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

Show the compile control for SDXL in the Train settings

Browse source 
The compile select was nested inside the DiT-only branch of the precision area, so
supports_compile=true for sdxl never rendered it. Lift it out of the ternary: the
precision selects stay family-specific, the compile control follows supports_compile.
Verified with a live Playwright pass (mxfp8 listed for DiT families on sm100, compile
select present for SDXL, DiT precision selector still hidden for SDXL).
This commit is contained in:
Daniel Han 2026-07-03 19:18:54 +00:00
commit 1521172b9d
1 changed files with 44 additions and 46 deletions
Download patch file Download diff file Expand all files Collapse all files

90
studio/frontend/src/features/images/train/diffusion-train-panel.tsx
View file

@ -812,52 +812,29 @@ export function DiffusionTrainPanel({
</div>
{isDiT ? (
<>
<div className="grid gap-1.5">
<Label className="text-xs">Base precision</Label>
<select
value={basePrecision}
onChange={(e) => {
precisionDirty.current = true;
setBasePrecision(e.target.value as typeof basePrecision);
}}
className={selectClass}
aria-label="Base precision"
>
{precisionModes.map((m) => (
<option key={m} value={m}>
{precisionLabel(m)}
</option>
))}
</select>
<p className="text-[11px] leading-snug text-muted-foreground">
How the frozen base weights are quantised. nf4 (4-bit) uses the least VRAM;
bf16 is fastest but needs the most. Auto picks this family&apos;s recommended
mode.
</p>
</div>
{supportsCompile && (
<div className="grid gap-1.5">
<Label className="text-xs">Compile transformer</Label>
<select
value={compileTransformer}
onChange={(e) =>
setCompileTransformer(e.target.value as typeof compileTransformer)
}
className={selectClass}
aria-label="Compile transformer"
>
<option value="auto">Auto</option>
<option value="on">On (faster after warmup)</option>
<option value="off">Off</option>
</select>
<p className="text-[11px] leading-snug text-muted-foreground">
torch.compile the transformer. Adds a one-time warmup, then speeds up each
step.
</p>
</div>
)}
</>
<div className="grid gap-1.5">
<Label className="text-xs">Base precision</Label>
<select
value={basePrecision}
onChange={(e) => {
precisionDirty.current = true;
setBasePrecision(e.target.value as typeof basePrecision);
}}
className={selectClass}
aria-label="Base precision"
>
{precisionModes.map((m) => (
<option key={m} value={m}>
{precisionLabel(m)}
</option>
))}
</select>
<p className="text-[11px] leading-snug text-muted-foreground">
How the frozen base weights are quantised. nf4 (4-bit) uses the least VRAM;
bf16 is fastest but needs the most. Auto picks this family&apos;s recommended
mode.
</p>
</div>
) : (
<div className="grid gap-1.5">
<Label className="text-xs">Precision</Label>
@ -876,6 +853,27 @@ export function DiffusionTrainPanel({
</p>
</div>
)}
{supportsCompile && (
<div className="grid gap-1.5">
<Label className="text-xs">Compile transformer</Label>
<select
value={compileTransformer}
onChange={(e) =>
setCompileTransformer(e.target.value as typeof compileTransformer)
}
className={selectClass}
aria-label="Compile transformer"
>
<option value="auto">Auto</option>
<option value="on">On (faster after warmup)</option>
<option value="off">Off</option>
</select>
<p className="text-[11px] leading-snug text-muted-foreground">
torch.compile the transformer blocks. Adds a one-time warmup, then speeds up
each step.
</p>
</div>
)}
</div>
</div>
);