vLLM: serve multiple LoRA adapters from a single base model deployment (multi-LoRA)

Steps

1. Start the server with --enable-lora and register one or more adapters via --lora-modules {name}={path}, e.g. `vllm serve meta-llama/Llama-3.2-3B-Instruct --enable-lora --lora-modules sql-lora=jeeejeee/llama32-3b-text2sql-spider`.
2. Set sizing flags that apply to all adapters: --max-loras, --max-lora-rank (match the largest rank among the adapters you're serving, don't over-set it), and --max-cpu-loras.
3. Confirm registration by calling `GET /v1/models` — each LoRA adapter is listed alongside the base model (with the JSON form of --lora-modules, its parent field points back to the base model id).
4. Send inference requests specifying the adapter name in the model field, e.g. `curl http://localhost:8000/v1/completions -d '{"model": "sql-lora", "prompt": "San Francisco is a", "max_tokens": 7}'` — base-model and multiple LoRA requests are served concurrently.
5. To add/remove adapters without restarting the server, set VLLM_ALLOW_RUNTIME_LORA_UPDATING=True and POST to /v1/load_lora_adapter and /v1/unload_lora_adapter with a lora_name/lora_path JSON body.