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

studio: sort TIGHT (CPU offload) GGUF variants after GPU-only fits

Browse source 
Sort order is now: recommended -> fits (largest first) -> tight/CPU
offload (largest first) -> OOM (smallest first). Previously tight
variants were mixed with fits variants.
This commit is contained in:
Daniel Han 2026-03-15 06:50:16 +00:00
commit 7b19cb418e
1 changed files with 10 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

15
studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
View file

@ -230,17 +230,22 @@ function GgufVariantExpander({
const sortedVariants = useMemo(() => {
if (!variants) return variants;
// Tier: 0 = fits (GPU only), 1 = tight (needs CPU offload), 2 = OOM
const tierOf = (v: GgufVariantDetail) => {
const f = getGgufFit(v.size_bytes);
return f === "fits" ? 0 : f === "tight" ? 1 : 2;
};
return [...variants].sort((a, b) => {
const aIsRec = a.quant === effectiveRecommended;
const bIsRec = b.quant === effectiveRecommended;
if (aIsRec !== bIsRec) return aIsRec ? -1 : 1;
const aOom = getGgufFit(a.size_bytes) === "oom";
const bOom = getGgufFit(b.size_bytes) === "oom";
if (aOom !== bOom) return aOom ? 1 : -1;
const aTier = tierOf(a);
const bTier = tierOf(b);
if (aTier !== bTier) return aTier - bTier;
// Non-OOM: largest first (best quality); OOM: smallest first (most likely to fit)
return aOom ? a.size_bytes - b.size_bytes : b.size_bytes - a.size_bytes;
// fits/tight: largest first (best quality); OOM: smallest first
return aTier === 2 ? a.size_bytes - b.size_bytes : b.size_bytes - a.size_bytes;
});
}, [variants, effectiveRecommended, getGgufFit]);