sprmarc/vllm-docker

vLLM Docker Container Image

vLLM is a fast and easy-to-use library for LLM inference and serving. This container image runs the OpenAI API server of vLLM.

Image URL: ghcr.io/substratusai/vllm

Support the project by adding a star! ❤️

Join us on Discord:

Quickstart

Deploy Mistral 7B Instruct using Docker:

docker run -d -p 8080:8080 --gpus=all \
  -e MODEL=mistralai/Mistral-7B-Instruct-v0.1 \
  ghcr.io/substratusai/vllm

Deploy Mistral 7B Instruct using K8s:

kubectl apply -f https://raw.githubusercontent.com/substratusai/vllm-docker/main/k8s-deployment.yaml

Configuration Options

The following configuration options are available by using environment variables:

Env Name	Description
MODEL	REQUIRED, The model ID to serve. This can be in the form of hf_org/model or utilize a path to point to a local model. Example value: mistralai/Mistral-7B-Instruct-v0.1
SERVED_MODEL_NAME	OPTIONAL, The model name used in the API. If not specified, the model name will be the same as the huggingface name.
GPU_MEMORY_UTILIZATION	OPTIONAL, the max memory allowed to be utilized, default is 0.90
PORT	OPTIONAL, the port to use for serving, default is 8080
QUANTIZATION	OPTIONAL, the quantization method. Choices: 'awq', 'squeezellm'
DTYPE	OPTIONAL, the data type for model weights. Needs to be "half" when "awq" is used
MAX_MODEL_LEN	OPTIONAL, model context length. By default this is automatically derived from the model. Needs to be set to something low when using awq
EXTRA_ARGS	OPTIONAL, Any additional command line arguments to pass along

Please see the vLLM source code of arg-utils.py for more details.

The container image automatically detects the number of GPUs and sets --tensor-parallel-size to be equal to number of GPUs available. The gpu-count.py script is used to detect number of GPUs.

Building

docker build -t ghcr.io/substratusai/vllm .

Helm Chart / K8s

Please see the vLLM helm chart that uses this image: substratusai/helm/vllm