The Triton backend for vLLM is designed to run supported models on a vLLM engine. You can learn more about Triton backends in the backend repo.

This is a Python-based backend. When using this backend, all requests are placed on the vLLM AsyncEngine as soon as they are received. Inflight batching and paged attention is handled by the vLLM engine.

Where can I ask general questions about Triton and Triton backends? Be sure to read all the information below as well as the general Triton documentation available in the main server repo. If you don't find your answer there you can ask questions on the main Triton issues page.

There are several ways to install and deploy the vLLM backend.

Pull a tritonserver:<xx.yy>-vllm-python-py3 container with vLLM backend from the NGC registry. <xx.yy> is the version of Triton that you want to use. Please note, that Triton's vLLM container has been introduced starting from 23.10 release.

docker pull nvcr.io/nvidia/tritonserver:<xx.yy>-vllm-python-py3

You can follow steps described in the Building With Docker guide and use the build.py script.

A sample command to build a Triton Server container with all options enabled is shown below. Feel free to customize flags according to your needs.

Please use NGC registry to get the latest version of the Triton vLLM container, which corresponds to the latest YY.MM (year.month) of Triton release.

# YY.MM is the version of Triton.
export TRITON_CONTAINER_VERSION=<YY.MM>
./build.py -v  --enable-logging
                --enable-stats
                --enable-tracing
                --enable-metrics
                --enable-gpu-metrics
                --enable-cpu-metrics
                --enable-gpu
                --filesystem=gcs
                --filesystem=s3
                --filesystem=azure_storage
                --endpoint=http
                --endpoint=grpc
                --endpoint=sagemaker
                --endpoint=vertex-ai
                --upstream-container-version=${TRITON_CONTAINER_VERSION}
                --backend=python:r${TRITON_CONTAINER_VERSION}
                --backend=vllm:r${TRITON_CONTAINER_VERSION}

You can install the vLLM backend directly into the NGC Triton container. In this case, please install vLLM first. You can do so by running pip install vllm==<vLLM_version>. Then, set up the vLLM backend in the container with the following commands:

mkdir -p /opt/tritonserver/backends/vllm
wget -P /opt/tritonserver/backends/vllm https://raw.githubusercontent.com/triton-inference-server/vllm_backend/main/src/model.py

You can see an example model_repository in the samples folder. You can use this as is and change the model by changing the model value in model.json. model.json represents a key-value dictionary that is fed to vLLM's AsyncLLMEngine when initializing the model. You can see supported arguments in vLLM's arg_utils.py. Specifically, here and here.

For multi-GPU support, EngineArgs like tensor_parallel_size can be specified in model.json.

Note: vLLM greedily consume up to 90% of the GPU's memory under default settings. The sample model updates this behavior by setting gpu_memory_utilization to 50%. You can tweak this behavior using fields like gpu_memory_utilization and other settings in model.json.

Once you have the model repository set up, it is time to launch the Triton server. We will use the pre-built Triton container with vLLM backend from NGC in this example.

docker run --gpus all -it --net=host --rm -p 8001:8001 --shm-size=1G --ulimit memlock=-1 --ulimit stack=67108864 -v ${PWD}:/work -w /work nvcr.io/nvidia/tritonserver:<xx.yy>-vllm-python-py3 tritonserver --model-repository ./model_repository

Replace <xx.yy> with the version of Triton that you want to use. Note that Triton's vLLM container was first published starting from 23.10 release.

After you start Triton you will see output on the console showing the server starting up and loading the model. When you see output like the following, Triton is ready to accept inference requests.

I1030 22:33:28.291908 1 grpc_server.cc:2513] Started GRPCInferenceService at 0.0.0.0:8001
I1030 22:33:28.292879 1 http_server.cc:4497] Started HTTPService at 0.0.0.0:8000
I1030 22:33:28.335154 1 http_server.cc:270] Started Metrics Service at 0.0.0.0:8002

After you start Triton with the sample model_repository, you can quickly run your first inference request with the generate endpoint.

Try out the command below.

$ curl -X POST localhost:8000/v2/models/vllm_model/generate -d '{"text_input": "What is Triton Inference Server?", "parameters": {"stream": false, "temperature": 0}}'

Upon success, you should see a response from the server like this one:

{"model_name":"vllm_model","model_version":"1","text_output":"What is Triton Inference Server?\n\nTriton Inference Server is a server that is used by many"}

In the samples folder, you can also find a sample client, client.py which uses Triton's asyncio gRPC client library to run inference on Triton.

You can check the vLLM version included in Triton Inference Server from Framework Containers Support Matrix. Note: The vLLM Triton Inference Server container has been introduced starting from 23.10 release.

You can use pip install ... within the container to upgrade vLLM version.

If you are running multiple instances of Triton server with a Python-based backend, you need to specify a different shm-region-prefix-name for each server. See here for more information.

You can read further in the vLLM Quick Deploy guide in the tutorials repository.