Skip to content

Latest commit

 

History

History
458 lines (344 loc) · 12.8 KB

README.md

File metadata and controls

458 lines (344 loc) · 12.8 KB
Raw

LLM Microservice

This microservice, designed for Language Model Inference (LLM), processes input consisting of a query string and associated reranked documents. It constructs a prompt based on the query and documents, which is then used to perform inference with a large language model. The service delivers the inference results as output.

A prerequisite for using this microservice is that users must have a LLM text generation service (etc., TGI, vLLM and Ray) already running. Users need to set the LLM service's endpoint into an environment variable. The microservice utilizes this endpoint to create an LLM object, enabling it to communicate with the LLM service for executing language model operations.

Overall, this microservice offers a streamlined way to integrate large language model inference into applications, requiring minimal setup from the user beyond initiating a TGI/vLLM/Ray service and configuring the necessary environment variables. This allows for the seamless processing of queries and documents to generate intelligent, context-aware responses.

Validated LLM Models

Model TGI-Gaudi vLLM-CPU vLLM-Gaudi Ray
Intel/neural-chat-7b-v3-3 ✓ ✓ ✓ ✓
Llama-2-7b-chat-hf ✓ ✓ ✓ ✓
Llama-2-70b-chat-hf ✓ - ✓ x
Meta-Llama-3-8B-Instruct ✓ ✓ ✓ ✓
Meta-Llama-3-70B-Instruct ✓ - ✓ x
Phi-3 x Limit 4K Limit 4K ✓

Clone OPEA GenAIComps

Clone this repository at your desired location and set an environment variable for easy setup and usage throughout the instructions.

git clone https://github.com/opea-project/GenAIComps.git

export OPEA_GENAICOMPS_ROOT=$(pwd)/GenAIComps

Prerequisites

You must create a user account with HuggingFace and obtain permission to use the gated LLM models by adhering to the guidelines provided on the respective model's webpage. The environment variables LLM_MODEL would be the HuggingFace model id and the HF_TOKEN is your HuggugFace account's "User Access Token".

🚀1. Start Microservice with Python (Option 1)

To start the LLM microservice, you need to install python packages first.

1.1 Install Requirements

# Install opea-comps
pip install opea-comps

# Install requirements from comps/llms
cd ${OPEA_GENAICOMPS_ROOT}/comps/llms

pip install -r requirements.txt

1.2 Start LLM Service with Python Script

1.2.1 Start the TGI Service

Install the requirements for TGI Service

cd ${OPEA_GENAICOMPS_ROOT}/comps/llms/text-generation/tgi

pip install -r requirements.txt

Execute the docker run command to initiate the backend, along with the Python script that launches the microservice.

export TGI_HOST_IP=$(hostname -I | awk '{print $1}')  # This sets IP of the current machine
export LLM_MODEL=${your_hf_llm_model}
export DATA_DIR=$HOME/data  # Location to download the model
export HF_TOKEN=${your_hf_api_token}

# Initiate the backend
docker run -d \
  -p 8008:80 \
  -e HF_TOKEN=${HF_TOKEN} \
  -v ${DATA_DIR}:/data \
  --name tgi_service \
  --shm-size 1g \
  ghcr.io/huggingface/text-generation-inference:1.4 \
  --model-id ${LLM_MODEL}

# Start the microservice with an endpoint as the above docker run command
export TGI_LLM_ENDPOINT="http://${TGI_HOST_IP}:8008"

python llm.py

1.2.2 Start the vLLM Service

Install the requirements for vLLM Service

cd ${OPEA_GENAICOMPS_ROOT}/comps/llms/text-generation/vllm/langchain

pip install -r requirements.txt

Execute the docker run command to initiate the backend, along with the Python script that launches the microservice.

export vLLM_HOST_IP=$(hostname -I | awk '{print $1}')  # This sets IP of the current machine
export LLM_MODEL=${your_hf_llm_model}
export DATA_DIR=$HOME/data  # Location to download the model
export HF_TOKEN=${your_hf_api_token}

# Build the image first as opea/vllm:cpu
bash ${OPEA_GENAICOMPS_ROOT}/comps/llms/text-generation/vllm/langchain/dependency/build_docker_vllm.sh cpu

# Initiate the backend
docker run -d -it \
  --name vllm_service \
  -p 8008:80 \
  -e HF_TOKEN=${HF_TOKEN} \
  -e VLLM_CPU_KVCACHE_SPACE=40 \
  -v ${DATA_DIR}:/data \
  opea/vllm:cpu \
  --model ${LLM_MODEL} \
  --port 80

# Start the microservice with an endpoint as the above docker run command
export vLLM_ENDPOINT="http://${vLLM_HOST_IP}:8008"

python llm.py

1.2.3 Start the Ray Service

Install the requirements for Ray Service

cd ${OPEA_GENAICOMPS_ROOT}/comps/llms/text-generation/vllm/ray

pip install -r requirements.txt

Execute the docker run command to initiate the backend, along with the Python script that launches the microservice.

export vLLM_RAY_HOST_IP=$(hostname -I | awk '{print $1}')  # This sets IP of the current machine
export LLM_MODEL=${your_hf_llm_model}
export DATA_DIR=$HOME/data  # Location to download the model
export HF_TOKEN=${your_hf_api_token}

# Build the image first as opea/vllm:cpu
bash ${OPEA_GENAICOMPS_ROOT}/comps/llms/text-generation/vllm/ray/dependency/build_docker_vllmray.sh

# Initiate the backend
docker run \
  --name="vllm-ray-service" \
  --runtime=habana \
  -v $DATA_DIR:/data \
  -e HABANA_VISIBLE_DEVICES=all \
  -e OMPI_MCA_btl_vader_single_copy_mechanism=none \
  --cap-add=sys_nice \
  --ipc=host \
  -p 8006:8000 \
  -e HF_TOKEN=$HF_TOKEN \
  opea/vllm_ray:habana \
  /bin/bash -c " \
    ray start --head && \
    python vllm_ray_openai.py \
    --port_number 8000 \
    --model_id_or_path $LLM_MODEL \
    --tensor_parallel_size 2 \
    --enforce_eager False"

# Start the microservice with an endpoint as the above docker run command
export vLLM_RAY_ENDPOINT="http://${vLLM_RAY_HOST_IP}:8006"

python llm.py

🚀2. Start Microservice with Docker (Option 2)

In order to start the microservices with docker, you need to build the docker images first for the microservice.

2.1 Build Docker Image

2.1.1 TGI

# Build the microservice docker
cd ${OPEA_GENAICOMPS_ROOT}

docker build \
  --build-arg https_proxy=$https_proxy \
  --build-arg http_proxy=$http_proxy \
  -t opea/llm-tgi:latest \
  -f comps/llms/text-generation/tgi/Dockerfile .

2.1.2 vLLM

# Build vllm docker
bash ${OPEA_GENAICOMPS_ROOT}/comps/llms/text-generation/vllm/langchain/dependency/build_docker_vllm.sh hpu

# Build the microservice docker
cd ${OPEA_GENAICOMPS_ROOT}

docker build \
  --build-arg https_proxy=$https_proxy \
  --build-arg http_proxy=$http_proxy \
  -t opea/llm-vllm:latest \
  -f comps/llms/text-generation/vllm/langchain/Dockerfile .

2.1.3 Ray

# Build the Ray Serve docker
bash ${OPEA_GENAICOMPS_ROOT}/comps/llms/text-generation/vllm/ray/dependency/build_docker_vllmray.sh

# Build the microservice docker
cd ${OPEA_GENAICOMPS_ROOT}

docker build \
  --build-arg https_proxy=$https_proxy \
  --build-arg http_proxy=$http_proxy \
  -t opea/llm-vllm-ray:latest \
  -f comps/llms/text-generation/vllm/ray/Dockerfile .

2.2 Start LLM Service with the built image

To start a docker container, you have two options:

  • A. Run Docker with CLI
  • B. Run Docker with Docker Compose

You can choose one as needed. If you start an LLM microservice with docker compose, the docker_compose_llm.yaml file will automatically start both endpoint and the microservice docker.

2.2.1 Setup Environment Variables

In order to start TGI and LLM services, you need to setup the following environment variables first.

export HF_TOKEN=${your_hf_api_token}
export TGI_LLM_ENDPOINT="http://${your_ip}:8008"
export LLM_MODEL=${your_hf_llm_model}
export DATA_DIR=$HOME/data

In order to start vLLM and LLM services, you need to setup the following environment variables first.

export HF_TOKEN=${your_hf_api_token}
export vLLM_LLM_ENDPOINT="http://${your_ip}:8008"
export LLM_MODEL=${your_hf_llm_model}

In order to start Ray serve and LLM services, you need to setup the following environment variables first.

export HF_TOKEN=${your_hf_api_token}
export RAY_Serve_ENDPOINT="http://${your_ip}:8008"
export LLM_MODEL=${your_hf_llm_model}
export CHAT_PROCESSOR="ChatModelLlama"

2.3 Run Docker with CLI (Option A)

2.3.1 TGI

Start TGI endpoint.

docker run -d \
  -p 8008:80 \
  -e HF_TOKEN=${HF_TOKEN} \
  -v ${DATA_DIR}:/data \
  --name tgi_service \
  --shm-size 1g \
  ghcr.io/huggingface/text-generation-inference:1.4 \
  --model-id ${LLM_MODEL}

Start TGI microservice

docker run -d \
  --name="llm-tgi-server" \
  -p 9000:9000 \
  --ipc=host \
  -e http_proxy=$http_proxy \
  -e https_proxy=$https_proxy \
  -e TGI_LLM_ENDPOINT=$TGI_LLM_ENDPOINT \
  -e HF_TOKEN=$HF_TOKEN \
  opea/llm-tgi:latest

2.3.2 vLLM

Start vllm endpoint.

bash ${OPEA_GENAICOMPS_ROOT}/comps/llms/text-generation/vllm/langchain/dependency/launch_vllm_service.sh

Start vllm microservice.

docker run \
  --name="llm-vllm-server" \
  -p 9000:9000 \
  --ipc=host \
  -e http_proxy=$http_proxy \
  -e https_proxy=$https_proxy \
  -e no_proxy=${no_proxy} \
  -e vLLM_LLM_ENDPOINT=$vLLM_LLM_ENDPOINT \
  -e HF_TOKEN=$HF_TOKEN \
  -e LLM_MODEL=$LLM_MODEL \
  opea/llm-vllm:latest

2.3.3 Ray Serve

Start Ray Serve endpoint.

bash ${OPEA_GENAICOMPS_ROOT}/comps/llms/text-generation/vllm/ray/dependency/launch_vllmray.sh

Start Ray Serve microservice.

docker run -d \
  --name="llm-ray-server" \
  -p 9000:9000 \
  --ipc=host \
  -e http_proxy=$http_proxy \
  -e https_proxy=$https_proxy \
  -e RAY_Serve_ENDPOINT=$RAY_Serve_ENDPOINT \
  -e HF_TOKEN=$HF_TOKEN \
  -e LLM_MODEL=$LLM_MODEL \
  opea/llm-ray:latest

2.4 Run Docker with Docker Compose (Option B)

2.4.1 TGI

cd ${OPEA_GENAICOMPS_ROOT}/comps/llms/text-generation/tgi
docker compose -f docker_compose_llm.yaml up -d

2.4.2 vLLM

cd ${OPEA_GENAICOMPS_ROOT}/comps/llms/text-generation/vllm/langchain
docker compose -f docker_compose_llm.yaml up -d

2.4.3 Ray Serve

cd ${OPEA_GENAICOMPS_ROOT}/comps/llms/text-generation/vllm/ray
docker compose -f docker_compose_llm.yaml up -d

🚀3. Consume LLM Service

3.1 Check Service Status

curl http://${your_ip}:9000/v1/health_check\
  -X GET \
  -H 'Content-Type: application/json'

3.2 Verify the LLM Service

3.2.1 Verify the TGI Service

curl http://${your_ip}:8008/generate \
  -X POST \
  -d '{"inputs":"What is Deep Learning?","parameters":{"max_new_tokens":17, "do_sample": true}}' \
  -H 'Content-Type: application/json'

3.2.2 Verify the vLLM Service

curl http://${your_ip}:8008/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
  "model": ${your_hf_llm_model},
  "prompt": "What is Deep Learning?",
  "max_tokens": 32,
  "temperature": 0
  }'

3.2.3 Verify the Ray Service

curl http://${your_ip}:8008/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
  "model": ${your_hf_llm_model},
  "messages": [
        {"role": "assistant", "content": "You are a helpful assistant."},
        {"role": "user", "content": "What is Deep Learning?"}
    ],
  "max_tokens": 32,
  "stream": true
  }'

3.3 Consume LLM Service

You can set the following model parameters according to your actual needs, such as max_tokens, streaming.

The streaming parameter determines the format of the data returned by the API. It will return text string with streaming=false, return text streaming flow with streaming=true.

# non-streaming mode
curl http://${your_ip}:9000/v1/chat/completions \
  -X POST \
  -H 'Content-Type: application/json' \
  -d '{
  "query":"What is Deep Learning?",
  "max_tokens":17,
  "top_k":10,
  "top_p":0.95,
  "typical_p":0.95,
  "temperature":0.01,
  "repetition_penalty":1.03,
  "streaming":false
  }'


# streaming mode
curl http://${your_ip}:9000/v1/chat/completions \
  -X POST \
  -H 'Content-Type: application/json' \
  -d '{
  "query":"What is Deep Learning?",
  "max_tokens":17,
  "top_k":10,
  "top_p":0.95,
  "typical_p":0.95,
  "temperature":0.01,
  "repetition_penalty":1.03,
  "streaming":true
  }'