vllm-tgis-adapter 0.9.0.post1

vLLM adapter for a TGIS-compatible grpc server

vllm-tgis-adapter

vLLM adapter for a TGIS-compatible grpc server.

Install

vllm-tgis-adapter is available on PyPi

pip install vllm-tgis-adapter
python -m vllm_tgis_adapter

HealthCheck CLI

Installing the adapter also install a grpc healthcheck cli that can be used to monitor the status of the grpc server:

$ grpc_healtheck
health check...status: SERVING

See usage with

grpc_healthcheck --help

Build

python -m build
pip install dist/*whl
python -m vllm_tgis_adapter

Inference

This will start serving a grpc server on port 8033. This can be queried with grpcurl:

bash examples/inference.sh

Docker

Image available at quay.io/opendatahub/vllm, built from opendatahub-io/vllm's Dockerfile.ubi

docker pull quay.io/opendatahub/vllm

Inference

See examples

Contributing

Set up pre-commit for linting/style/misc fixes:

pip install pre-commit
pre-commit install
# to run on all files
pre-commit run --all-files

This project uses nox to manage test automation and uv for venv management:

pip install nox uv
nox --list  # list available sessions
nox -s tests-3.10 # run tests session for a specific python version
nox -s build-3.11 # build the wheel package
nox -s lint-3.11 -- --mypy # run linting with type checks

Testing without a GPU

The standard vllm built requires an Nvidia GPU. When this is not available, it is possible to compile vllm from source with CPU support:

git clone https://github.com/vllm-project/vllm
cd vllm

uv venv
source .venv/bin/activate

export UV_EXTRA_INDEX_URL=https://download.pytorch.org/whl/cpu \
    UV_INDEX_STRATEGY=unsafe-best-match\

.github/scripts/install_vllm_build_deps.py pyproject.toml

env \
    VLLM_TARGET_DEVICE=cpu \
    python setup.py bdist_wheel
export VLLM_VERSION_OVERRIDE=$PWD/dist/*whl
# the nox session can now be run with the custom built vllm cpu version

making it possible to run the tests on most hardware. Please note that the uv extra index url is required in order to install the torch CPU version.