pllmParties and offline work
Learn what each party does before and during a private inference request.
The client holds plaintext and authorizes each request. The preparation service creates one-time protocol material before a request. The inference service consumes that material while evaluating the model. A provider implements a component. An operator controls a deployed process and its policy.
For public-weight inference, preparation commits to the model, transformer body, stages, quantization, and retry limit. It expands seeds supplied by the client into input masks, output masks, and one-time tickets. It computes corrections, sends them to inference, and waits for confirmation. During inference, only the client and inference service exchange tickets and masked activations.
Reserved rows can be used only once. Cancellation, failure, replay, or early completion invalidates every unused row in that reservation. See preparation and deployment.