ui: model status and load progress via /models/sse feed

[ServeurpersoCom]

Overview

When you pick a model that is not loaded yet, you now see what is happening instead of a silent wait: the model row shows a live status and a progress bar as it loads, and if you just send a message to a cold model, the same progress shows right in the chat while it warms up. It updates itself in real time, so the interface always reflects the real state of your models.

Additional information

Video

LoadingProgress.mp4

Implementation

Replace the post-operation status polling with a single persistent /models/sse feed that becomes the source of truth for model state in router mode. The feed drives live row status, staged load progress and load/unload completion, surfaced both in the model selector and inline in chat while a cold model warms up. The stage labels and the tail share that each phase owns are centralized as constants, so the per-stage reports unify into one global percentage: text_model fills the bar while each later phase, spec then mmproj, takes a configurable trailing slice (set to 10%), tune the share and the names in one place. It is one reconnecting reader with one awaiter per model, no per-operation streams. Dispatch is event-kind based so it extends cleanly: the upcoming dedicated download process slots in by adding a downloading status and download_* events on the same feed, with its own byte-based progress alongside the stage-based load progress, no rework of the backbone.

Built on the server-side feed work by @ngxson : real-time load progress via /models/sse (#24828), mtmd load progress callback (#24865) and the spec model fix with the stages list (#24870); the dedicated download process (#24834) plugs into the same feed.

Ref: #24822

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES

[ServeurpersoCom]

We can configure the displayed text and the percentage for draft/mmproj inside model-loading.ts :

/**
 * Labels shown while a model loads, keyed by the stage reported on /models/sse.
 */
export const MODEL_LOAD_STAGE_LABELS: Record<ApiModelLoadStage, string> = {
        text_model: 'Loading weights',
        spec_model: 'Loading draft',
        mmproj_model: 'Loading projector'
};

/**
 * Share of the bar reserved for each load phase after text_model.
 * text_model fills the rest, so a plain model reaches 100% on its own.
 */
export const MODEL_LOAD_TAIL_SHARE = 0.1;

And it will be easy to make it work for downloading and single model mode when the backend does.

[allozaur]

Wonderful work. One architectural nitpick and we good to go @ServeurpersoCom