Inference Platforms
Inference platforms and Semantic Router solve different routing problems. Semantic Router chooses a model or model pool from request meaning and policy. The inference platform deploys that model and chooses a replica according to capacity, locality, and health.
Choose an integration
| Platform | Start with | Typical ownership |
|---|---|---|
| vLLM Production Stack | Production Stack | vLLM model services, discovery, and replica routing. |
| AIBrix | AIBrix | Model deployment, autoscaling, and replica-level traffic management. |
| llm-d | llm-d | InferencePool endpoint selection and distributed inference patterns. |
| NVIDIA Dynamo | Dynamo | Dynamo graphs, workers, and frontend lifecycle. |
Use the platform your infrastructure team already supports. These guides do not replace the platform's release-specific installation, sizing, or upgrade documentation.
Keep the two layers aligned
For every model pool, align:
- the Semantic Router provider name and
provider_model_id; - the platform's served model identity;
- the stable Service, Gateway, or frontend address; and
- the modality, context, tool, and protocol capabilities declared by routing policy.
Use Service DNS or a managed gateway address rather than a Kubernetes
ClusterIP. Keep replica scheduling out of semantic policy, and keep prompt
meaning out of the replica scheduler.
Validate the complete path
- Verify direct generation from every model pool.
- Validate the canonical Router configuration.
- Verify a request through each public virtual model.
- Confirm both the semantic selection and the serving replica.
- Exercise unavailable-model behavior explicitly; do not assume the Router or platform provides cross-model fallback unless you configured and tested it.
For the surrounding deployment choices, return to Choose a Deployment.