> ## Documentation Index > Fetch the complete documentation index at: https://infercrane.mintlify.site/llms.txt > Use this file to discover all available pages before exploring further. # vLLM > How InferCrane manages and routes requests to vLLM runtimes. # vLLM vLLM is InferCrane's default qualified runtime profile. InferCrane owns deployment state and lifecycle; vLLM owns model execution. A supervised [vLLM Router](https://github.com/vllm-project/semantic-router) process distributes requests across healthy standalone replicas. ## Connect existing workers Each worker must expose the OpenAI-compatible vLLM API and be reachable from the InferCrane gateway. ```bash theme={"theme":{"light":"github-light-default","dark":"vesper"}} infercrane target add gpu-a \ --url http://gpu-a:8000 \ --runtime vllm \ --upstream-model Qwen/Qwen3-8B infercrane apply Qwen/Qwen3-8B \ --name qwen-prod \ --targets gpu-a \ --idempotency-key qwen-prod-initial \ --wait ``` The reconciler verifies health and served-model identity before publishing a route. An unhealthy or mismatched worker does not enter the request path. ## Responsibility boundary | InferCrane | vLLM | | ------------------------------------- | ----------------------------- | | Desired and observed deployment state | Model loading and execution | | Durable operations and revisions | OpenAI-compatible worker API | | Health/model reconciliation | Token generation and KV cache | | Safe routing generations | Per-replica runtime metrics | InferCrane does not implement an inference engine or distributed KV cache. SGLang and custom OCI workloads use the same versioned runtime boundary and have independent qualification states. Follow an OpenAI request from alias resolution to a healthy vLLM replica.