Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
// SPDX-License-Identifier: AGPL-3.0-only
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
// Separate module with injected deps: loaded-models-api.ts reaches the chat store and
// cannot be imported by node --test.
// Extension-qualified so node --test resolves it.
import { unloadResident } from "../api-monitor/unload-resident.ts";
export type ResidentChatModel = {
checkpoint: string;
/** Status reports the load path, the store the repo id. */
aliases: string[];
};
export type EjectChatModelDeps = {
readResident: () => Promise<ResidentChatModel | null>;
unload: (modelPath: string) => Promise<void>;
matches: (left: string, right: string) => boolean;
cachedRow?: boolean;
readCached?: () => Promise<string[]>;
};
export type EjectChatModelResult = {
unloadedAliases: string[];
stillResident: string | null;
replacedBy: string | null;
};
/**
* Read, unload, re-read: /unload naming a replaced model is a 200 no-op, and the read is
* scoped so a model switched in before the click is never unloaded.
*/
export async function ejectChatModel(
target: string,
deps: EjectChatModelDeps,
): Promise<EjectChatModelResult> {
const seen: { resident: ResidentChatModel | null } = { resident: null };
const { unloadedAliases, stillResident } = await unloadResident({
readResident: async () => {
const resident = await deps.readResident();
seen.resident = resident;
if (!resident) return null;
const namesTarget = resident.aliases.some((alias) =>
deps.matches(target, alias),
);
return namesTarget ? resident : null;
},
unload: deps.unload,
});
if (unloadedAliases.length > 0) {
return { unloadedAliases, stillResident, replacedBy: null };
}
// Transformers keeps a replaced model in backend.models (reported as `loaded`), so check
// what the runtime really holds before calling the row somebody else's.
const holdsTarget = (names: string[] | undefined) =>
names?.some((name) => deps.matches(target, name)) ?? false;
// Re-read even for a cached row, since another tab may have evicted it; unloading a name nobody
// holds would still stop the active model's queues; a failed read keeps the row's own word.
const cachedNow = await deps.readCached?.().catch(() => undefined);
if (cachedNow === undefined ? deps.cachedRow : holdsTarget(cachedNow)) {
// A cached row is never the active model, so the scoped read can never
// match it. Name it directly: the Transformers backend can still hold it.
await deps.unload(target);
// /unload answers 200 for a name no longer held, so confirm with a re-read.
const survived = holdsTarget(await deps.readCached?.());
return {
unloadedAliases: survived ? [] : [target],
stillResident: survived ? target : null,
replacedBy: null,
};
}
// Nothing unloaded: /unload naming a replaced model is a 200 no-op, so do not fire it.
return {
unloadedAliases: [],
stillResident: null,
replacedBy: seen.resident?.checkpoint ?? null,
};
}