Deployment

A deployment assigns each protocol role to a process and operator. It records endpoints, authentication, separate credentials, provider identity, storage, capacity, expiration, observability, shutdown behavior, and trust boundaries. Review privacy and threat models before changing that placement.

The client must keep plaintext prompts, masks, private scales, model state, and decoding. Client-to-inference, client-to-preparation, and preparation-to-inference connections use different credentials. Changing the URL of an ordinary model API does not make its unmodified SDK private.

A loopback pllm benchmark run checks that the services work together on one machine. It does not show that production operators are independent. Deployment assurance must refer to the exact plan and environment being evaluated. The serving integration guide lists the current gateway and provider-role lifecycle commands.

Python SDK example

from pllm.deployment import Deployment

deployment = Deployment.local(root=".pllm/local")
assert deployment.to_spec() == {"kind": "local", "root": ".pllm/local"}

API: pllm.deployment.Deployment

This creates configuration only. Generic remote role deployment is not supported.