import { type ModelPool } from "./modelPool.js";
import { type LogOptions } from "./serveLog.js";
/** mlx_lm's own name for whatever a server was started with. A client
 *  that cannot spell a repo id (Harbor allows a self-hosted model name one
 *  slash) can ask for this instead. It names the one model a single-model
 *  server serves, and nothing when several are served, since it would be
 *  a guess which. */
export declare const DEFAULT_MODEL_ALIAS = "default_model";
/** The served model a request for `requested` goes to: itself when it is
 *  served, the one served model for `default_model`, and undefined
 *  otherwise. */
export declare function servedModelFor(models: string[], requested: string): string | undefined;
export type FrontDoor = {
    port: number;
    /** Ends every request in progress on a model, and restarts its process
     *  when a cancelled request would otherwise keep running. Returns how
     *  many requests were ended. */
    cancel: (model: string) => Promise<number>;
    /** From now on, every request gets a 503. */
    refuseNew: () => void;
    close: () => Promise<void>;
};
/** Where the door prints what it saw, and how much of it. Absent for a door
 *  that logs nothing, which is what the tests of the forwarding itself use. */
export type DoorLogging = LogOptions & {
    log: (line: string) => void;
};
/** The request with its reply length held to `limit`, the server's
 *  --max-tokens. mlx_lm.server reads a request's own `max_tokens` over its
 *  flag, so the flag alone only caps requests that name no length. Capping
 *  here makes it the ceiling the flag promises. A request that names no
 *  length is left alone, since the process then applies the flag itself. */
export declare function capMaxTokens(body: Record<string, unknown>, limit: number | undefined): Record<string, unknown>;
/** Listen on `port` (0 for any) and forward each request to the process of
 *  the model its body names. A request for a model that is not served gets
 *  a 404 and reaches no process. With `logging`, every request is printed
 *  as it ends. With `maxTokens`, a request asking for a longer reply is
 *  held to it. */
export declare function startFrontDoor(port: number, pool: ModelPool, logging?: DoorLogging, maxTokens?: number, onShutdown?: () => void): Promise<FrontDoor>;