Multinode Examples

For general TensorRT-LLM features and engine configuration, see the Reference Guide.

Recommended Path

For multinode TensorRT-LLM deployments, start from the checked-in Kubernetes recipes under recipes/. Those manifests are the supported entrypoints for launching multi-node workers, frontend services, and related routing components.

The main TRT-LLM recipe entrypoints are:

• DeepSeek-R1 WideEP on GB200
• Qwen3-235B-A22B-FP8 aggregated
• Qwen3-235B-A22B-FP8 disaggregated
• Qwen3-32B-FP8 aggregated
• Qwen3-32B-FP8 disaggregated
• GPT-OSS-120B aggregated
• GPT-OSS-120B disaggregated
• Nemotron-3-Super-FP8 disaggregated

For model-level setup, prerequisites, and hardware notes, use the recipe README files:

• DeepSeek-R1 recipes
• Qwen3-235B-A22B-FP8 recipes
• Qwen3-32B-FP8 recipes
• GPT-OSS-120B recipes
• Kimi-K2.5 recipes

Quick Start

At a high level, the Kubernetes workflow is:

1. Install the Dynamo platform on Kubernetes. See the Kubernetes Deployment Guide.
2. Create a namespace and any required secrets such as a Hugging Face token.
3. Apply the recipe’s model cache and model download manifests when the recipe includes them.
4. Apply the recipe’s deploy.yaml.
5. Port-forward the frontend service and send test requests to /v1/models or /v1/chat/completions.

Example flow:

$
export NAMESPACE=dynamo-demo
$
kubectl create namespace ${NAMESPACE}
$
$
kubectl create secret generic hf-token-secret \
>
  --from-literal=HF_TOKEN="your-token-here" \
>
  -n ${NAMESPACE}
$
$
# Example: deploy DeepSeek-R1 TRT-LLM WideEP on GB200.
$
kubectl apply -f recipes/deepseek-r1/model-cache/model-cache.yaml -n ${NAMESPACE}
$
kubectl apply -f recipes/deepseek-r1/model-cache/model-download.yaml -n ${NAMESPACE}
$
kubectl wait --for=condition=Complete job/model-download -n ${NAMESPACE} --timeout=7200s
$
kubectl apply -f recipes/deepseek-r1/trtllm/disagg/wide_ep/gb200/deploy.yaml -n ${NAMESPACE}

After the deployment is ready, port-forward the frontend service named by the recipe and send a test request:

$
kubectl port-forward svc/<frontend-service> 8000:8000 -n ${NAMESPACE}
$
$
curl http://localhost:8000/v1/models

Notes

• The TRT-LLM engine config files used by launch and deploy flows live under examples/backends/trtllm/engine_configs/.
• If you need to customize model parallelism, replica counts, or routing mode, edit the recipe-local manifest rather than introducing a separate scheduler-specific guide.
• For the current catalog of supported recipes, see recipes/README.md.