vLLM 0.32

Apply these vLLM rules after the Python baseline. They cover the 0.32 release line as checked against 0.32.0. Treat model code, engine workers, caches, and administrative interfaces as one trusted service; put user and tenant policy at its ingress.

Serving and identity

• Run on a supported Python in vLLM's declared 3.11–3.14 range and match the selected CUDA, ROCm, XPU, CPU, or other wheel to its published accelerator and PyTorch stack. Rebuild against the intended stack when replacing a bundled binary dependency.
• Put vllm serve behind an authenticated TLS gateway. Allow only required route and method pairs, enforce caller and tenant quotas there, bind --host to a private interface or protected --uds, and prevent direct access to the backend.
• Configure --api-key or VLLM_API_KEY as an additional backend check. Its middleware covers only /v1, /v2, /inference, and /cohere, skips OPTIONS, and grants no application roles. Block or independently authorize /invocations, /generative_scoring, /pooling, /classify, /score, /rerank, /tokenize, /detokenize, /metrics, and operational routes such as /pause, /abort_requests, and /update_weights.
• Authorize every exposed inference route, model, adapter, and operation from gateway-authenticated identity. The Chat Completions user field is ignored, and client X-Request-Id or session_id is correlation data, not ownership evidence. Keep /v1/load_lora_adapter and /v1/unload_lora_adapter for administrators even though they fall under the API-key prefix.
• Add /v1/systemone