vllm serve meta-llama/Llama-3.1-8B-Instruct --served-model-name my-llama

Makes the server answer to the name my-llama in /v1/models and requests, while actually running Llama-3.1-8B-Instruct underneath. Clients then pass "model": "my-llama" in their requests. Useful for stable names across model upgrades or for hiding which weights you run.