jan 18, 2024 · e6494c0 · e6494c0
1 parent 906e597
commit e6494c0
Show file tree

Hide file tree

Showing 15 changed files with 309 additions and 28 deletions.
diff --git a/.github/workflows/config/llama-2-7b-chat-hf-vllm-fp32.yaml b/.github/workflows/config/llama-2-7b-chat-hf-vllm-fp32.yaml
@@ -0,0 +1,27 @@
+port: 8000
+name: llama-2-7b-chat-hf-vllm
+route_prefix: /llama-2-7b-chat-hf-vllm
+cpus_per_worker: 24
+gpus_per_worker: 0
+deepspeed: false
+vllm:
+  enabled: true
+  precision: fp32
+workers_per_group: 2
+device: "cpu"
+ipex:
+  enabled: false
+  precision: bf16
+model_description:
+  model_id_or_path: meta-llama/Llama-2-7b-chat-hf
+  tokenizer_name_or_path: meta-llama/Llama-2-7b-chat-hf
+  chat_processor: ChatModelLLama
+  prompt:
+    intro: ''
+    human_id: '[INST] {msg} [/INST]
+
+      '
+    bot_id: ''
+    stop_words: []
+  config:
+    use_auth_token: ''
diff --git a/.github/workflows/workflow_inference.yml b/.github/workflows/workflow_inference.yml
@@ -34,7 +34,7 @@ jobs:
     name: inference test
     strategy:
       matrix:
-        model: [ gpt-j-6b, gpt2, bloom-560m, opt-125m, mpt-7b, mistral-7b-v0.1, mpt-7b-bigdl, neural-chat-7b-v3-1, CodeLlama-7b-hf, falcon-7b ]
+        model: [ gpt-j-6b, gpt2, bloom-560m, opt-125m, mpt-7b, mistral-7b-v0.1, mpt-7b-bigdl, neural-chat-7b-v3-1, CodeLlama-7b-hf, falcon-7b, llama-2-7b-chat-hf-vllm ]
         isPR:
           - ${{inputs.ci_type == 'pr'}}
 
@@ -45,6 +45,7 @@ jobs:
           - { model: "gpt-j-6b"}
           - { model: "mistral-7b-v0.1"}
           - { model: "mpt-7b-bigdl"}
+          - { model: "llama-2-7b-chat-hf-vllm"}
           - dtuner_model: nathan0/mpt-7b-deltatuner-model
             model: mpt-7b
 
@@ -64,13 +65,15 @@ jobs:
     steps:
       - name: Checkout
         uses: actions/checkout@v2
-      
+
       - name: Determine Target
         id: "target"
         run: |
           target="inference"
           if [[ ${{ matrix.model }} == "mpt-7b-bigdl" ]]; then
             target="${target}_bigdl_cpu"
+          elif [[ ${{ matrix.model }} == "llama-2-7b-chat-hf-vllm" ]]; then
+            target="${target}_vllm"
           fi
           echo "target is ${target}"
           echo "target=$target" >> $GITHUB_OUTPUT
@@ -79,6 +82,8 @@ jobs:
         run: |
           if [[ ${{ matrix.model }} == "mpt-7b-bigdl" ]]; then
             DF_SUFFIX=".bigdl-cpu"
+          elif [[ ${{ matrix.model }} == "llama-2-7b-chat-hf-vllm" ]]; then
+            DF_SUFFIX=".vllm"
           else
             DF_SUFFIX=".cpu_and_deepspeed"
           fi
@@ -106,12 +111,16 @@ jobs:
           TARGET=${{steps.target.outputs.target}}
           if [[ ${{ matrix.model }} == "mpt-7b-bigdl" ]]; then
             docker exec "${TARGET}" bash -c "python inference/serve.py --config_file inference/models/bigdl/mpt-7b-bigdl.yaml --simple"
+          elif [[ ${{ matrix.model }} == "llama-2-7b-chat-hf-vllm" ]]; then
+            docker exec "${TARGET}" bash -c "python inference/serve.py --config_file .github/workflows/config/llama-2-7b-chat-hf-vllm-fp32.yaml --simple"
           else
             docker exec "${TARGET}" bash -c "python inference/serve.py --simple --models ${{ matrix.model }}"
           fi
+          echo Non-streaming query:
           docker exec "${TARGET}" bash -c "python examples/inference/api_server_simple/query_single.py --model_endpoint http://127.0.0.1:8000/${{ matrix.model }}"
+          echo Streaming query:
           docker exec "${TARGET}" bash -c "python examples/inference/api_server_simple/query_single.py --model_endpoint http://127.0.0.1:8000/${{ matrix.model }} --streaming_response"
-      
+
       - name: Run Inference Test with Deltatuner
         if: ${{ matrix.dtuner_model }}
         run: |
@@ -125,7 +134,7 @@ jobs:
           TARGET=${{steps.target.outputs.target}}
           if [[ ${{ matrix.model }} =~ ^(gpt2|falcon-7b|mpt-7b.*)$ ]]; then
             echo ${{ matrix.model }} is not supported!
-          else
+          elif [[ ! ${{ matrix.model }} == "llama-2-7b-chat-hf-vllm" ]]; then
             docker exec "${TARGET}" bash -c "python .github/workflows/config/update_inference_config.py --config_file inference/models/\"${{ matrix.model }}\".yaml --output_file \"${{ matrix.model }}\".yaml.deepspeed --deepspeed"
             docker exec "${TARGET}" bash -c "python inference/serve.py --config_file \"${{ matrix.model }}\".yaml.deepspeed --simple"
             docker exec "${TARGET}" bash -c "python examples/inference/api_server_simple/query_single.py --model_endpoint http://127.0.0.1:8000/${{ matrix.model }}"
@@ -143,16 +152,16 @@ jobs:
             docker exec "${TARGET}" bash -c "python examples/inference/api_server_simple/query_single.py --model_endpoint http://127.0.0.1:8000/${{ matrix.model }}"
             docker exec "${TARGET}" bash -c "python examples/inference/api_server_simple/query_single.py --model_endpoint http://127.0.0.1:8000/${{ matrix.model }} --streaming_response"
           fi
-      
+
       - name: Run Inference Test with REST API
         run: |
           TARGET=${{steps.target.outputs.target}}
           if [[ ${{ matrix.model }} == "mpt-7b-bigdl" ]]; then
             docker exec "${TARGET}" bash -c "python inference/serve.py --config_file inference/models/bigdl/mpt-7b-bigdl.yaml"
-          else
+          elif [[ ! ${{ matrix.model }} == "llama-2-7b-chat-hf-vllm" ]]; then
             docker exec "${TARGET}" bash -c "python inference/serve.py --models ${{ matrix.model }}"
+            docker exec "${TARGET}" bash -c "python examples/inference/api_server_openai/query_http_requests.py --model_name ${{ matrix.model }}"
           fi
-          docker exec "${TARGET}" bash -c "python examples/inference/api_server_openai/query_http_requests.py --model_name ${{ matrix.model }}"
 
       - name: Stop Ray
         run: |
@@ -161,7 +170,7 @@ jobs:
           if [[ ! -z "$cid" ]]; then
             docker exec "${TARGET}" bash -c "ray stop"
           fi
-      
+
       - name: Stop Container
         if: success() || failure()
         run: |
@@ -173,4 +182,4 @@ jobs:
         run: echo "to be continued"
 
 
-      
+
diff --git a/.github/workflows/workflow_orders_on_pr.yml b/.github/workflows/workflow_orders_on_pr.yml
@@ -20,7 +20,7 @@ jobs:
 
   call-lint:
     uses: ./.github/workflows/workflow_lint.yml
-    
+
   call-tests:
     needs: call-lint
     uses: ./.github/workflows/workflow_tests.yml

diff --git a/dev/docker/Dockerfile.vllm b/dev/docker/Dockerfile.vllm
@@ -0,0 +1,42 @@
+# syntax=docker/dockerfile:1
+FROM ubuntu:22.04
+
+ENV LANG C.UTF-8
+
+WORKDIR /root/llm-on-ray
+
+RUN --mount=type=cache,target=/var/cache/apt apt-get update -y \
+    && apt-get install -y build-essential cmake wget curl git vim htop ssh net-tools \
+    && apt-get clean \
+    && rm -rf /var/lib/apt/lists/*
+
+ENV CONDA_DIR /opt/conda
+RUN wget --quiet https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda.sh && \
+    /bin/bash ~/miniconda.sh -b -p /opt/conda
+ENV PATH $CONDA_DIR/bin:$PATH
+
+# setup env
+SHELL ["/bin/bash", "--login", "-c"]
+
+RUN --mount=type=cache,target=/opt/conda/pkgs conda init bash && \
+    unset -f conda && \
+    export PATH=$CONDA_DIR/bin/:${PATH} && \
+    conda config --add channels intel && \
+    conda install -y -c conda-forge python==3.9 gxx=12.3 gxx_linux-64=12.3
+
+COPY ./pyproject.toml .
+COPY ./dev/scripts/install-vllm-cpu.sh .
+
+RUN mkdir ./finetune && mkdir ./inference
+
+RUN --mount=type=cache,target=/root/.cache/pip pip install -e .[cpu] -f https://developer.intel.com/ipex-whl-stable-cpu \
+    -f https://download.pytorch.org/whl/torch_stable.html
+
+# Install vllm-cpu
+# Activate base first for loading g++ envs ($CONDA_PREFIX/etc/conda/activate.d/*)
+RUN --mount=type=cache,target=/root/.cache/pip \
+    source /opt/conda/bin/activate base && ./install-vllm-cpu.sh
+
+# TODO: workaround, remove this when fixed in vllm-cpu upstream
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install xformers
diff --git a/dev/scripts/install-vllm-cpu.sh b/dev/scripts/install-vllm-cpu.sh
@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+
+# Check tools
+[[ -n $(which g++) ]] || { echo "GNU C++ Compiler (g++) is not found!";  exit 1; }
+[[ -n $(which pip) ]] || { echo "pip command is not found!";  exit 1; }
+
+# g++ version should be >=12.3
+version_greater_equal()
+{
+    printf '%s\n%s\n' "$2" "$1" | sort --check=quiet --version-sort
+}
+gcc_version=$(g++ -dumpversion)
+echo
+echo Current GNU C++ Compiler version: $gcc_version
+echo
+version_greater_equal "${gcc_version}" 12.3.0 || { echo "GNU C++ Compiler 12.3.0 or above is required!"; exit 1; }
+
+# Install from source
+MAX_JOBS=8 pip install -v git+https://github.com/bigPYJ1151/vllm@PR_Branch \
+    -f https://download.pytorch.org/whl/torch_stable.html
diff --git a/dev/scripts/start-ray-cluster.sh b/dev/scripts/start-ray-cluster.sh
@@ -3,9 +3,12 @@
 set -eo pipefail
 
 # Setup oneapi envs before starting Ray
-source /opt/intel/oneapi/setvars.sh
-
-export CCL_ZE_IPC_EXCHANGE=sockets
+if [[ -e "/opt/intel/oneapi/setvars.sh" ]]; then
+  source /opt/intel/oneapi/setvars.sh
+  export CCL_ZE_IPC_EXCHANGE=sockets
+else
+  echo "/opt/intel/oneapi/setvars.sh doesn't exist, not loading."
+fi
 
 # Setup Ray cluster
 RAY_SERVE_ENABLE_EXPERIMENTAL_STREAMING=1 ray start --head --node-ip-address 127.0.0.1 --ray-debugger-external

diff --git a/docs/vllm.md b/docs/vllm.md
@@ -0,0 +1,43 @@
+# Setting up vLLM For Intel CPU
+
+__NOTICE: The support for vLLM is experimental and subject to change.__
+
+## Install vLLM for Intel CPU
+
+vLLM for CPU currently supports Intel® 4th Gen Xeon® Scalable Performance processor (formerly codenamed Sapphire Rapids) for best performance and is runnable with FP32 precision for other Xeon processors.
+
+Please run the following script to install vLLM for CPU into your current environment. Currently a GNU C++ compiler with >=12.3 version is required to build and install.
+
+```bash
+$ dev/scripts/install-vllm-cpu.sh
+```
+
+## Setup
+
+Please follow [Deploying and Serving LLMs on Intel CPU/GPU/Gaudi](serve.md) document to setup other environments.
+
+## Run
+
+#### Serving
+
+To serve model with vLLM, run the following:
+
+```bash
+$ python serve.py --config_file inference/models/vllm/llama-2-7b-chat-hf-vllm.yaml --simple --keep_serve_terminal
+```
+
+In the above example, `vllm` property is set to `true` in the config file for enabling vLLM.
+
+#### Querying
+
+To start a non-streaming query, run the following:
+
+```bash
+$ python examples/inference/api_server_simple/query_single.py --model_endpoint http://127.0.0.1:8000/llama-2-7b-chat-hf
+```
+
+To start a streaming query, run the following:
+
+```bash
+$ python examples/inference/api_server_simple/query_single.py --model_endpoint http://127.0.0.1:8000/llama-2-7b-chat-hf --streaming_response
+```
diff --git a/inference/deepspeed_predictor.py b/inference/deepspeed_predictor.py
@@ -18,7 +18,7 @@
     InferenceConfig,
     DEVICE_CPU,
     DEVICE_XPU,
-    IPEX_PRECISION_BF16,
+    PRECISION_BF16,
 )
 
 
@@ -139,7 +139,7 @@ def init_model(self, local_rank: int):
             pipe.model = ipex.optimize_transformers(
                 pipe.model.eval(),
                 dtype=torch.bfloat16
-                if self.infer_conf.ipex.precision == IPEX_PRECISION_BF16
+                if self.infer_conf.ipex.precision == PRECISION_BF16
                 else torch.float32,
                 inplace=True,
             )

diff --git a/inference/inference_config.py b/inference/inference_config.py
@@ -3,8 +3,8 @@
 from pydantic_yaml import parse_yaml_raw_as
 from typing import List, Dict, Union
 
-IPEX_PRECISION_BF16 = "bf16"
-IPEX_PRECISION_FP32 = "fp32"
+PRECISION_BF16 = "bf16"
+PRECISION_FP32 = "fp32"
 
 DEVICE_CPU = "cpu"
 DEVICE_HPU = "hpu"
@@ -32,7 +32,18 @@ class Ipex(BaseModel):
     @validator("precision")
     def _check_precision(cls, v: str):
         if v:
-            assert v in [IPEX_PRECISION_BF16, IPEX_PRECISION_FP32]
+            assert v in [PRECISION_BF16, PRECISION_FP32]
+        return v
+
+
+class Vllm(BaseModel):
+    enabled: bool = False
+    precision: str = "bf16"
+
+    @validator("precision")
+    def _check_precision(cls, v: str):
+        if v:
+            assert v in [PRECISION_BF16, PRECISION_FP32]
         return v
 
 
@@ -89,6 +100,7 @@ class InferenceConfig(BaseModel):
     gpus_per_worker: int = 0
     hpus_per_worker: int = 0
     deepspeed: bool = False
+    vllm: Vllm = Vllm()
     workers_per_group: int = 2
     device: str = DEVICE_CPU
     ipex: Ipex = Ipex()

diff --git a/inference/models/vllm/llama-2-7b-chat-hf-vllm.yaml b/inference/models/vllm/llama-2-7b-chat-hf-vllm.yaml
@@ -0,0 +1,27 @@
+port: 8000
+name: llama-2-7b-chat-hf
+route_prefix: /llama-2-7b-chat-hf
+cpus_per_worker: 24
+gpus_per_worker: 0
+deepspeed: false
+vllm:
+  enabled: true
+  precision: bf16
+workers_per_group: 2
+device: "cpu"
+ipex:
+  enabled: false
+  precision: bf16
+model_description:
+  model_id_or_path: meta-llama/Llama-2-7b-chat-hf
+  tokenizer_name_or_path: meta-llama/Llama-2-7b-chat-hf
+  chat_processor: ChatModelLLama
+  prompt:
+    intro: ''
+    human_id: '[INST] {msg} [/INST]
+
+      '
+    bot_id: ''
+    stop_words: []
+  config:
+    use_auth_token: ''
diff --git a/inference/predictor.py b/inference/predictor.py
@@ -3,6 +3,7 @@
 from transformers import AutoTokenizer, StoppingCriteriaList
 from inference.inference_config import InferenceConfig
 from utils import StoppingCriteriaSub
+from typing import List, AsyncGenerator, Union
 
 
 class Predictor:
@@ -72,11 +73,20 @@ def configure_tokenizer(self, model_name):
             tokenizer.pad_token = tokenizer.eos_token
             model.generation_config.pad_token_id = model.generation_config.eos_token_id
 
-    def generate(self, prompt, **config):
+    def generate(self, prompts: Union[str, List[str]], **config) -> Union[str, List[str]]:
         pass
 
-    def streaming_generate(self, prompt, streamer, **config):
+    async def generate_async(
+        self, prompts: Union[str, List[str]], **config
+    ) -> Union[str, List[str]]:
+        pass
+
+    # output is streamed into streamer
+    def streaming_generate(self, prompt: str, streamer, **config) -> None:
         pass
 
     def get_streamer(self):
         pass
+
+    async def stream_results(self, results_generator) -> AsyncGenerator[str, None]:
+        pass