ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Trim verbose comments

Browse source
This commit is contained in:
oobabooga 2026-06-24 19:21:29 -03:00
commit e2cd5ab90d
2 changed files with 19 additions and 32 deletions
Download patch file Download diff file Expand all files Collapse all files

19
studio/backend/core/inference/diffusion.py
View file

@ -3,19 +3,12 @@
"""Local diffusion (text-to-image) backend.
A small torch-only singleton that loads a diffusers pipeline and generates
images. Single-file GGUF checkpoints are dequantised on-device via
``diffusers.GGUFQuantizationConfig``; the rest of the pipeline (VAE, text
encoders, scheduler) comes from the matching base repo (see
``diffusion_families``). torch/diffusers are imported lazily inside the methods
so this module imports cleanly in a no-torch runtime.
Loading runs on a background thread (``begin_load``) so the route returns
immediately and the frontend can poll ``load_progress`` for a download bar.
This backend owns no GPU-handoff policy: it assumes it is the heavy GPU user
while loaded. Coordinating with the chat backend lives in the GPU arbiter the
routes call, not here.
A torch-only singleton: it dequantises a single-file GGUF on-device via
``GGUFQuantizationConfig`` and pulls the rest of the pipeline (VAE, text
encoders, scheduler) from the matching base repo. torch/diffusers are imported
lazily so this stays importable in a no-torch runtime. ``begin_load`` runs on a
background thread; poll ``load_progress`` for the download bar. GPU-handoff
policy lives in the arbiter the routes call, not here.
"""
from __future__ import annotations

32
studio/frontend/src/features/images/images-page.tsx
View file

@ -70,9 +70,8 @@ const MODELS: ModelOption[] = [
{ id: MODEL.repo_id, name: MODEL.label, description: "Text-to-image · GGUF", isGguf: true },
];
// Common aspect ratios (landscape-oriented; the Flip button gives the portrait
// mirror). Picking one locks the W:H proportion; the sliders size it, and the
// generation snaps to Z-Image's ~1-megapixel-trained 16px grid.
// Common aspect ratios (landscape; Flip gives the portrait mirror). Picking one
// locks the W:H proportion; the sliders set the size.
const ASPECT_RATIOS: Record<string, [number, number]> = {
"1:1": [1, 1],
"3:2": [3, 2],
@ -100,10 +99,9 @@ function matchAspect(width: number, height: number): { key: string; portrait: bo
return { key: found ? found[0] : "custom", portrait: height > width };
}
// The gallery is persisted on the backend (durable across reloads); this module
// cache only holds the last-fetched records + their object/data URLs so a tab
// switch re-renders instantly without a refetch flash. Object URLs live for the
// app's lifetime (revoked only on delete), so they stay valid across remounts.
// Module cache of the backend-persisted gallery, so a tab switch re-renders
// instantly. Object URLs are revoked only on delete (not unmount), so they stay
// valid across remounts.
const galleryCache: {
images: GalleryImage[];
selectedId: string | null;
@ -114,10 +112,8 @@ const galleryCache: {
inflight: Set<string>;
} = { images: [], selectedId: null, quant: null, srcById: new Map(), inflight: new Set() };
// Export filename: app name, a compact sortable timestamp, and the seed. A batch
// shares the seed + timestamp, so a "_<n>" suffix is added past the first image.
// e.g. Unsloth_20260624-143005_123.png / Unsloth_20260624-143005_123_1.png
// (the full recipe is embedded in the PNG regardless).
// Export filename, e.g. Unsloth_20260624-143005_123.png. Batch siblings share
// the seed + timestamp, so they get a "_<n>" suffix past the first one.
function exportFilename(image: GalleryImage): string {
const d = new Date(image.created_at * 1000);
const p = (n: number) => String(n).padStart(2, "0");
@ -148,9 +144,9 @@ const LOAD_TOAST_CLASSNAMES = {
description: "mt-0 w-full",
} as const;
// Render the chat ModelLoadDescription for a progress poll. The base text-
// encoder/VAE repo downloads alongside the GGUF, so the total can exceed the
// picked quant's size — that's the real one-time download.
// Render the chat ModelLoadDescription for a progress poll. The base repo
// (text-encoder/VAE) downloads alongside the GGUF, so the total exceeds the
// picked quant's size.
function loadToastDescription(p: DiffusionLoadProgress) {
// "Downloading" only when bytes actually remain to fetch — a cached model (or
// the pre-estimate window, total still 0) shouldn't claim a download.
@ -457,11 +453,9 @@ export function ImagesPage() {
toast.success("Settings restored to inputs");
}, []);
// Size controls: a locked aspect ratio keeps the paired dimension in step as
// you drag a slider; "custom" frees both. Flip swaps width and height and
// keeps the lock (the ratio just applies in the other orientation).
// h/w multiplier for the locked ratio [a,b] (a:b = long:short): landscape puts
// the long side on width (h = w*b/a), portrait puts it on height (h = w*a/b).
// A locked ratio keeps the paired dimension in step while dragging; "custom"
// frees both; Flip swaps W/H. ratioHW is the h/w for ratio [a,b] (long:short):
// landscape h = w*b/a, portrait h = w*a/b.
const ratioHW = (a: number, b: number) => (portrait ? a / b : b / a);
const changeAspect = (key: string) => {
setAspect(key);