vLLM

vLLM Sidecar

Run Dynamo beside a stock vLLM engine through native gRPC.
View as Markdown

Experimental. The vLLM sidecar, launchers, packaging, and feature coverage can change without notice.

dynamo-vllm-sidecar is a CPU-only Dynamo worker that connects to vLLM’s native gRPC service. It preserves the upstream engine process and argument surface while using Dynamo for request handling and distributed serving. See the Sidecar Backends page for the common architecture.

Readiness

Deployment pathAggregatedDisaggregated
Local launcherValidated on one GPUValidated on two GPUs with NIXL
Kubernetes exampleValidatedValidated with NIXL

This table covers launch topology only. The vLLM feature matrix describes the in-process backend; sidecar feature parity is still under evaluation. See the vLLM sidecar README for current protocol limitations.

Launch Locally

From a Dynamo source checkout, build or install Dynamo so dynamo-vllm-sidecar is on PATH. Install a vLLM build that provides vllm-rs and its native gRPC server.

Start Dynamo’s local discovery services, then run the aggregated launcher:

$docker compose -f dev/docker-compose.yml up -d
$./lib/sidecar/vllm/launch/agg.sh --model Qwen/Qwen3-0.6B

To run separate prefill and decode engines on two GPUs:

$./lib/sidecar/vllm/launch/disagg.sh --model Qwen/Qwen3-0.6B

Each launcher starts the Dynamo frontend, the vLLM engine process or processes, and the matching sidecar workers. It binds the native gRPC endpoints to loopback.

Verify the frontend:

$curl localhost:8000/v1/chat/completions \
>  -H "Content-Type: application/json" \
>  -d '{
>    "model": "Qwen/Qwen3-0.6B",
>    "messages": [{"role": "user", "content": "Hello"}],
>    "max_tokens": 32
>  }'

Deploy on Kubernetes

No published sidecar image is available yet. Follow the Kubernetes quick start to build dynamo-sidecar, which contains all three engine-specific sidecar executables. The vLLM manifests run dynamo-vllm-sidecar as the container command and pair it with a stock upstream vLLM image. The source tree includes aggregated and disaggregated manifests.