SGLang 0.5

Apply the Python 3 prompt first. These rules cover the SGLang 0.5 line, checked against the 0.5.18 tag: the Python serving runtime, its native and OpenAI-compatible APIs, and the Model Gateway source shipped with each release. SGLang requires Python 3.10 or newer; use a maintained interpreter with compatible dependency wheels. Verify separately versioned gateway packages and alternative serving backends against their actual build.

HTTP exposure and authentication

• Keep inference workers on loopback or a private interface behind an authenticated, TLS-protected gateway that allows only required inference routes; --host defaults to 127.0.0.1, while binding 0.0.0.0 exposes additional interfaces.
• Configure distinct --api-key and --admin-api-key values. Normal HTTP routes require only the inference key when set; an admin key alone leaves normal routes open. ADMIN_OPTIONAL routes prefer the admin key, fall back to the inference key, and allow access when neither is configured.
• Block /server_info and /get_server_info from inference clients at both worker and gateway boundaries. The Python handler serializes startup settings containing credentials, and the gateway places these routes in its public route group; key configuration alone does not make them safe to expose. Rotate credentials if those responses were exposed.
• Deny inference clients all weight updates, tensor-based LoRA loading, weight reads and transfers, profiling, internal-state changes, and HiCache administration. In particular, keep /update_weights_from_tensor and