Skip to content

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.

Clean up Hub runtime/config coupling (#9886)

Eyera9dfb40780e1f
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index e19d58383b2..f8284711745 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -3,6 +3,7 @@
import {
applyModelLoadConfigToRuntime,
+ clearModelConfigHandoff,
currentRuntimePerModelConfig,
type DeletedModelRef,
type ExternalConnectionRef,
@@ -14,9 +15,11 @@ import {
type PerModelConfig,
isServedByMlx,
loadedContextFields,
+ modelConfigHandoffForDestination,
resolveInitialConfig,
SidebarModelConfig,
useActiveModelConfig,
+ useModelConfigHandoffStore,
} from "@/features/model-picker";
import { ProjectComposer, Thread } from "@/components/assistant-ui/thread";
import { usePlatformStore } from "@/config/env";
@@ -2255,6 +2258,24 @@ export function ChatPage({
// Controlled, so the chord can open the switcher and not just its trigger.
const [projectPickerOpen, setProjectPickerOpen] = useState(false);
const [modelSelectorLocked, setModelSelectorLocked] = useState(false);
+ const modelConfigRequest = useModelConfigHandoffStore((state) =>
+ modelConfigHandoffForDestination(state.request, {
+ active,
+ newChatId: search.new,
+ threadId: search.thread,
+ compareId: search.compare,
+ projectId: search.project,
+ }),
+ );
+ const handleModelConfigRequestAdopted = useCallback(
+ (requestId: string) => {
+ setSettingsOpen(false);
+ setModelSelectorLocked(false);
+ setModelSelectorOpen(true);
+ clearModelConfigHandoff(requestId);
+ },
+ [setSettingsOpen],
+ );
const viewBeforeCompareRef = useRef<ChatSearch | null>(null);
// Latest non-compare view, so exiting compare can restore it even when compare was opened from a
// path that does not set viewBeforeCompareRef.
@@ -3954,13 +3975,17 @@ export function ChatPage({
activeGgufVariant={activeGgufVariant}
activeModelConfig={activeModelConfig}
activeLoadedContextLength={loadedContextLength}
+ configRequest={modelConfigRequest}
+ onConfigRequestAdopted={handleModelConfigRequestAdopted}
onValueChange={handleCheckpointChange}
onEject={handleEject}
onFoldersChange={refreshLocalModels}
onModelsChange={refreshModelLists}
deleteDisabled={modelOperationInProgress}
variant="ghost"
- open={active && modelSelectorOpen}
+ open={
+ active && (modelSelectorOpen || modelConfigRequest !== null)
+ }
onOpenChange={handleModelSelectorOpenChange}
triggerDataTour="chat-model-selector"
contentDataTour="chat-model-selector-popover"
@@ -4250,7 +4275,7 @@ export function ChatPage({
</div>
<ChatSettingsPanel
- open={active && settingsOpen}
+ open={active && modelConfigRequest === null && settingsOpen}
onOpenChange={(open) => {
setSettingsOpen(open);
}}
diff --git a/studio/frontend/src/features/chat/lib/resident-model-match.ts b/studio/frontend/src/features/chat/lib/resident-model-match.ts
index c3a77559e4f..d309c99c012 100644
--- a/studio/frontend/src/features/chat/lib/resident-model-match.ts
+++ b/studio/frontend/src/features/chat/lib/resident-model-match.ts
@@ -36,8 +36,8 @@ export function residentModelMatchesPick(
if (!status.active_model) {
return false;
}
- // A standalone file has no quant to choose between, and its row carries no label
- // (settingsGgufVariantForRow) while the backend derives one from the filename.
+ // A standalone file has no quant to choose between, and its picker entry carries
+ // no variant while the backend derives one from the filename.
const picksItsOwnVariant = !(
!pick.ggufVariant && isStandaloneGgufPath(pick.loadPath ?? pick.id)
);
diff --git a/studio/frontend/src/features/hub/catalog/dataset-download-section.tsx b/studio/frontend/src/features/hub/catalog/dataset-download-section.tsx
index f3268aabda9..fdb74de1dd1 100644
--- a/studio/frontend/src/features/hub/catalog/dataset-download-section.tsx
+++ b/studio/frontend/src/features/hub/catalog/dataset-download-section.tsx
@@ -8,12 +8,8 @@ import {
} from "@/components/ui/tooltip";
import { useHttpPartialsResumable, useRepoDownload } from "../download-manager";
import { deleteCachedDataset } from "../inventory";
-import { cn } from "@/lib/utils";
-import { TrainIcon } from "../components/train-icon";
-import { HUB_POST_DOWNLOAD_ACTIONS_VISIBLE } from "../lib/hub-feature-flags";
import { DotTag } from "./dot-tag";
import { PathInfoButton } from "./path-info-button";
-import { HugeiconsIcon } from "@hugeicons/react";
import { useState } from "react";
import { useHfTokenStore } from "../stores/hf-token-store";
import { formatBytes } from "../lib/format";
@@ -36,7 +32,6 @@ export function DatasetDownloadSection({
partialResumable = false,
cachePath,
knownBytes,
- onTrain,
onChange,
}: {
repoId: string;
@@ -46,7 +41,6 @@ export function DatasetDownloadSection({
partialResumable?: boolean;
cachePath?: string | null;
knownBytes?: number | null;
- onTrain?: () => void;
onChange?: () => void;
}) {
const hfToken = useHfTokenStore((s) => s.token);
@@ -162,23 +156,8 @@ export function DatasetDownloadSection({
)}
</div>
</div>
- {/* Train CTA hidden until Hub→train picker ships; divider pairs with it. */}
- {(!isDownloaded || downloading || HUB_POST_DOWNLOAD_ACTIONS_VISIBLE) && (
- <CardDivider />
- )}
- {isDownloaded && !downloading ? (
- <button
- type="button"
- onClick={() => onTrain?.()}
- className={cn(
- "hub-action-btn w-28",
- !HUB_POST_DOWNLOAD_ACTIONS_VISIBLE && "hidden",
- )}
- >
- <HugeiconsIcon icon={TrainIcon} strokeWidth={1.75} />
- Train
- </button>
- ) : (
+ {(!isDownloaded || downloading) && <CardDivider />}
+ {(!isDownloaded || downloading) && (
<DownloadActionButton
downloading={downloadAction.downloading}
cancelling={downloadAction.cancelling}
diff --git a/studio/frontend/src/features/hub/catalog/download-card.tsx b/studio/frontend/src/features/hub/catalog/download-card.tsx
index ab79e98b4a6..4ba98dfac66 100644
--- a/studio/frontend/src/features/hub/catalog/download-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/download-card.tsx
@@ -23,7 +23,7 @@ import {
ArrowReloadHorizontalIcon,
Delete02Icon,
Download01Icon,
- Settings02Icon,
+ PlayIcon,
} from "@hugeicons/core-free-icons";
import { HugeiconsIcon } from "@hugeicons/react";
import {
@@ -96,36 +96,36 @@ export function CardDivider() {
);
}
-export function CardSettingsButton({
+export function ModelRunActionButton({
label,
onClick,
+ loading = false,
}: {
label: string;
onClick: () => void;
+ loading?: boolean;
}) {
return (
- <Tooltip>
- <TooltipTrigger asChild>
- <button
- type="button"
- aria-label={label}
- onClick={(e) => {
- e.stopPropagation();
- onClick();
- }}
- className="inline-flex size-7 shrink-0 cursor-pointer items-center justify-center rounded-full text-muted-foreground opacity-0 transition-[opacity,background-color,color] duration-150 hover:bg-foreground/[0.06] hover:text-foreground focus-visible:opacity-100 group-hover/dl:opacity-100 dark:hover:bg-white/[0.08]"
- >
- <HugeiconsIcon
- icon={Settings02Icon}
- strokeWidth={1.75}
- className="size-4"
- />
- </button>
- </TooltipTrigger>
- <TooltipContent side="top" className="tooltip-compact">
- {label}
- </TooltipContent>
- </Tooltip>
+ <button
+ type="button"
+ aria-label={loading ? `${label}. Opening configuration.` : label}
+ aria-busy={loading}
+ disabled={loading}
+ onClick={onClick}
+ className="hub-run-action-btn w-28"
+ >
+ {loading ? (
+ <>
+ <Spinner />
+ Opening…
+ </>
+ ) : (
+ <>
+ <HugeiconsIcon icon={PlayIcon} strokeWidth={1.75} />
+ Run
+ </>
+ )}
+ </button>
);
}
diff --git a/studio/frontend/src/features/hub/catalog/download-section.tsx b/studio/frontend/src/features/hub/catalog/download-section.tsx
index 31b32646ea0..dee6adb1d44 100644
--- a/studio/frontend/src/features/hub/catalog/download-section.tsx
+++ b/studio/frontend/src/features/hub/catalog/download-section.tsx
@@ -2,6 +2,7 @@
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
import type { ModelInventoryFormat } from "../inventory";
+import type { HubModelRunSelection } from "../lib/model-run-selection";
import { GgufDownloadCard } from "./gguf-download-card";
import { SafetensorsDownloadCard } from "./safetensors-download-card";
@@ -13,7 +14,6 @@ export function DownloadSection({
partialTransport = null,
partialResumable = false,
modelFormat,
- canRun = true,
isActive,
activeQuant,
preferredGgufFile = null,
@@ -25,10 +25,8 @@ export function DownloadSection({
systemRamGb,
cachePath,
knownBytes,
- onLoad,
- onUseInChat,
- onEject,
- onTrain,
+ onRun,
+ runPending = false,
onChange,
showMemoryBar = true,
mediaRuntime = false,
@@ -40,7 +38,6 @@ export function DownloadSection({
partialTransport?: string | null;
partialResumable?: boolean;
modelFormat?: ModelInventoryFormat | null;
- canRun?: boolean;
isActive: boolean;
activeQuant: string | null;
preferredGgufFile?: string | null;
@@ -52,10 +49,8 @@ export function DownloadSection({
systemRamGb?: number;
cachePath?: string | null;
knownBytes?: number | null;
- onLoad: (opts: { ggufVariant?: string; expectedBytes?: number }) => void;
- onUseInChat?: () => void;
- onEject?: () => void;
- onTrain?: () => void;
+ onRun?: (selection: HubModelRunSelection) => void;
+ runPending?: boolean;
onChange?: () => void;
/** False for diffusion / audio / video GGUFs, which do not load through
* llama.cpp and so have nothing the KV estimator can say about them. */
@@ -69,7 +64,6 @@ export function DownloadSection({
isActive={isActive}
activeQuant={activeQuant}
preferredFile={preferredGgufFile}
-
preferredFileIntent={preferredGgufFileIntent}
isLoadingThisModel={isLoadingThisModel}
gpuGb={gpuGb}
@@ -77,9 +71,8 @@ export function DownloadSection({
systemRamGb={systemRamGb}
cachePath={cachePath}
isPartial={isPartial}
- onLoad={onLoad}
- onUseInChat={onUseInChat}
- onEject={onEject}
+ onRun={onRun}
+ runPending={runPending}
onChange={onChange}
showMemoryBar={showMemoryBar}
mediaRuntime={mediaRuntime}
@@ -94,15 +87,12 @@ export function DownloadSection({
partialTransport={partialTransport}
partialResumable={partialResumable}
modelFormat={modelFormat}
- canRun={canRun}
isActive={isActive}
isLoadingThisModel={isLoadingThisModel}
cachePath={cachePath}
knownBytes={knownBytes}
- onLoad={onLoad}
- onUseInChat={onUseInChat}
- onEject={onEject}
- onTrain={onTrain}
+ onRun={onRun ? () => onRun({}) : undefined}
+ runPending={runPending}
onChange={onChange}
/>
);
diff --git a/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx b/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
index 3bbfca9e990..ffde621fad7 100644
--- a/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/gguf-download-card.tsx
@@ -39,8 +39,6 @@ import {
MoreVerticalIcon,
PinIcon,
PinOffIcon,
- PlayIcon,
- RemoveCircleIcon,
} from "@hugeicons/core-free-icons";
import { HugeiconsIcon } from "@hugeicons/react";
import {
@@ -69,11 +67,11 @@ import {
ggufVariantTransferLabel,
sortDownloadableGgufVariants,
} from "../lib/gguf-variant-sort";
-import { HUB_GGUF_RUN_ACTIONS_VISIBLE } from "../lib/hub-feature-flags";
import {
ggufVariantsMatch,
normalizeGgufVariantIdentity,
} from "../lib/model-identity";
+import type { HubModelRunSelection } from "../lib/model-run-selection";
import { useHfTokenStore } from "../stores/hf-token-store";
import { DotTag } from "./dot-tag";
import { DownloadStopIndicator } from "./download-cancel-indicator";
@@ -81,6 +79,7 @@ import {
CardDivider,
DeleteConfirmDialog,
DownloadCard,
+ ModelRunActionButton,
UpdateConfirmDialog,
} from "./download-card";
import {
@@ -145,21 +144,17 @@ const CHIP_BASE =
"inline-flex h-5 shrink-0 items-center justify-center whitespace-nowrap rounded-full border px-2 text-ui-11p5 font-medium tabular-nums leading-3 shadow-[inset_0_1px_0_rgba(255,255,255,0.04)]";
const CHIP_DEFAULT =
"border-foreground/15 bg-muted text-foreground/85 dark:border-border/60 dark:bg-white/[0.04] dark:text-foreground/85";
-const CHIP_ACTIVE =
- "border-control-accent/40 bg-control-accent/10 text-control-accent";
function QuantBadge({
quant,
fit,
showFit = true,
- active = false,
variant = "trigger",
tooltipMode = "eager",
}: {
quant: string;
fit: GgufFitClass;
showFit?: boolean;
- active?: boolean;
variant?: "trigger" | "menu";
tooltipMode?: "eager" | "lazy" | "none";
}) {
@@ -179,7 +174,7 @@ function QuantBadge({
// `shrink` overrides CHIP_BASE's shrink-0 so a long file-path quant
// label can shrink and truncate instead of overflowing the row.
"min-w-0 max-w-full shrink gap-1.5 cursor-help",
- active ? CHIP_ACTIVE : CHIP_DEFAULT,
+ CHIP_DEFAULT,
)}
>
{showFit && (
@@ -194,14 +189,9 @@ function QuantBadge({
) : (
// Trigger quant label is the row's primary identity and is short
// (e.g. "Q4_K_M"); keep it `shrink-0` + `whitespace-nowrap` so it never
- // collapses to "q…" when the Update/Run actions crowd the row. The info
+ // collapses to "q…" when trailing actions crowd the row. The info
// group's `overflow-hidden` sacrifices the trailing status tags instead.
- <span
- className={cn(
- "inline-flex shrink-0 cursor-help items-center gap-1.5 whitespace-nowrap text-ui-12p5 font-medium tracking-tight tabular-nums",
- active ? "text-control-accent" : "text-foreground",
- )}
- >
+ <span className="inline-flex shrink-0 cursor-help items-center gap-1.5 whitespace-nowrap text-ui-12p5 font-medium tracking-tight tabular-nums text-foreground">
{showFit && (
<HugeiconsIcon
icon={InformationCircleIcon}
@@ -276,9 +266,9 @@ function createGgufVariantMenuItems(
}));
}
-// Shared options menu: used on every variant row, the run bar, and the
-// single-model (non-GGUF) run bar. Omit `quant` for a repo-level model. The
-// identifier uses llama.cpp's repo:quant syntax so it pastes into `-hf`.
+// Shared options menu for downloaded variant rows and the selected-variant
+// action strip. Omit `quant` for a repo-level model. The identifier uses
+// llama.cpp's repo:quant syntax so it pastes into `-hf`.
export function QuantOptionsMenu({
repoId,
quant,
@@ -296,7 +286,7 @@ export function QuantOptionsMenu({
downloaded: boolean;
canDelete: boolean;
onDelete: (quant?: string) => void;
- // Hidden in the run bar; pinning belongs to the On Device list.
+ // Hidden in the selected-variant action strip; pinning belongs to On Device.
showPin?: boolean;
buttonClassName?: string;
iconClassName?: string;
@@ -449,7 +439,7 @@ const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
repoId,
item,
selected,
- loaded,
+ mutationBlocked,
liveActive,
showFitInfo,
onSelect,
@@ -458,7 +448,7 @@ const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
repoId: string;
item: GgufVariantMenuItem;
selected: boolean;
- loaded: boolean;
+ mutationBlocked: boolean;
liveActive: boolean;
showFitInfo: boolean;
onSelect: (quant: string) => void;
@@ -477,7 +467,8 @@ const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
},
[selectVariant],
);
- const canDelete = (item.downloaded || item.partial) && !loaded && !liveActive;
+ const canDelete =
+ (item.downloaded || item.partial) && !mutationBlocked && !liveActive;
return (
<div
@@ -493,7 +484,7 @@ const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
: "hover:bg-foreground/[0.05] dark:hover:bg-foreground/[0.06]",
)}
>
- {/* Status (On device / Loaded / Partial) sits beside the quant on the
+ {/* Status (On device / Partial) sits beside the quant on the
left so the model's identity reads as one unit; only the size pins
right. No per-row "GGUF" tag: every row here is a GGUF quant and the
trigger already labels it, so repeating it only stole the room the
@@ -503,13 +494,10 @@ const GgufVariantMenuRow = memo(function GgufVariantMenuRow({
quant={item.label}
fit={item.fit}
showFit={showFitInfo}
- active={loaded}
variant="menu"
tooltipMode="lazy"
/>
- {item.downloaded && (
- <DotTag tone="success" label={loaded ? "Loaded" : "On device"} />
- )}
+ {item.downloaded && <DotTag tone="success" label="On device" />}
{!item.downloaded && item.partial && (
<Tooltip>
<TooltipTrigger asChild={true}>
@@ -567,8 +555,8 @@ export function GgufDownloadCard({
cachePath,
preferLocalCache = false,
isPartial = false,
- onLoad,
- onEject,
+ onRun,
+ runPending = false,
onChange,
showMemoryBar = true,
mediaRuntime = false,
@@ -587,10 +575,8 @@ export function GgufDownloadCard({
cachePath?: string | null;
preferLocalCache?: boolean;
isPartial?: boolean;
- onLoad: (opts: { ggufVariant?: string; expectedBytes?: number }) => void;
- /** Accepted for API parity; the run bar ejects instead of opening chat. */
- onUseInChat?: () => void;
- onEject?: () => void;
+ onRun?: (selection: HubModelRunSelection) => void;
+ runPending?: boolean;
onChange?: () => void;
/** False for diffusion / audio / video GGUFs. They load through a different
* planner onto a single torch device rather than the aggregate inference
@@ -774,12 +760,6 @@ export function GgufDownloadCard({
const ctaDisabled = isLoadingThisModel || !selected;
const selectedIsActive =
isActive && activeQuant && ggufVariantsMatch(selected?.quant, activeQuant);
- const isGgufRunCta =
- !!selected?.downloaded &&
- !cancelling &&
- !downloadingThisVariant &&
- !isLoadingThisModel &&
- !selectedIsActive;
// No verdict beats a wrong one: a media repo's fit is the diffusion planner's question, and
// this card only knows how to answer llama.cpp's. The picker still badges those rows.
const showFitInfo = !mediaRuntime && (Boolean(gpuGb) || Boolean(systemRamGb));
@@ -829,7 +809,7 @@ export function GgufDownloadCard({
? true
: downloadingThisVariant
? false
- : ctaDisabled && !selectedIsActive,
+ : ctaDisabled,
isPartial: Boolean(selected?.partial),
partialTransport: selected?.partial_transport ?? null,
partialResumable: selected?.partial_resumable === true,
@@ -897,6 +877,29 @@ export function GgufDownloadCard({
}, [updateTarget, updateTargetVariant, repoId]);
const variantListUnavailable = !sortedVariants || sortedVariants.length === 0;
const showVariantLoadingState = loading && variantListUnavailable;
+ const showUpdateAction =
+ selected?.downloaded === true &&
+ online &&
+ updateAvailable &&
+ !selectedIsActive &&
+ !isLoadingThisModel &&
+ !runPending &&
+ !cancelling &&
+ !downloadAction.starting &&
+ !downloadingThisVariant;
+ const showDownloadAction =
+ !selected?.downloaded ||
+ downloadingThisVariant ||
+ cancelling ||
+ downloadAction.starting;
+ const showRunAction =
+ Boolean(onRun) &&
+ selected?.downloaded === true &&
+ selected.partial !== true &&
+ !downloadingThisVariant &&
+ !cancelling &&
+ !downloadAction.starting &&
+ !isLoadingThisModel;
// Keep showing download progress while the variant list is unavailable, so a
// remount never hides an in-flight download behind the variant status card.
@@ -990,15 +993,21 @@ export function GgufDownloadCard({
</>
}
>
- <Popover open={open} onOpenChange={setOpen}>
+ <Popover
+ open={runPending ? false : open}
+ onOpenChange={(nextOpen) => {
+ if (!runPending) setOpen(nextOpen);
+ }}
+ >
<PopoverTrigger asChild={true}>
<button
type="button"
+ disabled={runPending}
onClick={(e) => {
e.preventDefault();
setOpen((o) => !o);
}}
- className="hub-menu-trigger flex h-9 min-w-0 flex-1 cursor-pointer items-center gap-2 rounded-full px-3 text-left transition-colors hover:bg-foreground/[0.04] data-[state=open]:bg-foreground/[0.06] dark:hover:bg-white/[0.04] dark:data-[state=open]:bg-white/[0.06]"
+ className="hub-menu-trigger flex h-9 min-w-0 flex-1 cursor-pointer items-center gap-2 rounded-full px-3 text-left transition-colors hover:bg-foreground/[0.04] data-[state=open]:bg-foreground/[0.06] disabled:cursor-wait disabled:opacity-60 disabled:hover:bg-transparent dark:hover:bg-white/[0.04] dark:data-[state=open]:bg-white/[0.06] dark:disabled:hover:bg-transparent"
>
{/* Quant label + status tags travel together as one left-aligned
group so the fit-info icon never floats orphaned from its tags.
@@ -1011,7 +1020,6 @@ export function GgufDownloadCard({
quant={selectedLabel ?? selected.quant}
fit={selectedFit ?? "oom"}
showFit={showFitInfo}
- active={Boolean(selectedIsActive)}
/>
) : (
<span className="text-ui-12p5 text-muted-foreground">
@@ -1019,10 +1027,7 @@ export function GgufDownloadCard({
</span>
)}
{selected?.downloaded && (
- <DotTag
- tone="success"
- label={selectedIsActive ? "Loaded" : "On device"}
- />
+ <DotTag tone="success" label="On device" />
)}
{selected && !selected.downloaded && selected.partial && (
<Tooltip>
@@ -1075,7 +1080,11 @@ export function GgufDownloadCard({
repoId={repoId}
item={item}
selected={item.key === selectedVariantKey}
- loaded={isActive && item.key === activeVariantKey}
+ mutationBlocked={
+ runPending ||
+ isLoadingThisModel ||
+ (isActive && item.key === activeVariantKey)
+ }
liveActive={liveActive}
showFitInfo={showFitInfo}
onSelect={handleSelectVariant}
@@ -1087,7 +1096,6 @@ export function GgufDownloadCard({
</PopoverContent>
</Popover>
- {/* TODO: inference settings gear hidden for now, work on it in a future PR. */}
{/* Options only resolve managed HF-cache repos, so skip local paths;
they also only apply to quants actually on disk. */}
{selected &&
@@ -1102,7 +1110,8 @@ export function GgufDownloadCard({
Boolean(selected.downloaded || selected.partial) &&
!selectedIsActive &&
!downloadingThisVariant &&
- !isLoadingThisModel
+ !isLoadingThisModel &&
+ !runPending
}
onDelete={(q) => q && handleDeleteVariant(q)}
showPin={false}
@@ -1111,109 +1120,86 @@ export function GgufDownloadCard({
/>
)}
- {!isGgufRunCta && <CardDivider />}
+ {(showUpdateAction || showDownloadAction || showRunAction) && (
+ <CardDivider />
+ )}
- {selected?.downloaded &&
- online &&
- updateAvailable &&
- !selectedIsActive &&
- !downloadingThisVariant && (
- <button
- type="button"
- onClick={() => selected && setUpdateTarget(selected.quant)}
- aria-label={`Update ${repoId}`}
- className="hub-action-btn ml-1 text-amber-700 dark:text-amber-300"
- >
- <HugeiconsIcon
- icon={ArrowReloadHorizontalIcon}
- strokeWidth={1.75}
- />
- Update
- </button>
- )}
+ {showUpdateAction && (
+ <button
+ type="button"
+ onClick={() => selected && setUpdateTarget(selected.quant)}
+ aria-label={`Update ${repoId}`}
+ className="hub-action-btn ml-1 text-amber-700 dark:text-amber-300"
+ >
+ <HugeiconsIcon
+ icon={ArrowReloadHorizontalIcon}
+ strokeWidth={1.75}
+ />
+ Update
+ </button>
+ )}
- <button
- type="button"
- disabled={downloadAction.disabled}
- onClick={() => {
- if (downloadingThisVariant) {
- downloadAction.onClick();
- return;
- }
- if (selectedIsActive) {
- onEject?.();
- return;
- }
- if (!selected) return;
- if (selected.downloaded) {
- onLoad({
+ {showDownloadAction && (
+ <button
+ type="button"
+ disabled={downloadAction.disabled}
+ onClick={downloadAction.onClick}
+ aria-label={downloadAction.ariaLabel}
+ className={cn(
+ "hub-action-btn w-24",
+ ctaDisabled &&
+ !downloadingThisVariant &&
+ !cancelling &&
+ "opacity-70",
+ (cancelling || downloadAction.starting) && "opacity-70",
+ downloadingThisVariant &&
+ !cancelling &&
+ "hover:bg-rose-500/10 hover:text-rose-600 dark:hover:text-rose-400",
+ )}
+ >
+ {cancelling ? (
+ <span className="inline-flex items-center gap-2 text-muted-foreground">
+ <Spinner />
+ Cancelling…
+ </span>
+ ) : downloadingThisVariant ? (
+ <span className="inline-flex items-center gap-2">
+ <DownloadStopIndicator mode={downloadAction.stopMode} />
+ {downloadAction.progressPercent != null
+ ? `${downloadAction.progressPercent}%`
+ : null}
+ </span>
+ ) : downloadAction.starting ? (
+ <span className="inline-flex items-center gap-2">
+ <Spinner />
+ Starting…
+ </span>
+ ) : (
+ <>
+ <HugeiconsIcon icon={Download01Icon} strokeWidth={1.75} />
+ {downloadAction.downloadLabel}
+ </>
+ )}
+ </button>
+ )}
+
+ {showRunAction && onRun && selected && (
+ <ModelRunActionButton
+ label={`Configure and run ${repoId} ${selectedLabel ?? selected.quant}`}
+ loading={runPending}
+ onClick={() =>
+ onRun({
ggufVariant: selected.quant,
- expectedBytes: selected.size_bytes,
- });
- } else {
- downloadAction.onClick();
+ ggufFilename: selected.filename,
+ expectedBytes:
+ selected.download_size_bytes &&
+ selected.download_size_bytes > 0
+ ? selected.download_size_bytes
+ : selected.size_bytes,
+ })
}
- }}
- aria-label={downloadAction.ariaLabel}
- className={cn(
- isGgufRunCta ? "hub-run-action-btn w-24" : "hub-action-btn w-24",
- isGgufRunCta && "ml-2",
- ctaDisabled &&
- !selectedIsActive &&
- !downloadingThisVariant &&
- !cancelling &&
- "opacity-70",
- (cancelling || downloadAction.starting) && "opacity-70",
- downloadingThisVariant &&
- !cancelling &&
- "hover:bg-rose-500/10 hover:text-rose-600 dark:hover:text-rose-400",
- !HUB_GGUF_RUN_ACTIONS_VISIBLE &&
- !downloadingThisVariant &&
- !cancelling &&
- !isLoadingThisModel &&
- (selectedIsActive || selected?.downloaded) &&
- "hidden",
- )}
- >
- {cancelling ? (
- <span className="inline-flex items-center gap-2 text-muted-foreground">
- <Spinner />
- Cancelling…
- </span>
- ) : downloadingThisVariant ? (
- <span className="inline-flex items-center gap-2">
- <DownloadStopIndicator mode={downloadAction.stopMode} />
- {downloadAction.progressPercent != null
- ? `${downloadAction.progressPercent}%`
- : null}
- </span>
- ) : downloadAction.starting ? (
- <span className="inline-flex items-center gap-2">
- <Spinner />
- Starting…
- </span>
- ) : isLoadingThisModel ? (
- <span className="inline-flex items-center gap-2">
- <Spinner />
- Loading…
- </span>
- ) : selectedIsActive ? (
- <>
- <HugeiconsIcon icon={RemoveCircleIcon} strokeWidth={1.75} />
- Eject
- </>
- ) : selected?.downloaded ? (
- <>
- <HugeiconsIcon icon={PlayIcon} strokeWidth={1.75} />
- Run
- </>
- ) : (
- <>
- <HugeiconsIcon icon={Download01Icon} strokeWidth={1.75} />
- {downloadAction.downloadLabel}
- </>
- )}
- </button>
+ />
+ )}
</DownloadCard>
{/* Only a quant actually on disk gets charted: an undownloaded one has no
weights to measure, and the fit badge already tiers those. */}
diff --git a/studio/frontend/src/features/hub/catalog/hub-model-settings-view.tsx b/studio/frontend/src/features/hub/catalog/hub-model-settings-view.tsx
deleted file mode 100644
index 8802447bef9..00000000000
--- a/studio/frontend/src/features/hub/catalog/hub-model-settings-view.tsx
+++ /dev/null
@@ -1,150 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-// Full-page settings for one model, opened from the Hub. The same controls exist in the
-// chat picker's popover, but a popover is a poor place to work through every knob.
-
-import {
- ModelConfigPage,
- type ModelPickTarget,
- modelConfigInstanceKey,
-} from "@/features/model-picker";
-import type { PerModelConfig } from "@/features/model-picker";
-import { cn } from "@/lib/utils";
-import {
- Globe02Icon,
-} from "@hugeicons/core-free-icons";
-import {
- ChevronLeftIcon,
-} from "lucide-react";
-import { HugeiconsIcon } from "@hugeicons/react";
-import { useEffect, useRef, useState } from "react";
-
-export function HubModelSettingsView({
- target,
- loadedConfig = null,
- loadedContextLength = null,
- onBack,
- onRun,
- compact = false,
-}: {
- target: ModelPickTarget;
- /** Non-null when this model is loaded, so the page can show live values. */
- loadedConfig?: PerModelConfig | null;
- loadedContextLength?: number | null;
- onBack: () => void;
- /** Apply + load with these settings. */
- onRun: (config: PerModelConfig) => void;
- compact?: boolean;
-}) {
- const scrollRef = useRef<HTMLDivElement | null>(null);
- const [scrolled, setScrolled] = useState(false);
- // Mirrors HubDetailView so this view sits at the Hub's measure.
- const measure = compact
- ? "mx-auto w-full max-w-[var(--hub-measure-compact)] px-5 sm:px-5"
- : "mx-auto w-full max-w-[var(--hub-measure)] px-5 sm:px-8";
-
- useEffect(() => {
- const el = scrollRef.current;
- if (!el) return;
- const onScroll = () => {
- const next = el.scrollTop > 0;
- setScrolled((current) => (current === next ? current : next));
- };
- onScroll();
- el.addEventListener("scroll", onScroll, { passive: true });
- return () => el.removeEventListener("scroll", onScroll);
- }, []);
-
- return (
- <div className="relative flex min-h-0 flex-1 flex-col">
- <div
- aria-hidden="true"
- data-scrolled={scrolled || undefined}
- className="hub-scroll-fade pointer-events-none absolute inset-x-0 top-0 z-10 h-7"
- />
- <div
- ref={scrollRef}
- data-hub-scroll="true"
- className={cn(
- "min-h-0 flex-1 overflow-x-hidden overflow-y-auto [overflow-anchor:none] [scrollbar-width:thin]",
- compact
- ? "mr-2 [scrollbar-gutter:stable]"
- : "[scrollbar-gutter:stable_both-edges]",
- )}
- >
- <div
- className="hub-detail-bar sticky top-0 z-20"
- data-scrolled={scrolled || undefined}
- >
- <div className={`${measure} py-3`}>
- <button
- type="button"
- onClick={onBack}
- className="-ml-1.5 inline-flex h-8 cursor-pointer select-none items-center gap-1.5 rounded-full pl-1.5 pr-2.5 text-ui-12p5 font-medium text-muted-foreground transition-colors hover:bg-foreground/[0.05] hover:text-foreground dark:hover:bg-white/[0.06]"
- >
- <ChevronLeftIcon
- strokeWidth={1.75}
- className="size-3.5"
- />
- Back to Hub
- </button>
- </div>
- </div>
-
- <div className={cn(measure, "pb-20")}>
- <header className="flex flex-col gap-1 pb-5">
- <h1 className="min-w-0 break-words text-ui-24 font-semibold leading-[1.1] tracking-[-0.022em] text-foreground">
- {target.displayName}
- </h1>
- <p className="min-w-0 break-all text-ui-12 text-muted-foreground">
- {target.id}
- {target.ggufVariant ? ` · ${target.ggufVariant}` : ""}
- </p>
- </header>
-
- <div className="mb-5 flex items-start gap-2.5 rounded-xl border border-border/60 bg-card px-4 py-3">
- <span className="mt-0.5 flex size-7 shrink-0 items-center justify-center rounded-lg border border-border/60 bg-muted/40">
- <HugeiconsIcon
- icon={Globe02Icon}
- strokeWidth={1.75}
- className="size-3.5"
- />
- </span>
- <p className="min-w-0 text-ui-12 leading-[1.5] text-muted-foreground">
- {/* Only what auto-switch reaches: it indexes GGUFs and skips Ollama. */}
- {(target.apiLoadable ?? target.isGguf)
- ? "Saved settings apply everywhere this model loads, including when an OpenAI-compatible API request asks for it."
- : "Saved settings apply everywhere Unsloth loads this model."}{" "}
- Turn on{" "}
- <span className="font-medium text-foreground">
- Remember for this model
- </span>{" "}
- below to keep them.
- </p>
- </div>
-
- <div className="rounded-xl border border-border/60 bg-card px-4 py-4">
- <ModelConfigPage
- // Keyed on the live config too, like the sidebar entry: ModelConfigPage
- // reads loadedConfig once, so opening before status hydrates would leave
- // the editor on saved values that Apply then writes back over.
- key={modelConfigInstanceKey(
- target.id,
- target.ggufVariant,
- loadedConfig,
- )}
- target={target}
- onRun={onRun}
- loadedConfig={loadedConfig}
- loadedContextLength={loadedContextLength}
- variant="page"
- // The page heading already names the model; "Run settings" would repeat it.
- showHeader={false}
- />
- </div>
- </div>
- </div>
- </div>
- );
-}
diff --git a/studio/frontend/src/features/hub/catalog/local-dataset-card.tsx b/studio/frontend/src/features/hub/catalog/local-dataset-card.tsx
index a4e6e5c9f5a..76f7ef4aef7 100644
--- a/studio/frontend/src/features/hub/catalog/local-dataset-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/local-dataset-card.tsx
@@ -1,9 +1,6 @@
// SPDX-License-Identifier: AGPL-3.0-only
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-import { TrainIcon } from "@/features/hub/components/train-icon";
-import { HUB_POST_DOWNLOAD_ACTIONS_VISIBLE } from "@/features/hub/lib/hub-feature-flags";
-import { HugeiconsIcon } from "@hugeicons/react";
import type { LocalSource } from "../types";
import { DotTag } from "./dot-tag";
import { PathInfoButton } from "./path-info-button";
@@ -12,14 +9,12 @@ interface LocalDatasetCardProps {
sourceLabel: string;
source: LocalSource;
path: string;
- onTrain?: () => void;
}
export function LocalDatasetCard({
sourceLabel,
source,
path,
- onTrain,
}: LocalDatasetCardProps) {
return (
<div className="hub-download-card">
@@ -37,23 +32,6 @@ export function LocalDatasetCard({
<PathInfoButton path={path} />
</div>
</div>
- {HUB_POST_DOWNLOAD_ACTIONS_VISIBLE && (
- <div
- aria-hidden="true"
- className="ml-1 mr-0 h-5 w-px shrink-0 bg-foreground/[0.06] opacity-100 transition-opacity duration-150 group-hover/dl:opacity-0 dark:bg-white/[0.04]"
- />
- )}
- {/* Train CTA hidden until Hub→train picker ships. */}
- {onTrain && HUB_POST_DOWNLOAD_ACTIONS_VISIBLE && (
- <button
- type="button"
- onClick={() => onTrain()}
- className="hub-action-btn w-28"
- >
- <HugeiconsIcon icon={TrainIcon} strokeWidth={1.75} />
- Train
- </button>
- )}
</div>
</div>
);
diff --git a/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx b/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx
index 9a66c9366c5..28da3bd5408 100644
--- a/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/local-on-device-card.tsx
@@ -2,30 +2,22 @@
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
import { ModelMemoryBarFor } from "@/components/model-memory-bar";
-import { useVramBudgetFraction } from "@/hooks/use-vram-budget-fraction";
import {
Popover,
PopoverContent,
PopoverTrigger,
} from "@/components/ui/popover";
-import { Spinner } from "@/components/ui/spinner";
import {
Tooltip,
TooltipContent,
TooltipTrigger,
} from "@/components/ui/tooltip";
+import { useVramBudgetFraction } from "@/hooks/use-vram-budget-fraction";
import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
import { cn } from "@/lib/utils";
-import {
- Alert02Icon,
- CubeIcon,
- PlayIcon,
- RemoveCircleIcon,
- Share05Icon,
-} from "@hugeicons/core-free-icons";
+import { Alert02Icon, CubeIcon, Share05Icon } from "@hugeicons/core-free-icons";
import { HugeiconsIcon } from "@hugeicons/react";
import { useCallback, useMemo, useState } from "react";
-import { TrainIcon } from "../components/train-icon";
import {
downloadManager,
jobKeyOf,
@@ -40,6 +32,7 @@ import {
deleteCachedModel,
} from "../inventory";
import { formatBytes } from "../lib/format";
+import type { HubModelRunSelection } from "../lib/model-run-selection";
import {
ggufFilenamesMatch,
@@ -47,37 +40,30 @@ import {
} from "../lib/gguf-filename";
import {
ggufVariantDisplayLabel,
+ resolveLocalGgufVariant,
sortLocalGgufVariants,
} from "../lib/gguf-variant-sort";
-import {
- HUB_GGUF_RUN_ACTIONS_VISIBLE,
- HUB_NON_GGUF_RUN_ACTIONS_VISIBLE,
- HUB_POST_DOWNLOAD_ACTIONS_VISIBLE,
-} from "../lib/hub-feature-flags";
import { ggufVariantsMatch } from "../lib/model-identity";
import { confirmExternalLink } from "../stores/external-link-confirm";
import { useHfTokenStore } from "../stores/hf-token-store";
+import { DeleteImpactSummary, useDeleteImpact } from "./delete-impact";
import { DotTag } from "./dot-tag";
import {
CardDeleteButton,
- CardSettingsButton,
+ CardDivider,
CardUpdateButton,
DeleteConfirmDialog,
+ ModelRunActionButton,
UpdateConfirmDialog,
} from "./download-card";
import { PathInfoButton } from "./path-info-button";
import { TransportConflictDialog } from "./transport-conflict-dialog";
-import { DeleteImpactSummary, useDeleteImpact } from "./delete-impact";
import { useCardDelete } from "./use-card-delete";
import { useGgufVariantFetchState } from "./use-gguf-variant-fetch-state";
-type LocalLoadOptions = {
- ggufVariant?: string;
- expectedBytes?: number;
-};
-
interface LocalOnDeviceCardProps {
modelId: string;
+ displayName: string;
repoId: string | null;
sourceLabel: string;
source: LocalModelInfo["source"];
@@ -96,11 +82,9 @@ interface LocalOnDeviceCardProps {
baseModelSummary?: string | null;
adapterType?: string | null;
trainingMethod?: string | null;
- canRun?: boolean;
isActive: boolean;
activeGgufVariant?: string | null;
isLoading: boolean;
- loadingPhase?: "downloading" | "starting";
preferredFile?: string | null;
preferredFileIntent?: number;
@@ -109,21 +93,9 @@ interface LocalOnDeviceCardProps {
gpuCount?: number;
systemRamGb?: number;
unsupportedReason?: string | null;
- onLoad: (opts?: LocalLoadOptions) => void;
- /** Accepted for API parity; the run bar ejects instead of opening chat. */
- onUseInChat: () => void;
- onEject?: () => void;
- onTrain?: () => void;
+ onRun?: (selection: HubModelRunSelection) => void;
+ runPending?: boolean;
onChange?: () => void;
- /**
- * Open settings for the quant this card is showing. ``quantIsUserPicked`` says whether
- * it came from this card's selector or was derived from the resident model, which
- * decides whether a fresher status read may override it.
- */
- onOpenSettings?: (
- ggufVariant: string | null,
- quantIsUserPicked: boolean,
- ) => void;
}
function formatAdapterLabel(
@@ -217,6 +189,7 @@ function BaseModelReference({
export function LocalOnDeviceCard({
modelId,
+ displayName,
repoId,
sourceLabel,
source,
@@ -231,11 +204,9 @@ export function LocalOnDeviceCard({
baseModelSummary,
adapterType,
trainingMethod,
- canRun = true,
isActive,
activeGgufVariant = null,
isLoading,
- loadingPhase,
preferredFile = null,
preferredFileIntent = 0,
@@ -243,11 +214,9 @@ export function LocalOnDeviceCard({
gpuCount,
systemRamGb,
unsupportedReason,
- onLoad,
- onEject,
- onTrain,
+ onRun,
+ runPending = false,
onChange,
- onOpenSettings,
}: LocalOnDeviceCardProps) {
const [deleteOpen, setDeleteOpen] = useState(false);
const [updateOpen, setUpdateOpen] = useState(false);
@@ -278,7 +247,10 @@ export function LocalOnDeviceCard({
// Update availability is derived from the GGUF variant metadata; offline rows
// keep the button hidden because there is no remote revision to fetch.
const online = useOnlineStatus();
- const deleteImpact = useDeleteImpact(deleteOpen && Boolean(repoId), repoId ?? "");
+ const deleteImpact = useDeleteImpact(
+ deleteOpen && Boolean(repoId),
+ repoId ?? "",
+ );
const { deleting, runDelete } = useCardDelete({
action: async () => {
if (!repoId) return;
@@ -324,7 +296,11 @@ export function LocalOnDeviceCard({
}));
const canDelete =
- source === "hf_cache" && !!repoId && !isActive && !isLoading;
+ source === "hf_cache" &&
+ !!repoId &&
+ !isActive &&
+ !isLoading &&
+ !runPending;
const variants = useMemo(() => {
const localVariants = currentVariantState.variants;
const remoteVariants = remoteVariantState.variants;
@@ -357,7 +333,6 @@ export function LocalOnDeviceCard({
variants
? sortLocalGgufVariants(variants, {
defaultVariant: currentVariantState.defaultVariant,
- activeGgufVariant: isActive ? activeGgufVariant : null,
gpuGb,
gpuCount,
systemRamGb,
@@ -367,8 +342,6 @@ export function LocalOnDeviceCard({
[
variants,
currentVariantState.defaultVariant,
- isActive,
- activeGgufVariant,
gpuGb,
gpuCount,
systemRamGb,
@@ -390,32 +363,12 @@ export function LocalOnDeviceCard({
)
? selectedVariantState.quant
: preferredQuant;
- const selectedQuant =
- selectedVariantOverride &&
- sortedVariants?.some((variant) =>
- ggufVariantsMatch(variant.quant, selectedVariantOverride),
- )
- ? selectedVariantOverride
- : (sortedVariants?.find(
- (variant) =>
- isActive && ggufVariantsMatch(variant.quant, activeGgufVariant),
- )?.quant ??
- sortedVariants?.find((variant) =>
- ggufVariantsMatch(variant.quant, currentVariantState.defaultVariant),
- )?.quant ??
- sortedVariants?.[0]?.quant ??
- null);
- // Only the first branch is a choice; the rest read the store, which can be stale.
- const quantIsUserPicked = Boolean(
- selectedVariantOverride &&
- sortedVariants?.some((variant) =>
- ggufVariantsMatch(variant.quant, selectedVariantOverride),
- ),
- );
- const selectedVariant =
- sortedVariants?.find((variant) =>
- ggufVariantsMatch(variant.quant, selectedQuant),
- ) ?? null;
+ const selectedVariant = resolveLocalGgufVariant(sortedVariants, {
+ selectedVariant: selectedVariantOverride,
+ activeVariant: isActive ? activeGgufVariant : null,
+ defaultVariant: currentVariantState.defaultVariant,
+ });
+ const selectedQuant = selectedVariant?.quant ?? null;
// True while a managed download/update for this repo+variant is in flight.
const updateJobActive = useDownloadManagerStore((s) =>
repoId
@@ -442,6 +395,7 @@ export function LocalOnDeviceCard({
!!repoId &&
!isActive &&
!isLoading &&
+ !runPending &&
!updateJobActive &&
updateAvailable;
// Update runs as a MANAGED download (same path as a normal download) so it
@@ -466,18 +420,6 @@ export function LocalOnDeviceCard({
void remoteVariantState.refresh();
});
};
- const selectedVariantIsActive =
- needsVariantSelection && selectedQuant
- ? isActive && ggufVariantsMatch(activeGgufVariant, selectedQuant)
- : isActive;
- const variantUnavailable =
- needsVariantSelection &&
- (currentVariantState.loading ||
- currentVariantState.error !== null ||
- selectedVariant === null);
- const variantActionPending =
- needsVariantSelection && currentVariantState.loading;
-
const formatLabel =
modelFormat === "gguf"
? "GGUF"
@@ -491,9 +433,11 @@ export function LocalOnDeviceCard({
const formatTone =
modelFormat === "adapter" ? "adapter" : isGguf ? "gguf" : "checkpoint";
const showOldCacheHint = source === "hf_cache" && !!unsupportedReason;
- const runActionsVisible = isGguf
- ? HUB_GGUF_RUN_ACTIONS_VISIBLE
- : HUB_NON_GGUF_RUN_ACTIONS_VISIBLE;
+ const selectedVariantReady =
+ !needsVariantSelection ||
+ (selectedVariant?.downloaded === true && selectedVariant.partial !== true);
+ const showRunAction =
+ Boolean(onRun) && selectedVariantReady && !updateJobActive && !isLoading;
return (
<div className="flex w-full flex-col gap-2">
@@ -515,17 +459,19 @@ export function LocalOnDeviceCard({
<div className="group/dl flex items-center">
<div className="relative flex h-9 min-w-0 flex-1 items-center pl-3 pr-2">
<span className="flex min-w-0 items-center gap-1.5 text-ui-12 text-muted-foreground">
- <DotTag
- tone="success"
- label={selectedVariantIsActive ? "Loaded" : "On device"}
- />
+ <DotTag tone="success" label="On device" />
<DotTag tone={formatTone} label={formatLabel} />
{needsVariantSelection && (
- <Popover open={variantOpen} onOpenChange={setVariantOpen}>
+ <Popover
+ open={runPending ? false : variantOpen}
+ onOpenChange={(nextOpen) => {
+ if (!runPending) setVariantOpen(nextOpen);
+ }}
+ >
<PopoverTrigger asChild={true}>
<button
type="button"
- disabled={currentVariantState.loading}
+ disabled={currentVariantState.loading || runPending}
className="inline-flex h-6 max-w-[170px] shrink-0 cursor-pointer items-center gap-1.5 rounded-[8px] border border-format-gguf/35 px-2 font-mono text-ui-10p5 leading-none text-format-gguf transition-colors hover:bg-format-gguf/8 disabled:cursor-not-allowed disabled:opacity-60"
>
<span className="truncate">
@@ -562,9 +508,6 @@ export function LocalOnDeviceCard({
variant.quant,
selectedQuant,
);
- const isLoaded =
- ggufVariantsMatch(variant.quant, activeGgufVariant) &&
- isActive;
return (
<button
key={variant.filename}
@@ -590,9 +533,6 @@ export function LocalOnDeviceCard({
{label}
</span>
<span className="flex shrink-0 items-center gap-1.5">
- {isLoaded && (
- <DotTag tone="success" label="Loaded" />
- )}
<span className="text-ui-10 text-muted-foreground tabular-nums">
{formatBytes(variant.size_bytes)}
</span>
@@ -623,15 +563,6 @@ export function LocalOnDeviceCard({
)}
</span>
<div className="ml-auto flex items-center gap-0.5">
- {onOpenSettings && (
- <CardSettingsButton
- label={`Settings for ${repoId}`}
- // The quant this card resolved, so settings edits what is on screen.
- onClick={() =>
- onOpenSettings(selectedQuant ?? null, quantIsUserPicked)
- }
- />
- )}
{canUpdate && (
<CardUpdateButton
label={`Update ${repoId}`}
@@ -648,103 +579,42 @@ export function LocalOnDeviceCard({
<PathInfoButton path={path} />
</div>
</div>
- {onTrain && HUB_POST_DOWNLOAD_ACTIONS_VISIBLE && (
- <div
- aria-hidden="true"
- className="ml-1 mr-0 h-5 w-px shrink-0 bg-foreground/[0.06] opacity-100 transition-opacity duration-150 group-hover/dl:opacity-0 dark:bg-white/[0.04]"
- />
- )}
- <div
- className={cn(
- "group/pair flex h-9 shrink-0 items-stretch gap-1.5",
- !runActionsVisible && "hidden",
- )}
- >
- {onTrain && HUB_POST_DOWNLOAD_ACTIONS_VISIBLE && (
- <button
- type="button"
- onClick={() => onTrain()}
- className="hub-action-btn w-24"
- >
- <HugeiconsIcon icon={TrainIcon} strokeWidth={1.75} />
- Train
- </button>
- )}
- <button
- type="button"
- disabled={isLoading || variantUnavailable || !canRun}
- onClick={() => {
- if (!canRun) return;
- if (selectedVariantIsActive) {
- onEject?.();
- return;
- }
- if (needsVariantSelection) {
- if (!selectedVariant) return;
- onLoad({
- ggufVariant: selectedVariant.quant,
- expectedBytes: selectedVariant.size_bytes,
- });
- return;
+ {showRunAction && onRun && (
+ <>
+ <CardDivider />
+ <ModelRunActionButton
+ label={`Configure and run ${displayName.trim() || "this model"}`}
+ loading={runPending}
+ onClick={() =>
+ onRun(
+ needsVariantSelection && selectedVariant
+ ? {
+ ggufVariant: selectedVariant.quant,
+ ggufFilename: selectedVariant.filename,
+ expectedBytes:
+ selectedVariant.download_size_bytes &&
+ selectedVariant.download_size_bytes > 0
+ ? selectedVariant.download_size_bytes
+ : selectedVariant.size_bytes,
+ }
+ : {},
+ )
}
- onLoad();
- }}
- className={cn(
- isLoading ||
- selectedVariantIsActive ||
- variantUnavailable ||
- !canRun
- ? "hub-action-btn w-24"
- : "hub-run-action-btn w-24",
- (isLoading || variantUnavailable || !canRun) && "opacity-70",
- )}
- >
- {isLoading ? (
- <>
- <Spinner />
- {loadingPhase === "downloading" ? "Preparing…" : "Loading…"}
- </>
- ) : selectedVariantIsActive ? (
- <>
- <HugeiconsIcon icon={RemoveCircleIcon} strokeWidth={1.75} />
- Eject
- </>
- ) : variantActionPending ? (
- <>
- <Spinner />
- Loading…
- </>
- ) : canRun ? (
- <>
- <HugeiconsIcon icon={PlayIcon} strokeWidth={1.75} />
- Run
- </>
- ) : (
- <>
- <HugeiconsIcon icon={Alert02Icon} strokeWidth={1.75} />
- No run
- </>
- )}
- </button>
- </div>
+ />
+ </>
+ )}
</div>
- {/* Below the action row, not inside it: the row is a horizontal flex
- container, so a full-width bar there becomes another flex item and
- squeezes the Run/Train buttons. */}
- {/* A direct .gguf path skips variant selection entirely, so selectedQuant
- is null for exactly the local files this is meant to cover. The path
- names the weights on its own there, and the backend resolves a
- direct file without needing a quant to match. */}
+ {/* Render the full-width memory bar below the horizontal card row so it
+ does not compete with row controls for width. */}
+ {/* A direct .gguf path skips variant selection, so selectedQuant is null.
+ The path itself still supplies the local artifact identity. */}
{showMemoryBar &&
(repoId || localGgufPath) &&
(selectedQuant || localGgufPath.toLowerCase().endsWith(".gguf")) ? (
<ModelMemoryBarFor
- // The card's own path is what Run opens, so it is the identity the
- // estimate has to use: a repo cached under several roots or revisions
- // can otherwise resolve a different snapshot and chart weights the
- // click does not load. It is also the only identity a custom or local
- // GGUF has, where repoId is null and the bar used to be suppressed
- // outright despite a perfectly good path being right here.
+ // The card's exact path identifies the downloaded artifact across
+ // cache roots and revisions. It is also the only identity available
+ // for a custom or local GGUF when repoId is null.
repoId={repoId || localGgufPath}
loadId={localGgufPath}
quant={selectedQuant ?? ""}
diff --git a/studio/frontend/src/features/hub/catalog/model-download-state.ts b/studio/frontend/src/features/hub/catalog/model-download-state.ts
new file mode 100644
index 00000000000..bebfce38a8d
--- /dev/null
+++ b/studio/frontend/src/features/hub/catalog/model-download-state.ts
@@ -0,0 +1,22 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { SelectedModelView } from "../types";
+
+export type ModelDownloadState = {
+ isDownloaded: boolean;
+ isPartial: boolean;
+ partialTransport: string | null;
+ partialResumable: boolean;
+};
+
+export function modelDownloadState(
+ model: SelectedModelView,
+): ModelDownloadState {
+ return {
+ isDownloaded: model.isDownloaded,
+ isPartial: model.isPartial ?? false,
+ partialTransport: model.partialTransport ?? null,
+ partialResumable: model.partialResumable === true,
+ };
+}
diff --git a/studio/frontend/src/features/hub/catalog/model-inspector.tsx b/studio/frontend/src/features/hub/catalog/model-inspector.tsx
index 0265ca9765a..e46203b0b03 100644
--- a/studio/frontend/src/features/hub/catalog/model-inspector.tsx
+++ b/studio/frontend/src/features/hub/catalog/model-inspector.tsx
@@ -18,6 +18,7 @@ import {
formatShortDate,
} from "@/features/hub/lib/format";
import { useHfTokenStore } from "@/features/hub/stores/hf-token-store";
+import { taskForMediaPick } from "@/features/model-picker/components/model-selector/audio-picker-policy";
import { Tick02Icon } from "@/lib/tick-icon";
import { cn, formatCompact } from "@/lib/utils";
import {
@@ -43,6 +44,11 @@ import { memo, useDeferredValue, useMemo } from "react";
import { selectActiveJob, useDownloadManagerStore } from "../download-manager";
import { useCopyFeedback } from "../hooks/use-copy-feedback";
import { useDatasetSize } from "../hooks/use-dataset-size";
+import {
+ type HubModelRunSelection,
+ isHubModelRunEligible,
+} from "../lib/model-run-selection";
+import { studioPageForTask } from "../lib/unsloth-support";
import {
formatLibrary,
formatLocalUpdated,
@@ -52,23 +58,14 @@ import {
import { confirmExternalLink } from "../stores/external-link-confirm";
import type { SelectedModelView } from "../types";
import { DatasetDownloadSection } from "./dataset-download-section";
-import { taskForMediaPick } from "@/features/model-picker/components/model-selector/audio-picker-policy";
-import { routableToMediaPage } from "../lib/local-path";
-import { studioPageForTask } from "../lib/unsloth-support";
import { DownloadSection } from "./download-section";
import { LocalDatasetCard } from "./local-dataset-card";
import { LocalOnDeviceCard } from "./local-on-device-card";
+import { modelDownloadState } from "./model-download-state";
import { ModelReadme } from "./model-readme";
import { OwnerAvatar } from "./owner-avatar";
import { AccessChip, CapabilityPill } from "./shared";
-// HF pipeline_tag values authoritative for embedding-only repos; capability
-// labels (code/vision/audio) can leak onto them via name or tags.
-const EMBEDDING_PIPELINE_TAGS: ReadonlySet<string> = new Set([
- "feature-extraction",
- "sentence-similarity",
-]);
-
function ViewRepositoryButton({
repoId,
isDataset,
@@ -389,7 +386,6 @@ export type ModelInspectorRuntime = {
isActive: boolean;
activeGgufVariant: string | null;
isLoadingThisModel: boolean;
- loadingPhase?: "downloading" | "starting";
minMemory: string | null;
vramInfo: {
est: number;
@@ -402,18 +398,13 @@ export type ModelInspectorRuntime = {
};
export type ModelInspectorActions = {
- onLoad: (opts: { ggufVariant?: string; expectedBytes?: number }) => void;
- onLoadLocal: (opts?: {
- ggufVariant?: string;
- expectedBytes?: number;
- }) => void;
- onUseInChat: () => void;
- onEject?: () => void;
- onTrain?: () => void;
onInventoryChange?: () => void;
onSearchHub?: (query: string) => void;
- /** Open settings with the quant the card resolved. */
- onOpenSettings?: (ggufVariant: string | null) => void;
+ onRun?: (
+ selection: HubModelRunSelection,
+ mediaPage: ReturnType<typeof studioPageForTask>,
+ ) => void;
+ runConfigPending?: boolean;
};
export const ModelInspector = memo(function ModelInspector({
@@ -432,7 +423,6 @@ export const ModelInspector = memo(function ModelInspector({
metadataUnavailable?: boolean;
selectionHiddenByFilters?: boolean;
preferredGgufFile?: string | null;
-
preferredGgufFileIntent?: number;
runtime: ModelInspectorRuntime;
actions: ModelInspectorActions;
@@ -441,7 +431,6 @@ export const ModelInspector = memo(function ModelInspector({
isActive,
activeGgufVariant,
isLoadingThisModel,
- loadingPhase,
minMemory,
vramInfo,
gpuGb,
@@ -449,17 +438,16 @@ export const ModelInspector = memo(function ModelInspector({
systemRamGb,
} = runtime;
const {
- onLoad,
- onLoadLocal,
- onUseInChat,
- onEject,
- onTrain,
onInventoryChange,
onSearchHub,
- onOpenSettings,
+ onRun,
+ runConfigPending = false,
} = actions;
const deviceType = usePlatformStore((s) => s.deviceType);
const chatOnly = usePlatformStore((s) => s.isChatOnly());
+ const chatOnlyMeasured = usePlatformStore(
+ (s) => s.isChatOnly() && !s.capabilitiesUnknown(),
+ );
const hfToken = useHfTokenStore((s) => s.token);
const datasetRepoId = isDataset && model?.hubRepoId ? model.hubRepoId : null;
const datasetSize = useDatasetSize(datasetRepoId, {
@@ -527,13 +515,14 @@ export const ModelInspector = memo(function ModelInspector({
<p className="max-w-sm text-ui-12p5 leading-5 text-muted-foreground">
{isDataset
? "Choose a dataset from the catalog to inspect its download state and details."
- : "Choose an item from the catalog to inspect its runtime fit, download state, and model card."}
+ : "Choose an item from the catalog to inspect its hardware fit, download state, and model card."}
</p>
</div>
</div>
);
}
+ const downloadState = modelDownloadState(model);
const updatedRaw = model.updatedAt
? formatRelativeShort(model.updatedAt)
: formatLocalUpdated(model.localUpdatedAt);
@@ -560,55 +549,23 @@ export const ModelInspector = memo(function ModelInspector({
const paramsLabel = model.totalParams
? formatCompact(model.totalParams)
: "N/A";
- const unslothSupported = unslothSupport.status !== "unsupported";
- // Embedding-only non-GGUF repos have no generative head, so keep them out of
- // the Run gate. Prefer the pipeline tag, else the capability heuristic.
- const isEmbeddingOnly =
- !model.isGguf &&
- model.capabilities.some((c) => c.key === "embedding") &&
- (EMBEDDING_PIPELINE_TAGS.has(model.pipelineTag?.toLowerCase() ?? "") ||
- !model.capabilities.some(
- (c) =>
- c.key === "conversational" ||
- c.key === "tools" ||
- c.key === "reasoning" ||
- c.key === "code" ||
- c.key === "vision" ||
- c.key === "audio",
- ));
- // An image / video model runs on its own page, which onLoad already routes to;
- // the chat gates below would leave it greyed out as if it were unusable. Only when the
- // row is one onLoad can actually route there: those pages resolve a routed `model` as a
- // Hub id, so a filesystem row fails routableToMediaPage and the click falls through to
- // the chat loader, which unloads the resident model for a load that can only fail.
- // Whether this model's runtime is llama.cpp at all. Deliberately NOT
- // runsOnMediaPage below: that one also asks whether the click can be ROUTED to
- // the page, which is a different question. A diffusion GGUF on a filesystem row
- // is not routable and still does not load through llama.cpp, so a memory bar
- // there would describe the wrong runtime either way.
- const runsOnMediaRuntime =
- studioPageForTask(
- taskForMediaPick(model.pipelineTag, model.task) ?? undefined,
- ) !== undefined;
- const runsOnMediaPage =
- studioPageForTask(
- taskForMediaPick(model.pipelineTag, model.task) ?? undefined,
- ) !== undefined &&
- routableToMediaPage(model.kind, model.localSource);
- // Chat-only hosts (no supported GPU / usable MLX) run inference only through
- // llama.cpp, so only GGUF is loadable.
- const canRunModel =
- !isDataset &&
- (runsOnMediaPage ||
- ((model.runtimeCapabilities?.canChat ?? true) &&
- !isEmbeddingOnly &&
- (model.isGguf || (!chatOnly && unslothSupported))));
- const canTrainModel =
- !isDataset &&
- (model.runtimeCapabilities?.canTrain ?? false) &&
- model.modelFormat !== "gguf" &&
- model.modelFormat !== "adapter" &&
- unslothSupported;
+ // Media models use a separate runtime, so the llama.cpp memory estimate does
+ // not describe their load.
+ const mediaPage = studioPageForTask(
+ taskForMediaPick(model.pipelineTag, model.task) ?? undefined,
+ );
+ const runsOnMediaRuntime = mediaPage !== undefined;
+ const runEligible = isHubModelRunEligible({
+ model,
+ isDataset,
+ mediaRuntime: runsOnMediaRuntime,
+ nonGgufRuntimeAvailable:
+ !chatOnlyMeasured && unslothSupport.status !== "unsupported",
+ });
+ const runAction =
+ runEligible && onRun
+ ? (selection: HubModelRunSelection) => onRun(selection, mediaPage)
+ : undefined;
const languages = parseLanguageTags(model.tags);
const datasetSizeBytes =
@@ -687,13 +644,9 @@ export const ModelInspector = memo(function ModelInspector({
<InspectorDownloadSlot>
<DatasetDownloadSection
repoId={model.hubRepoId}
- isDownloaded={model.isDownloaded}
- isPartial={model.isPartial ?? false}
- partialTransport={model.partialTransport ?? null}
- partialResumable={model.partialResumable === true}
+ {...downloadState}
cachePath={model.path}
knownBytes={model.cachedBytes}
- onTrain={onTrain}
onChange={onInventoryChange}
/>
</InspectorDownloadSlot>
@@ -704,7 +657,6 @@ export const ModelInspector = memo(function ModelInspector({
sourceLabel={model.sourceLabel}
source={model.localSource ?? "custom"}
path={model.path}
- onTrain={onTrain}
/>
</InspectorDownloadSlot>
)}
@@ -714,6 +666,7 @@ export const ModelInspector = memo(function ModelInspector({
<LocalOnDeviceCard
showMemoryBar={!runsOnMediaRuntime}
modelId={model.id}
+ displayName={model.title}
repoId={model.hubRepoId}
sourceLabel={model.sourceLabel}
source={model.localSource ?? "custom"}
@@ -727,30 +680,23 @@ export const ModelInspector = memo(function ModelInspector({
baseModelSummary={model.baseModelSummary}
adapterType={model.adapterType}
trainingMethod={model.trainingMethod}
- canRun={canRunModel}
isActive={isActive}
activeGgufVariant={activeGgufVariant}
isLoading={isLoadingThisModel}
- loadingPhase={loadingPhase}
gpuGb={gpuGb}
gpuCount={gpuCount}
systemRamGb={systemRamGb}
-
preferredFile={preferredGgufFile}
preferredFileIntent={preferredGgufFileIntent}
unsupportedReason={
- unslothSupport.status === "unsupported" && !unslothSupport.supportedIn
+ unslothSupport.status === "unsupported" &&
+ !unslothSupport.supportedIn
? (unslothSupport.reason ?? "Unsupported format")
: null
}
- onLoad={onLoadLocal}
- onUseInChat={onUseInChat}
- onEject={onEject}
- onTrain={
- model.isDownloaded && canTrainModel ? onTrain : undefined
- }
+ onRun={runAction}
+ runPending={runConfigPending}
onChange={onInventoryChange}
- onOpenSettings={onOpenSettings}
/>
) : (
<DownloadSection
@@ -758,16 +704,11 @@ export const ModelInspector = memo(function ModelInspector({
mediaRuntime={runsOnMediaRuntime}
repoId={model.isLocal ? (model.hubRepoId ?? model.id) : model.id}
isGguf={model.isGguf}
- isDownloaded={model.isDownloaded}
- isPartial={model.isPartial ?? false}
- partialTransport={model.partialTransport ?? null}
- partialResumable={model.partialResumable === true}
+ {...downloadState}
modelFormat={model.modelFormat}
- canRun={canRunModel}
isActive={isActive}
activeQuant={isActive ? (activeGgufVariant ?? null) : null}
preferredGgufFile={preferredGgufFile}
-
preferredGgufFileIntent={preferredGgufFileIntent}
isLoadingThisModel={isLoadingThisModel}
gpuGb={gpuGb}
@@ -775,12 +716,8 @@ export const ModelInspector = memo(function ModelInspector({
systemRamGb={systemRamGb}
cachePath={model.path}
knownBytes={model.cachedBytes}
- onLoad={model.isLocal ? onLoadLocal : onLoad}
- onUseInChat={onUseInChat}
- onEject={onEject}
- onTrain={
- model.isDownloaded && canTrainModel ? onTrain : undefined
- }
+ onRun={runAction}
+ runPending={runConfigPending}
onChange={onInventoryChange}
/>
)}
diff --git a/studio/frontend/src/features/hub/catalog/models-catalog-lists.tsx b/studio/frontend/src/features/hub/catalog/models-catalog-lists.tsx
index e4402ca1145..2219e5daf7e 100644
--- a/studio/frontend/src/features/hub/catalog/models-catalog-lists.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-catalog-lists.tsx
@@ -267,15 +267,12 @@ export function DownloadedList({
onClearFilters,
scrollElement,
columns = 1,
- activeCheckpoint,
- activeGgufVariant,
isDataset,
inventoryTokens,
deviceType,
compact = false,
sort,
onInventoryChange,
- onOpenModelSettings,
}: {
cachedRows: CachedInventoryRow[];
localRows: LocalInventoryRow[];
@@ -288,8 +285,6 @@ export function DownloadedList({
onClearFilters?: () => void;
scrollElement: HTMLDivElement | null;
columns?: number;
- activeCheckpoint: string | null;
- activeGgufVariant: string | null;
isDataset: boolean;
inventoryTokens: readonly string[];
deviceType: string | null;
@@ -297,7 +292,6 @@ export function DownloadedList({
compact?: boolean;
sort: InventorySort;
onInventoryChange?: () => void;
- onOpenModelSettings?: (row: CachedInventoryRow | LocalInventoryRow) => void;
}) {
// Pinned repos surface first regardless of the active sort, which still orders within groups.
const pinnedIds = usePinnedModelsStore((s) => s.pinned);
@@ -400,15 +394,12 @@ export function DownloadedList({
<InventoryRow
row={item.row}
selected={selectedId === item.row.id}
- activeCheckpoint={activeCheckpoint}
- activeGgufVariant={activeGgufVariant}
isDataset={isDataset}
dimmed={!inventoryRowMatches(item.row, inventoryTokens)}
deviceType={deviceType}
compact={compact}
onSelect={onSelect}
onChange={onInventoryChange}
- onOpenSettings={onOpenModelSettings}
/>
);
diff --git a/studio/frontend/src/features/hub/catalog/models-catalog-rows.tsx b/studio/frontend/src/features/hub/catalog/models-catalog-rows.tsx
index cfa37450d0d..7ae5b1c319d 100644
--- a/studio/frontend/src/features/hub/catalog/models-catalog-rows.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-catalog-rows.tsx
@@ -21,7 +21,6 @@ import {
ggufVariantDisplayLabel,
useHfTokenStore,
} from "@/features/hub";
-import { modelIdsMatch } from "../lib/model-identity";
import {
ModelRowMenu,
pinKey,
@@ -241,7 +240,6 @@ export function StatChip({
function CatalogRow({
selected,
- active,
onClick,
tooltip,
label,
@@ -249,7 +247,6 @@ function CatalogRow({
variant = "flat",
}: {
selected: boolean;
- active?: boolean;
tooltip?: ReactNode;
onClick: () => void;
label: string;
@@ -262,7 +259,6 @@ function CatalogRow({
const button = (
<div
data-selected={selected || undefined}
- data-active={active || undefined}
onPointerEnter={arm}
onFocusCapture={arm}
className={cn(
@@ -423,14 +419,12 @@ export function buildRowStatusTooltip({
export const DiscoverModelRow = memo(function DiscoverModelRow({
row,
selected,
- active,
deviceType,
isDataset,
onSelect,
}: {
row: DiscoverRow;
selected: boolean;
- active: boolean;
deviceType: string | null;
isDataset: boolean;
onSelect: (id: string) => void;
@@ -467,7 +461,6 @@ export const DiscoverModelRow = memo(function DiscoverModelRow({
return (
<CatalogRow
selected={selected}
- active={active}
tooltip={tooltip}
label={row.repo}
onClick={handleClick}
@@ -538,47 +531,18 @@ export const DiscoverModelRow = memo(function DiscoverModelRow({
);
});
-function cachedRowActive(
- row: CachedInventoryRow,
- activeCheckpoint: string | null,
- activeGgufVariant: string | null,
-): boolean {
- if (!modelIdsMatch(activeCheckpoint, row.loadId)) return false;
- if (row.modelFormat === "gguf") {
- return row.capabilities.requiresVariant ? activeGgufVariant !== null : true;
- }
- return activeGgufVariant === null;
-}
-
-function localRowActive(
- row: LocalInventoryRow,
- activeCheckpoint: string | null,
- activeGgufVariant: string | null,
-): boolean {
- if (!modelIdsMatch(activeCheckpoint, row.loadId)) return false;
- if (row.modelFormat === "gguf") {
- return row.capabilities.requiresVariant ? activeGgufVariant !== null : true;
- }
- return activeGgufVariant === null;
-}
-
export const InventoryRow = memo(function InventoryRow({
row,
selected,
- activeCheckpoint,
- activeGgufVariant,
isDataset,
dimmed,
deviceType,
compact = false,
onSelect,
onChange,
- onOpenSettings,
}: {
row: CachedInventoryRow | LocalInventoryRow;
selected: boolean;
- activeCheckpoint: string | null;
- activeGgufVariant: string | null;
isDataset: boolean;
dimmed: boolean;
deviceType: string | null;
@@ -586,8 +550,6 @@ export const InventoryRow = memo(function InventoryRow({
compact?: boolean;
onSelect: (id: string) => void;
onChange?: () => void;
- /** Open this model's settings page. Omitted for datasets. */
- onOpenSettings?: (row: CachedInventoryRow | LocalInventoryRow) => void;
}) {
const rowModelId =
row.kind === "cache"
@@ -616,10 +578,6 @@ export const InventoryRow = memo(function InventoryRow({
deviceType,
]);
const handleClick = useCallback(() => onSelect(row.id), [onSelect, row.id]);
- const active =
- row.kind === "cache"
- ? cachedRowActive(row, activeCheckpoint, activeGgufVariant)
- : localRowActive(row, activeCheckpoint, activeGgufVariant);
const title = row.kind === "cache" ? row.repo : row.title;
const subLabel = row.owner;
@@ -735,18 +693,13 @@ export const InventoryRow = memo(function InventoryRow({
const rowPinned =
cacheDeletableRepoId != null &&
pinnedKeys.includes(pinKey(cacheDeletableRepoId));
- // Settings applies to any downloaded model, not just deletable ones, so the menu renders
- // when either action applies. `deletableRepoId` keeps the delete closures' narrowing.
- const settingsAction =
- !isDataset && onOpenSettings ? { onOpen: () => onOpenSettings(row) } : undefined;
const deletableRepoId = canDelete ? cacheDeletableRepoId : null;
const deleteAction =
- deletableRepoId || settingsAction ? (
+ deletableRepoId ? (
<ModelRowMenu
ariaLabel={`More options for ${deletableRepoId ?? rowModelId}`}
buttonClassName="pointer-events-auto hub-modal-pe-guard size-8 opacity-0 transition-opacity group-hover/row:opacity-100 focus-visible:opacity-100 data-[state=open]:opacity-100 [@media(pointer:coarse)]:opacity-100"
iconClassName="size-4"
- settings={settingsAction}
pin={
isDataset || !deletableRepoId
? undefined
@@ -811,7 +764,6 @@ export const InventoryRow = memo(function InventoryRow({
<CatalogRow
variant="flat"
selected={selected}
- active={active}
tooltip={tooltip}
label={title}
onClick={handleClick}
@@ -882,7 +834,6 @@ export const InventoryRow = memo(function InventoryRow({
<CatalogRow
variant="card"
selected={selected}
- active={active}
tooltip={tooltip}
label={title}
onClick={handleClick}
diff --git a/studio/frontend/src/features/hub/catalog/models-catalog.tsx b/studio/frontend/src/features/hub/catalog/models-catalog.tsx
index 0d94d335ddc..2913cad48a2 100644
--- a/studio/frontend/src/features/hub/catalog/models-catalog.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-catalog.tsx
@@ -18,7 +18,6 @@ import {
import type {
CachedInventoryRow,
DiscoverRow,
- InventoryRow,
LocalInventoryRow,
ModelsTab,
} from "../types";
@@ -46,8 +45,6 @@ export interface ModelsCatalogState {
inventoryError: boolean;
inventoryWarning: boolean;
query: string;
- activeCheckpoint: string | null;
- activeGgufVariant: string | null;
searchError: string | null;
searchFailure?: HubFailure | null;
online: boolean;
@@ -74,7 +71,6 @@ export interface ModelsCatalogHandlers {
onRetry: () => void;
onInventoryChange?: () => void;
onSwitchDevice?: () => void;
- onOpenModelSettings?: (row: InventoryRow) => void;
}
function assignRef<T>(ref: RefObject<T | null>, value: T | null) {
@@ -113,8 +109,6 @@ export const ModelsCatalog = memo(function ModelsCatalog({
inventoryError,
inventoryWarning,
query,
- activeCheckpoint,
- activeGgufVariant,
searchError,
searchFailure,
online,
@@ -134,7 +128,6 @@ export const ModelsCatalog = memo(function ModelsCatalog({
onRetry,
onInventoryChange,
onSwitchDevice,
- onOpenModelSettings,
} = handlers;
const [scrolled, setScrolled] = useState(false);
const [streamingActive, setStreamingActive] = useState(false);
@@ -476,8 +469,6 @@ export const ModelsCatalog = memo(function ModelsCatalog({
typeFilterActive={typeFilterActive}
onClearFilters={onClearFilters}
scrollElement={downloadedScrollEl}
- activeCheckpoint={activeCheckpoint}
- activeGgufVariant={activeGgufVariant}
isDataset={isDataset}
inventoryTokens={inventoryTokens}
deviceType={deviceType}
@@ -485,7 +476,6 @@ export const ModelsCatalog = memo(function ModelsCatalog({
columns={discoverView === "two" ? 2 : 1}
sort={inventorySort}
onInventoryChange={onInventoryChange}
- onOpenModelSettings={onOpenModelSettings}
/>
</div>
) : (
diff --git a/studio/frontend/src/features/hub/catalog/models-header.tsx b/studio/frontend/src/features/hub/catalog/models-header.tsx
index 09b7b85471b..47798d0501f 100644
--- a/studio/frontend/src/features/hub/catalog/models-header.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-header.tsx
@@ -1,11 +1,6 @@
// SPDX-License-Identifier: AGPL-3.0-only
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-import {
- Tooltip,
- TooltipContent,
- TooltipTrigger,
-} from "@/components/ui/tooltip";
import { useSettingsDialogStore } from "@/features/settings";
import {
ChipIcon,
@@ -13,7 +8,6 @@ import {
Database02Icon,
PackageIcon,
RamMemoryIcon,
- RemoveCircleIcon,
} from "@hugeicons/core-free-icons";
import type { IconSvgElement } from "@hugeicons/react";
import { HugeiconsIcon } from "@hugeicons/react";
@@ -47,10 +41,7 @@ export function ModelsHeader({
gpuSharedLabel,
ramLabel,
coreLabel,
- activeCheckpoint,
- activeGgufVariant,
onTitleClick,
- onEject,
}: {
cachedCount: number;
localCount: number;
@@ -59,10 +50,7 @@ export function ModelsHeader({
gpuSharedLabel: string | null;
ramLabel: string;
coreLabel: string;
- activeCheckpoint: string | null;
- activeGgufVariant: string | null;
onTitleClick: () => void;
- onEject: () => void;
}) {
const openSettings = useSettingsDialogStore((s) => s.openDialog);
const gpuMemoryValue = gpuSharedLabel
@@ -75,8 +63,8 @@ export function ModelsHeader({
onTitleClick={onTitleClick}
subtitle={
isDataset
- ? "Discover, download, and train on datasets locally."
- : "Discover, download, and run inference models locally."
+ ? "Discover and download datasets locally."
+ : "Discover and download models locally."
}
/>
@@ -100,43 +88,6 @@ export function ModelsHeader({
/>
<StatPill icon={RamMemoryIcon} label="RAM" value={ramLabel} />
<StatPill icon={CpuIcon} label="CPU" value={coreLabel} />
-
- {activeCheckpoint && (
- <div className="hub-tag-soft ml-1 inline-flex items-center gap-1.5 px-2 py-1 text-ui-11p5">
- <span
- className="size-1.5 rounded-full bg-emerald-500"
- aria-hidden="true"
- />
- <Tooltip>
- <TooltipTrigger asChild={true}>
- <span className="max-w-[120px] cursor-default truncate font-medium text-primary">
- {activeCheckpoint}
- {activeGgufVariant ? ` · ${activeGgufVariant}` : ""}
- </span>
- </TooltipTrigger>
- <TooltipContent
- side="bottom"
- sideOffset={6}
- className="tooltip-compact"
- >
- {activeCheckpoint}
- {activeGgufVariant ? ` · ${activeGgufVariant}` : ""}
- </TooltipContent>
- </Tooltip>
- <button
- type="button"
- onClick={onEject}
- className="-mr-0.5 ml-0.5 inline-flex cursor-pointer items-center gap-1 rounded-md px-1.5 text-ui-11 text-muted-foreground transition-colors hover:text-foreground"
- >
- <HugeiconsIcon
- icon={RemoveCircleIcon}
- strokeWidth={1.75}
- className="size-3"
- />
- Eject
- </button>
- </div>
- )}
</div>
</header>
);
diff --git a/studio/frontend/src/features/hub/catalog/safetensors-download-card.tsx b/studio/frontend/src/features/hub/catalog/safetensors-download-card.tsx
index 2e709478b8d..d3685ca266b 100644
--- a/studio/frontend/src/features/hub/catalog/safetensors-download-card.tsx
+++ b/studio/frontend/src/features/hub/catalog/safetensors-download-card.tsx
@@ -1,31 +1,18 @@
// SPDX-License-Identifier: AGPL-3.0-only
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-import { Spinner } from "@/components/ui/spinner";
import {
Tooltip,
TooltipContent,
TooltipTrigger,
} from "@/components/ui/tooltip";
-import { cn } from "@/lib/utils";
-import {
- Alert02Icon,
- PlayIcon,
- RemoveCircleIcon,
-} from "@hugeicons/core-free-icons";
-import { HugeiconsIcon } from "@hugeicons/react";
import { useEffect, useState } from "react";
-import { TrainIcon } from "../components/train-icon";
import { useHttpPartialsResumable, useRepoDownload } from "../download-manager";
import { useOnlineStatus } from "../hooks/use-online-status";
import { deleteCachedModel } from "../inventory";
import type { ModelInventoryFormat } from "../inventory";
import { fetchModelSize } from "../lib/dataset-size";
import { formatBytes } from "../lib/format";
-import {
- HUB_NON_GGUF_RUN_ACTIONS_VISIBLE,
- HUB_POST_DOWNLOAD_ACTIONS_VISIBLE,
-} from "../lib/hub-feature-flags";
import { fingerprintToken } from "../lib/token-fingerprint";
import { useHfTokenStore } from "../stores/hf-token-store";
import { DotTag } from "./dot-tag";
@@ -34,6 +21,7 @@ import {
DeleteConfirmDialog,
DownloadActionButton,
DownloadCard,
+ ModelRunActionButton,
} from "./download-card";
import { QuantOptionsMenu } from "./gguf-download-card";
import { useCardDelete } from "./use-card-delete";
@@ -60,14 +48,12 @@ export function SafetensorsDownloadCard({
partialTransport = null,
partialResumable = false,
modelFormat,
- canRun = true,
isActive,
isLoadingThisModel,
cachePath,
knownBytes,
- onLoad,
- onEject,
- onTrain,
+ onRun,
+ runPending = false,
onChange,
}: {
repoId: string;
@@ -76,17 +62,13 @@ export function SafetensorsDownloadCard({
partialTransport?: string | null;
partialResumable?: boolean;
modelFormat?: ModelInventoryFormat | null;
- canRun?: boolean;
isActive: boolean;
isLoadingThisModel: boolean;
/** Owning cache dir, threaded into delete so it targets this copy. */
cachePath?: string | null;
knownBytes?: number | null;
- onLoad: (opts: { ggufVariant?: string; expectedBytes?: number }) => void;
- /** Accepted for API parity; the run bar ejects instead of opening chat. */
- onUseInChat?: () => void;
- onEject?: () => void;
- onTrain?: () => void;
+ onRun?: () => void;
+ runPending?: boolean;
onChange?: () => void;
}) {
const hfToken = useHfTokenStore((s) => s.token);
@@ -167,16 +149,23 @@ export function SafetensorsDownloadCard({
partialResumable,
partialsResumable,
});
- const showActionPair = isDownloaded && !downloading && (canRun || !!onTrain);
- const showUnavailableAction =
- isDownloaded && !downloading && !canRun && !onTrain;
- const trainActionVisible = !!onTrain && HUB_POST_DOWNLOAD_ACTIONS_VISIBLE;
+ const showDownloadAction =
+ !isDownloaded || downloading || cancelling || downloadAction.starting;
+ const showRunAction =
+ Boolean(onRun) &&
+ isDownloaded &&
+ !isPartial &&
+ !downloading &&
+ !cancelling &&
+ !downloadAction.starting &&
+ !isLoadingThisModel;
const canDelete =
(isDownloaded || isPartial) &&
!downloading &&
!repoPeerActive &&
!isActive &&
- !isLoadingThisModel;
+ !isLoadingThisModel &&
+ !runPending;
// Same preview the On Device and picker rows run: without it this card kept an enabled Delete
// for a companion base an installed image GGUF still needs, and the refusal arrived as a 400
@@ -218,13 +207,8 @@ export function SafetensorsDownloadCard({
>
<div className="relative flex h-9 min-w-0 flex-1 items-center pl-3 pr-2">
<span className="flex items-center gap-1.5 text-ui-12 text-muted-foreground">
- {(isActive || isDownloaded) && (
- <DotTag
- tone="success"
- label={isActive ? "Loaded" : "On device"}
- />
- )}
- {!isDownloaded && !isActive && isPartial && !downloading && (
+ {isDownloaded && <DotTag tone="success" label="On device" />}
+ {!isDownloaded && isPartial && !downloading && (
<Tooltip>
<TooltipTrigger asChild={true}>
<span className="inline-flex">
@@ -248,9 +232,8 @@ export function SafetensorsDownloadCard({
)}
</span>
<div className="ml-auto flex items-center gap-0.5">
- {/* TODO: inference settings gear hidden for now, work on it in a future PR. */}
{/* Same 3-dots menu as GGUF, at repo level (no quant); pinning is
- omitted in the run bar. Managed HF-cache repos only. */}
+ omitted here. Managed HF-cache repos only. */}
{(isDownloaded || (isPartial && !downloading)) &&
!/^([/\\~.]|[A-Za-z]:)/.test(repoId) && (
<QuantOptionsMenu
@@ -266,81 +249,12 @@ export function SafetensorsDownloadCard({
)}
</div>
</div>
- {/* Info/actions hairline; dropped for the run action row (no divider before
- Run, as in the GGUF card's Run CTA), restored when the Train pair ships. */}
- {(!showActionPair || trainActionVisible) && <CardDivider />}
- {showActionPair ? (
- <div
- className={cn(
- "group/pair flex h-9 shrink-0 items-stretch gap-1.5",
- !HUB_NON_GGUF_RUN_ACTIONS_VISIBLE && "hidden",
- )}
- >
- {trainActionVisible && (
- <button
- type="button"
- onClick={onTrain}
- className="hub-action-btn w-24"
- >
- <HugeiconsIcon icon={TrainIcon} strokeWidth={1.75} />
- Train
- </button>
- )}
- <button
- type="button"
- disabled={isLoadingThisModel || !canRun}
- onClick={() => {
- if (!canRun) return;
- if (isActive) {
- onEject?.();
- return;
- }
- onLoad({});
- }}
- className={cn(
- isLoadingThisModel || isActive || !canRun
- ? "hub-action-btn w-24"
- : "hub-run-action-btn w-24",
- (isLoadingThisModel || !canRun) && "opacity-70",
- )}
- >
- {isLoadingThisModel ? (
- <>
- <Spinner />
- Loading…
- </>
- ) : isActive ? (
- <>
- <HugeiconsIcon icon={RemoveCircleIcon} strokeWidth={1.75} />
- Eject
- </>
- ) : canRun ? (
- <>
- <HugeiconsIcon icon={PlayIcon} strokeWidth={1.75} />
- Run
- </>
- ) : (
- <>
- <HugeiconsIcon icon={Alert02Icon} strokeWidth={1.75} />
- No run
- </>
- )}
- </button>
- </div>
- ) : showUnavailableAction ? (
- <button
- type="button"
- disabled={true}
- className="hub-action-btn w-28 opacity-70"
- >
- <HugeiconsIcon icon={Alert02Icon} strokeWidth={1.75} />
- Unavailable
- </button>
- ) : (
+ {(showDownloadAction || showRunAction) && <CardDivider />}
+ {showDownloadAction && (
<DownloadActionButton
downloading={downloadAction.downloading}
cancelling={downloadAction.cancelling}
- loading={isLoadingThisModel || downloadAction.starting}
+ loading={downloadAction.starting}
isPartial={downloadAction.isPartial}
partialResumable={downloadAction.partialResumable}
stopMode={downloadAction.stopMode}
@@ -350,6 +264,13 @@ export function SafetensorsDownloadCard({
className={repoPeerActive ? "opacity-70" : undefined}
/>
)}
+ {showRunAction && onRun && (
+ <ModelRunActionButton
+ label={`Configure and run ${repoId}`}
+ onClick={onRun}
+ loading={runPending}
+ />
+ )}
</DownloadCard>
</div>
);
diff --git a/studio/frontend/src/features/hub/hooks/use-selected-model-view.ts b/studio/frontend/src/features/hub/hooks/use-selected-model-view.ts
index 5b9c582f1d1..68b88a29edc 100644
--- a/studio/frontend/src/features/hub/hooks/use-selected-model-view.ts
+++ b/studio/frontend/src/features/hub/hooks/use-selected-model-view.ts
@@ -12,7 +12,6 @@ import type {
CachedInventoryRow,
DiscoverRow,
LocalInventoryRow,
- SelectedResourceRef,
SelectedModelView,
} from "../types";
@@ -27,54 +26,6 @@ function detectViewCapabilities(
return detectCapabilities(tags, pipelineTag, modelId);
}
-function cachedResource(row: CachedInventoryRow): SelectedResourceRef {
- return {
- repoId: row.repoId,
- localPath: row.cachePath ?? null,
- source: "hub_cache",
- cacheState: row.partial ? "partial" : "cached",
- runId: row.loadId,
- trainId: row.loadId,
- };
-}
-
-function localResource(
- row: LocalInventoryRow,
- repoId: string | null = row.repoId,
-): SelectedResourceRef {
- const cacheState = row.partial
- ? "partial"
- : row.source === "hf_cache"
- ? "cached"
- : "local";
- const id =
- row.source === "hf_cache" &&
- row.activeCache !== false &&
- repoId &&
- !row.partial
- ? repoId
- : row.loadId;
- return {
- repoId,
- localPath: row.path,
- source: row.source,
- cacheState,
- runId: id,
- trainId: id,
- };
-}
-
-function remoteResource(row: DiscoverRow): SelectedResourceRef {
- return {
- repoId: row.result.id,
- localPath: null,
- source: "huggingface",
- cacheState: row.isPartialOnDevice ? "partial" : "remote",
- runId: row.result.id,
- trainId: row.result.id,
- };
-}
-
function localFormatLabel(row: LocalInventoryRow): string {
if (row.modelFormat === "gguf") return "local GGUF";
if (row.modelFormat === "adapter") return "local adapter";
@@ -103,14 +54,10 @@ export function useSelectedModelView({
selectedLocalRow &&
selectedLocalRow.source !== "hf_cache"
) {
- const resource = localResource(
- selectedLocalRow,
- selectedDiscoverRow.result.id,
- );
return {
- id: resource.runId,
+ id: selectedLocalRow.loadId,
+ loadId: selectedLocalRow.loadId,
kind: "local",
- resource,
displayId: selectedDiscoverRow.id,
hubRepoId: selectedDiscoverRow.result.id,
owner: selectedDiscoverRow.owner,
@@ -136,10 +83,10 @@ export function useSelectedModelView({
: null,
baseModelHubId: selectedDiscoverRow.result.baseModel ?? null,
isDownloaded: !selectedLocalRow.partial,
+ runtimeCanChat: selectedLocalRow.capabilities.canChat,
isPartial: selectedLocalRow.partial ?? false,
partialTransport: selectedLocalRow.partialTransport ?? null,
partialResumable: selectedLocalRow.partialResumable === true,
- runtimeCapabilities: selectedLocalRow.capabilities,
capabilities: selectedDiscoverRow.capabilities,
license: detectLicense(selectedDiscoverRow.result.tags),
pipelineTag: selectedDiscoverRow.result.pipelineTag,
@@ -161,14 +108,14 @@ export function useSelectedModelView({
const onDevicePath =
selectedCachedRow?.cachePath ?? selectedLocalRow?.path ?? null;
- const resource = selectedCachedRow
- ? cachedResource(selectedCachedRow)
+ const isResolvedPartial = selectedCachedRow
+ ? Boolean(selectedCachedRow.partial)
: selectedLocalRow?.source === "hf_cache"
- ? localResource(selectedLocalRow, selectedDiscoverRow.result.id)
- : remoteResource(selectedDiscoverRow);
- const isResolvedPartial = resource.cacheState === "partial";
- const isResolvedOnDevice =
- resource.cacheState === "cached" || resource.cacheState === "local";
+ ? Boolean(selectedLocalRow.partial)
+ : selectedDiscoverRow.isPartialOnDevice;
+ const isResolvedOnDevice = selectedCachedRow
+ ? !selectedCachedRow.partial
+ : selectedLocalRow?.source === "hf_cache" && !selectedLocalRow.partial;
const resolvedModelFormat =
selectedCachedRow?.modelFormat ??
(selectedLocalRow?.modelFormat &&
@@ -179,8 +126,8 @@ export function useSelectedModelView({
: null);
return {
id: selectedDiscoverRow.id,
+ loadId: selectedCachedRow?.loadId ?? selectedLocalRow?.loadId ?? null,
kind: "discover",
- resource,
displayId: selectedDiscoverRow.id,
hubRepoId: selectedDiscoverRow.result.id,
owner: selectedDiscoverRow.owner,
@@ -208,6 +155,10 @@ export function useSelectedModelView({
: null,
baseModelHubId: selectedDiscoverRow.result.baseModel ?? null,
isDownloaded: isResolvedOnDevice,
+ runtimeCanChat:
+ selectedCachedRow?.capabilities.canChat ??
+ selectedLocalRow?.capabilities.canChat ??
+ false,
isPartial: isResolvedPartial,
partialTransport:
selectedCachedRow?.partialTransport ??
@@ -215,8 +166,6 @@ export function useSelectedModelView({
null,
partialResumable:
(selectedCachedRow ?? selectedLocalRow)?.partialResumable === true,
- runtimeCapabilities:
- selectedCachedRow?.capabilities ?? selectedLocalRow?.capabilities,
capabilities: selectedDiscoverRow.capabilities,
license: detectLicense(selectedDiscoverRow.result.tags),
pipelineTag: selectedDiscoverRow.result.pipelineTag,
@@ -241,7 +190,6 @@ export function useSelectedModelView({
}
if (selectedCachedRow) {
- const resource = cachedResource(selectedCachedRow);
const cachedSummary = isDatasetMode
? "Cached dataset, ready to use."
: selectedCachedRow.isGguf
@@ -263,8 +211,8 @@ export function useSelectedModelView({
const mergedBaseModel = selectedHfResult?.baseModel ?? null;
return {
id: selectedCachedRow.repoId,
+ loadId: selectedCachedRow.loadId,
kind: "cache",
- resource,
displayId: selectedCachedRow.repoId,
hubRepoId: selectedCachedRow.repoId,
owner: selectedCachedRow.owner,
@@ -282,10 +230,10 @@ export function useSelectedModelView({
baseModelSource: mergedBaseModel ? "huggingface" : null,
baseModelHubId: mergedBaseModel,
isDownloaded: !selectedCachedRow.partial,
+ runtimeCanChat: selectedCachedRow.capabilities.canChat,
isPartial: selectedCachedRow.partial ?? false,
partialTransport: selectedCachedRow.partialTransport ?? null,
partialResumable: selectedCachedRow.partialResumable === true,
- runtimeCapabilities: selectedCachedRow.capabilities,
capabilities: detectViewCapabilities(
mergedTags,
mergedPipelineTag,
@@ -315,7 +263,6 @@ export function useSelectedModelView({
if (selectedLocalRow) {
const localHubRepoId =
selectedLocalRow.source === "hf_cache" ? selectedLocalRow.repoId : null;
- const resource = localResource(selectedLocalRow, localHubRepoId);
const localDisplayId = selectedLocalRow.repoId ?? selectedLocalRow.loadId;
const isPartialHubCache =
selectedLocalRow.source === "hf_cache" &&
@@ -343,8 +290,8 @@ export function useSelectedModelView({
if (isPartialHubCache && selectedLocalRow.repoId) {
return {
id: selectedLocalRow.repoId,
+ loadId: selectedLocalRow.loadId,
kind: "cache",
- resource,
displayId: selectedLocalRow.repoId,
hubRepoId: selectedLocalRow.repoId,
owner: selectedLocalRow.owner,
@@ -359,10 +306,10 @@ export function useSelectedModelView({
requiresVariant: selectedLocalRow.capabilities.requiresVariant,
modelFormat: selectedLocalRow.modelFormat,
isDownloaded: false,
+ runtimeCanChat: selectedLocalRow.capabilities.canChat,
isPartial: true,
partialTransport: selectedLocalRow.partialTransport ?? null,
partialResumable: selectedLocalRow.partialResumable === true,
- runtimeCapabilities: selectedLocalRow.capabilities,
capabilities: detectViewCapabilities(
mergedTags,
mergedPipelineTag,
@@ -392,8 +339,8 @@ export function useSelectedModelView({
return {
id: selectedLocalRow.loadId,
+ loadId: selectedLocalRow.loadId,
kind: "local",
- resource,
displayId: localDisplayId,
hubRepoId: localHubRepoId,
owner: selectedLocalRow.owner,
@@ -415,7 +362,7 @@ export function useSelectedModelView({
adapterType: selectedLocalRow.adapterType ?? null,
trainingMethod: selectedLocalRow.trainingMethod ?? null,
isDownloaded: true,
- runtimeCapabilities: selectedLocalRow.capabilities,
+ runtimeCanChat: selectedLocalRow.capabilities.canChat,
capabilities: detectViewCapabilities(
mergedTags,
mergedPipelineTag,
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index ee382baace6..9b681bdc591 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -4,35 +4,28 @@
import { usePlatformStore } from "@/config/env";
import {
applyActiveModelStatusToStore,
+ clearNewChatDraft,
getInferenceStatus,
isExternalModelId,
isSpeechOnlyStatus,
- listGgufVariants,
resolveInferenceCheckpointId,
- useChatModelRuntime,
useChatRuntimeStore,
} from "@/features/chat";
-import { useOnlineStatus } from "@/features/hub/hooks/use-online-status";
import { useHubInfiniteScroll } from "@/features/hub";
+import { useOnlineStatus } from "@/features/hub/hooks/use-online-status";
import {
- type ModelPickTarget,
- type PerModelConfig,
- adoptLegacyConfigKey,
- applyModelLoadConfigToRuntime,
- applyPerModelConfigToRuntime,
- currentRuntimePerModelConfig,
+ clearModelConfigHandoff,
+ createModelConfigHandoffRequestId,
hfModelFitsDevice,
loadScopedGpu,
- resolveInitialConfig,
- useActiveModelConfig,
+ requestModelConfigHandoff,
} from "@/features/model-picker";
import { loadOpenAIAutoSwitchSettings } from "@/features/settings";
-import { taskForMediaPick } from "@/features/model-picker/components/model-selector/audio-picker-policy";
-import { diffusionRouteSearch } from "@/lib/diffusion-route-search";
import { useDebouncedValue } from "@/hooks/use-debounced-value";
import { useGpuInfo, useInferenceGpuInfo } from "@/hooks/use-gpu-info";
import { useVramBudgetFraction } from "@/hooks/use-vram-budget-fraction";
-import { toast } from "@/lib/toast";
+import { diffusionRouteSearch } from "@/lib/diffusion-route-search";
+import { subscribeModelLifecycle } from "@/lib/model-lifecycle-events";
import { cn } from "@/lib/utils";
import { useNavigate, useSearch } from "@tanstack/react-router";
import {
@@ -43,10 +36,11 @@ import {
useRef,
useState,
} from "react";
+import { toast } from "sonner";
import { ExternalLinkConfirmDialog } from "./catalog/external-link-confirm-dialog";
+import { FreeUpSpaceDialog } from "./catalog/free-up-space-dialog";
import { HubDetailView } from "./catalog/hub-detail-view";
import { HubFeed } from "./catalog/hub-feed";
-import { HubModelSettingsView } from "./catalog/hub-model-settings-view";
import { HubTopBar } from "./catalog/hub-top-bar";
import {
ModelsCatalog,
@@ -64,7 +58,6 @@ import {
ResultListHeader,
} from "./catalog/models-table";
import { ModelsToolbar } from "./catalog/models-toolbar";
-import { FreeUpSpaceDialog } from "./catalog/free-up-space-dialog";
import { OnDeviceFoldersDialog } from "./catalog/on-device-folders-dialog";
import { OwnerScopeToggle } from "./catalog/owner-scope-toggle";
import { useDiscoverSearch } from "./hooks/use-discover-search";
@@ -81,15 +74,7 @@ import { useHubModelVram } from "./hooks/use-hub-model-vram";
import { useModelsSelection } from "./hooks/use-models-selection";
import { resolveSelectionUrlSync } from "./lib/selection-resolution";
import { useHubInventory } from "./inventory";
-import { LOCAL_MODEL_SOURCE } from "./inventory/constants";
-import { settingsGgufVariantForRow } from "./inventory/settings-identity";
import { adoptResidentModelStatus } from "./lib/adopt-inference-status";
-import { subscribeResidentStatusRefresh } from "./lib/resident-status-refresh";
-import {
- type RefreshSupersession,
- registerRefresh,
- supersedingRefresh,
-} from "./lib/superseded-refresh";
import {
CHANNEL_TO_SECTION,
type ChannelId,
@@ -104,19 +89,25 @@ import {
isHiddenModelId,
} from "./lib/hidden-models";
import { inventoryRowMatches, tokenizeQuery } from "./lib/inventory-search";
-import { looksLikeLocalPath, routableToMediaPage } from "./lib/local-path";
+import { routableToMediaPage } from "./lib/local-path";
+import { residentModelIdMatches } from "./lib/model-identity";
import {
- ggufVariantsMatch,
- modelIdsMatch,
- residentModelIdMatches,
-} from "./lib/model-identity";
+ createHubModelConfigHandoff,
+ type HubModelRunSelection,
+} from "./lib/model-run-selection";
import {
type ModelTypeFilter,
matchesModelType,
} from "./lib/model-type-filter";
import { resolveOwnerProviderLogo } from "./lib/provider-logos";
-import { studioPageForTask } from "./lib/unsloth-support";
+import { subscribeResidentStatusRefresh } from "./lib/resident-status-refresh";
+import {
+ type RefreshSupersession,
+ registerRefresh,
+ supersedingRefresh,
+} from "./lib/superseded-refresh";
import { fingerprintToken } from "./lib/token-fingerprint";
+import { studioPageForTask } from "./lib/unsloth-support";
import {
buildDiscoverRows,
detectResultFormat,
@@ -135,30 +126,61 @@ import type {
ModelsTab,
ResourceTypeFilter,
SelectedModelView,
- SelectedResourceRef,
} from "./types";
-// What per-model settings are keyed by, which is not always what the loader is handed: a repo
-// cached outside the active HF cache loads by snapshot path while the picker keys it by repo
-// id. The row decides that, not the view it is shown in; `hub_cache` marks exactly those.
-function modelConfigIdentity(
- kind: SelectedModelView["kind"],
- resource: SelectedResourceRef,
-): string {
- if (kind !== "cache" && resource.source !== "hub_cache") {
- return resource.runId;
- }
- return resource.repoId ?? resource.runId;
-}
-
const MODELS_TAB_STORAGE_KEY = "unsloth.hub.modelsTab";
const ALL_MODELS_VIEW_STORAGE_KEY = "unsloth.hub.allModelsView";
const INVENTORY_SORT_STORAGE_KEY = "unsloth.hub.inventorySort";
const OWNER_SCOPE_STORAGE_KEY = "unsloth.hub.ownerScope";
+const RUN_CONFIG_REFRESH_TIMEOUT_MS = 5_000;
// Iconless models (no provider logo, e.g. Ornith, Inkling) show once they clear this many likes.
const MIN_ICONLESS_MODEL_LIKES = 30;
+async function waitForRunConfigRefresh(
+ refresh: Promise<void>,
+ signal: AbortSignal,
+): Promise<void> {
+ if (signal.aborted) return;
+ let timeoutId: number | undefined;
+ let onAbort: (() => void) | undefined;
+ try {
+ await Promise.race([
+ refresh,
+ new Promise<void>((resolve) => {
+ timeoutId = window.setTimeout(resolve, RUN_CONFIG_REFRESH_TIMEOUT_MS);
+ }),
+ new Promise<void>((resolve) => {
+ onAbort = resolve;
+ signal.addEventListener("abort", onAbort, { once: true });
+ }),
+ ]);
+ } finally {
+ if (timeoutId !== undefined) window.clearTimeout(timeoutId);
+ if (onAbort) signal.removeEventListener("abort", onAbort);
+ }
+}
+
+class RunConfigOpenCoordinator {
+ private active: AbortController | null = null;
+
+ begin(): AbortController {
+ this.cancel();
+ const controller = new AbortController();
+ this.active = controller;
+ return controller;
+ }
+
+ cancel(): void {
+ this.active?.abort();
+ this.active = null;
+ }
+
+ finish(controller: AbortController): void {
+ if (this.active === controller) this.active = null;
+ }
+}
+
/** Discover browsing scope: the whole Hub (default) or only the unsloth org. */
export type OwnerScope = "unsloth" | "all";
@@ -191,8 +213,6 @@ const FEED_LIST_CHANNEL_ID: ChannelId = "unsloth-latest";
type DiscoverMode = "feed" | "channel-list" | "search";
-type ModelLoadOptions = { ggufVariant?: string; expectedBytes?: number };
-
// Focused list heading stays "Models"/"Datasets" regardless of filters; only search relabels it.
function buildFocusedHeading({
query,
@@ -365,10 +385,20 @@ function selectedRepoMatchesRuntime(
ggufVariant: string | null,
): boolean {
if (!selectedModel || !runtimeId) return false;
- if (!modelIdsMatch(runtimeId, selectedModel.resource.runId)) return false;
+ if (
+ !residentModelIdMatches(
+ runtimeId,
+ selectedModel.id,
+ selectedModel.loadId,
+ selectedModel.hubRepoId,
+ selectedModel.path,
+ selectedModel.displayId,
+ )
+ ) {
+ return false;
+ }
if (selectedModel.modelFormat === "gguf") {
- const localPath =
- selectedModel.resource.localPath ?? selectedModel.path ?? "";
+ const localPath = selectedModel.path ?? "";
return ggufVariant !== null || localPath.toLowerCase().endsWith(".gguf");
}
return ggufVariant === null;
@@ -413,41 +443,23 @@ export function ModelsPage() {
const preferredGgufFile = hubSearch.file ?? null;
const preferredGgufFileIntent = hubSearch.intent ?? 0;
- const { selectModel, loadingModel, loadProgress, ejectModel } =
- useChatModelRuntime();
const checkpoint = useChatRuntimeStore((s) => s.params.checkpoint);
const residentCheckpoint = useChatRuntimeStore((s) => s.residentCheckpoint);
- // Resident, not merely picked. An image or video load evicts the chat model
- // and leaves the pick alone, so the cards kept saying "Loaded" for weights the
- // backend had already released. `undefined` is "no status read yet", which
- // stays as it was rather than flashing "On device" on every launch.
const activeCheckpoint =
checkpoint && !isExternalModelId(checkpoint) && residentCheckpoint !== null
? checkpoint
: null;
const activeGgufVariant = useChatRuntimeStore((s) => s.activeGgufVariant);
- const activeLoadedContextLength = useChatRuntimeStore(
- (s) => s.loadedContextLength,
- );
+ const loadingModel = useChatRuntimeStore((s) => s.loadingModelPick);
const [initialResidentStatusSettled, setInitialResidentStatusSettled] =
useState(false);
- // Live settings of the loaded model, so its page shows what it is running with.
- const { config: activeModelConfig } = useActiveModelConfig();
- // Shared with the chat model selector: list only models sized for this device.
const fitOnDeviceOnly = useChatRuntimeStore((s) => s.fitOnDeviceOnly);
const setFitOnDeviceOnly = useChatRuntimeStore((s) => s.setFitOnDeviceOnly);
- // Drops a response that lands after a newer read started, or after unmount.
const residentStatusSeq = useRef(0);
- // The newest read started, so a dropped response can resolve with that instead of nothing. The
- // settings handlers' own guard counts settings opens only: a focus refresh supersedes without bumping.
const residentStatusSupersession = useRef<RefreshSupersession>({
latest: null,
});
- // /status cannot say whether an empty answer is an idle eviction that will reload or a real
- // unload, and only this endpoint knows. Read alongside every status read rather than cached,
- // since the idle timeout is editable while this page stays mounted. Last answer wins on
- // failure, not the default: "disarmed" clears the checkpoint, so a blip would discard a selection.
const idleUnloadArmed = useRef(false);
const readIdleUnloadArmed = useCallback(
(): Promise<boolean> =>
@@ -459,60 +471,61 @@ export function ModelsPage() {
.catch(() => idleUnloadArmed.current),
[],
);
- // Returns the read so a caller that needs the answer first can wait for it.
- const refreshResidentModelStatus = useCallback((): Promise<void> => {
- const seq = ++residentStatusSeq.current;
- const read = Promise.all([getInferenceStatus(), readIdleUnloadArmed()])
- .then(([status, idleUnloadArmed]) => {
- // A newer read owns the store, so this writes nothing; resolving with the refresh that took
- // it over keeps an awaiting caller off the pre-switch store the await was there to replace.
- if (seq !== residentStatusSeq.current)
- return supersedingRefresh(residentStatusSupersession.current, seq);
- const store = useChatRuntimeStore.getState();
- adoptResidentModelStatus(
- {
- // The loadable identifier: a GGUF off disk loads by path, and two files sharing a stem collapse.
- // Null for a speech model: this page is the other writer of params.checkpoint,
- // so adopting one here made it the chat model just as the mount sync did.
- checkpointId: isSpeechOnlyStatus(status)
- ? null
- : resolveInferenceCheckpointId(status),
- speechOnly: isSpeechOnlyStatus(status),
- ggufVariant: status.gguf_variant ?? null,
- },
- {
- checkpoint: store.params.checkpoint,
- checkpointIsExternal: isExternalModelId(store.params.checkpoint),
- activeGgufVariant: store.activeGgufVariant,
- modelLoading: store.modelLoading,
- idleUnloadArmed,
- },
- {
- setCheckpoint: (checkpointId, ggufVariant) => {
- store.setCheckpoint(checkpointId, ggufVariant);
+ const refreshResidentModelStatus = useCallback(
+ (
+ { preserveIdleUnloaded = true }: {
+ preserveIdleUnloaded?: boolean;
+ } = {},
+ ): Promise<void> => {
+ const seq = ++residentStatusSeq.current;
+ const read = Promise.all([getInferenceStatus(), readIdleUnloadArmed()])
+ .then(([status, idleUnloadArmed]) => {
+ if (seq !== residentStatusSeq.current)
+ return supersedingRefresh(residentStatusSupersession.current, seq);
+ const store = useChatRuntimeStore.getState();
+ adoptResidentModelStatus(
+ {
+ // The loadable identifier: a GGUF off disk loads by path, and two files sharing a stem collapse.
+ // Null for a speech model: this page is the other writer of params.checkpoint,
+ // so adopting one here made it the chat model just as the mount sync did.
+ checkpointId: isSpeechOnlyStatus(status)
+ ? null
+ : resolveInferenceCheckpointId(status),
+ speechOnly: isSpeechOnlyStatus(status),
+ ggufVariant: status.gguf_variant ?? null,
},
- clearCheckpoint: () => {
- store.clearCheckpoint();
+ {
+ checkpoint: store.params.checkpoint,
+ checkpointIsExternal: isExternalModelId(store.params.checkpoint),
+ activeGgufVariant: store.activeGgufVariant,
+ modelLoading: store.modelLoading,
+ idleUnloadArmed: preserveIdleUnloaded && idleUnloadArmed,
},
- // The one entry point that has applied no status yet, so settings would read defaults.
- applyStatus: (previous) => {
- applyActiveModelStatusToStore(status, {
- previousCheckpoint: previous.checkpoint ?? undefined,
- previousGgufVariant: previous.ggufVariant,
- adoptingExistingServerModel:
- previous.checkpoint === null || previous.checkpoint === "",
- });
+ {
+ setCheckpoint: (checkpointId, ggufVariant) => {
+ store.setCheckpoint(checkpointId, ggufVariant);
+ },
+ clearCheckpoint: () => {
+ store.clearCheckpoint();
+ },
+ applyStatus: (previous) => {
+ applyActiveModelStatusToStore(status, {
+ previousCheckpoint: previous.checkpoint ?? undefined,
+ previousGgufVariant: previous.ggufVariant,
+ adoptingExistingServerModel:
+ previous.checkpoint === null || previous.checkpoint === "",
+ });
+ },
},
- },
- );
- })
- .catch(() => undefined);
- // In start order, synchronously with the sequence number, so a dropped response finds the newest read.
- registerRefresh(residentStatusSupersession.current, seq, read);
- return read;
- }, [readIdleUnloadArmed]);
-
- // Mount, then whenever this tab could have missed an API-driven switch. Re-reading is safe.
+ );
+ })
+ .catch(() => undefined);
+ registerRefresh(residentStatusSupersession.current, seq, read);
+ return read;
+ },
+ [readIdleUnloadArmed],
+ );
+
useEffect(() => {
let active = true;
void refreshResidentModelStatus().finally(() => {
@@ -523,12 +536,20 @@ export function ModelsPage() {
);
return () => {
active = false;
- // A response still in flight adopts nothing once the Hub is gone.
residentStatusSeq.current += 1;
unsubscribe();
};
}, [refreshResidentModelStatus]);
+ useEffect(
+ () =>
+ subscribeModelLifecycle(({ runtime }) => {
+ if (runtime === "chat" || runtime === "stt") return;
+ void refreshResidentModelStatus({ preserveIdleUnloaded: false });
+ }),
+ [refreshResidentModelStatus],
+ );
+
const { tab, setTab: setModelsTab } = useModelsTabState();
const [query, setQuery] = useState("");
const [sortBy, setSortBy] = useState<HfSortKey>(
@@ -1184,17 +1205,31 @@ export function ModelsPage() {
accessToken: apiHfToken,
online,
});
+ const [runConfigOpenCoordinator] = useState(
+ () => new RunConfigOpenCoordinator(),
+ );
+ const [runConfigOpening, setRunConfigOpening] = useState<{
+ modelId: string;
+ controller: AbortController;
+ } | null>(null);
+ useEffect(() => {
+ return () => runConfigOpenCoordinator.cancel();
+ }, [runConfigOpenCoordinator, selectedId]);
const handleSelect = useCallback(
(id: string) => {
+ runConfigOpenCoordinator.cancel();
+ setRunConfigOpening(null);
void navigate({
to: "/hub",
search: (prev) => ({ ...prev, model: id, file: undefined }),
});
},
- [navigate],
+ [navigate, runConfigOpenCoordinator],
);
const handleCloseDetail = useCallback(() => {
+ runConfigOpenCoordinator.cancel();
+ setRunConfigOpening(null);
// From split view, "Back to Hub" returns to the main hub feed (not the filtered list): leave
// split mode, reset discover state, and clear the inline preview and channel.
if (allModelsView === "split") {
@@ -1218,7 +1253,7 @@ export function ModelsPage() {
to: "/hub",
search: (prev) => ({ ...prev, model: undefined }),
});
- }, [navigate, allModelsView]);
+ }, [navigate, allModelsView, runConfigOpenCoordinator]);
const handleQueryChange = useCallback(
(next: string) => {
if (next.trim() === "") {
@@ -1340,7 +1375,14 @@ export function ModelsPage() {
const isLoadingThisModel = useMemo(() => {
if (!loadingModel || !selectedModel) return false;
- return modelIdsMatch(loadingModel.id, selectedModel.resource.runId);
+ return residentModelIdMatches(
+ loadingModel.id,
+ selectedModel.id,
+ selectedModel.loadId,
+ selectedModel.hubRepoId,
+ selectedModel.path,
+ selectedModel.displayId,
+ );
}, [loadingModel, selectedModel]);
const { vramInfo, minMemory } = useHubModelVram(selectedModel, gpu);
@@ -1360,322 +1402,6 @@ export function ModelsPage() {
gpu.cpuCore > 0 && gpu.cpuThread > 0
? `${gpu.cpuCore}/${gpu.cpuThread}`
: "Unavailable";
-
- const openNewChat = useCallback(() => {
- void navigate({ to: "/chat", search: { new: crypto.randomUUID() } });
- }, [navigate]);
- const runSelectedModel = useCallback(
- (opts: ModelLoadOptions, isDownloaded: boolean) => {
- if (!selectedModel) return;
- const runId = selectedModel.resource.runId;
- // An image / video model is run by its own page, not by chat: loading it here evicted
- // the resident chat model for a llama.cpp load that could only fail. Same resolution
- // and destination the chat picker uses, so both surfaces route a pick identically.
- // `task` is not optional here: only CachedModelRepo carries pipelineTag, so every
- // cached GGUF repo (the reported MiniMax-H3 case) reports its modality on `task`.
- const mediaPage = studioPageForTask(
- taskForMediaPick(selectedModel.pipelineTag, selectedModel.task) ?? undefined,
- );
- // The target pages read a routed `model` as a Hub id, so a runId that is a PATH would
- // arrive as a repo that does not exist -- prefer the Hub id, which loads the same copy
- // since the loader reuses whichever cache root holds it. That covers a filesystem row
- // (left on today's route, and the backend preflight now refuses it by name) and a
- // cached repo the inventory pinned to its snapshot directory, whose symlinked entries
- // the pages' containment check rejects anyway.
- const routeId = runId && !looksLikeLocalPath(runId) ? runId : selectedModel.hubRepoId;
- if (
- mediaPage &&
- routableToMediaPage(selectedModel.kind, selectedModel.localSource) &&
- routeId
- ) {
- void navigate({
- to: `/${mediaPage}`,
- // `quant` is consumed verbatim as a gguf filename, so a label rides `ggufQuant`.
- search: diffusionRouteSearch(routeId, {
- ggufVariant: opts.ggufVariant ?? null,
- }),
- });
- return;
- }
- const configIdentity = modelConfigIdentity(
- selectedModel.kind,
- selectedModel.resource,
- );
- // A cached repo used to be keyed by its snapshot path (its runId); move that record over first.
- adoptLegacyConfigKey(configIdentity, runId, opts.ggufVariant);
- const resolvedConfig = resolveInitialConfig(
- configIdentity,
- opts.ggufVariant,
- );
- const rememberedConfig = resolvedConfig.remembered
- ? resolvedConfig.config
- : null;
- const previousConfig = currentRuntimePerModelConfig({
- includeMaxSeqLength: true,
- });
- const hasAppliedConfig = applyModelLoadConfigToRuntime(rememberedConfig);
- void selectModel({
- id: runId,
- ggufVariant: opts.ggufVariant,
- isDownloaded,
- expectedBytes: opts.expectedBytes,
- keepSpeculative: hasAppliedConfig,
- throwOnError: true,
- previousConfig,
- // The runtime store is not enough: applyPerModelConfigToRuntime has no field
- // for the launch flags, and /load only inherits them from the SAME resident
- // model, so a cold launch or a switch from another model ran without the
- // arguments this model was remembered with.
- ...(rememberedConfig ? { config: rememberedConfig } : {}),
- })
- .then(() => {
- // Read fresh: the load is async, so the checkpoint may have changed.
- const store = useChatRuntimeStore.getState();
- if (!modelIdsMatch(store.params.checkpoint, runId)) {
- store.setCheckpoint(runId, opts.ggufVariant ?? null);
- }
- })
- .catch(() => undefined);
- openNewChat();
- },
- [navigate, openNewChat, selectModel, selectedModel],
- );
- const handleLoad = useCallback(
- (opts: ModelLoadOptions) =>
- runSelectedModel(opts, selectedModel?.isDownloaded ?? true),
- [runSelectedModel, selectedModel],
- );
-
- // Full-page per-model settings. Local state, not a URL param: a deep link would need the row re-resolved.
- const [settingsTarget, setSettingsTarget] = useState<ModelPickTarget | null>(
- null,
- );
- // Opening settings is where a stale read costs something: the editor is seeded once and Apply
- // reloads with what it seeded. Resolving a target is openModelSettings' own read.
- useEffect(() => {
- if (settingsTarget) void refreshResidentModelStatus();
- }, [settingsTarget, refreshResidentModelStatus]);
- // Bumped per open so a slow lookup for an abandoned row cannot land on the chosen one.
- const settingsOpenSeq = useRef(0);
- const openModelSettings = useCallback(
- async (row: CachedInventoryRow | LocalInventoryRow) => {
- const openSeq = ++settingsOpenSeq.current;
- // Before anything reads the store: the status effect watches settingsTarget. Every path out
- // seeds the editor from the store and Apply reloads with that, so a stale read misconfigures.
- await refreshResidentModelStatus();
- if (settingsOpenSeq.current !== openSeq) return;
- // loadId is what the loader accepts; repoId is only a display/API alias.
- const id = row.loadId;
- // Every name this row answers to.
- const rowAliases =
- row.kind === "local"
- ? [id, row.repoId, row.path]
- : [id, row.repoId, row.cachePath];
- // Cached repo rows carry no quant, and a null variant keys the config to `repo::` while the loader reads `repo::Q4_K_M`.
- let ggufVariant = settingsGgufVariantForRow(row);
- if (!ggufVariant && row.isGguf && row.capabilities.requiresVariant) {
- // A local row only carries a repo id inside the HF cache, so a plain folder of quants has
- // none while still needing one; the listing scans a path in the same position.
- const repoId =
- row.kind === "cache" ? row.repoId : (row.repoId ?? row.path ?? null);
- if (repoId) {
- try {
- const res = await listGgufVariants(repoId, hfApiToken(hfToken), {
- preferLocalCache: true,
- localPath:
- row.kind === "local" ? row.path : (row.cachePath ?? null),
- });
- const downloaded = res.variants.filter((v) => v.downloaded);
- // Re-read after this await too: the lookup hits the network, and a switch during it would
- // leave the branch below on the displaced model's quant. Against the server, not the store:
- // nothing pushes an API switch into this tab.
- await refreshResidentModelStatus();
- if (settingsOpenSeq.current !== openSeq) return;
- const settled = useChatRuntimeStore.getState();
- const settledCheckpoint =
- settled.params.checkpoint &&
- !isExternalModelId(settled.params.checkpoint)
- ? settled.params.checkpoint
- : null;
- const settledIsActive = rowAliases.some((alias) =>
- modelIdsMatch(alias, settledCheckpoint),
- );
- ggufVariant =
- // Loaded quant, then repo default, then whatever is on disk, mirroring LocalOnDeviceCard.
- (settledIsActive
- ? downloaded.find((v) =>
- ggufVariantsMatch(v.quant, settled.activeGgufVariant),
- )?.quant
- : undefined) ??
- downloaded.find((v) =>
- ggufVariantsMatch(v.quant, res.default_variant),
- )?.quant ??
- downloaded[0]?.quant ??
- null;
- } catch {
- ggufVariant = null;
- }
- }
- if (!ggufVariant) {
- // A model needing a quant cannot be configured without one: the picker matches variants exactly.
- toast.error("Couldn't determine which quant to configure.", {
- description:
- "Settings for this model are per quant. Check the connection or the model's cache, then try again.",
- });
- return;
- }
- }
- // The lookup is async: without this, whichever call finished last would win.
- if (settingsOpenSeq.current !== openSeq) {
- return;
- }
- // A repo in a previous cache loads by snapshot path, so `id` ends in the revision hash.
- const configId = row.kind === "cache" ? row.repoId : id;
- const leaf = configId.split(/[\\/]/).filter(Boolean).pop() ?? configId;
- setSettingsTarget({
- id,
- configId,
- displayName: ggufVariant ? `${leaf} · ${ggufVariant}` : leaf,
- ggufVariant,
- isGguf: row.isGguf,
- apiLoadable:
- row.isGguf &&
- (row.kind !== "local" || row.source !== LOCAL_MODEL_SOURCE.OLLAMA),
- meta: {
- source: "local",
- isLora: row.modelFormat === "adapter",
- ggufVariant: ggufVariant ?? undefined,
- isGguf: row.isGguf,
- // A partial opens settings too; claiming complete would skip download progress.
- isDownloaded: !row.partial,
- // Not on inventory rows; ModelConfigPage reads the GGUF header itself.
- contextLength: null,
- },
- });
- },
- [activeCheckpoint, activeGgufVariant, hfToken, refreshResidentModelStatus],
- );
- // Applying loads the model with exactly these settings, already persisted by ModelConfigPage.
- const runSettingsTarget = useCallback(
- (config: PerModelConfig) => {
- const target = settingsTarget;
- if (!target) return;
- const previousConfig = currentRuntimePerModelConfig({
- includeMaxSeqLength: true,
- });
- applyPerModelConfigToRuntime(config);
- setSettingsTarget(null);
- void selectModel({
- id: target.id,
- source: "local",
- ggufVariant: target.ggufVariant ?? undefined,
- isGguf: target.isGguf,
- // A partial row opens settings too; claiming complete would skip download progress.
- isDownloaded: target.meta.isDownloaded,
- isLora: target.meta.isLora,
- keepSpeculative: true,
- forceReload: true,
- // The submitted config, not only its echo in the runtime store: the store
- // does not carry llamaExtraArgs, so without this the load omits the field
- // and the route keeps the resident server's old list. Applying an edit, or
- // clearing the box, would then do nothing on this page.
- config,
- previousConfig,
- }).catch(() => undefined);
- },
- [selectModel, settingsTarget],
- );
- const handleLoadLocal = useCallback(
- (opts: ModelLoadOptions = {}) => runSelectedModel(opts, true),
- [runSelectedModel],
- );
- const handleTrain = useCallback(() => {
- // Hub → train integration ships in a later PR.
- }, []);
- // Opened from the on-device card, which passes in the quant it already resolved.
- const openSelectedModelSettings = useCallback(
- async (ggufVariant: string | null, quantIsUserPicked = false) => {
- if (!selectedModel) return;
- // Shared with openModelSettings: another row's pending lookup must not land here.
- const openSeq = ++settingsOpenSeq.current;
- // Unconditional, as in openModelSettings: the editor seeds from the store and Apply reloads it.
- await refreshResidentModelStatus();
- if (settingsOpenSeq.current !== openSeq) return;
- let variant = ggufVariant;
- // The card derived this quant from the store's active variant. A user-picked quant stands.
- if (!quantIsUserPicked) {
- const settled = useChatRuntimeStore.getState();
- const settledCheckpoint =
- settled.params.checkpoint &&
- !isExternalModelId(settled.params.checkpoint)
- ? settled.params.checkpoint
- : null;
- // Every name this model answers to, as the row menu path matches them.
- const aliases = [
- selectedModel.resource.runId,
- selectedModel.resource.repoId,
- selectedModel.resource.localPath,
- ];
- if (
- settled.activeGgufVariant &&
- aliases.some((alias) => modelIdsMatch(alias, settledCheckpoint))
- ) {
- variant = settled.activeGgufVariant;
- }
- }
- // The card passes null while its lookup is pending or failed, so guard as openModelSettings does.
- if (!variant && selectedModel.isGguf && selectedModel.requiresVariant) {
- toast.error("Couldn't determine which quant to configure.", {
- description:
- "Settings for this model are per quant. Check the connection or the model's cache, then try again.",
- });
- return;
- }
- const id = selectedModel.resource.runId;
- const configId = modelConfigIdentity(
- selectedModel.kind,
- selectedModel.resource,
- );
- // As in runSelectedModel: the editor seeds from this key, so the move must happen first.
- adoptLegacyConfigKey(configId, id, variant);
- const leaf = configId.split(/[\\/]/).filter(Boolean).pop() ?? configId;
- setSettingsTarget({
- id,
- configId,
- displayName: variant ? `${leaf} · ${variant}` : leaf,
- ggufVariant: variant,
- isGguf: selectedModel.isGguf,
- apiLoadable:
- selectedModel.isGguf &&
- selectedModel.localSource !== LOCAL_MODEL_SOURCE.OLLAMA,
- meta: {
- source: "local",
- isLora: selectedModel.modelFormat === "adapter",
- ggufVariant: variant ?? undefined,
- isGguf: selectedModel.isGguf,
- isDownloaded: selectedModel.isDownloaded,
- contextLength: null,
- },
- });
- },
- [selectedModel, refreshResidentModelStatus],
- );
- // Whether the settings page is open on the loaded model, so it can show the live launch config.
- // A GGUF off disk loads by path but is reported by its public id, so the row's path and its
- // settings identity are both offered as aliases; a loose .gguf carries no variant.
- const settingsTargetIsStandaloneFile =
- settingsTarget !== null &&
- settingsTarget.ggufVariant == null &&
- settingsTarget.id.toLowerCase().endsWith(".gguf");
- const settingsTargetIsResident =
- settingsTarget !== null &&
- residentModelIdMatches(
- activeCheckpoint,
- settingsTarget.id,
- settingsTarget.configId,
- ) &&
- (settingsTargetIsStandaloneFile ||
- ggufVariantsMatch(activeGgufVariant, settingsTarget.ggufVariant));
const handleSearchHub = useCallback(
(next: string) => {
const trimmed = next.trim();
@@ -1701,13 +1427,77 @@ export function ModelsPage() {
},
[navigate, setModelsTab, setOwnerScope],
);
+ const handleRun = useCallback(
+ async (
+ selection: HubModelRunSelection,
+ mediaPage: ReturnType<typeof studioPageForTask>,
+ ) => {
+ if (!selectedModel) return;
+ if (mediaPage) {
+ runConfigOpenCoordinator.cancel();
+ setRunConfigOpening(null);
+ if (
+ !selectedModel.hubRepoId ||
+ !routableToMediaPage(selectedModel.kind, selectedModel.localSource)
+ ) {
+ return;
+ }
+ void navigate({
+ to: `/${mediaPage}`,
+ search: diffusionRouteSearch(selectedModel.hubRepoId, selection),
+ });
+ return;
+ }
+ const controller = runConfigOpenCoordinator.begin();
+ setRunConfigOpening({ modelId: selectedModel.id, controller });
+ try {
+ await waitForRunConfigRefresh(
+ refreshResidentModelStatus(),
+ controller.signal,
+ );
+ if (controller.signal.aborted) return;
+ const requestId = createModelConfigHandoffRequestId();
+ const request = createHubModelConfigHandoff({
+ requestId,
+ model: selectedModel,
+ selection,
+ });
+ if (!request) {
+ toast.error("Couldn't open run settings.", {
+ description:
+ "The model or selected quantization is no longer available to run.",
+ });
+ return;
+ }
+ clearNewChatDraft();
+ const chatRuntime = useChatRuntimeStore.getState();
+ chatRuntime.setActiveThreadId(null);
+ chatRuntime.setActiveProjectId(null);
+ chatRuntime.setIncognito(false);
+ requestModelConfigHandoff(request);
+ void navigate({ to: "/chat", search: { new: requestId } }).catch(() => {
+ clearModelConfigHandoff(requestId);
+ });
+ } finally {
+ runConfigOpenCoordinator.finish(controller);
+ setRunConfigOpening((current) =>
+ current?.controller === controller ? null : current,
+ );
+ }
+ },
+ [
+ navigate,
+ refreshResidentModelStatus,
+ runConfigOpenCoordinator,
+ selectedModel,
+ ],
+ );
const inspectorRuntime = useMemo(
() => ({
isActive,
activeGgufVariant,
isLoadingThisModel,
- loadingPhase: loadProgress?.phase,
minMemory,
vramInfo,
gpuGb: inferenceGpu.available ? inferenceGpu.memoryTotalGb : undefined,
@@ -1721,7 +1511,6 @@ export function ModelsPage() {
isActive,
activeGgufVariant,
isLoadingThisModel,
- loadProgress?.phase,
minMemory,
vramInfo,
inferenceGpu.available,
@@ -1733,24 +1522,17 @@ export function ModelsPage() {
const inspectorActions = useMemo(
() => ({
- onLoad: handleLoad,
- onLoadLocal: handleLoadLocal,
- onUseInChat: openNewChat,
- onEject: () => void ejectModel(),
- onTrain: handleTrain,
onInventoryChange: refreshInventory,
onSearchHub: handleSearchHub,
- onOpenSettings: openSelectedModelSettings,
+ onRun: handleRun,
+ runConfigPending: runConfigOpening?.modelId === selectedModel?.id,
}),
[
- handleLoad,
- handleLoadLocal,
- openNewChat,
- ejectModel,
- handleTrain,
+ handleRun,
handleSearchHub,
refreshInventory,
- openSelectedModelSettings,
+ runConfigOpening,
+ selectedModel?.id,
],
);
@@ -1769,8 +1551,6 @@ export function ModelsPage() {
inventoryError,
inventoryWarning,
query,
- activeCheckpoint,
- activeGgufVariant,
searchError,
searchFailure,
online,
@@ -1796,8 +1576,6 @@ export function ModelsPage() {
inventoryError,
inventoryWarning,
query,
- activeCheckpoint,
- activeGgufVariant,
searchError,
searchFailure,
online,
@@ -1834,7 +1612,6 @@ export function ModelsPage() {
onRetry: handleRetrySearch,
onInventoryChange: refreshInventory,
onSwitchDevice: handleSwitchDevice,
- onOpenModelSettings: openModelSettings,
}),
[
handleSelect,
@@ -1843,7 +1620,6 @@ export function ModelsPage() {
handleRetrySearch,
refreshInventory,
handleSwitchDevice,
- openModelSettings,
],
);
@@ -1984,8 +1760,8 @@ export function ModelsPage() {
const detailOpen = urlModel !== null;
const splitMode = allModelsView === "split";
- // Unreachable under an opaque overlay: the detail view (full-page only) or settings.
- const catalogCovered = (detailOpen && !splitMode) || settingsTarget !== null;
+ // Unreachable under the full-page detail overlay.
+ const catalogCovered = detailOpen && !splitMode;
return (
<div className="hub-page flex min-h-0 min-w-0 flex-1 basis-0 flex-col overflow-hidden bg-background">
@@ -1998,10 +1774,7 @@ export function ModelsPage() {
gpuSharedLabel={gpuSharedLabel}
ramLabel={ramLabel}
coreLabel={coreLabel}
- activeCheckpoint={activeCheckpoint}
- activeGgufVariant={activeGgufVariant}
onTitleClick={handleResetToDiscover}
- onEject={() => void ejectModel()}
/>
<ModelsToolbar
tab={tab}
@@ -2064,7 +1837,6 @@ export function ModelsPage() {
detailOpen ? (
<div
className="hub-canvas z-20 flex min-h-0 flex-col max-lg:absolute max-lg:inset-0 lg:relative lg:min-w-0 lg:flex-1"
- inert={settingsTarget !== null || undefined}
>
<HubDetailView
model={selectedModel}
@@ -2089,7 +1861,6 @@ export function ModelsPage() {
detailOpen && (
<div
className="hub-canvas absolute inset-0 z-20 flex min-h-0 flex-col"
- inert={settingsTarget !== null || undefined}
>
<HubDetailView
model={selectedModel}
@@ -2106,22 +1877,6 @@ export function ModelsPage() {
</div>
)
)}
-
- {/* Above the detail overlay (z-30), so settings do not stack behind a preview. */}
- {settingsTarget && (
- <div className="hub-canvas absolute inset-0 z-30 flex min-h-0 flex-col">
- <HubModelSettingsView
- target={settingsTarget}
- loadedConfig={settingsTargetIsResident ? activeModelConfig : null}
- loadedContextLength={
- settingsTargetIsResident ? activeLoadedContextLength : null
- }
- onBack={() => setSettingsTarget(null)}
- onRun={runSettingsTarget}
- compact={splitMode}
- />
- </div>
- )}
</div>
<OnDeviceFoldersDialog
diff --git a/studio/frontend/src/features/hub/hub.css b/studio/frontend/src/features/hub/hub.css
index 7c2e8966664..b05394d52fe 100644
--- a/studio/frontend/src/features/hub/hub.css
+++ b/studio/frontend/src/features/hub/hub.css
@@ -48,8 +48,8 @@
}
@layer base {
- /* Shared measure for every centered Hub column (top bar, catalog, detail,
- settings). Container-relative, so the column tracks the width actually
+ /* Shared measure for every centered Hub column (top bar, catalog, detail).
+ Container-relative, so the column tracks the width actually
available instead of pinning to 1100px and leaving wide empty gutters.
The floor keeps the old measure on ~1280px laptops, the cap stops rows
stretching on ultrawides, and `max-width` never forces overflow, so the
@@ -594,30 +594,17 @@
box-shadow: none;
}
- /* On-device rows reuse the result-row card but carry selection/active state:
- a neutral tint for the inspected row and a success tint with a matching
- success edge for the model currently loaded. Discover rows never set these
- attributes. */
+ /* On-device rows reuse the result-row card with a neutral tint for the
+ inspected row. */
.hub-page .hub-result-row[data-selected="true"] {
background-color: color-mix(in srgb, var(--foreground) 5%, var(--card));
}
- .hub-page .hub-result-row[data-active="true"] {
- background-color: color-mix(in srgb, var(--status-success) 7%, var(--card));
- box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16),
- inset 2px 0 0 0 var(--status-success);
- }
-
html.dark .hub-page .hub-result-row[data-selected="true"] {
background-color: color-mix(in srgb, var(--foreground) 11%, var(--background));
box-shadow: none;
}
- html.dark .hub-page .hub-result-row[data-active="true"] {
- background-color: color-mix(in srgb, var(--status-success) 14%, var(--background));
- box-shadow: inset 2px 0 0 0 var(--status-success);
- }
-
/* Recent-searches suggestion panel. The surface comes from
.menu-soft-surface; these rules add the row + control hover affordances
using the same neutral foreground mixes as the option menus and list
@@ -738,14 +725,6 @@
background-color: color-mix(in srgb, var(--foreground) 6%, transparent);
}
- /* Currently-loaded checkpoint: a brand-green left stripe plus a faint
- green wash, orthogonal to the neutral selected tint so a row that is
- both loaded and selected still reads as loaded. */
- html:not(.dark) .hub-page .catalog-row[data-active="true"] {
- background-color: color-mix(in srgb, var(--status-success) 7%, transparent);
- box-shadow: inset 2px 0 0 0 var(--status-success);
- }
-
html.dark .hub-page .catalog-row {
background-color: transparent;
border-color: transparent;
@@ -760,11 +739,6 @@
html.dark .hub-page .catalog-row[data-selected="true"] {
background-color: color-mix(in srgb, var(--foreground) 7%, transparent);
}
-
- html.dark .hub-page .catalog-row[data-active="true"] {
- background-color: color-mix(in srgb, var(--status-success) 12%, transparent);
- box-shadow: inset 2px 0 0 0 var(--status-success);
- }
}
@layer utilities {
@@ -1110,7 +1084,7 @@
margin-block: 0;
}
- /* Inspector action pills (Download / Run / Train / Cancel) for the hub
+ /* Inspector action pills (Download / Update / Cancel) for the hub
inspector card. Centralized so all three sections (GGUF, safetensors,
datasets) stay in lock-step on type scale, padding, hover surface, and
the 16px icon size — mirrors the pattern used for `.sidebar-nav-btn`.
@@ -1158,12 +1132,6 @@
flex-shrink: 0;
}
- /* Primary "Run" action — confirms loading/running a model. Pill shape
- (rounded-full) so it reads as the unmistakable CTA in a row of
- ghost actions and matches the live-model status-success token used
- elsewhere in the app. Same h-9 geometry, 13px
- medium type, tracking-tight, and 16px icon slot as .hub-action-btn
- so the height aligns with adjacent ghost buttons. */
.hub-run-action-btn {
display: inline-flex;
height: 2.25rem;
@@ -1182,8 +1150,9 @@
letter-spacing: -0.025em;
color: var(--status-success-foreground);
transition: background-color 150ms, transform 150ms;
- box-shadow: inset 0 1px 0
- color-mix(in srgb, var(--status-success-foreground) 18%, transparent),
+ box-shadow:
+ inset 0 1px 0
+ color-mix(in srgb, var(--status-success-foreground) 18%, transparent),
0 1px 2px color-mix(in srgb, var(--status-success) 22%, transparent);
}
@@ -1196,8 +1165,13 @@
}
.hub-run-action-btn:disabled {
- cursor: not-allowed;
- opacity: 0.5;
+ cursor: wait;
+ opacity: 0.78;
+ }
+
+ .hub-run-action-btn:focus-visible {
+ outline: 2px solid var(--ring);
+ outline-offset: 2px;
}
.hub-run-action-btn svg {
diff --git a/studio/frontend/src/features/hub/index.ts b/studio/frontend/src/features/hub/index.ts
index a828daa1b06..70e557c2c59 100644
--- a/studio/frontend/src/features/hub/index.ts
+++ b/studio/frontend/src/features/hub/index.ts
@@ -78,7 +78,6 @@ export {
localSourceLabel,
normalizeCapabilities,
normalizeModelFormat,
- normalizeRuntime,
normalizeTimestamp,
partialSetFromRows,
removeScanFolder,
diff --git a/studio/frontend/src/features/hub/inventory/index.ts b/studio/frontend/src/features/hub/inventory/index.ts
index b88a796baf5..1fbe6734c26 100644
--- a/studio/frontend/src/features/hub/inventory/index.ts
+++ b/studio/frontend/src/features/hub/inventory/index.ts
@@ -48,7 +48,6 @@ export {
localSourceLabel,
normalizeCapabilities,
normalizeModelFormat,
- normalizeRuntime,
} from "./view-models";
export {
epochMillisecondsToSeconds,
diff --git a/studio/frontend/src/features/hub/inventory/settings-identity.ts b/studio/frontend/src/features/hub/inventory/settings-identity.ts
deleted file mode 100644
index bf054f6cf58..00000000000
--- a/studio/frontend/src/features/hub/inventory/settings-identity.ts
+++ /dev/null
@@ -1,21 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-// How an inventory row maps onto the identity its saved settings are keyed by.
-
-import type { CachedInventoryRow, LocalInventoryRow } from "./types";
-
-/**
- * The GGUF variant a settings page keys this row's config by, before any per-repo lookup. A
- * standalone `.gguf` has no quant to choose between, but the inventory still labels it from its
- * filename, and adopting that label would key its settings to `<path>:Q4_K_M` while every other
- * surface uses the bare path, leaving two configs for one file.
- */
-export function settingsGgufVariantForRow(
- row: CachedInventoryRow | LocalInventoryRow,
-): string | null {
- if (row.kind === "local" && row.path.toLowerCase().endsWith(".gguf")) {
- return null;
- }
- return row.formatVariant?.trim() || null;
-}
diff --git a/studio/frontend/src/features/hub/inventory/types.ts b/studio/frontend/src/features/hub/inventory/types.ts
index b634bc88c49..a694cac7ac5 100644
--- a/studio/frontend/src/features/hub/inventory/types.ts
+++ b/studio/frontend/src/features/hub/inventory/types.ts
@@ -10,12 +10,6 @@ export type ModelInventoryFormat =
| "adapter"
| "checkpoint"
| "unknown";
-export type ModelInventoryRuntime =
- | "llama_cpp"
- | "transformers"
- | "adapter"
- | "unknown";
-
export interface ModelInventoryCapabilities {
canTrain: boolean;
canChat: boolean;
@@ -43,7 +37,6 @@ export interface CachedInventoryRow {
repo: string;
isGguf: boolean;
modelFormat: ModelInventoryFormat;
- runtime: ModelInventoryRuntime;
formatVariant?: string | null;
capabilities: ModelInventoryCapabilities;
bytes: number;
@@ -88,7 +81,6 @@ export interface LocalInventoryRow {
path: string;
isGguf: boolean;
modelFormat: ModelInventoryFormat;
- runtime: ModelInventoryRuntime;
formatVariant?: string | null;
capabilities: ModelInventoryCapabilities;
baseModel?: string | null;
diff --git a/studio/frontend/src/features/hub/inventory/use-hub-inventory.ts b/studio/frontend/src/features/hub/inventory/use-hub-inventory.ts
index 6bbdc09fcdb..54fc61f1c32 100644
--- a/studio/frontend/src/features/hub/inventory/use-hub-inventory.ts
+++ b/studio/frontend/src/features/hub/inventory/use-hub-inventory.ts
@@ -266,12 +266,6 @@ function liveDownloadInventoryRows(
inventory_id: cachedInventoryId(modelFormat, job.repoId),
load_id: job.repoId,
model_format: modelFormat,
- runtime:
- modelFormat === "gguf"
- ? "llama_cpp"
- : modelFormat === "safetensors"
- ? "transformers"
- : "unknown",
size_bytes: job.displayBytes,
// live jobs already use epoch milliseconds
last_modified: job.startedAt,
@@ -547,7 +541,6 @@ export function useHubInventory(
isGguf: false,
loadId: ds.id,
modelFormat: "unknown" as const,
- runtime: "unknown" as const,
formatVariant: null,
capabilities: defaultCapabilities("unknown"),
updatedAt: normalizeTimestamp(ds.updated_at),
diff --git a/studio/frontend/src/features/hub/inventory/view-models.ts b/studio/frontend/src/features/hub/inventory/view-models.ts
index 2951a8a4306..5bb72c15510 100644
--- a/studio/frontend/src/features/hub/inventory/view-models.ts
+++ b/studio/frontend/src/features/hub/inventory/view-models.ts
@@ -9,7 +9,6 @@ import type {
BackendModelCapabilities,
LocalModelInfo,
ModelInventoryFormat,
- ModelInventoryRuntime,
} from "./api";
import type {
ModelInventoryCapabilities,
@@ -88,26 +87,6 @@ export function normalizeModelFormat(
return fallback;
}
-export function normalizeRuntime(
- value: string | null | undefined,
- modelFormat: ModelInventoryFormat,
-): ModelInventoryRuntime {
- if (
- value === "llama_cpp" ||
- value === "transformers" ||
- value === "adapter" ||
- value === "unknown"
- ) {
- return value;
- }
- if (modelFormat === "gguf") return "llama_cpp";
- if (modelFormat === "adapter") return "adapter";
- if (modelFormat === "safetensors" || modelFormat === "checkpoint") {
- return "transformers";
- }
- return "unknown";
-}
-
export function defaultCapabilities(
modelFormat: ModelInventoryFormat,
partial = false,
@@ -188,7 +167,6 @@ export function buildCachedInventoryRow(
inventory_id?: string | null;
load_id?: string | null;
model_format?: ModelInventoryFormat | null;
- runtime?: string | null;
format_variant?: string | null;
capabilities?: BackendModelCapabilities | null;
last_modified?: number | null;
@@ -225,10 +203,6 @@ export function buildCachedInventoryRow(
repo: row.repo_id.includes("/") ? repoOf(row.repo_id) : row.repo_id,
isGguf: modelFormat === "gguf",
modelFormat,
- runtime: normalizeRuntime(
- inferredFromEndpoint ? null : row.runtime,
- modelFormat,
- ),
formatVariant: row.format_variant ?? null,
capabilities,
bytes: row.size_bytes,
@@ -309,7 +283,6 @@ export function buildLocalInventoryRows(
path: model.path,
isGguf: modelFormat === "gguf",
modelFormat,
- runtime: normalizeRuntime(model.runtime, modelFormat),
formatVariant: model.format_variant ?? null,
capabilities,
baseModel,
diff --git a/studio/frontend/src/features/hub/lib/adopt-inference-status.ts b/studio/frontend/src/features/hub/lib/adopt-inference-status.ts
index dfe4502d06f..7ea393e6779 100644
--- a/studio/frontend/src/features/hub/lib/adopt-inference-status.ts
+++ b/studio/frontend/src/features/hub/lib/adopt-inference-status.ts
@@ -1,10 +1,8 @@
// SPDX-License-Identifier: AGPL-3.0-only
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-// Adopting the resident model into the chat runtime store from the Hub. Landing straight on /hub
-// is the one entry point where nothing has applied /api/inference/status yet, and pinning only
-// the checkpoint leaves every other useActiveModelConfig field at its default, which the settings
-// page would pass on as the live config. So adoption applies the whole status, as chat does.
+// Landing straight on /hub is the one entry point where nothing else applies
+// /api/inference/status. Cache mutation guards need the same resident state as Chat.
import { ggufVariantsMatch, modelIdsMatch } from "./model-identity.ts";
@@ -14,7 +12,7 @@ export interface ResidentAdoptionState {
/** Whether that checkpoint names an external provider's model. */
checkpointIsExternal: boolean;
activeGgufVariant: string | null;
- /** ``modelLoading``: a load this tab started still owns the store. */
+ /** ``modelLoading``: a model load still owns the store. */
modelLoading: boolean;
/** Whether the idle-unload loop will actually free the model. From
* ``/openai-auto-switch``; ``/status`` says nothing about it. */
@@ -36,8 +34,7 @@ export interface ResidentStatusFacts {
export interface ResidentAdoptionActions {
/** Re-pin ``params.checkpoint`` onto the resident model. */
setCheckpoint: (checkpointId: string, ggufVariant: string | null) => void;
- /** Drop a checkpoint the server has genuinely unloaded. Optional, so a caller
- * that only wants the pinning half can leave it out. */
+ /** Drop a checkpoint the server has genuinely unloaded. */
clearCheckpoint?: () => void;
/**
* Apply the rest of the status. Receives the store values from BEFORE ``setCheckpoint`` ran,
@@ -59,23 +56,20 @@ export function adoptResidentModelStatus(
actions: ResidentAdoptionActions,
): boolean {
const { checkpointId } = status;
- // An external selection has no local mirror, so the resident GGUF's settings would
- // describe a model the user is not talking to.
+ // An external selection has no local mirror.
if (state.checkpointIsExternal) {
return false;
}
- // A load this tab started applies its own status when it settles, and owns the params
- // meanwhile. Adopting underneath it would fight both.
+ // A load applies its own status when it settles and owns the store meanwhile.
if (state.modelLoading) {
return false;
}
if (!checkpointId) {
// An empty status means one of two things and /status cannot say which. Armed, the idle
// loop frees the model but keeps a stash the next request reloads, so clearing would drop
- // a selection that is coming back. Disarmed, nothing brings it back, and leaving the row
- // resident seeds the settings editor from a launch config nothing is running.
+ // a selection that is coming back. Disarmed, nothing brings it back.
// A speech model is neither: an Audio load took the slot, not the idle loop, and no
- // stash reloads the chat model. Holding the pick left the Hub calling it Loaded.
+ // stash reloads the chat model. Keeping the prior pick would leave stale Hub residency.
if (state.idleUnloadArmed && !status.speechOnly) {
return false;
}
diff --git a/studio/frontend/src/features/hub/lib/gguf-variant-sort.ts b/studio/frontend/src/features/hub/lib/gguf-variant-sort.ts
index 7fe155f36c1..479a47f65da 100644
--- a/studio/frontend/src/features/hub/lib/gguf-variant-sort.ts
+++ b/studio/frontend/src/features/hub/lib/gguf-variant-sort.ts
@@ -92,17 +92,46 @@ export function sortDownloadableGgufVariants(
export function sortLocalGgufVariants(
variants: readonly GgufVariantDetail[],
options: GgufFitInput & {
- activeGgufVariant?: string | null;
defaultVariant?: string | null;
},
): GgufVariantDetail[] {
+ const defaultVariant = options.defaultVariant?.trim();
return [...variants].sort((a, b) => {
- const aActive = ggufVariantsMatch(a.quant, options.activeGgufVariant);
- const bActive = ggufVariantsMatch(b.quant, options.activeGgufVariant);
- if (aActive !== bActive) return aActive ? -1 : 1;
- const aDefault = ggufVariantsMatch(a.quant, options.defaultVariant);
- const bDefault = ggufVariantsMatch(b.quant, options.defaultVariant);
- if (aDefault !== bDefault) return aDefault ? -1 : 1;
+ if (defaultVariant) {
+ const aDefault = ggufVariantsMatch(a.quant, defaultVariant);
+ const bDefault = ggufVariantsMatch(b.quant, defaultVariant);
+ if (aDefault !== bDefault) return aDefault ? -1 : 1;
+ }
return compareGgufVariantFitAndSize(a, b, options);
});
}
+
+export function resolveLocalGgufVariant<T extends { quant: string }>(
+ variants: readonly T[] | null | undefined,
+ options: {
+ selectedVariant?: string | null;
+ activeVariant?: string | null;
+ defaultVariant?: string | null;
+ },
+): T | null {
+ if (!variants || variants.length === 0) {
+ return null;
+ }
+ for (const candidate of [
+ options.selectedVariant,
+ options.activeVariant,
+ options.defaultVariant,
+ ]) {
+ // Blank candidates must not match a blank variant key.
+ if (!candidate?.trim()) {
+ continue;
+ }
+ const match = variants.find((variant) =>
+ ggufVariantsMatch(variant.quant, candidate),
+ );
+ if (match) {
+ return match;
+ }
+ }
+ return variants[0] ?? null;
+}
diff --git a/studio/frontend/src/features/hub/lib/hub-feature-flags.ts b/studio/frontend/src/features/hub/lib/hub-feature-flags.ts
deleted file mode 100644
index 6c74f71d1bf..00000000000
--- a/studio/frontend/src/features/hub/lib/hub-feature-flags.ts
+++ /dev/null
@@ -1,16 +0,0 @@
-// SPDX-License-Identifier: AGPL-3.0-only
-// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-
-// Hub feature flags for staged rollout. JSX/wiring stay in place so flipping
-// a flag here is the only edit needed to re-enable.
-
-// Post-download New Chat / Use in chat / Train CTAs; hidden until the
-// Hub-aware chat and train pickers ship. Run CTAs are gated separately below.
-export const HUB_POST_DOWNLOAD_ACTIONS_VISIBLE = false;
-
-export const HUB_GGUF_RUN_ACTIONS_VISIBLE = true;
-
-// Post-download Run CTA for non-GGUF models (MLX repos classify as safetensors).
-// Run has no dependency on the Hub-aware chat/train pickers, so it enables
-// independently of HUB_POST_DOWNLOAD_ACTIONS_VISIBLE, which still gates those.
-export const HUB_NON_GGUF_RUN_ACTIONS_VISIBLE = true;
diff --git a/studio/frontend/src/features/hub/lib/model-run-selection.ts b/studio/frontend/src/features/hub/lib/model-run-selection.ts
new file mode 100644
index 00000000000..1a80bb59db5
--- /dev/null
+++ b/studio/frontend/src/features/hub/lib/model-run-selection.ts
@@ -0,0 +1,172 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { ModelConfigHandoffRequest } from "@/features/model-picker";
+import type { SelectedModelView } from "../types";
+import { EMBEDDING_TAGS } from "./hf-model-meta.ts";
+import { routableToMediaPage } from "./local-path.ts";
+
+export interface HubModelRunSelection {
+ ggufVariant?: string;
+ ggufFilename?: string;
+ expectedBytes?: number;
+}
+
+function isPresent(value: string | null | undefined): value is string {
+ return Boolean(value?.trim());
+}
+
+function hasSupportedFormat(model: SelectedModelView): boolean {
+ if (model.modelFormat === "gguf") {
+ return model.isGguf;
+ }
+ return (
+ (model.modelFormat === "safetensors" ||
+ model.modelFormat === "checkpoint" ||
+ model.modelFormat === "adapter") &&
+ !model.isGguf
+ );
+}
+
+const GENERATIVE_CAPABILITIES = new Set([
+ "conversational",
+ "tools",
+ "reasoning",
+ "code",
+ "vision",
+ "audio",
+ "diffusion",
+]);
+
+function isEmbeddingOnly(model: SelectedModelView): boolean {
+ if (
+ model.isGguf ||
+ !model.capabilities.some((capability) => capability.key === "embedding")
+ ) {
+ return false;
+ }
+ if (EMBEDDING_TAGS.has(model.pipelineTag?.trim().toLowerCase() ?? "")) {
+ return true;
+ }
+ return !model.capabilities.some((capability) =>
+ GENERATIVE_CAPABILITIES.has(capability.key),
+ );
+}
+
+function hasCompleteInventoryModel(model: SelectedModelView): boolean {
+ return model.isDownloaded && !model.isPartial && isPresent(model.loadId);
+}
+
+function hasRunnableChatModel(model: SelectedModelView): boolean {
+ return (
+ model.runtimeCanChat &&
+ hasCompleteInventoryModel(model) &&
+ !isEmbeddingOnly(model) &&
+ hasSupportedFormat(model)
+ );
+}
+
+function hasValidSelection(
+ model: SelectedModelView,
+ selection: HubModelRunSelection,
+): boolean {
+ const hasGgufVariant = isPresent(selection.ggufVariant);
+ const hasGgufFilename = isPresent(selection.ggufFilename);
+ if (model.modelFormat !== "gguf") {
+ return !(hasGgufVariant || hasGgufFilename);
+ }
+ return !model.requiresVariant || hasGgufVariant;
+}
+
+function createHandoffMeta(
+ model: SelectedModelView,
+ selection: HubModelRunSelection,
+ usesLocalIdentity: boolean,
+): ModelConfigHandoffRequest["meta"] {
+ const meta: ModelConfigHandoffRequest["meta"] = {
+ source: usesLocalIdentity ? "local" : "hub",
+ isLora: model.modelFormat === "adapter",
+ loadId: model.loadId,
+ isDownloaded: true,
+ isGguf: model.modelFormat === "gguf",
+ pipelineTag: model.task ?? model.pipelineTag ?? null,
+ };
+ if (isPresent(selection.ggufVariant)) {
+ meta.ggufVariant = selection.ggufVariant;
+ }
+ if (isPresent(selection.ggufFilename)) {
+ meta.ggufFilename = selection.ggufFilename;
+ }
+ if (
+ selection.expectedBytes != null &&
+ Number.isFinite(selection.expectedBytes) &&
+ selection.expectedBytes > 0
+ ) {
+ meta.expectedBytes = selection.expectedBytes;
+ }
+ return meta;
+}
+
+export function isHubModelRunEligible({
+ model,
+ isDataset,
+ mediaRuntime,
+ nonGgufRuntimeAvailable,
+}: {
+ model: SelectedModelView | null;
+ isDataset: boolean;
+ mediaRuntime: boolean;
+ nonGgufRuntimeAvailable: boolean;
+}): boolean {
+ if (!model || isDataset) {
+ return false;
+ }
+
+ if (mediaRuntime) {
+ return (
+ hasCompleteInventoryModel(model) &&
+ isPresent(model.hubRepoId) &&
+ routableToMediaPage(model.kind, model.localSource)
+ );
+ }
+
+ if (!hasRunnableChatModel(model)) {
+ return false;
+ }
+
+ return model.modelFormat === "gguf" || nonGgufRuntimeAvailable;
+}
+
+export function createHubModelConfigHandoff({
+ requestId,
+ model,
+ selection,
+}: {
+ requestId: string;
+ model: SelectedModelView;
+ selection: HubModelRunSelection;
+}): ModelConfigHandoffRequest | null {
+ if (!isPresent(requestId)) {
+ return null;
+ }
+ if (!hasRunnableChatModel(model)) {
+ return null;
+ }
+
+ if (!hasValidSelection(model, selection)) {
+ return null;
+ }
+
+ const usesLocalIdentity = model.isLocal && model.localSource !== "hf_cache";
+ const id = usesLocalIdentity ? model.id : (model.hubRepoId ?? model.id);
+ if (!isPresent(id)) {
+ return null;
+ }
+
+ return {
+ requestId,
+ id,
+ displayName: model.title,
+ meta: createHandoffMeta(model, selection, usesLocalIdentity),
+ };
+}
diff --git a/studio/frontend/src/features/hub/lib/resident-status-refresh.ts b/studio/frontend/src/features/hub/lib/resident-status-refresh.ts
index 2fb0fe14466..7178c3c5a50 100644
--- a/studio/frontend/src/features/hub/lib/resident-status-refresh.ts
+++ b/studio/frontend/src/features/hub/lib/resident-status-refresh.ts
@@ -1,11 +1,8 @@
// SPDX-License-Identifier: AGPL-3.0-only
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-// When the Hub has to re-read /api/inference/status. An API request can auto-switch the resident
-// model at any moment and nothing else on /hub reads status, so a mount-only read leaves every
-// "loaded" marker (and the settings page's live config) pinned to whatever was resident when the
-// Hub opened. A background timer would keep asking a server that usually has not changed, so
-// re-read only on the moments this tab could have missed a switch.
+// An API request can switch the resident model while the Hub is open. Re-read
+// only when this tab could have missed a switch, rather than polling.
/** The event targets to listen on; injected so this is testable off a browser. */
export interface ResidentStatusRefreshTargets {
diff --git a/studio/frontend/src/features/hub/lib/superseded-refresh.ts b/studio/frontend/src/features/hub/lib/superseded-refresh.ts
index 2c68333939e..fd7dc53c672 100644
--- a/studio/frontend/src/features/hub/lib/superseded-refresh.ts
+++ b/studio/frontend/src/features/hub/lib/superseded-refresh.ts
@@ -1,17 +1,8 @@
// SPDX-License-Identifier: AGPL-3.0-only
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-// A caller that awaits a status refresh is waiting for the store to be settled, not for one
-// particular HTTP response. Refreshes are sequenced so the newest read owns the store, and a
-// response that lands after a newer read started writes nothing. Resolving that dropped
-// response's promise anyway releases its caller onto the store the newer read has not written
-// yet, which is exactly the pre-read store the await was there to replace: the Hub's settings
-// handlers would then resolve a quant from the model an API switch already displaced and bake
-// it into the settings target, and Apply reloads and persists under whatever the target names.
-// A settings open is guarded against a newer settings open, but nothing bumps that guard for a
-// focus or visibilitychange refresh, so those are the reads that can supersede one unnoticed.
-//
-// So a dropped response resolves with the refresh that superseded it instead of with nothing.
+// A caller waiting for resident state must wait for the newest refresh, not only
+// for an older response that was dropped after a focus or visibility refresh.
/** The newest refresh started, so a superseded one can hand its caller that promise. */
export interface RefreshSupersession {
diff --git a/studio/frontend/src/features/hub/types.ts b/studio/frontend/src/features/hub/types.ts
index 64e160e8a12..046f726276a 100644
--- a/studio/frontend/src/features/hub/types.ts
+++ b/studio/frontend/src/features/hub/types.ts
@@ -4,7 +4,6 @@
import type {
BaseModelSource,
LocalSource,
- ModelInventoryCapabilities,
ModelInventoryFormat,
} from "@/features/hub/inventory";
import type { HfModelResult } from "@/features/hub/hooks/use-hub-model-search";
@@ -39,27 +38,10 @@ export interface DiscoverRow {
capabilities: Capability[];
}
-export type SelectedResourceSource = "huggingface" | "hub_cache" | LocalSource;
-
-export type SelectedResourceCacheState =
- | "remote"
- | "cached"
- | "local"
- | "partial";
-
-export interface SelectedResourceRef {
- repoId: string | null;
- localPath: string | null;
- source: SelectedResourceSource;
- cacheState: SelectedResourceCacheState;
- runId: string;
- trainId: string;
-}
-
export interface SelectedModelView {
id: string;
+ loadId: string | null;
kind: "discover" | "cache" | "local";
- resource: SelectedResourceRef;
displayId: string;
hubRepoId: string | null;
owner: string;
@@ -79,10 +61,10 @@ export interface SelectedModelView {
adapterType?: string | null;
trainingMethod?: string | null;
isDownloaded: boolean;
+ runtimeCanChat: boolean;
isPartial?: boolean;
partialTransport?: string | null;
partialResumable?: boolean;
- runtimeCapabilities?: ModelInventoryCapabilities;
capabilities: Capability[];
license: string | null;
pipelineTag?: string;
diff --git a/studio/frontend/src/features/model-picker/components/model-selector.tsx b/studio/frontend/src/features/model-picker/components/model-selector.tsx
index 2f81ff30250..2c46a2180bd 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector.tsx
@@ -34,7 +34,14 @@ import {
useState,
} from "react";
import {
- isOllamaLinkPath,
+ type ModelConfigHandoffRequest,
+ modelConfigTarget,
+ modelConfigTargetIsResident,
+ modelConfigTargetMatchesSelection,
+} from "../model-config/model-config-handoff";
+import { modelConfigInstanceKey } from "../model-config/config-signature";
+import {
+ ggufVariantsMatch,
modelDisplayName,
} from "../model-config/model-identity";
import {
@@ -153,6 +160,8 @@ interface ModelSelectorProps {
activeLoadedContextLength?: number | null;
selectedConfig?: PerModelConfig | null;
selectedGgufVariant?: string | null;
+ configRequest?: ModelConfigHandoffRequest | null;
+ onConfigRequestAdopted?: (requestId: string) => void;
onValueChange?: (value: string, meta: ModelSelectorChangeMeta) => void;
/** Optional task-specific resolver for companion assets a GGUF row alone cannot describe. */
resolveDownloadFootprint?: ModelDownloadFootprintResolver;
@@ -358,11 +367,14 @@ function ModelSelectorContent({
loraModels,
externalModels,
value,
+ loaded,
activeGgufVariant,
activeModelConfig,
activeLoadedContextLength,
selectedConfig,
selectedGgufVariant,
+ configRequest,
+ onConfigRequestAdopted,
onSelect,
resolveDownloadFootprint,
onEject,
@@ -383,11 +395,14 @@ function ModelSelectorContent({
loraModels: LoraModelOption[];
externalModels: ExternalModelOption[];
value?: string;
+ loaded?: boolean;
activeGgufVariant?: string | null;
activeModelConfig?: PerModelConfig | null;
activeLoadedContextLength?: number | null;
selectedConfig?: PerModelConfig | null;
selectedGgufVariant?: string | null;
+ configRequest?: ModelConfigHandoffRequest | null;
+ onConfigRequestAdopted?: (requestId: string) => void;
onSelect: (id: string, meta: ModelSelectorChangeMeta) => void;
resolveDownloadFootprint?: ModelDownloadFootprintResolver;
onEject?: () => void;
@@ -445,6 +460,9 @@ function ModelSelectorContent({
const [configTarget, setConfigTarget] = useState<ModelPickTarget | null>(
null,
);
+ const [adoptedConfigRequestId, setAdoptedConfigRequestId] = useState<
+ string | null
+ >(null);
// The picker remounts on each open but this section state does not, so re-derive the default
// section on the open edge.
@@ -504,21 +522,51 @@ function ModelSelectorContent({
}
}
- const visibleConfigTarget = open ? configTarget : null;
const openConfigPage = (id: string, meta: ModelSelectorChangeMeta) => {
- const leaf = id.includes("/") ? id.slice(id.lastIndexOf("/") + 1) : id;
- const isGguf = meta.isGguf ?? Boolean(meta.ggufVariant);
- setConfigTarget({
- id,
- displayName: meta.ggufVariant ? `${leaf} · ${meta.ggufVariant}` : leaf,
- ggufVariant: meta.ggufVariant ?? null,
- isGguf,
- // Ollama's models sit under a link dir the resolver skips, so mirroring their settings would
- // advertise an impossible load.
- apiLoadable: isGguf && !isOllamaLinkPath(id),
- meta,
- });
+ setConfigTarget(modelConfigTarget(id, meta));
};
+ const requestedConfigTarget = useMemo(
+ () =>
+ open && configRequest
+ ? modelConfigTarget(
+ configRequest.id,
+ configRequest.meta,
+ configRequest.displayName,
+ )
+ : null,
+ [configRequest, open],
+ );
+ const visibleConfigTarget = open
+ ? (requestedConfigTarget ?? configTarget)
+ : null;
+ const visibleConfigMatchesSelection =
+ visibleConfigTarget !== null &&
+ modelConfigTargetMatchesSelection(visibleConfigTarget, value);
+ const visibleConfigIsResident =
+ visibleConfigTarget !== null &&
+ modelConfigTargetIsResident({
+ target: visibleConfigTarget,
+ selectedId: value,
+ activeGgufVariant,
+ loaded,
+ });
+ const visibleLoadedConfig = visibleConfigIsResident
+ ? (activeModelConfig ?? null)
+ : null;
+ if (
+ configRequest &&
+ requestedConfigTarget &&
+ adoptedConfigRequestId !== configRequest.requestId
+ ) {
+ setAdoptedConfigRequestId(configRequest.requestId);
+ setConfigTarget(requestedConfigTarget);
+ }
+ useEffect(() => {
+ if (!configRequest || adoptedConfigRequestId !== configRequest.requestId) {
+ return;
+ }
+ onConfigRequestAdopted?.(configRequest.requestId);
+ }, [adoptedConfigRequestId, configRequest, onConfigRequestAdopted]);
const handlePick = (id: string, meta: ModelSelectorChangeMeta) => {
if (meta.source === "external") {
onSelect(id, meta);
@@ -535,6 +583,11 @@ function ModelSelectorContent({
<PopoverContent
align="start"
alignOffset={10}
+ aria-label={
+ visibleConfigTarget
+ ? `Run settings for ${visibleConfigTarget.displayName}`
+ : undefined
+ }
data-tour={dataTour}
onKeyDown={handlePickerEntryKeyDown}
className={cn(
@@ -569,38 +622,39 @@ function ModelSelectorContent({
{visibleConfigTarget ? (
<div className="min-h-0 w-full overflow-y-auto px-4 pt-4 pb-4">
<ModelConfigPage
- key={`${visibleConfigTarget.id}::${visibleConfigTarget.ggufVariant ?? ""}`}
- target={visibleConfigTarget}
- onBack={() => setConfigTarget(null)}
- onRun={(config, isDiffusion) =>
- onSelect(visibleConfigTarget.id, {
- ...visibleConfigTarget.meta,
- config,
- isDiffusion,
- forceReload: true,
- })
- }
- loadedConfig={
- value === visibleConfigTarget.id &&
- (activeGgufVariant ?? null) ===
- (visibleConfigTarget.ggufVariant ?? null)
- ? (activeModelConfig ?? null)
- : null
- }
- loadedContextLength={
- value === visibleConfigTarget.id &&
- (activeGgufVariant ?? null) ===
- (visibleConfigTarget.ggufVariant ?? null)
- ? (activeLoadedContextLength ?? null)
- : null
- }
- initialConfig={
- value === visibleConfigTarget.id &&
- (selectedGgufVariant ?? null) ===
- (visibleConfigTarget.ggufVariant ?? null)
- ? (selectedConfig ?? null)
- : null
- }
+ key={modelConfigInstanceKey(
+ visibleConfigTarget.configId ?? visibleConfigTarget.id,
+ visibleConfigTarget.ggufVariant,
+ visibleLoadedConfig,
+ )}
+ target={visibleConfigTarget}
+ onBack={() => setConfigTarget(null)}
+ onRun={(config, isDiffusion) =>
+ onSelect(
+ visibleConfigTarget.configId ?? visibleConfigTarget.id,
+ {
+ ...visibleConfigTarget.meta,
+ config,
+ isDiffusion,
+ forceReload: true,
+ },
+ )
+ }
+ loadedConfig={visibleLoadedConfig}
+ loadedContextLength={
+ visibleConfigIsResident
+ ? (activeLoadedContextLength ?? null)
+ : null
+ }
+ initialConfig={
+ visibleConfigMatchesSelection &&
+ ggufVariantsMatch(
+ selectedGgufVariant,
+ visibleConfigTarget.ggufVariant,
+ )
+ ? (selectedConfig ?? null)
+ : null
+ }
/>
</div>
) : (
@@ -662,6 +716,8 @@ export function ModelSelector({
activeLoadedContextLength,
selectedConfig,
selectedGgufVariant,
+ configRequest,
+ onConfigRequestAdopted,
onValueChange,
resolveDownloadFootprint,
onEject,
@@ -828,11 +884,14 @@ export function ModelSelector({
loraModels={loraModels}
externalModels={externalModels}
value={selected}
+ loaded={loaded}
activeGgufVariant={activeGgufVariant}
activeModelConfig={activeModelConfig}
activeLoadedContextLength={activeLoadedContextLength}
selectedConfig={selectedConfig}
selectedGgufVariant={selectedGgufVariant}
+ configRequest={configRequest}
+ onConfigRequestAdopted={onConfigRequestAdopted}
onSelect={handleSelect}
resolveDownloadFootprint={resolveDownloadFootprint}
onEject={onEject ? handleEject : undefined}
diff --git a/studio/frontend/src/features/model-picker/components/model-selector/model-row-menu.tsx b/studio/frontend/src/features/model-picker/components/model-selector/model-row-menu.tsx
index fe08f9e1fed..c72a1d33d64 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector/model-row-menu.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector/model-row-menu.tsx
@@ -1,9 +1,8 @@
// SPDX-License-Identifier: AGPL-3.0-only
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
-// Condensed row actions for model rows: everything except the run-settings gear collapses into
-// one dots menu (pin, update, delete) so rows do not grow an icon strip. Mirrors the sidebar
-// chat rows' MoreVertical menu pattern.
+// Condensed row actions for model rows so pin, update, and delete do not grow into an icon strip.
+// Mirrors the sidebar chat rows' MoreVertical menu pattern.
import {
DropdownMenu,
@@ -30,7 +29,6 @@ import {
MoreVerticalIcon,
PinIcon,
PinOffIcon,
- Settings02Icon,
} from "@hugeicons/core-free-icons";
import { HugeiconsIcon } from "@hugeicons/react";
import { RefreshCw } from "lucide-react";
@@ -79,17 +77,11 @@ interface ModelRowMenuCachePath {
variant?: string;
}
-/** The model's settings page: load config plus what the API will apply. */
-interface ModelRowMenuSettings {
- onOpen: () => void;
-}
-
export function ModelRowMenu({
ariaLabel,
buttonClassName,
iconClassName,
cachePath,
- settings,
pin,
update,
del,
@@ -99,7 +91,6 @@ export function ModelRowMenu({
iconClassName?: string;
/** Enables "Reveal in Finder" for cached repos. */
cachePath?: ModelRowMenuCachePath;
- settings?: ModelRowMenuSettings;
pin?: ModelRowMenuPin;
update?: ModelRowMenuUpdate;
del?: ModelRowMenuDelete;
@@ -180,7 +171,7 @@ export function ModelRowMenu({
});
}, [cachePathRepoId, cachePathVariant]);
- if (!pin && !update && !del && !cachePath && !settings) return null;
+ if (!pin && !update && !del && !cachePath) return null;
return (
<>
@@ -209,21 +200,6 @@ export function ModelRowMenu({
sideOffset={2}
className="unsloth-plus-menu menu-flat-destructive w-48"
>
- {settings && (
- <DropdownMenuItem
- onSelect={(e) => {
- e.stopPropagation();
- settings.onOpen();
- }}
- >
- <HugeiconsIcon
- icon={Settings02Icon}
- strokeWidth={1.75}
- className="size-icon"
- />
- <span>Settings</span>
- </DropdownMenuItem>
- )}
{pin && (
<DropdownMenuItem
onSelect={(e) => {
diff --git a/studio/frontend/src/features/model-picker/components/sidebar-model-config.tsx b/studio/frontend/src/features/model-picker/components/sidebar-model-config.tsx
index c8ee44b0811..b786da8f192 100644
--- a/studio/frontend/src/features/model-picker/components/sidebar-model-config.tsx
+++ b/studio/frontend/src/features/model-picker/components/sidebar-model-config.tsx
@@ -44,8 +44,8 @@ export function SidebarModelConfig({
onReload,
}: SidebarModelConfigProps) {
// A standalone .gguf has no quant to choose between, but the loader labels it from its filename and
- // /status echoes that back. Keying settings by it would write "<path>:Q4_K_M" while every other
- // surface uses the bare path. Same rule as settingsGgufVariantForRow.
+ // /status echoes that back. Keying settings by it would write "<path>:Q4_K_M" while all other
+ // settings entry points use the bare path.
const settingsGgufVariant = isStandaloneGgufPath(modelId)
? null
: ggufVariant;
diff --git a/studio/frontend/src/features/model-picker/index.ts b/studio/frontend/src/features/model-picker/index.ts
index e1d67ae6747..ec924421eda 100644
--- a/studio/frontend/src/features/model-picker/index.ts
+++ b/studio/frontend/src/features/model-picker/index.ts
@@ -40,6 +40,17 @@ export type {
ModelSelectorChangeMeta,
} from "./components/model-selector";
export { modelConfigInstanceKey } from "./model-config/config-signature";
+export {
+ clearModelConfigHandoff,
+ createModelConfigHandoffRequestId,
+ modelConfigHandoffForDestination,
+ modelConfigTarget,
+ modelConfigTargetIsResident,
+ modelConfigTargetMatchesSelection,
+ requestModelConfigHandoff,
+ useModelConfigHandoffStore,
+ type ModelConfigHandoffRequest,
+} from "./model-config/model-config-handoff";
export {
applyModelLoadConfigToRuntime,
applyPerModelConfigToRuntime,
diff --git a/studio/frontend/src/features/model-picker/model-config/model-config-handoff.ts b/studio/frontend/src/features/model-picker/model-config/model-config-handoff.ts
new file mode 100644
index 00000000000..beddfa8700e
--- /dev/null
+++ b/studio/frontend/src/features/model-picker/model-config/model-config-handoff.ts
@@ -0,0 +1,170 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { create } from "zustand";
+import type {
+ ModelPickTarget,
+ ModelSelectorChangeMeta,
+} from "../components/model-selector/types";
+import {
+ ggufVariantsMatch,
+ isOllamaLinkPath,
+ isStandaloneGgufPath,
+ modelDisplayName,
+ residentModelIdMatches,
+} from "./model-identity";
+import { adoptLegacyConfigKey } from "./per-model-config";
+
+export interface ModelConfigHandoffRequest {
+ requestId: string;
+ id: string;
+ displayName?: string;
+ meta: ModelSelectorChangeMeta;
+}
+
+interface ModelConfigHandoffState {
+ request: ModelConfigHandoffRequest | null;
+ submit: (request: ModelConfigHandoffRequest) => void;
+ clear: (requestId: string) => void;
+}
+
+interface RandomUuidSource {
+ randomUUID?: () => string;
+}
+
+function fallbackRequestId(): string {
+ const random = Math.random().toString(36).slice(2, 10).padEnd(8, "0");
+ return `${Date.now()}-${random}`;
+}
+
+export function createModelConfigHandoffRequestId(
+ cryptoSource: RandomUuidSource | null = globalThis.crypto ?? null,
+): string {
+ if (typeof cryptoSource?.randomUUID !== "function") {
+ return fallbackRequestId();
+ }
+ try {
+ return cryptoSource.randomUUID();
+ } catch {
+ return fallbackRequestId();
+ }
+}
+
+export function modelConfigTarget(
+ id: string,
+ meta: ModelSelectorChangeMeta,
+ displayName?: string,
+): ModelPickTarget {
+ const name = displayName?.trim() || modelDisplayName(id);
+ const isGguf = meta.isGguf ?? Boolean(meta.ggufVariant);
+ const separateLoadId =
+ meta.source === "hub" && Boolean(meta.loadId?.trim()) && meta.loadId !== id
+ ? meta.loadId
+ : null;
+ const loadId = separateLoadId ?? id;
+ return {
+ id: loadId,
+ displayName: meta.ggufVariant ? `${name} · ${meta.ggufVariant}` : name,
+ ggufVariant: meta.ggufVariant ?? null,
+ isGguf,
+ apiLoadable: isGguf && !isOllamaLinkPath(id) && !isOllamaLinkPath(loadId),
+ ...(separateLoadId ? { configId: id } : {}),
+ meta,
+ };
+}
+
+export function modelConfigTargetMatchesSelection(
+ target: ModelPickTarget,
+ selectedId: string | null | undefined,
+): boolean {
+ return residentModelIdMatches(
+ selectedId,
+ target.configId,
+ target.id,
+ target.meta.loadId,
+ );
+}
+
+export function modelConfigTargetIsResident({
+ target,
+ selectedId,
+ activeGgufVariant,
+ loaded = true,
+}: {
+ target: ModelPickTarget;
+ selectedId: string | null | undefined;
+ activeGgufVariant: string | null | undefined;
+ loaded?: boolean;
+}): boolean {
+ if (!loaded) {
+ return false;
+ }
+ if (!modelConfigTargetMatchesSelection(target, selectedId)) {
+ return false;
+ }
+ if (
+ target.isGguf &&
+ target.ggufVariant == null &&
+ (isStandaloneGgufPath(target.id) || isOllamaLinkPath(target.id))
+ ) {
+ return true;
+ }
+ return ggufVariantsMatch(activeGgufVariant, target.ggufVariant);
+}
+
+export function modelConfigHandoffForDestination(
+ request: ModelConfigHandoffRequest | null,
+ destination: {
+ active: boolean;
+ newChatId?: string | null;
+ threadId?: string | null;
+ compareId?: string | null;
+ projectId?: string | null;
+ },
+): ModelConfigHandoffRequest | null {
+ if (!request) {
+ return null;
+ }
+ if (
+ !destination.active ||
+ destination.threadId ||
+ destination.compareId ||
+ destination.projectId ||
+ request.requestId !== destination.newChatId
+ ) {
+ return null;
+ }
+ return request;
+}
+
+export const useModelConfigHandoffStore = create<ModelConfigHandoffState>(
+ (set) => ({
+ request: null,
+ submit: (request) => set({ request }),
+ clear: (requestId) =>
+ set((state) =>
+ state.request?.requestId === requestId ? { request: null } : state,
+ ),
+ }),
+);
+
+export function requestModelConfigHandoff(
+ request: ModelConfigHandoffRequest,
+): void {
+ if (
+ request.meta.source === "hub" &&
+ request.meta.loadId?.trim() &&
+ request.meta.loadId !== request.id
+ ) {
+ adoptLegacyConfigKey(
+ request.id,
+ request.meta.loadId,
+ request.meta.ggufVariant,
+ );
+ }
+ useModelConfigHandoffStore.getState().submit(request);
+}
+
+export function clearModelConfigHandoff(requestId: string): void {
+ useModelConfigHandoffStore.getState().clear(requestId);
+}
diff --git a/studio/frontend/src/features/model-picker/model-config/model-identity.ts b/studio/frontend/src/features/model-picker/model-config/model-identity.ts
index c5e006d0958..b0ec746d845 100644
--- a/studio/frontend/src/features/model-picker/model-config/model-identity.ts
+++ b/studio/frontend/src/features/model-picker/model-config/model-identity.ts
@@ -13,10 +13,12 @@ export {
isNativeFileLabel,
isOllamaLinkPath,
isStandaloneGgufPath,
+ ggufVariantsMatch,
modelDisplayName,
normalizeGgufVariantIdentity,
normalizeModelIdentity,
publicModelId,
+ residentModelIdMatches,
} from "@/features/hub/lib/model-identity";
const MODEL_STORAGE_KEY_PREFIX = "v2:";
diff --git a/studio/frontend/tests/appearance-accent-vars.test.ts b/studio/frontend/tests/appearance-accent-vars.test.ts
index b8b77b00987..80731cae34c 100644
--- a/studio/frontend/tests/appearance-accent-vars.test.ts
+++ b/studio/frontend/tests/appearance-accent-vars.test.ts
@@ -252,21 +252,6 @@ test("a narrow valid band between custom and elevated surfaces is not skipped",
assert.ok(ratio(corrected, "#212121") >= 2.5);
});
-test("loaded-model stripes stay on the semantic success color", () => {
- const hubCss = readFileSync(
- new URL("../src/features/hub/hub.css", import.meta.url),
- "utf8",
- );
- const activeBlocks = [
- ...hubCss.matchAll(/\[data-active="true"\]\s*\{([^}]+)\}/g),
- ];
- assert.equal(activeBlocks.length, 4);
- for (const [, declarations = ""] of activeBlocks) {
- assert.match(declarations, /box-shadow:[^;]+var\(--status-success\)/);
- assert.doesNotMatch(declarations, /var\(--primary\)/);
- }
-});
-
test("resize-handle glows follow the primary token", () => {
const source = readFileSync(
new URL("../src/components/ui/resizable.tsx", import.meta.url),
diff --git a/studio/frontend/tests/chat-model-residency.test.ts b/studio/frontend/tests/chat-model-residency.test.ts
index b1389103e54..044ab7a3b91 100644
--- a/studio/frontend/tests/chat-model-residency.test.ts
+++ b/studio/frontend/tests/chat-model-residency.test.ts
@@ -140,14 +140,6 @@ test("the picker's Loaded badge asks residency, not the selection", () => {
);
});
-test("the hub cards ask residency before saying Loaded", () => {
- const hub = readFileSync(
- new URL("../src/features/hub/hub-page.tsx", import.meta.url),
- "utf8",
- );
- assert.match(hub, /residentCheckpoint !== null\s*\n?\s*\? checkpoint/);
-});
-
// Nothing in the chat runtime polls /status: refresh runs on mount and when the
// model lists change, never on a timer. So an eviction caused by the Images
// page was never observed and residentCheckpoint stayed undefined, which reads
diff --git a/studio/frontend/tests/hub-local-gguf-selection.test.ts b/studio/frontend/tests/hub-local-gguf-selection.test.ts
new file mode 100644
index 00000000000..56cbc536521
--- /dev/null
+++ b/studio/frontend/tests/hub-local-gguf-selection.test.ts
@@ -0,0 +1,145 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import { registerBundlerResolver } from "./helpers/kit.ts";
+
+registerBundlerResolver();
+
+const { resolveLocalGgufVariant, sortLocalGgufVariants } = await import(
+ "../src/features/hub/lib/gguf-variant-sort.ts"
+);
+
+const variants = [{ quant: "Q4_K_M" }, { quant: "Q8_0" }, { quant: "Q6_K" }];
+
+test("an explicit Hub quant selection takes priority", () => {
+ assert.equal(
+ resolveLocalGgufVariant(variants, {
+ selectedVariant: "q6_k",
+ activeVariant: "Q8_0",
+ defaultVariant: "Q4_K_M",
+ })?.quant,
+ "Q6_K",
+ );
+});
+
+test("the resident quant takes priority over the repository default", () => {
+ assert.equal(
+ resolveLocalGgufVariant(variants, {
+ activeVariant: "q8_0",
+ defaultVariant: "Q4_K_M",
+ })?.quant,
+ "Q8_0",
+ );
+});
+
+test("selection falls back through default, first variant, and empty state", () => {
+ assert.equal(
+ resolveLocalGgufVariant(variants, {
+ activeVariant: "missing",
+ defaultVariant: "q4_k_m",
+ })?.quant,
+ "Q4_K_M",
+ );
+ assert.equal(resolveLocalGgufVariant(variants, {})?.quant, "Q4_K_M");
+ assert.equal(resolveLocalGgufVariant([], {}), null);
+ assert.equal(resolveLocalGgufVariant(null, {}), null);
+});
+
+test("blank preferences preserve the first-variant fallback", () => {
+ const withBlankQuant = [{ quant: "Q4_K_M" }, { quant: "" }];
+ assert.equal(resolveLocalGgufVariant(withBlankQuant, {})?.quant, "Q4_K_M");
+ assert.equal(
+ resolveLocalGgufVariant(withBlankQuant, {
+ selectedVariant: null,
+ activeVariant: null,
+ defaultVariant: null,
+ })?.quant,
+ "Q4_K_M",
+ );
+ assert.equal(
+ resolveLocalGgufVariant(withBlankQuant, {
+ selectedVariant: " ",
+ activeVariant: "",
+ })?.quant,
+ "Q4_K_M",
+ );
+ assert.equal(
+ resolveLocalGgufVariant([{ quant: "" }, { quant: "Q8_0" }], {
+ selectedVariant: "Q8_0",
+ })?.quant,
+ "Q8_0",
+ );
+});
+
+const localVariants = [
+ { quant: "", filename: "blank-key.gguf", size_bytes: 100 * 1024 ** 3 },
+ { quant: "Q4_K_M", filename: "model-Q4_K_M.gguf", size_bytes: 4 * 1024 ** 3 },
+ { quant: "Q8_0", filename: "model-Q8_0.gguf", size_bytes: 8 * 1024 ** 3 },
+];
+
+test("sorting and selection ignore blank defaults", () => {
+ for (const defaultVariant of [undefined, null, "", " \t\n", "missing"]) {
+ const sorted = sortLocalGgufVariants(localVariants, {
+ defaultVariant,
+ gpuGb: 24,
+ });
+ assert.deepEqual(sorted, [
+ localVariants[2],
+ localVariants[1],
+ localVariants[0],
+ ]);
+ assert.equal(
+ resolveLocalGgufVariant(sorted, { defaultVariant }),
+ localVariants[2],
+ );
+ }
+ assert.deepEqual(
+ localVariants.map((variant) => variant.quant),
+ ["", "Q4_K_M", "Q8_0"],
+ );
+});
+
+test("blank selections preserve resident and default preferences after sorting", () => {
+ const defaultVariant = " q4_k_m ";
+ const sorted = sortLocalGgufVariants(localVariants, {
+ defaultVariant,
+ gpuGb: 24,
+ });
+ assert.equal(sorted[0], localVariants[1]);
+ for (const selectedVariant of [undefined, null, "", " \t\n", "missing"]) {
+ assert.equal(
+ resolveLocalGgufVariant(sorted, {
+ selectedVariant,
+ activeVariant: " q8_0 ",
+ defaultVariant,
+ }),
+ localVariants[2],
+ );
+ assert.equal(
+ resolveLocalGgufVariant(sorted, {
+ selectedVariant,
+ activeVariant: " ",
+ defaultVariant,
+ }),
+ localVariants[1],
+ );
+ }
+});
+
+test("a blank-quant variant remains eligible for the ranked fallback", () => {
+ const blank = {
+ quant: "",
+ filename: "blank-key.gguf",
+ size_bytes: 4 * 1024 ** 3,
+ };
+ const oversized = {
+ quant: "Q8_0",
+ filename: "model-Q8_0.gguf",
+ size_bytes: 100 * 1024 ** 3,
+ };
+ const sorted = sortLocalGgufVariants([oversized, blank], { gpuGb: 24 });
+ assert.equal(resolveLocalGgufVariant(sorted, {}), blank);
+});
diff --git a/studio/frontend/tests/hub-media-run-gate.test.ts b/studio/frontend/tests/hub-media-run-gate.test.ts
index 25373ed3d40..1ef8fb60389 100644
--- a/studio/frontend/tests/hub-media-run-gate.test.ts
+++ b/studio/frontend/tests/hub-media-run-gate.test.ts
@@ -4,37 +4,98 @@
import assert from "node:assert/strict";
import test from "node:test";
-import { routableToMediaPage } from "../src/features/hub/lib/local-path.ts";
+import type { SelectedModelView } from "../src/features/hub/types.ts";
import { studioPageForTask } from "../src/features/hub/lib/unsloth-support.ts";
+import { registerBundlerResolver } from "./helpers/kit.ts";
-// What the inspector must decide before enabling Run, and what runSelectedModel requires
-// before navigating. They have to agree: the handler falls through to the chat loader for a
-// row it cannot route, and that unloads the resident model for a load that can only fail.
-const runsOnMediaPage = (
- task: string | null | undefined,
+registerBundlerResolver();
+
+const { isHubModelRunEligible } = await import(
+ "../src/features/hub/lib/model-run-selection.ts"
+);
+const { taskForMediaPick } = await import(
+ "../src/features/model-picker/components/model-selector/audio-picker-policy.ts"
+);
+
+function mediaModel(
+ task: SelectedModelView["pipelineTag"],
kind: "discover" | "cache" | "local",
- localSource?: string | null,
-) => studioPageForTask(task) !== undefined && routableToMediaPage(kind, localSource);
+ localSource?: SelectedModelView["localSource"],
+): SelectedModelView {
+ const loadId = "/cache/models--Org--Model/snapshots/revision";
+ return {
+ id: kind === "local" ? loadId : "Org/Model",
+ loadId,
+ kind,
+ displayId: "Org/Model",
+ hubRepoId: "Org/Model",
+ owner: "Org",
+ title: "Model",
+ summary: "Model",
+ sourceLabel: "Hub cache",
+ path: loadId,
+ localSource,
+ isLocal: kind === "local",
+ isGguf: false,
+ modelFormat: "safetensors",
+ isDownloaded: true,
+ runtimeCanChat: false,
+ capabilities: [],
+ license: null,
+ pipelineTag: task,
+ };
+}
+
+function mediaRunEligible(model: SelectedModelView): boolean {
+ const task = taskForMediaPick(model.pipelineTag, model.task);
+ return isHubModelRunEligible({
+ model,
+ isDataset: false,
+ mediaRuntime: studioPageForTask(task ?? undefined) !== undefined,
+ nonGgufRuntimeAvailable: false,
+ });
+}
// The backend tags a local non-GGUF diffusers checkpoint text-to-image (_local_model_task),
// and it does that for every local row whatever its source, so these rows are real.
test("a filesystem diffusion row never counts as running on a media page", () => {
- assert.equal(runsOnMediaPage("text-to-image", "local", "models_dir"), false);
- assert.equal(runsOnMediaPage("text-to-image", "local", "lmstudio"), false);
- assert.equal(runsOnMediaPage("text-to-image", "local", "ollama"), false);
- assert.equal(runsOnMediaPage("text-to-video", "local", "custom"), false);
- assert.equal(runsOnMediaPage("text-to-image", "local", undefined), false);
+ assert.equal(
+ mediaRunEligible(mediaModel("text-to-image", "local", "models_dir")),
+ false,
+ );
+ assert.equal(
+ mediaRunEligible(mediaModel("text-to-image", "local", "lmstudio")),
+ false,
+ );
+ assert.equal(
+ mediaRunEligible(mediaModel("text-to-image", "local", "ollama")),
+ false,
+ );
+ assert.equal(
+ mediaRunEligible(mediaModel("text-to-video", "local", "custom")),
+ false,
+ );
+ assert.equal(mediaRunEligible(mediaModel("text-to-image", "local")), false);
});
-test("hub-backed diffusion rows stay runnable on their page", () => {
+test("complete Hub-backed diffusion rows stay runnable on their page", () => {
// An hf_cache row is a complete Hub snapshot, so it routes like a cached repo.
- assert.equal(runsOnMediaPage("text-to-image", "local", "hf_cache"), true);
- assert.equal(runsOnMediaPage("text-to-image", "cache"), true);
- assert.equal(runsOnMediaPage("image-text-to-video", "discover"), true);
+ assert.equal(
+ mediaRunEligible(mediaModel("text-to-image", "local", "hf_cache")),
+ true,
+ );
+ assert.equal(mediaRunEligible(mediaModel("text-to-image", "cache")), true);
+ assert.equal(
+ mediaRunEligible(mediaModel("image-text-to-video", "discover")),
+ true,
+ );
});
-test("a chat task is unaffected by the media gate", () => {
- assert.equal(runsOnMediaPage("text-generation", "cache"), false);
- assert.equal(runsOnMediaPage(null, "discover"), false);
- assert.equal(runsOnMediaPage("text-generation", "local", "hf_cache"), false);
+test("chat tasks are not eligible through the media route", () => {
+ assert.equal(mediaRunEligible(mediaModel("text-generation", "cache")), false);
+ assert.equal(mediaRunEligible(mediaModel(undefined, "discover")), false);
+ assert.equal(
+ mediaRunEligible(mediaModel("text-generation", "local", "hf_cache")),
+ false,
+ );
});
diff --git a/studio/frontend/tests/hub-model-run-selection.test.ts b/studio/frontend/tests/hub-model-run-selection.test.ts
new file mode 100644
index 00000000000..c8b9fe47775
--- /dev/null
+++ b/studio/frontend/tests/hub-model-run-selection.test.ts
@@ -0,0 +1,481 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import type { SelectedModelView } from "../src/features/hub/types.ts";
+import { registerBundlerResolver } from "./helpers/kit.ts";
+
+registerBundlerResolver();
+
+const { createHubModelConfigHandoff, isHubModelRunEligible } = await import(
+ "../src/features/hub/lib/model-run-selection.ts"
+);
+const { modelConfigTarget } = await import(
+ "../src/features/model-picker/model-config/model-config-handoff.ts"
+);
+
+function selectedModel(
+ overrides: Partial<SelectedModelView> = {},
+): SelectedModelView {
+ return {
+ id: "Org/Model",
+ loadId: "/cache/models--Org--Model/snapshots/revision",
+ kind: "cache",
+ displayId: "Org/Model",
+ hubRepoId: "Org/Model",
+ owner: "Org",
+ title: "Model",
+ summary: "Model",
+ sourceLabel: "Hub cache",
+ path: "/cache/models--Org--Model/snapshots/revision",
+ isLocal: false,
+ isGguf: false,
+ modelFormat: "safetensors",
+ isDownloaded: true,
+ runtimeCanChat: true,
+ capabilities: [],
+ license: null,
+ ...overrides,
+ };
+}
+
+function localModel(
+ id: string,
+ overrides: Partial<SelectedModelView> = {},
+): SelectedModelView {
+ return selectedModel({
+ id,
+ loadId: id,
+ kind: "local",
+ displayId: id,
+ hubRepoId: null,
+ path: id,
+ localSource: "custom",
+ isLocal: true,
+ ...overrides,
+ });
+}
+
+function eligible(
+ model: SelectedModelView,
+ overrides: Partial<{
+ isDataset: boolean;
+ mediaRuntime: boolean;
+ nonGgufRuntimeAvailable: boolean;
+ }> = {},
+): boolean {
+ return isHubModelRunEligible({
+ model,
+ isDataset: false,
+ mediaRuntime: false,
+ nonGgufRuntimeAvailable: true,
+ ...overrides,
+ });
+}
+
+test("only complete chat-loadable models in supported formats are eligible", () => {
+ assert.equal(eligible(selectedModel()), true);
+ assert.equal(
+ eligible(
+ selectedModel({
+ isGguf: true,
+ modelFormat: "gguf",
+ requiresVariant: true,
+ }),
+ { nonGgufRuntimeAvailable: false },
+ ),
+ true,
+ );
+ assert.equal(eligible(selectedModel(), { isDataset: true }), false);
+ assert.equal(
+ eligible(selectedModel(), { nonGgufRuntimeAvailable: false }),
+ false,
+ );
+ assert.equal(eligible(selectedModel({ runtimeCanChat: false })), false);
+ assert.equal(eligible(selectedModel({ isDownloaded: false })), false);
+ assert.equal(eligible(selectedModel({ isPartial: true })), false);
+ assert.equal(eligible(selectedModel({ loadId: null })), false);
+
+ assert.equal(eligible(selectedModel({ modelFormat: "checkpoint" })), true);
+ assert.equal(eligible(selectedModel({ modelFormat: "adapter" })), true);
+ assert.equal(
+ eligible(selectedModel({ modelFormat: "checkpoint" }), {
+ nonGgufRuntimeAvailable: false,
+ }),
+ false,
+ );
+ assert.equal(
+ eligible(selectedModel({ modelFormat: "adapter" }), {
+ nonGgufRuntimeAvailable: false,
+ }),
+ false,
+ );
+ assert.equal(eligible(selectedModel({ modelFormat: "unknown" })), false);
+});
+
+test("complete Hub-backed media models are eligible for their dedicated runtime", () => {
+ const mediaModel = selectedModel({
+ pipelineTag: "text-to-image",
+ runtimeCanChat: false,
+ });
+
+ assert.equal(
+ eligible(mediaModel, {
+ mediaRuntime: true,
+ nonGgufRuntimeAvailable: false,
+ }),
+ true,
+ );
+ assert.equal(
+ eligible(mediaModel, { mediaRuntime: true, isDataset: true }),
+ false,
+ );
+ assert.equal(
+ eligible(selectedModel({ isDownloaded: false }), { mediaRuntime: true }),
+ false,
+ );
+ assert.equal(
+ eligible(selectedModel({ isPartial: true }), { mediaRuntime: true }),
+ false,
+ );
+ assert.equal(
+ eligible(selectedModel({ hubRepoId: null }), { mediaRuntime: true }),
+ false,
+ );
+ assert.equal(
+ eligible(localModel("/models/image-model"), { mediaRuntime: true }),
+ false,
+ );
+ assert.equal(
+ eligible(
+ localModel("/inactive-cache/image-model", {
+ hubRepoId: "Org/Image-Model",
+ localSource: "hf_cache",
+ runtimeCanChat: false,
+ }),
+ { mediaRuntime: true },
+ ),
+ true,
+ );
+});
+
+test("embedding-only non-GGUF models never enter the Chat run flow", () => {
+ const embedding = { key: "embedding" as const, label: "Embeddings" };
+
+ for (const pipelineTag of [
+ "feature-extraction",
+ "sentence-similarity",
+ " text-embeddings-inference ",
+ ]) {
+ assert.equal(
+ eligible(selectedModel({ pipelineTag, capabilities: [embedding] })),
+ false,
+ pipelineTag,
+ );
+ }
+ assert.equal(eligible(selectedModel({ capabilities: [embedding] })), false);
+ assert.equal(
+ eligible(
+ selectedModel({
+ capabilities: [
+ embedding,
+ { key: "conversational", label: "Conversational" },
+ ],
+ }),
+ ),
+ true,
+ );
+ assert.equal(
+ eligible(
+ selectedModel({
+ isGguf: true,
+ modelFormat: "gguf",
+ pipelineTag: "feature-extraction",
+ capabilities: [embedding],
+ }),
+ ),
+ true,
+ );
+});
+
+test("checkpoint handoffs use the whole-model Chat identity", () => {
+ const request = createHubModelConfigHandoff({
+ requestId: "request-checkpoint",
+ model: selectedModel({ modelFormat: "checkpoint" }),
+ selection: {},
+ });
+
+ assert.equal(request?.id, "Org/Model");
+ assert.equal(request?.meta.isLora, false);
+ assert.equal(request?.meta.isGguf, false);
+ assert.equal(
+ createHubModelConfigHandoff({
+ requestId: "request-checkpoint-variant",
+ model: selectedModel({ modelFormat: "checkpoint" }),
+ selection: { ggufVariant: "Q4_K_M" },
+ }),
+ null,
+ );
+});
+
+test("adapter handoffs preserve loader identity and LoRA intent", () => {
+ const loadId = String.raw`C:\Models\adapter`;
+ const localRequest = createHubModelConfigHandoff({
+ requestId: "request-adapter",
+ model: localModel(loadId, {
+ modelFormat: "adapter",
+ }),
+ selection: {},
+ });
+
+ assert.deepEqual(localRequest, {
+ requestId: "request-adapter",
+ id: loadId,
+ displayName: "Model",
+ meta: {
+ source: "local",
+ isLora: true,
+ loadId,
+ isDownloaded: true,
+ isGguf: false,
+ pipelineTag: null,
+ },
+ });
+
+ const cachedLoadId =
+ "/inactive-cache/models--Org--Adapter/snapshots/revision";
+ const cachedRequest = createHubModelConfigHandoff({
+ requestId: "request-cached-adapter",
+ model: localModel(cachedLoadId, {
+ hubRepoId: "Org/Adapter",
+ localSource: "hf_cache",
+ modelFormat: "adapter",
+ }),
+ selection: {},
+ });
+
+ assert.ok(cachedRequest);
+ assert.equal(cachedRequest.id, "Org/Adapter");
+ assert.equal(cachedRequest.meta.source, "hub");
+ assert.equal(cachedRequest.meta.loadId, cachedLoadId);
+ assert.equal(cachedRequest.meta.isLora, true);
+
+ const target = modelConfigTarget(
+ cachedRequest.id,
+ cachedRequest.meta,
+ cachedRequest.displayName,
+ );
+ assert.equal(target.id, cachedLoadId);
+ assert.equal(target.configId, "Org/Adapter");
+ assert.equal(target.meta.isLora, true);
+});
+
+test("MLX safetensors follow runtime availability", () => {
+ const models = [
+ selectedModel({ libraryName: " MLX " }),
+ selectedModel({ tags: ["transformers", "MLX"] }),
+ selectedModel({
+ id: "mlx-community/Model",
+ hubRepoId: "mlx-community/Model",
+ }),
+ selectedModel({
+ id: "Org/Model-MLX-4bit",
+ hubRepoId: "Org/Model-MLX-4bit",
+ }),
+ localModel(String.raw`C:\Models\mlx-community\Qwen2`),
+ ];
+
+ for (const model of models) {
+ assert.equal(eligible(model), true);
+ assert.equal(eligible(model, { nonGgufRuntimeAvailable: false }), false);
+ assert.notEqual(
+ createHubModelConfigHandoff({
+ requestId: "request-mlx",
+ model,
+ selection: {},
+ }),
+ null,
+ );
+ }
+});
+
+test("GGUF remains runnable when its conversion retains an MLX identity", () => {
+ const models = [
+ selectedModel({
+ id: "mlx-community/Model-MLX-GGUF",
+ hubRepoId: "mlx-community/Model-MLX-GGUF",
+ isGguf: true,
+ modelFormat: "gguf",
+ requiresVariant: true,
+ tags: ["mlx", "gguf"],
+ libraryName: "mlx",
+ }),
+ localModel(String.raw`C:\Models\mlx-community\Model.gguf`, {
+ isGguf: true,
+ modelFormat: "gguf",
+ requiresVariant: true,
+ tags: ["mlx", "gguf"],
+ libraryName: "mlx",
+ }),
+ ];
+
+ for (const model of models) {
+ assert.equal(eligible(model), true);
+ assert.notEqual(
+ createHubModelConfigHandoff({
+ requestId: "request-gguf",
+ model,
+ selection: {
+ ggufVariant: "Q4_K_M",
+ ggufFilename: "model-Q4_K_M.gguf",
+ },
+ }),
+ null,
+ );
+ }
+});
+
+test("managed cache handoffs retain separate public, loader, and GGUF identities", () => {
+ const loadId = String.raw`C:\Users\alice\.cache\models--Org--Model\snapshots\revision`;
+ const ggufVariant = " builds/Q4 K M ";
+ const ggufFilename = " builds/My Model Q4 K M.gguf ";
+ const request = createHubModelConfigHandoff({
+ requestId: "request-1",
+ model: selectedModel({
+ id: "cache:gguf:Org%2FModel",
+ loadId,
+ isGguf: true,
+ modelFormat: "gguf",
+ requiresVariant: true,
+ task: "text-generation",
+ }),
+ selection: {
+ ggufVariant,
+ ggufFilename,
+ expectedBytes: 4096,
+ },
+ });
+
+ assert.deepEqual(request, {
+ requestId: "request-1",
+ id: "Org/Model",
+ displayName: "Model",
+ meta: {
+ source: "hub",
+ isLora: false,
+ loadId,
+ isDownloaded: true,
+ isGguf: true,
+ pipelineTag: "text-generation",
+ ggufVariant,
+ ggufFilename,
+ expectedBytes: 4096,
+ },
+ });
+});
+
+test("matched Discover rows keep the Chat picker's local configuration identity", () => {
+ const loadId = "/mnt/c/models/Org/Model";
+ const request = createHubModelConfigHandoff({
+ requestId: "request-2",
+ model: selectedModel({
+ id: loadId,
+ loadId,
+ kind: "local",
+ hubRepoId: "Org/Model",
+ isLocal: true,
+ localSource: "custom",
+ }),
+ selection: {},
+ });
+
+ assert.equal(request?.id, loadId);
+ assert.equal(request?.meta.source, "local");
+ assert.equal(request?.meta.loadId, loadId);
+});
+
+test("opaque local loader identities remain unchanged on every desktop path style", () => {
+ const loadIds = [
+ String.raw`C:\Models\Family\Model.gguf`,
+ String.raw`\\server\models\Model.gguf`,
+ "/mnt/c/Models/Family/Model.gguf",
+ "/home/alice/models/Model.gguf",
+ "/Users/alice/Models/Model.gguf",
+ "ollama-manifest:library/model:latest",
+ ];
+
+ for (const loadId of loadIds) {
+ const request = createHubModelConfigHandoff({
+ requestId: loadId,
+ model: selectedModel({
+ id: loadId,
+ loadId,
+ kind: "local",
+ hubRepoId: null,
+ path: loadId,
+ isLocal: true,
+ localSource: loadId.startsWith("ollama-") ? "ollama" : "custom",
+ isGguf: true,
+ modelFormat: "gguf",
+ }),
+ selection: {},
+ });
+
+ assert.equal(request?.id, loadId);
+ assert.equal(request?.meta.loadId, loadId);
+ }
+});
+
+test("opaque local identities retain the inventory display name", () => {
+ const request = createHubModelConfigHandoff({
+ requestId: "ollama-display",
+ model: selectedModel({
+ id: "ollama-manifest:%2Fhome%2Fhz%2F.ollama%2Fmodels%2Fmanifests%2Flibrary%2Fllama",
+ loadId:
+ "ollama-manifest:%2Fhome%2Fhz%2F.ollama%2Fmodels%2Fmanifests%2Flibrary%2Fllama",
+ title: "Llama 3.2",
+ kind: "local",
+ hubRepoId: null,
+ isLocal: true,
+ localSource: "ollama",
+ isGguf: true,
+ modelFormat: "gguf",
+ }),
+ selection: {},
+ });
+
+ assert.equal(request?.displayName, "Llama 3.2");
+});
+
+test("invalid format and variant combinations fail closed", () => {
+ assert.equal(
+ createHubModelConfigHandoff({
+ requestId: "missing-variant",
+ model: selectedModel({
+ isGguf: true,
+ modelFormat: "gguf",
+ requiresVariant: true,
+ }),
+ selection: {},
+ }),
+ null,
+ );
+ assert.equal(
+ createHubModelConfigHandoff({
+ requestId: "unexpected-variant",
+ model: selectedModel(),
+ selection: { ggufVariant: "Q4_K_M" },
+ }),
+ null,
+ );
+ assert.equal(
+ createHubModelConfigHandoff({
+ requestId: "partial",
+ model: selectedModel({ isPartial: true }),
+ selection: {},
+ }),
+ null,
+ );
+});
diff --git a/studio/frontend/tests/hub-resident-status.test.ts b/studio/frontend/tests/hub-resident-status.test.ts
index 9c201697476..fd4ef28fc26 100644
--- a/studio/frontend/tests/hub-resident-status.test.ts
+++ b/studio/frontend/tests/hub-resident-status.test.ts
@@ -38,9 +38,8 @@ function refusing(messages: {
};
}
-test("landing on the Hub applies the whole status, not just the checkpoint", () => {
- // Nothing else on /hub hydrates the runtime store, so pinning only the checkpoint
- // leaves useActiveModelConfig at its defaults and the settings page offers those.
+test("landing on the Hub synchronizes the full resident status", () => {
+ // Cache mutation guards read the same resident store as Chat.
const { calls, actions } = spies();
const adopted = adoptResidentModelStatus(RESIDENT, emptyStore(), actions);
assert.equal(adopted, true);
@@ -107,7 +106,7 @@ test("nothing is adopted when no model is loaded", () => {
});
test("an external-provider selection is left alone", () => {
- // No local mirror, so the resident GGUF's settings would describe the wrong model.
+ // An external selection has no local runtime mirror.
const { calls, actions } = spies();
const adopted = adoptResidentModelStatus(
RESIDENT,
@@ -134,8 +133,7 @@ test("a load in flight is not fought", () => {
});
test("an empty status drops the checkpoint when idle unload is disarmed", () => {
- // Nothing will bring the model back, so it really is gone. Leaving the row resident would
- // seed the settings editor from a launch config nothing is running.
+ // Nothing will bring the model back, so it really is gone.
const cleared: string[] = [];
const adopted = adoptResidentModelStatus(
{ checkpointId: null, ggufVariant: null },
@@ -174,8 +172,6 @@ test("an empty status leaves the checkpoint pinned while idle unload is armed",
});
test("a speech model in the slot clears the pick even while idle unload is armed", () => {
- // Not the idle eviction the rule above is about: an Audio load took the single slot and
- // no stash reloads the chat model, so holding the pick left the Hub calling it Loaded.
const cleared: string[] = [];
const adopted = adoptResidentModelStatus(
{ checkpointId: null, ggufVariant: null, speechOnly: true },
@@ -217,7 +213,7 @@ test("an empty status leaves an external pick alone", () => {
}),
refusing({
setCheckpoint: "an external pick is not a local model",
- applyStatus: "the resident GGUF's status describes another model",
+ applyStatus: "the resident GGUF belongs to another runtime",
}),
);
assert.equal(adopted, false);
@@ -268,10 +264,7 @@ test("a tab going hidden does not read", () => {
assert.equal(reads, 1);
});
-test("an auto-switch under a mounted Hub stops hiding the live config", () => {
- // End to end: an API request swaps the resident model under a mounted Hub. Without a second
- // read settingsTargetIsResident says it is not resident, so the editor seeds from saved
- // values that Apply then reloads over the API's choice.
+test("an auto-switch under a mounted Hub updates the mutation guard", () => {
const store = emptyStore({
checkpoint: "unsloth/Qwen3-8B-GGUF",
activeGgufVariant: "Q4_K_M",
@@ -295,8 +288,7 @@ test("an auto-switch under a mounted Hub stops hiding the live config", () => {
);
};
- // hub-page.tsx's settingsTargetIsResident, for the model the API just loaded.
- const settingsTargetIsResident = () =>
+ const selectedModelIsResident = () =>
residentModelIdMatches(store.checkpoint, serverStatus.checkpointId) &&
ggufVariantsMatch(store.activeGgufVariant, serverStatus.ggufVariant);
@@ -304,12 +296,12 @@ test("an auto-switch under a mounted Hub stops hiding the live config", () => {
subscribeResidentStatusRefresh(readStatusAndAdopt, targets);
assert.equal(
- settingsTargetIsResident(),
+ selectedModelIsResident(),
false,
"precondition: the mount-time read predates the switch",
);
targets.fire("window", "focus");
- assert.equal(settingsTargetIsResident(), true);
+ assert.equal(selectedModelIsResident(), true);
// A load this tab started owns the store until it settles, so a mid-switch refresh
// must not re-pin the model being moved away from.
diff --git a/studio/frontend/tests/hub-runtime-boundary.test.ts b/studio/frontend/tests/hub-runtime-boundary.test.ts
new file mode 100644
index 00000000000..3bd66fed69b
--- /dev/null
+++ b/studio/frontend/tests/hub-runtime-boundary.test.ts
@@ -0,0 +1,352 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import assert from "node:assert/strict";
+import { existsSync, readFileSync, readdirSync } from "node:fs";
+import path from "node:path";
+import { test } from "node:test";
+import { fileURLToPath } from "node:url";
+
+const HERE = path.dirname(fileURLToPath(import.meta.url));
+const FRONTEND_ROOT = path.join(HERE, "..");
+const HUB_ROOT = path.join(FRONTEND_ROOT, "src/features/hub");
+
+function sourceFiles(directory: string): string[] {
+ return readdirSync(directory, { withFileTypes: true }).flatMap((entry) => {
+ const target = path.join(directory, entry.name);
+ if (entry.isDirectory()) return sourceFiles(target);
+ return /\.(?:ts|tsx)$/.test(entry.name) ? [target] : [];
+ });
+}
+
+const HUB_SOURCE_FILES = sourceFiles(HUB_ROOT);
+const HUB_SOURCES = HUB_SOURCE_FILES.map((file) =>
+ readFileSync(file, "utf8"),
+).join("\n");
+const HUB_RENDER_SOURCES = `${HUB_SOURCES}\n${readFileSync(
+ path.join(HUB_ROOT, "hub.css"),
+ "utf8",
+)}`;
+
+test("obsolete Hub action and settings modules stay removed", () => {
+ const removedModules = [
+ "catalog/hub-model-settings-view.tsx",
+ "inventory/settings-identity.ts",
+ "lib/hub-feature-flags.ts",
+ ];
+
+ for (const modulePath of removedModules) {
+ assert.equal(
+ existsSync(path.join(HUB_ROOT, modulePath)),
+ false,
+ modulePath,
+ );
+ }
+});
+
+test("the Hub no longer owns model runtime actions or model configuration", () => {
+ const removedIdentifiers = [
+ "HubModelSettingsView",
+ "useChatModelRuntime",
+ "ModelConfigPage",
+ "PerModelConfig",
+ "modelConfigIdentity",
+ "settingsGgufVariantForRow",
+ "openSelectedModelSettings",
+ "onOpenModelSettings",
+ "CardSettingsButton",
+ "settingsTarget",
+ "runSelectedModel",
+ "openNewChat",
+ "loadingPhase",
+ "runtimeCapabilities",
+ "SelectedResourceRef",
+ ];
+
+ for (const identifier of removedIdentifiers) {
+ assert.equal(HUB_SOURCES.includes(identifier), false, identifier);
+ }
+});
+
+test("residency remains a cache mutation safety input", () => {
+ const hubPage = readFileSync(path.join(HUB_ROOT, "hub-page.tsx"), "utf8");
+ const adoptionSource = readFileSync(
+ path.join(HUB_ROOT, "lib/adopt-inference-status.ts"),
+ "utf8",
+ );
+ const ggufCard = readFileSync(
+ path.join(HUB_ROOT, "catalog/gguf-download-card.tsx"),
+ "utf8",
+ );
+ const localCard = readFileSync(
+ path.join(HUB_ROOT, "catalog/local-on-device-card.tsx"),
+ "utf8",
+ );
+ const safetensorsCard = readFileSync(
+ path.join(HUB_ROOT, "catalog/safetensors-download-card.tsx"),
+ "utf8",
+ );
+
+ assert.match(hubPage, /getInferenceStatus\(\)/);
+ assert.match(hubPage, /adoptResidentModelStatus\(/);
+ assert.match(
+ hubPage,
+ /Promise\.all\(\[getInferenceStatus\(\), readIdleUnloadArmed\(\)\]\)\s*\.then\(\(\[status, idleUnloadArmed\]\) => \{/,
+ );
+ assert.match(
+ hubPage,
+ /adoptResidentModelStatus\([\s\S]*?modelLoading: store\.modelLoading,\s*idleUnloadArmed: preserveIdleUnloaded && idleUnloadArmed,/,
+ );
+ assert.match(hubPage, /\.catch\(\(\) => idleUnloadArmed\.current\)/);
+ assert.match(
+ hubPage,
+ /applyStatus: \(previous\) => \{\s*applyActiveModelStatusToStore\(status, \{/,
+ );
+ assert.match(
+ hubPage,
+ /adoptingExistingServerModel:\s*previous\.checkpoint === null \|\| previous\.checkpoint === "",/,
+ );
+ assert.match(
+ hubPage,
+ /checkpointId: isSpeechOnlyStatus\(status\)\s*\? null\s*: resolveInferenceCheckpointId\(status\),/,
+ );
+ assert.match(hubPage, /speechOnly: isSpeechOnlyStatus\(status\),/);
+ assert.doesNotMatch(hubPage, /setCheckpoint\(status\.active_model/);
+ assert.match(
+ adoptionSource,
+ /state\.idleUnloadArmed && !status\.speechOnly/,
+ );
+ assert.match(
+ hubPage,
+ /subscribeResidentStatusRefresh\(\s*refreshResidentModelStatus/,
+ );
+ assert.match(
+ hubPage,
+ /subscribeModelLifecycle\(\(\{ runtime \}\) => \{\s*if \(runtime === "chat" \|\| runtime === "stt"\) return;\s*void refreshResidentModelStatus\(\{ preserveIdleUnloaded: false \}\);\s*\}\)/,
+ );
+ assert.match(
+ hubPage,
+ /idleUnloadArmed: preserveIdleUnloaded && idleUnloadArmed/,
+ );
+ assert.ok((hubPage.match(/residentModelIdMatches\(/g) ?? []).length >= 2);
+ assert.ok((hubPage.match(/selectedModel\.loadId/g) ?? []).length >= 2);
+ assert.match(hubPage, /runtime=\{inspectorRuntime\}/);
+
+ const selectedView = readFileSync(
+ path.join(HUB_ROOT, "hooks/use-selected-model-view.ts"),
+ "utf8",
+ );
+ assert.match(selectedView, /loadId: selectedCachedRow\.loadId/);
+ assert.match(
+ selectedView,
+ /loadId: selectedCachedRow\?\.loadId \?\? selectedLocalRow\?\.loadId \?\? null/,
+ );
+
+ assert.match(
+ ggufCard,
+ /mutationBlocked=\{\s*runPending \|\|\s*isLoadingThisModel \|\|\s*\(isActive && item\.key === activeVariantKey\)/,
+ );
+ assert.match(
+ ggufCard,
+ /canDelete=\{[\s\S]*?!selectedIsActive[\s\S]*?!isLoadingThisModel[\s\S]*?\}/,
+ );
+ assert.match(
+ ggufCard,
+ /const showUpdateAction =[\s\S]*?!selectedIsActive[\s\S]*?!isLoadingThisModel[\s\S]*?;/,
+ );
+ assert.match(
+ localCard,
+ /const canUpdate =[\s\S]*?!isActive &&[\s\S]*?!isLoading &&[\s\S]*?updateAvailable;/,
+ );
+ assert.match(
+ localCard,
+ /const canDelete =[\s\S]*?!isActive &&[\s\S]*?!isLoading[\s\S]*?;/,
+ );
+ assert.match(
+ safetensorsCard,
+ /const canDelete =[\s\S]*?!isActive &&[\s\S]*?!isLoadingThisModel &&[\s\S]*?!runPending;/,
+ );
+});
+
+test("the Hub routes Run to its owning workflow without owning runtime actions", () => {
+ const hubPage = readFileSync(path.join(HUB_ROOT, "hub-page.tsx"), "utf8");
+ const modelInspector = readFileSync(
+ path.join(HUB_ROOT, "catalog/model-inspector.tsx"),
+ "utf8",
+ );
+ const ggufCard = readFileSync(
+ path.join(HUB_ROOT, "catalog/gguf-download-card.tsx"),
+ "utf8",
+ );
+ const localCard = readFileSync(
+ path.join(HUB_ROOT, "catalog/local-on-device-card.tsx"),
+ "utf8",
+ );
+ const safetensorsCard = readFileSync(
+ path.join(HUB_ROOT, "catalog/safetensors-download-card.tsx"),
+ "utf8",
+ );
+
+ assert.match(HUB_RENDER_SOURCES, /hub-run-action-btn/);
+ assert.match(HUB_RENDER_SOURCES, />\s*Run\s*</);
+ assert.match(
+ hubPage,
+ /clearNewChatDraft\(\);[\s\S]*?setActiveThreadId\(null\);[\s\S]*?setActiveProjectId\(null\);[\s\S]*?setIncognito\(false\);[\s\S]*?requestModelConfigHandoff\(request\)/,
+ );
+ assert.match(hubPage, /requestModelConfigHandoff\(request\)/);
+ assert.match(hubPage, /to: "\/chat", search: \{ new: requestId \}/);
+ assert.match(
+ modelInspector,
+ /const mediaPage = studioPageForTask\([\s\S]*?taskForMediaPick\(model\.pipelineTag, model\.task\)/,
+ );
+ assert.match(
+ hubPage,
+ /routableToMediaPage\(selectedModel\.kind, selectedModel\.localSource\)/,
+ );
+ assert.match(
+ hubPage,
+ /to: `\/\$\{mediaPage\}`,[\s\S]*?diffusionRouteSearch\(selectedModel\.hubRepoId, selection\)/,
+ );
+ assert.match(
+ hubPage,
+ /const controller = runConfigOpenCoordinator\.begin\(\);[\s\S]*?await waitForRunConfigRefresh\(\s*refreshResidentModelStatus\(\),\s*controller\.signal,\s*\);\s*if \(controller\.signal\.aborted\) return;/,
+ );
+ assert.match(hubPage, /const RUN_CONFIG_REFRESH_TIMEOUT_MS = 5_000;/);
+ assert.match(
+ hubPage,
+ /const handleCloseDetail = useCallback\(\(\) => \{\s*runConfigOpenCoordinator\.cancel\(\);\s*setRunConfigOpening\(null\);/,
+ );
+ assert.match(HUB_RENDER_SOURCES, /Opening…/);
+ assert.match(HUB_RENDER_SOURCES, /aria-busy=\{loading\}/);
+ assert.match(ggufCard, /open=\{runPending \? false : open\}/);
+ assert.match(ggufCard, /disabled=\{runPending\}/);
+ assert.match(localCard, /open=\{runPending \? false : variantOpen\}/);
+ assert.match(
+ localCard,
+ /disabled=\{currentVariantState\.loading \|\| runPending\}/,
+ );
+ assert.match(
+ localCard,
+ /label=\{`Configure and run \$\{displayName\.trim\(\) \|\| "this model"\}`\}/,
+ );
+ assert.doesNotMatch(
+ localCard,
+ /label=\{`Configure and run \$\{repoId \?\? modelId\}`\}/,
+ );
+ assert.match(
+ safetensorsCard,
+ /const showRunAction =[^;]*?!isLoadingThisModel[^;]*?;/s,
+ );
+ assert.doesNotMatch(
+ safetensorsCard,
+ /const showRunAction =[^;]*?repoPeerActive[^;]*?;/s,
+ );
+ assert.match(
+ ggufCard,
+ /const showRunAction =[^;]*?!isLoadingThisModel[^;]*?;/s,
+ );
+ assert.match(
+ localCard,
+ /const showRunAction =[^;]*?!isLoading[^;]*?;/s,
+ );
+ assert.match(
+ modelInspector,
+ /s\.isChatOnly\(\) && !s\.capabilitiesUnknown\(\)/,
+ );
+ assert.match(
+ modelInspector,
+ /nonGgufRuntimeAvailable:\s*!chatOnlyMeasured &&\s*unslothSupport\.status !== "unsupported"/,
+ );
+ assert.match(modelInspector, /const runEligible = isHubModelRunEligible\(/);
+ assert.match(
+ modelInspector,
+ /\? \(selection: HubModelRunSelection\) => onRun\(selection, mediaPage\)/,
+ );
+ assert.match(
+ ggufCard,
+ /const showRunAction =[^;]*?!downloadingThisVariant[^;]*?;/s,
+ );
+ assert.match(
+ modelInspector,
+ /<ModelStatusChips[\s\S]*?chatOnly=\{chatOnly\}/,
+ );
+ assert.doesNotMatch(
+ HUB_RENDER_SOURCES,
+ />\s*(?:Eject|Loaded|Use in Chat|New Chat|Train)\s*</,
+ );
+ assert.doesNotMatch(
+ HUB_SOURCES,
+ /\bonUseInChat\b|\bonEject\b|\bonTrain\b|\bejectModel\b|\bselectModel\b|\brunId\b/,
+ );
+
+ const filesWithOnLoad = HUB_SOURCE_FILES.filter((file) =>
+ /\bonLoad(?:Local)?\b/.test(readFileSync(file, "utf8")),
+ ).map((file) => path.relative(HUB_ROOT, file).split(path.sep).join("/"));
+
+ // This is an image lifecycle callback, not a model runtime action.
+ assert.deepEqual(filesWithOnLoad.sort(), ["catalog/owner-avatar.tsx"]);
+});
+
+test("the Run handoff opens configuration immediately and remains nonce-scoped", () => {
+ const chatPage = readFileSync(
+ path.join(FRONTEND_ROOT, "src/features/chat/chat-page.tsx"),
+ "utf8",
+ );
+ const modelSelector = readFileSync(
+ path.join(
+ FRONTEND_ROOT,
+ "src/features/model-picker/components/model-selector.tsx",
+ ),
+ "utf8",
+ );
+
+ assert.match(chatPage, /modelConfigHandoffForDestination\(state\.request,/);
+ assert.match(
+ chatPage,
+ /active,\s*newChatId: search\.new,\s*threadId: search\.thread,\s*compareId: search\.compare,\s*projectId: search\.project,/,
+ );
+ assert.match(chatPage, /modelSelectorOpen \|\| modelConfigRequest !== null/);
+ assert.match(modelSelector, /requestedConfigTarget \?\? configTarget/);
+ assert.match(
+ modelSelector,
+ /aria-label=\{\s*visibleConfigTarget\s*\? `Run settings for \$\{visibleConfigTarget\.displayName\}`\s*:\s*undefined\s*\}/,
+ );
+ assert.match(
+ modelSelector,
+ /modelConfigInstanceKey\(\s*visibleConfigTarget\.configId \?\? visibleConfigTarget\.id,\s*visibleConfigTarget\.ggufVariant,\s*visibleLoadedConfig,/,
+ );
+ assert.match(
+ modelSelector,
+ /onRun=\{\(config, isDiffusion\) =>\s*onSelect\(\s*visibleConfigTarget\.configId \?\? visibleConfigTarget\.id,\s*\{/,
+ );
+ assert.match(
+ modelSelector,
+ /setAdoptedConfigRequestId\(configRequest\.requestId\);\s*setConfigTarget\(requestedConfigTarget\)/,
+ );
+ assert.match(
+ modelSelector,
+ /onConfigRequestAdopted\?\.\(configRequest\.requestId\)/,
+ );
+ assert.match(
+ chatPage,
+ /setSettingsOpen\(false\);\s*setModelSelectorLocked\(false\);\s*setModelSelectorOpen\(true\);\s*clearModelConfigHandoff\(requestId\)/,
+ );
+ assert.match(
+ chatPage,
+ /open=\{active && modelConfigRequest === null && settingsOpen\}/,
+ );
+});
+
+test("the shared model row menu no longer offers Hub model settings", () => {
+ const rowMenu = readFileSync(
+ path.join(
+ HERE,
+ "..",
+ "src/features/model-picker/components/model-selector/model-row-menu.tsx",
+ ),
+ "utf8",
+ );
+
+ assert.doesNotMatch(rowMenu, /ModelRowMenuSettings|Settings02Icon/);
+ assert.doesNotMatch(rowMenu, />\s*Settings\s*</);
+});
diff --git a/studio/frontend/tests/hub-superseded-refresh.test.ts b/studio/frontend/tests/hub-superseded-refresh.test.ts
index 390d6334e6a..79fd574302e 100644
--- a/studio/frontend/tests/hub-superseded-refresh.test.ts
+++ b/studio/frontend/tests/hub-superseded-refresh.test.ts
@@ -15,7 +15,7 @@ interface Status {
ggufVariant: string;
}
-/** What the store holds before an API request switches the resident model under the tab. */
+/** What the store holds before an API request switches the resident model. */
const STALE: Status = {
checkpoint: "unsloth/Qwen3-8B-GGUF",
ggufVariant: "Q8_0",
@@ -39,8 +39,8 @@ const flush = () => new Promise((r) => setTimeout(r, 0));
/**
* hub-page.tsx's refreshResidentModelStatus, with the status read held open so a test can
- * choose the order responses land in. `coalesce` is the fix: whether a dropped response
- * resolves with the refresh that superseded it or with nothing.
+ * choose the order responses land in. `coalesce` controls whether a dropped response
+ * waits for the refresh that superseded it.
*/
function hubPageRefresh(coalesce: boolean) {
let seq = 0;
@@ -88,19 +88,16 @@ async function settledEarly(promise: Promise<void>): Promise<boolean> {
return done;
}
-test("a settings open waits out the focus refresh that superseded its own read", async () => {
- // openModelSettings awaits a status read before it resolves a quant and builds the settings
- // target. A window focus refresh started while that read was out takes the sequence number,
- // so the settings read's response is dropped without writing the store.
+test("the initial safety read waits out a newer focus refresh", async () => {
const hub = hubPageRefresh(true);
- const settingsRead = hub.refresh();
+ const initialRead = hub.refresh();
hub.refresh();
hub.deliver(0);
assert.equal(
- await settledEarly(settingsRead),
+ await settledEarly(initialRead),
false,
- "a dropped response must not release the settings open",
+ "a dropped response must not release the safety gate",
);
assert.deepEqual(
hub.store,
@@ -109,51 +106,47 @@ test("a settings open waits out the focus refresh that superseded its own read",
);
hub.deliver(1);
- await settingsRead;
+ await initialRead;
assert.deepEqual(hub.store, SWITCHED);
});
-test("without it the dropped response opens settings on the displaced model", async () => {
- // The failure this guards: the settings handler's own sequence guard only counts settings
- // opens, so a focus refresh supersedes it silently and it proceeds on the pre-switch store,
- // naming the displaced model's quant in a target that Apply then loads and saves under.
+test("without coalescing the dropped response releases stale state", async () => {
const hub = hubPageRefresh(false);
- const settingsRead = hub.refresh();
+ const initialRead = hub.refresh();
hub.refresh();
hub.deliver(0);
- assert.equal(await settledEarly(settingsRead), true);
+ assert.equal(await settledEarly(initialRead), true);
assert.deepEqual(hub.store, STALE);
});
test("every dropped response in a chain waits for the one read that wins", async () => {
- // Focus and visibilitychange fire as a pair, so more than one refresh can pile up behind
- // the settings read, and each one in turn is dropped.
+ // Focus and visibilitychange can fire as a pair, so several reads may overlap.
const hub = hubPageRefresh(true);
- const settingsRead = hub.refresh();
+ const initialRead = hub.refresh();
hub.refresh();
hub.refresh();
hub.deliver(0);
hub.deliver(1);
- assert.equal(await settledEarly(settingsRead), false);
+ assert.equal(await settledEarly(initialRead), false);
assert.deepEqual(hub.store, STALE);
hub.deliver(2);
- await settingsRead;
+ await initialRead;
assert.deepEqual(hub.store, SWITCHED);
});
test("responses that land out of order still leave the store on the newest read", async () => {
const hub = hubPageRefresh(true);
- const settingsRead = hub.refresh();
+ const initialRead = hub.refresh();
hub.refresh();
// The newest read answers first and writes the store; the older response is dropped and
// finds its superseder already settled.
hub.deliver(1);
hub.deliver(0);
- await settingsRead;
+ await initialRead;
assert.deepEqual(hub.store, SWITCHED);
});
diff --git a/studio/frontend/tests/llama-extra-args-panel-hydration.test.ts b/studio/frontend/tests/llama-extra-args-panel-hydration.test.ts
index db10a23c6ea..b768267730a 100644
--- a/studio/frontend/tests/llama-extra-args-panel-hydration.test.ts
+++ b/studio/frontend/tests/llama-extra-args-panel-hydration.test.ts
@@ -94,18 +94,6 @@ test("a background auto-load hydrates a server-only override", () => {
assert.match(ADAPTER, /candidate\.kind === "gguf" &&\s*\n?\s*!isDiffusion/);
});
-const HUB = readFileSync(
- path.join(HERE, "..", "src/features/hub/hub-page.tsx"),
- "utf8",
-);
-
-test("applying from the Hub settings page carries the arguments into the load", () => {
- // applyPerModelConfigToRuntime does not store llamaExtraArgs, so a selection made
- // without the config left the field undefined: the load omitted it, the route kept
- // the resident server's old list, and an edit or a clear did nothing.
- assert.match(HUB, /forceReload: true,\n(\s*\/\/.*\n)*\s*config,/);
-});
-
test("a collapsed section stops objecting once nothing is left to object to", () => {
// Reset with Advanced collapsed clears the list, and the row is not mounted to
// withdraw the verdict it left standing, so Load stayed disabled.
@@ -337,15 +325,11 @@ const CHAT_PAGE = readFileSync(
"utf8",
);
-test("a launch that only applies the remembered config still carries its arguments", () => {
+test("a Chat launch that applies remembered config carries its arguments", () => {
// applyPerModelConfigToRuntime has no field for the launch flags, and /load only
// inherits them from the SAME resident model, so a cold launch or a switch from
- // another model ran without the arguments this model was remembered with. Both
- // paths that load through the runtime alone now pass the config itself.
- assert.match(
- HUB,
- /\.\.\.\(rememberedConfig \? \{ config: rememberedConfig \} : \{\}\),/,
- );
+ // another model ran without the arguments this model was remembered with. Chat
+ // passes the config itself.
assert.match(
CHAT_PAGE,
/const remembered = rememberedConfigFor\(selection\);/,
@@ -354,7 +338,4 @@ test("a launch that only applies the remembered config still carries its argumen
CHAT_PAGE,
/\.\.\.\(remembered \? \{ config: remembered \} : \{\}\),/,
);
- // Nothing is invented when there is no remembered config: the field stays absent,
- // which is what lets /load keep a resident model's own flags.
- assert.doesNotMatch(HUB, /config: rememberedConfig \?\? null/);
});
diff --git a/studio/frontend/tests/local-path.test.ts b/studio/frontend/tests/local-path.test.ts
index aefaf19405e..61db3a2d90f 100644
--- a/studio/frontend/tests/local-path.test.ts
+++ b/studio/frontend/tests/local-path.test.ts
@@ -3,10 +3,7 @@
import assert from "node:assert/strict";
import test from "node:test";
-import {
- looksLikeLocalPath,
- routableToMediaPage,
-} from "../src/features/hub/lib/local-path.ts";
+import { looksLikeLocalPath } from "../src/features/hub/lib/local-path.ts";
test("recognizes local paths across supported operating systems", () => {
for (const path of [
@@ -33,16 +30,3 @@ test("does not classify Hugging Face repository identifiers as paths", () => {
assert.equal(looksLikeLocalPath("HuggingFaceH4/ultrachat_200k"), false);
assert.equal(looksLikeLocalPath("unsloth/Qwen3-4B"), false);
});
-
-test("an HF-cache row routes to the media pages, a filesystem row does not", () => {
- // Inventory dedup can leave a complete hf_cache LOCAL row as the only row for a repo, and
- // it carries the Hub id. Excluding it by kind sent a diffusion GGUF to the chat loader,
- // where the backend refuses it, instead of to Images or Video.
- assert.equal(routableToMediaPage("local", "hf_cache"), true);
- assert.equal(routableToMediaPage("local", "models_dir"), false);
- assert.equal(routableToMediaPage("local", "lmstudio"), false);
- assert.equal(routableToMediaPage("local", "ollama"), false);
- assert.equal(routableToMediaPage("local", null), false);
- assert.equal(routableToMediaPage("cache", null), true);
- assert.equal(routableToMediaPage("discover", null), true);
-});
diff --git a/studio/frontend/tests/model-config-handoff.test.ts b/studio/frontend/tests/model-config-handoff.test.ts
new file mode 100644
index 00000000000..0959882c210
--- /dev/null
+++ b/studio/frontend/tests/model-config-handoff.test.ts
@@ -0,0 +1,285 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import type { ModelConfigHandoffRequest } from "../src/features/model-picker/model-config/model-config-handoff.ts";
+import {
+ installLocalStorageFake,
+ registerBundlerResolver,
+} from "./helpers/kit.ts";
+
+registerBundlerResolver();
+const { store } = installLocalStorageFake();
+
+const {
+ clearModelConfigHandoff,
+ createModelConfigHandoffRequestId,
+ modelConfigHandoffForDestination,
+ modelConfigTarget,
+ modelConfigTargetIsResident,
+ modelConfigTargetMatchesSelection,
+ requestModelConfigHandoff,
+ useModelConfigHandoffStore,
+} = await import(
+ "../src/features/model-picker/model-config/model-config-handoff.ts"
+);
+const { DEFAULT_PER_MODEL_CONFIG, resolveInitialConfig, savePerModelConfig } =
+ await import("../src/features/model-picker/model-config/per-model-config.ts");
+
+const FALLBACK_REQUEST_ID_RE = /^\d+-[a-z0-9]+$/;
+
+function request(
+ requestId: string,
+ id: string,
+ loadId: string,
+): ModelConfigHandoffRequest {
+ return {
+ requestId,
+ id,
+ meta: {
+ source: "hub",
+ isLora: false,
+ isDownloaded: true,
+ loadId,
+ },
+ };
+}
+
+test("handoff request IDs work without native randomUUID", () => {
+ assert.match(createModelConfigHandoffRequestId(null), FALLBACK_REQUEST_ID_RE);
+ assert.match(
+ createModelConfigHandoffRequestId({
+ randomUUID: () => {
+ throw new Error("unavailable");
+ },
+ }),
+ FALLBACK_REQUEST_ID_RE,
+ );
+ assert.equal(
+ createModelConfigHandoffRequestId({ randomUUID: () => "native-id" }),
+ "native-id",
+ );
+});
+
+test("only the matching consumer can clear the latest handoff", () => {
+ const first = request("first", "org/first", "/cache/first");
+ const second = request("second", "org/second", "/cache/second");
+
+ requestModelConfigHandoff(first);
+ requestModelConfigHandoff(second);
+ clearModelConfigHandoff(first.requestId);
+ assert.equal(useModelConfigHandoffStore.getState().request, second);
+
+ clearModelConfigHandoff(second.requestId);
+ assert.equal(useModelConfigHandoffStore.getState().request, null);
+});
+
+test("only the matching unoccupied active Chat destination receives a handoff", () => {
+ const pending = request("run-request", "org/model", "/cache/model");
+ assert.equal(
+ modelConfigHandoffForDestination(pending, {
+ active: true,
+ newChatId: "run-request",
+ }),
+ pending,
+ );
+
+ const blockedDestinations = [
+ { active: false, newChatId: "run-request" },
+ { active: true, newChatId: "another-request" },
+ { active: true, newChatId: "run-request", threadId: "thread" },
+ { active: true, newChatId: "run-request", compareId: "compare" },
+ { active: true, newChatId: "run-request", projectId: "project" },
+ ];
+ for (const destination of blockedDestinations) {
+ assert.equal(modelConfigHandoffForDestination(pending, destination), null);
+ }
+ assert.equal(
+ modelConfigHandoffForDestination(null, {
+ active: true,
+ newChatId: "run-request",
+ }),
+ null,
+ );
+});
+
+test("a Windows model path stays opaque while its display name is portable", () => {
+ const id = String.raw`C:\Users\models\Llama-3.gguf`;
+ const meta = {
+ source: "local" as const,
+ isLora: false,
+ isDownloaded: true,
+ isGguf: true,
+ loadId: id,
+ };
+
+ const target = modelConfigTarget(id, meta);
+
+ assert.equal(target.id, id);
+ assert.equal(target.configId, undefined);
+ assert.equal(target.meta.loadId, id);
+ assert.equal(target.displayName, "Llama-3");
+ assert.equal(target.isGguf, true);
+ assert.equal(target.apiLoadable, true);
+});
+
+test("a GGUF handoff preserves its public identity and exact variant", () => {
+ const id = "unsloth/Llama-GGUF";
+ const loadId = "/cache/models--unsloth--Llama-GGUF/snapshots/revision";
+ const target = modelConfigTarget(id, {
+ source: "hub",
+ isLora: false,
+ loadId,
+ ggufVariant: "Q4_K_M",
+ ggufFilename: "llama-q4_k_m.gguf",
+ isDownloaded: true,
+ isGguf: true,
+ });
+
+ assert.equal(target.id, loadId);
+ assert.equal(target.configId, id);
+ assert.equal(target.displayName, "Llama-GGUF · Q4_K_M");
+ assert.equal(target.ggufVariant, "Q4_K_M");
+ assert.equal(target.meta.ggufFilename, "llama-q4_k_m.gguf");
+});
+
+test("configuration residency follows public, loader, variant, and live state", () => {
+ const target = modelConfigTarget("Org/Model", {
+ source: "hub",
+ isLora: false,
+ loadId: String.raw`C:\Users\models\snapshots\revision`,
+ ggufVariant: "Q4_K_M",
+ isDownloaded: true,
+ isGguf: true,
+ });
+
+ assert.equal(modelConfigTargetMatchesSelection(target, "org/model"), true);
+ assert.equal(
+ modelConfigTargetMatchesSelection(
+ target,
+ String.raw`c:\users\models\snapshots\revision`,
+ ),
+ true,
+ );
+ assert.equal(
+ modelConfigTargetIsResident({
+ target,
+ selectedId: "org/model",
+ activeGgufVariant: "q4_k_m",
+ loaded: true,
+ }),
+ true,
+ );
+ assert.equal(
+ modelConfigTargetIsResident({
+ target,
+ selectedId: "org/model",
+ activeGgufVariant: "Q8_0",
+ loaded: true,
+ }),
+ false,
+ );
+ assert.equal(
+ modelConfigTargetIsResident({
+ target,
+ selectedId: "org/model",
+ activeGgufVariant: "Q4_K_M",
+ loaded: false,
+ }),
+ false,
+ );
+});
+
+test("only local GGUF files may ignore a derived active variant", () => {
+ const localPath = String.raw`C:\Models\Llama.gguf`;
+ const localTarget = modelConfigTarget(localPath, {
+ source: "local",
+ isLora: false,
+ loadId: localPath,
+ isDownloaded: true,
+ isGguf: true,
+ });
+ const hubTarget = modelConfigTarget("Org/Model.gguf", {
+ source: "hub",
+ isLora: false,
+ loadId: "Org/Model.gguf",
+ isDownloaded: true,
+ isGguf: true,
+ });
+
+ assert.equal(
+ modelConfigTargetIsResident({
+ target: localTarget,
+ selectedId: String.raw`c:\models\llama.gguf`,
+ activeGgufVariant: "Q4_K_M",
+ }),
+ true,
+ );
+ assert.equal(
+ modelConfigTargetIsResident({
+ target: hubTarget,
+ selectedId: "Org/Model.gguf",
+ activeGgufVariant: "Q4_K_M",
+ }),
+ false,
+ );
+});
+
+test("an Ollama load identity never advertises API-loadable settings", () => {
+ const id = "ollama-manifest:library/model:latest";
+ const target = modelConfigTarget(
+ id,
+ {
+ source: "local",
+ isLora: false,
+ loadId: id,
+ isDownloaded: true,
+ isGguf: true,
+ },
+ "Llama 3.2",
+ );
+
+ assert.equal(target.apiLoadable, false);
+ assert.equal(target.configId, undefined);
+ assert.equal(target.displayName, "Llama 3.2");
+ assert.equal(
+ modelConfigTargetIsResident({
+ target,
+ selectedId: id,
+ activeGgufVariant: "Q4_K_M",
+ loaded: true,
+ }),
+ true,
+ );
+});
+
+test("submitting a cached handoff adopts legacy path-keyed settings", () => {
+ store.clear();
+ const id = "unsloth/Llama-GGUF";
+ const loadId = "/cache/models--unsloth--Llama-GGUF/snapshots/legacy";
+ const ggufVariant = "Q4_K_M";
+ savePerModelConfig(loadId, ggufVariant, {
+ ...DEFAULT_PER_MODEL_CONFIG,
+ maxSeqLength: 32768,
+ });
+
+ requestModelConfigHandoff({
+ requestId: "legacy-settings",
+ id,
+ meta: {
+ source: "hub",
+ isLora: false,
+ loadId,
+ ggufVariant,
+ isDownloaded: true,
+ isGguf: true,
+ },
+ });
+
+ const adopted = resolveInitialConfig(id, ggufVariant);
+ assert.equal(adopted.remembered, true);
+ assert.equal(adopted.config.maxSeqLength, 32768);
+ assert.equal(resolveInitialConfig(loadId, ggufVariant).remembered, false);
+});
diff --git a/studio/frontend/tests/model-config-run-selection.test.ts b/studio/frontend/tests/model-config-run-selection.test.ts
new file mode 100644
index 00000000000..e1f1f210432
--- /dev/null
+++ b/studio/frontend/tests/model-config-run-selection.test.ts
@@ -0,0 +1,219 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import assert from "node:assert/strict";
+import { readFileSync } from "node:fs";
+import test from "node:test";
+import ts from "typescript";
+
+import { isExternalModelId } from "../src/features/chat/external-providers.ts";
+import { isLocalModelPath } from "../src/features/chat/utils/model-download-staging.ts";
+import type {
+ ModelPickTarget,
+ ModelSelectorChangeMeta,
+} from "../src/features/model-picker/components/model-selector/types.ts";
+import {
+ installLocalStorageFake,
+ registerBundlerResolver,
+} from "./helpers/kit.ts";
+
+registerBundlerResolver();
+installLocalStorageFake();
+const { modelConfigTarget } = await import(
+ "../src/features/model-picker/model-config/model-config-handoff.ts"
+);
+const { DEFAULT_PER_MODEL_CONFIG } = await import(
+ "../src/features/model-picker/model-config/per-model-config.ts"
+);
+
+function readSource(relativePath: string): ts.SourceFile {
+ return ts.createSourceFile(
+ relativePath,
+ readFileSync(new URL(relativePath, import.meta.url), "utf8"),
+ ts.ScriptTarget.Latest,
+ true,
+ ts.ScriptKind.TSX,
+ );
+}
+
+function findNode<T extends ts.Node>(
+ root: ts.Node,
+ predicate: (node: ts.Node) => node is T,
+): T {
+ let found: T | undefined;
+ function visit(node: ts.Node): void {
+ if (predicate(node)) {
+ found = node;
+ }
+ if (!found) {
+ ts.forEachChild(node, visit);
+ }
+ }
+ visit(root);
+ assert.ok(found, "Expected source node was not found");
+ return found;
+}
+
+// Execute the shipped callback and persistence guard without mounting the UI.
+const selector = readSource(
+ "../src/features/model-picker/components/model-selector.tsx",
+);
+const configPage = findNode(
+ selector,
+ (node): node is ts.JsxSelfClosingElement =>
+ ts.isJsxSelfClosingElement(node) &&
+ node.tagName.getText() === "ModelConfigPage",
+);
+const onRun = findNode(
+ configPage,
+ (node): node is ts.JsxAttribute =>
+ ts.isJsxAttribute(node) && node.name.getText() === "onRun",
+);
+assert.ok(onRun.initializer && ts.isJsxExpression(onRun.initializer));
+assert.ok(onRun.initializer.expression);
+const runConfiguredModel = new Function(
+ "visibleConfigTarget",
+ "onSelect",
+ "config",
+ `return (${onRun.initializer.expression.getText()})(config, false);`,
+) as (
+ target: ModelPickTarget,
+ onSelect: (id: string, meta: ModelSelectorChangeMeta) => void,
+ config: typeof DEFAULT_PER_MODEL_CONFIG,
+) => void;
+
+const runtime = readSource(
+ "../src/features/chat/hooks/use-chat-model-runtime.ts",
+);
+const declarations = [
+ "modelId",
+ "loadPath",
+ "ggufVariant",
+ "nativePathToken",
+ "indexedLocalPick",
+].map((name) => {
+ const declaration = findNode(
+ runtime,
+ (node): node is ts.VariableDeclaration =>
+ ts.isVariableDeclaration(node) && node.name.getText() === name,
+ );
+ return `const ${declaration.getText()};`;
+});
+const recordCall = findNode(
+ runtime,
+ (node): node is ts.CallExpression =>
+ ts.isCallExpression(node) &&
+ node.expression.getText() === "recordLastLocalModelLoad",
+);
+const recordGuard = recordCall.parent.parent.parent;
+assert.ok(ts.isIfStatement(recordGuard));
+type Selection = ModelSelectorChangeMeta & { id: string };
+type RememberedModel = {
+ id: string;
+ kind: "gguf" | "model";
+ ggufVariant: string | null;
+};
+const finishLoad = new Function(
+ "selection",
+ "loadResponse",
+ "isGguf",
+ "isLora",
+ "recordLastLocalModelLoad",
+ "isLocalModelPath",
+ "isExternalModelId",
+ `${declarations.join("\n")}\n${recordGuard.getText()}\nreturn loadPath;`,
+) as (
+ selection: Selection,
+ response: { is_gguf: boolean; is_lora: boolean },
+ isGguf: boolean,
+ isLora: boolean,
+ record: (model: RememberedModel) => void,
+ isLocalPath: typeof isLocalModelPath,
+ isExternalId: typeof isExternalModelId,
+) => string;
+
+const cases = [
+ {
+ name: "secondary GGUF cache",
+ id: "Org/Model-GGUF",
+ loadId: "/cache/models--Org--Model-GGUF/snapshots/revision",
+ isGguf: true,
+ ggufVariant: "Q4_K_M",
+ },
+ {
+ name: "Windows GGUF cache",
+ id: "Org/Model-GGUF",
+ loadId: String.raw`C:\Cache\models--Org--Model-GGUF\snapshots\revision`,
+ isGguf: true,
+ ggufVariant: "Q8_0",
+ },
+ {
+ name: "secondary safetensors cache",
+ id: "Org/Model",
+ loadId: "/cache/models--Org--Model/snapshots/revision",
+ isGguf: false,
+ },
+ { name: "active Hub cache", id: "Org/Model", isGguf: false },
+ {
+ name: "indexed local GGUF",
+ id: "/models/model.gguf",
+ isGguf: true,
+ source: "local" as const,
+ },
+ {
+ name: "native file picker",
+ id: "/models/picked.gguf",
+ isGguf: true,
+ source: "local" as const,
+ nativePathToken: "signed-path-token",
+ },
+];
+
+for (const entry of cases) {
+ test(`configuration Run preserves loading and remembering for ${entry.name}`, () => {
+ const { name: _name, id, ...fields } = entry;
+ const meta: ModelSelectorChangeMeta = {
+ source: "hub",
+ isLora: false,
+ isDownloaded: true,
+ ...fields,
+ };
+ const target = modelConfigTarget(id, meta);
+ const config = { ...DEFAULT_PER_MODEL_CONFIG, maxSeqLength: 8192 };
+ const calls: Selection[] = [];
+ runConfiguredModel(
+ target,
+ (selectedId, selectedMeta) => {
+ calls.push({ id: selectedId, ...selectedMeta });
+ },
+ config,
+ );
+
+ assert.deepEqual(calls, [
+ { id, ...meta, config, isDiffusion: false, forceReload: true },
+ ]);
+ const remembered: RememberedModel[] = [];
+ const loadPath = finishLoad(
+ calls[0],
+ { is_gguf: entry.isGguf, is_lora: false },
+ entry.isGguf,
+ false,
+ (model) => remembered.push(model),
+ isLocalModelPath,
+ isExternalModelId,
+ );
+ assert.equal(loadPath, meta.loadId ?? id);
+ assert.deepEqual(
+ remembered,
+ meta.nativePathToken
+ ? []
+ : [
+ {
+ id,
+ kind: entry.isGguf ? "gguf" : "model",
+ ggufVariant: meta.ggufVariant ?? null,
+ },
+ ],
+ );
+ });
+}
diff --git a/studio/frontend/tests/model-identity.test.ts b/studio/frontend/tests/model-identity.test.ts
index 2219d622550..6bf6e0de947 100644
--- a/studio/frontend/tests/model-identity.test.ts
+++ b/studio/frontend/tests/model-identity.test.ts
@@ -4,11 +4,6 @@
import assert from "node:assert/strict";
import test from "node:test";
-import { settingsGgufVariantForRow } from "../src/features/hub/inventory/settings-identity.ts";
-import type {
- CachedInventoryRow,
- LocalInventoryRow,
-} from "../src/features/hub/inventory/types.ts";
import {
isOllamaLinkPath,
isStandaloneGgufPath,
@@ -109,7 +104,7 @@ test("a resident path-loaded model is matched by the id /status reports", () =>
),
true,
);
- // A repo in an inactive cache keeps the repo id as its settings identity.
+ // A repo in an inactive cache can be reconciled through its public id.
assert.equal(
residentModelIdMatches(
"unsloth/Qwen3-8B-GGUF",
@@ -118,6 +113,24 @@ test("a resident path-loaded model is matched by the id /status reports", () =>
),
true,
);
+ const pinnedSnapshot =
+ "/mnt/old-cache/models--unsloth--Qwen3-8B-GGUF/snapshots/abc123";
+ assert.equal(
+ residentModelIdMatches(
+ pinnedSnapshot,
+ "unsloth/Qwen3-8B-GGUF",
+ "/mnt/old-cache/models--unsloth--Qwen3-8B-GGUF",
+ ),
+ false,
+ );
+ assert.equal(
+ residentModelIdMatches(
+ pinnedSnapshot,
+ pinnedSnapshot,
+ "unsloth/Qwen3-8B-GGUF",
+ ),
+ true,
+ );
// The raw identifier is still matched literally.
assert.equal(
residentModelIdMatches(
@@ -199,35 +212,6 @@ test("Ollama link paths are recognised the way the resolver excludes them", () =
assert.equal(isOllamaLinkPath(null), false);
});
-test("a standalone gguf keeps one settings identity across surfaces", () => {
- const loose = {
- kind: "local",
- path: "/srv/models/Qwen3-8B-Q4_K_M.gguf",
- // What hub/services/models/common.py emits for a single scanned file.
- formatVariant: "Q4_K_M",
- } as LocalInventoryRow;
- // The Chat picker opens the same file with no variant, so adopting the filename label would split the config.
- assert.equal(settingsGgufVariantForRow(loose), null);
-
- // A GGUF directory still has a variant slot for the quant lookup to fill.
- const repoDir = {
- kind: "local",
- path: "/srv/models/Qwen3-8B-GGUF",
- formatVariant: null,
- } as LocalInventoryRow;
- assert.equal(settingsGgufVariantForRow(repoDir), null);
- const lmStudioDir = {
- kind: "local",
- path: "/srv/lmstudio/Qwen3-8B-GGUF",
- formatVariant: "Q8_0",
- } as LocalInventoryRow;
- assert.equal(settingsGgufVariantForRow(lmStudioDir), "Q8_0");
-
- // Cached repo rows are unaffected (cache_inventory.py never sets one).
- const cached = { kind: "cache", formatVariant: null } as CachedInventoryRow;
- assert.equal(settingsGgufVariantForRow(cached), null);
-});
-
// The backfill matches on the folded identity, unambiguous only because storage holds one record per model.
test("importing the legacy load settings never doubles up a model", () => {
// The legacy casing names the model the v2 record already holds, so the import must leave it alone.
diff --git a/studio/frontend/tests/model-picker-partial-downloads.test.ts b/studio/frontend/tests/model-picker-partial-downloads.test.ts
index 2139db77555..e3b5c6ea51e 100644
--- a/studio/frontend/tests/model-picker-partial-downloads.test.ts
+++ b/studio/frontend/tests/model-picker-partial-downloads.test.ts
@@ -10,6 +10,24 @@ import assert from "node:assert/strict";
import { readFileSync } from "node:fs";
import test from "node:test";
import { fileURLToPath } from "node:url";
+import { createElement } from "react";
+import { renderToStaticMarkup } from "react-dom/server";
+
+import type {
+ CachedInventoryRow,
+ DiscoverRow,
+ LocalInventoryRow,
+ SelectedModelView,
+} from "../src/features/hub/types.ts";
+import { downloadActionLabel } from "../src/features/hub/catalog/use-download-card-state.ts";
+import { modelDownloadState } from "../src/features/hub/catalog/model-download-state.ts";
+import { registerStoreStubResolver } from "./helpers/kit.ts";
+
+registerStoreStubResolver();
+
+const { useSelectedModelView } = await import(
+ "../src/features/hub/hooks/use-selected-model-view.ts"
+);
function read(path: string): string {
return readFileSync(fileURLToPath(new URL(path, import.meta.url)), "utf-8");
@@ -152,19 +170,201 @@ test("complete and partial are alternatives, never both dots on one row", () =>
);
});
-test("selecting a partial opens its download instead of claiming the weights", () => {
+test("selecting a picker partial opens its download instead of claiming the weights", () => {
// isDownloaded is what the load path reads. Hard-coding true on these rows sent a torn snapshot
// straight to a load that fails on the missing shards -- the reason they were hidden at all.
assert.ok(
PICKERS.includes("isDownloaded: !isPartial,"),
"the pick reports what is actually on disk",
);
- // Hub reaches the same answer from the same field.
- assert.ok(
- read("../src/features/hub/hub-page.tsx").includes(
- "isDownloaded: !row.partial",
- ),
- );
+});
+
+test("Hub selections preserve download completeness and continuation state", () => {
+ const capabilities = {
+ canTrain: false,
+ canChat: false,
+ canDelete: true,
+ canDownload: true,
+ requiresVariant: false,
+ supportsLora: false,
+ supportsVision: false,
+ };
+ const cached: CachedInventoryRow = {
+ kind: "cache",
+ id: "cache:safetensors:Org%2FModel",
+ loadId: "Org/Model",
+ repoId: "Org/Model",
+ owner: "Org",
+ repo: "Model",
+ isGguf: false,
+ modelFormat: "safetensors",
+ capabilities,
+ bytes: 128,
+ partial: true,
+ partialTransport: "http",
+ partialResumable: true,
+ };
+ const localHfCache: LocalInventoryRow = {
+ kind: "local",
+ id: "hf_cache:safetensors:Org%2FModel",
+ loadId: "Org/Model",
+ repoId: "Org/Model",
+ owner: "Org",
+ title: "Model",
+ source: "hf_cache",
+ sourceLabel: "Hugging Face cache",
+ path: "/cache/models--Org--Model",
+ isGguf: false,
+ modelFormat: "safetensors",
+ capabilities,
+ updatedAt: 1,
+ partial: true,
+ partialTransport: "xet",
+ partialResumable: false,
+ };
+ const discover: DiscoverRow = {
+ id: "Org/Model",
+ owner: "Org",
+ repo: "Model",
+ result: {
+ id: "Org/Model",
+ downloads: 0,
+ likes: 0,
+ isGguf: false,
+ },
+ isAvailableOnDevice: false,
+ isPartialOnDevice: true,
+ summary: "Model",
+ capabilities: [],
+ };
+ const base = {
+ selectedDiscoverRow: null,
+ selectedCachedRow: null,
+ selectedLocalRow: null,
+ selectedHfResult: null,
+ isDatasetMode: false,
+ } satisfies Parameters<typeof useSelectedModelView>[0];
+
+ const cases: Array<{
+ name: string;
+ input: Parameters<typeof useSelectedModelView>[0];
+ kind: SelectedModelView["kind"];
+ downloaded: boolean;
+ partial: boolean;
+ transport: string | null;
+ resumable: boolean;
+ action: "Download" | "Resume" | "Continue";
+ }> = [
+ {
+ name: "direct cache row",
+ input: { ...base, selectedCachedRow: cached },
+ kind: "cache",
+ downloaded: false,
+ partial: true,
+ transport: "http",
+ resumable: true,
+ action: "Resume",
+ },
+ {
+ name: "discovery row backed by a cache row",
+ input: { ...base, selectedDiscoverRow: discover, selectedCachedRow: cached },
+ kind: "discover",
+ downloaded: false,
+ partial: true,
+ transport: "http",
+ resumable: true,
+ action: "Resume",
+ },
+ {
+ name: "discovery row backed by a local HF-cache row",
+ input: {
+ ...base,
+ selectedDiscoverRow: discover,
+ selectedLocalRow: localHfCache,
+ },
+ kind: "discover",
+ downloaded: false,
+ partial: true,
+ transport: "xet",
+ resumable: false,
+ action: "Continue",
+ },
+ {
+ name: "direct local HF-cache row",
+ input: { ...base, selectedLocalRow: localHfCache },
+ kind: "cache",
+ downloaded: false,
+ partial: true,
+ transport: "xet",
+ resumable: false,
+ action: "Continue",
+ },
+ {
+ name: "discovery partial awaiting its inventory row",
+ input: { ...base, selectedDiscoverRow: discover },
+ kind: "discover",
+ downloaded: false,
+ partial: true,
+ transport: null,
+ resumable: false,
+ action: "Continue",
+ },
+ {
+ name: "complete cache row",
+ input: {
+ ...base,
+ selectedCachedRow: {
+ ...cached,
+ partial: false,
+ partialTransport: null,
+ partialResumable: false,
+ },
+ },
+ kind: "cache",
+ downloaded: true,
+ partial: false,
+ transport: null,
+ resumable: false,
+ action: "Download",
+ },
+ ];
+
+ for (const entry of cases) {
+ const result = { current: null as SelectedModelView | null };
+ function Harness() {
+ result.current = useSelectedModelView(entry.input);
+ return null;
+ }
+ renderToStaticMarkup(createElement(Harness));
+ assert.ok(result.current, entry.name);
+ assert.equal(result.current.kind, entry.kind, entry.name);
+ assert.equal(result.current.isDownloaded, entry.downloaded, entry.name);
+ assert.equal(result.current.isPartial, entry.partial, entry.name);
+ assert.equal(result.current.partialTransport, entry.transport, entry.name);
+ assert.equal(result.current.partialResumable, entry.resumable, entry.name);
+ const downloadState = modelDownloadState(result.current);
+ assert.deepEqual(
+ downloadState,
+ {
+ isDownloaded: entry.downloaded,
+ isPartial: entry.partial,
+ partialTransport: entry.transport,
+ partialResumable: entry.resumable,
+ },
+ entry.name,
+ );
+ assert.equal(
+ downloadActionLabel(
+ downloadState.isPartial,
+ downloadState.partialResumable,
+ ),
+ entry.action,
+ entry.name,
+ );
+ }
+
+ const inspector = read("../src/features/hub/catalog/model-inspector.tsx");
+ assert.equal(inspector.split("{...downloadState}").length - 1, 2);
});
test("listing a partial never makes it auto-loadable", () => {
@@ -263,7 +463,7 @@ test("configure carries the same rule, because Run replays its metadata", () =>
);
assert.match(
selector,
- /onRun=\{\(config, isDiffusion\) =>\n\s*onSelect\(visibleConfigTarget\.id, \{\n\s*\.\.\.visibleConfigTarget\.meta,/,
+ /onRun=\{\(config, isDiffusion\) =>\s*onSelect\(\s*visibleConfigTarget\.configId \?\? visibleConfigTarget\.id,\s*\{\s*\.\.\.visibleConfigTarget\.meta,/,
);
});
diff --git a/tests/studio/test_model_picker_contracts.py b/tests/studio/test_model_picker_contracts.py
index 32804efb62d..0df204db67c 100644
--- a/tests/studio/test_model_picker_contracts.py
+++ b/tests/studio/test_model_picker_contracts.py
@@ -476,7 +476,7 @@ def test_diffusion_capability_labeled_image_generation():
def test_active_model_config_round_trips_gpu_fields():
"""The active model's config must carry the GPU Memory knobs (GGUF only) so a
- sidebar/hub-gear reload cannot silently reset manual GPU settings, and "Remember
+ sidebar reload cannot silently reset manual GPU settings, and "Remember
settings" cannot persist a GPU-less config over a saved one."""
src = _read("features/model-picker/hooks/use-active-model-config.ts")
for field in (
@@ -494,11 +494,8 @@ def test_active_model_config_round_trips_gpu_fields():
assert "export function gpuFieldsSignature" in shared
assert "gpuFieldsSignature(config)," in shared
assert "export function modelConfigInstanceKey" in shared
- for rel in (
- "features/model-picker/components/sidebar-model-config.tsx",
- "features/hub/catalog/hub-model-settings-view.tsx",
- ):
- assert "modelConfigInstanceKey(" in _read(rel), rel
+ sidebar = _read("features/model-picker/components/sidebar-model-config.tsx")
+ assert "modelConfigInstanceKey(" in sidebar
# apply-per-model-config re-exports it, so its own callers are unchanged.
reexport = _read("features/model-picker/model-config/apply-per-model-config.ts")
assert "export { gpuFieldsSignature };" in reexport
@@ -780,11 +777,7 @@ def test_native_picked_gguf_template_read_through_lease():
assert '${nativePathToken ?? ""}' in hook
-def test_model_load_guard_is_cross_instance():
- """The in-flight load guard must consult the shared store pick (not only the
- per-hook ref) and ejectModel must refuse while any instance is loading:
- three live useChatModelRuntime instances exist (chat page, hub page, hub
- gear dialog)."""
+def test_model_load_guard_uses_shared_store_state():
src = _read("features/chat/hooks/use-chat-model-runtime.ts")
assert "useChatRuntimeStore.getState().loadingModelPick" in src
assert "clearLoadingModelPick" in src
@@ -2748,15 +2741,6 @@ def test_monitor_stats_exclude_model_lifecycle_rows():
assert 'entry.kind != "lifecycle"' in backend, "the rule this mirrors"
-def test_api_reach_copy_is_limited_to_gguf_models():
- """The Hub opens this page for every downloaded model, but ModelConfigPage mirrors
- settings to the server only when target.isGguf, because API auto-switch indexes
- GGUFs only."""
- src = " ".join(_read("features/hub/catalog/hub-model-settings-view.tsx").split())
- assert "{(target.apiLoadable ?? target.isGguf)" in src
- assert "Saved settings apply everywhere Unsloth loads this model." in src
-
-
def test_backfill_includes_a_standalone_gguf_with_no_variant():
"""A standalone .gguf picked directly has no quant to choose between, so it is stored
with a null variant."""
@@ -2810,16 +2794,6 @@ def test_backfill_skips_future_schema_local_records():
assert "storedConfigVersion(raw) > STORAGE_SCHEMA_VERSION" in listing[:900]
-def test_detail_settings_need_a_resolved_quant():
- """The on-device card passes a null variant while its own lookup is pending or after it
- failed."""
- src = " ".join(_read("features/hub/hub-page.tsx").split())
- # `variant`, not the argument: a resident quant may have replaced it, so judge what is saved.
- guard = "if (!variant && selectedModel.isGguf && selectedModel.requiresVariant) {"
- assert guard in src
- assert src.count("Couldn't determine which quant to configure.") == 2
-
-
def test_a_failed_detail_fetch_is_retried():
"""A terminal row's updated_at never advances and selectedIsMissing stays true, so a
fetch that failed had nothing left to re-run the effect and the payload stayed
@@ -2830,24 +2804,9 @@ def test_a_failed_detail_fetch_is_retried():
assert "if (attemptsRef.current.count >= DETAIL_FETCH_ATTEMPTS) { return; }" in src
-def test_ollama_models_are_not_advertised_as_api_loadable():
- """local_model_resolver skips Ollama's scanner, so an Ollama GGUF is never in the
- auto-switch index and no OpenAI request can resolve it."""
- types_src = " ".join(_read("features/model-picker/components/model-selector/types.ts").split())
- assert "apiLoadable?: boolean;" in types_src
- hub = " ".join(_read("features/hub/hub-page.tsx").split())
- assert "row.source !== LOCAL_MODEL_SOURCE.OLLAMA" in hub
- assert "apiLoadable:" in hub
- backend = _read_backend("core/inference/local_model_resolver.py")
- assert (
- "Ollama's\n scanner is skipped" in backend or "scanner is skipped" in backend
- ), "the rule this mirrors"
-
-
-def test_cached_repo_settings_are_keyed_by_the_repo_id():
- """A repo cached outside the active HF cache reports load_id = the snapshot path
- (hub/services/cache_inventory.py), while the chat picker and the auto-switch index
- key it by repo_id."""
+def test_model_config_keeps_storage_and_load_identities_separate():
+ """The config target can store settings under a stable repo id while probes use
+ the concrete load id supplied by the picker."""
config_page = " ".join(_read("features/model-picker/components/model-config-page.tsx").split())
assert "const configId = target.configId ?? target.id;" in config_page
for call in (
@@ -2861,17 +2820,6 @@ def test_cached_repo_settings_are_keyed_by_the_repo_id():
assert "useDefaultChatTemplate( target.id," in config_page
assert "model_path: target.id," in config_page
- hub = " ".join(_read("features/hub/hub-page.tsx").split())
- assert 'if (kind !== "cache" && resource.source !== "hub_cache") {' in hub
- assert "return resource.repoId ?? resource.runId;" in hub
- # Both openers and the Hub's own load resolve through it.
- assert hub.count("modelConfigIdentity(") == 3
- assert 'const configId = row.kind === "cache" ? row.repoId : id;' in hub
- assert hub.count("configId,") >= 2
-
- backend = _read_backend("hub/tests/test_model_services.py")
- assert 'fields["load_id"] == str(snapshot)' in backend, "the rule this mirrors"
-
def test_backfill_splits_a_quant_suffix_the_way_the_backend_does():
"""The backfill compared server keys under an identity taken by splitting on the last
@@ -2909,14 +2857,6 @@ def test_backfill_splits_a_quant_suffix_the_way_the_backend_does():
assert (WORKDIR / "studio" / "frontend" / "tests" / "model-identity.test.ts").is_file()
-def test_the_detail_card_also_gates_ollama_out_of_the_api_promise():
- """Settings opens from two places in the Hub."""
- hub = " ".join(_read("features/hub/hub-page.tsx").split())
- assert hub.count("LOCAL_MODEL_SOURCE.OLLAMA") == 2
- assert "selectedModel.localSource !== LOCAL_MODEL_SOURCE.OLLAMA" in hub
- assert "row.source !== LOCAL_MODEL_SOURCE.OLLAMA" in hub
-
-
def test_the_settings_page_judges_the_config_storage_actually_keeps():
"""savePerModelConfig normalizes before deciding, and the runtime hands this page
Speculative Decoding "auto", which canonicalizes to null."""
@@ -2939,8 +2879,8 @@ def test_the_settings_page_judges_the_config_storage_actually_keeps():
def test_the_chat_picker_marks_ollama_targets_unloadable_by_the_api():
"""A settings target opened from the Chat model picker carried no apiLoadable, so the
`??"""
- picker = " ".join(_read("features/model-picker/components/model-selector.tsx").split())
- assert "apiLoadable: isGguf && !isOllamaLinkPath(id)," in picker
+ handoff = " ".join(_read("features/model-picker/model-config/model-config-handoff.ts").split())
+ assert "apiLoadable: isGguf && !isOllamaLinkPath(id) && !isOllamaLinkPath(loadId)," in handoff
sidebar = " ".join(_read("features/model-picker/components/sidebar-model-config.tsx").split())
assert "apiLoadable: isGguf && !isOllamaLinkPath(modelId)," in sidebar
# The same classification gates the backfill, or an older config still reaches the server.
@@ -2996,29 +2936,11 @@ def test_the_backfill_fills_in_fields_rather_than_skipping_known_keys():
assert "BEGIN IMMEDIATE" in db
-def test_the_hub_settings_page_matches_a_resident_path_loaded_model():
- """A GGUF loaded from an inactive HF cache or straight off disk loads by path, but
- /status reports the clean public id, so comparing it to settingsTarget.id said "not
- loaded" and the page showed saved or default values instead of the live launch
- config."""
- hub = " ".join(_read("features/hub/hub-page.tsx").split())
- assert (
- "residentModelIdMatches( activeCheckpoint, settingsTarget.id, settingsTarget.configId, )"
- in hub
- )
- assert "loadedConfig={settingsTargetIsResident ? activeModelConfig : null}" in hub
- assert "settingsTargetIsResident ? activeLoadedContextLength : null" in hub
- # The loadable identifier, as every other status reader records it -- except for a
- # speech model, which chat cannot adopt at all. speechOnly rides beside the null so
- # the helper can tell it from the empty slot the idle-unload rule is about.
- assert (
- "checkpointId: isSpeechOnlyStatus(status) ? null "
- ": resolveInferenceCheckpointId(status), "
- "speechOnly: isSpeechOnlyStatus(status)," in hub
- )
- assert "setCheckpoint(status.active_model" not in hub
+def test_public_model_identity_matches_the_backend_for_path_loaded_models():
+ """Path-loaded models use the backend's public identity rule wherever a
+ resident model is reconciled with a picker entry."""
chat = " ".join(_read("features/chat/lib/apply-inference-status-to-store.ts").split())
- assert "return status.model_identifier ?? status.active_model;" in chat, "the rule this mirrors"
+ assert "return status.model_identifier ?? status.active_model;" in chat
# The alias is the backend's own public id rule, not a private heuristic.
identity = _read("features/hub/lib/model-identity.ts")
assert "export function publicModelId(" in identity
@@ -3029,35 +2951,9 @@ def test_the_hub_settings_page_matches_a_resident_path_loaded_model():
assert "def public_model_id(" in backend, "the rule this mirrors"
-def test_the_hub_hydrates_the_live_settings_before_it_offers_them():
- """The Hub builds activeModelConfig out of the chat runtime store, and landing straight
- on /hub is the one entry point where nothing has applied /api/inference/status yet:
- useChatModelRuntime has no mount sync and the chat page is a different route."""
- hub = " ".join(_read("features/hub/hub-page.tsx").split())
- assert "adoptResidentModelStatus(" in hub
- assert "applyActiveModelStatusToStore(status, {" in hub
- assert "previousCheckpoint: previous.checkpoint ?? undefined," in hub
- assert "previousGgufVariant: previous.ggufVariant," in hub
- assert "modelLoading: store.modelLoading," in hub
-
- adopt = " ".join(_read("features/hub/lib/adopt-inference-status.ts").split())
- # Unconditional: a persisted checkpoint rehydrates without the fields saying how it launched.
- assert "actions.applyStatus(previous); return true;" in adopt
- # Never fight the owning load, nor describe an external model with the resident's settings.
- assert "if (state.checkpointIsExternal) { return false; }" in adopt
- assert "if (state.modelLoading) { return false; }" in adopt
-
- # Same call the chat runtime's own refresh makes, which is the rule this mirrors.
- runtime = " ".join(_read("features/chat/hooks/use-chat-model-runtime.ts").split())
- assert "applyActiveModelStatusToStore(statusRes, {" in runtime
-
-
-def test_the_hub_settings_editor_reseeds_when_the_live_config_lands():
+def test_the_sidebar_settings_editor_reseeds_when_the_live_config_lands():
"""ModelConfigPage reads loadedConfig in a useState initializer, so it seeds once per
mounted instance."""
- view = " ".join(_read("features/hub/catalog/hub-model-settings-view.tsx").split())
- assert "key={modelConfigInstanceKey( target.id, target.ggufVariant, loadedConfig, )}" in view
- # Same key the sidebar entry uses; that parity is the point.
sidebar = " ".join(_read("features/model-picker/components/sidebar-model-config.tsx").split())
assert "key={modelConfigInstanceKey(modelId, settingsGgufVariant, loadedConfig)}" in sidebar
@@ -3079,46 +2975,7 @@ def test_the_hub_settings_editor_reseeds_when_the_live_config_lands():
assert "const [initial] = useState(resolveInitial);" in page, "the rule this mirrors"
-def test_a_standalone_gguf_has_one_settings_key():
- """The inventory labels a single scanned .gguf from its filename, so the Hub row menu
- keyed its settings to `<path>:Q4_K_M` while the Chat picker, the detail card and the
- backfill all use the bare path: two surfaces, two configs."""
- hub = " ".join(_read("features/hub/hub-page.tsx").split())
- assert "let ggufVariant = settingsGgufVariantForRow(row);" in hub
- assert "row.formatVariant" not in hub, "the row's raw label is not a settings key"
-
- helper = " ".join(_read("features/hub/inventory/settings-identity.ts").split())
- assert 'row.kind === "local" && row.path.toLowerCase().endsWith(".gguf")' in helper
-
- common = _read_backend("hub/services/models/common.py")
- # The rule this mirrors: a variant is derived only for a single scanned file.
- # gguf_variant_key, not the hub's extract_quant_label: for a standalone file there is no
- # directory to qualify, so the key IS the quant token -- and it keeps the bpw modifier, which
- # is what makes it agree with the loader's own _extract_quant_label (the equality the sibling
- # test below depends on). The hub label drops that modifier.
- assert "gguf_variant_key(gguf_files[0].name)" in common
- assert "if scan_path.is_file() and len(gguf_files) == 1" in common
-
-
-def test_a_standalone_gguf_is_resident_despite_its_derived_quant():
- """A loose .gguf keys its settings by the bare path with no variant, but the loader
- derives one from the filename (llama_cpp sets _hf_variant from _extract_quant_label)
- and /status reports it, so an equality between the two could never hold and the
- settings page withheld the live launch config from the very file that was loaded."""
- hub = " ".join(_read("features/hub/hub-page.tsx").split())
- assert "const settingsTargetIsStandaloneFile =" in hub
- assert 'settingsTarget.id.toLowerCase().endsWith(".gguf")' in hub
- assert (
- "(settingsTargetIsStandaloneFile || ggufVariantsMatch(activeGgufVariant, settingsTarget.ggufVariant))"
- in hub
- )
- backend = _read_backend("core/inference/llama_cpp.py")
- assert (
- "self._hf_variant = _extract_quant_label(gguf_path)" in backend
- ), "the derived label this accounts for"
-
-
-def test_a_standalone_gguf_has_one_settings_identity_everywhere():
+def test_a_standalone_gguf_has_one_settings_identity_in_the_picker():
"""A loose .gguf has no quant to choose between, but llama_cpp falls back to
_extract_quant_label(gguf_path) when a load names no variant, and /status echoes
that as gguf_variant."""
@@ -3132,15 +2989,11 @@ def test_a_standalone_gguf_has_one_settings_identity_everywhere():
# The label still shows the quant; only the identity drops it.
assert "displayName: ggufVariant ? `${leaf} · ${ggufVariant}` : leaf," in sidebar
- # One rule, one definition: the Hub row applies the same test.
identity = _read("features/hub/lib/model-identity.ts")
assert "export function isStandaloneGgufPath(" in identity
# The suffix, and something that names a file on this machine: see
# test_a_repo_id_ending_in_gguf_keeps_its_quant for why the suffix is not enough.
assert "GGUF_SUFFIX_RE.test(modelId)" in identity
- row_identity = _read("features/hub/inventory/settings-identity.ts")
- assert 'row.path.toLowerCase().endsWith(".gguf")' in row_identity
-
# The precedence that makes the bare path win, asserted on the real function rather than on
# inference.py's text: #8702 moved this ladder to utils/openai_auto_switch_settings.py
# unchanged, and the grep that used to live here went red for a pure refactor.
@@ -3171,30 +3024,9 @@ def test_monitor_unload_clears_only_the_model_it_freed():
assert "store.clearCheckpoint();" in page
-def test_settings_open_reads_status_before_resolving_the_quant():
- """A cache row carries no quant, so opening its settings resolves one from the store's
- active variant."""
- page = " ".join(_read("features/hub/hub-page.tsx").split())
- assert "const refreshResidentModelStatus = useCallback((): Promise<void> => {" in page
- assert (
- "await refreshResidentModelStatus(); if (settingsOpenSeq.current !== openSeq) return;"
- in page
- )
- # Three: both handlers read on entry, and openModelSettings reads again after the
- # variant lookup, whose network round trip is its own window for a switch to land.
- assert page.count("await refreshResidentModelStatus();") == 3
-
-
-def test_a_local_quant_folder_resolves_its_variants_by_path():
+def test_a_local_quant_folder_lists_its_variants_by_path():
"""A local row carries a repo id only inside the HF cache, so a plain folder of
- quants has none while still being marked as needing one, and the row menu's
- Settings could then only reach the error toast."""
- hub = " ".join(_read("features/hub/hub-page.tsx").split())
- assert (
- 'const repoId = row.kind === "cache" ? row.repoId : (row.repoId ?? row.path ?? null);'
- in hub
- )
- # The on-device card already lists by path, so both surfaces choose from one set of quants.
+ quants has none while still needing a local variant listing."""
card = " ".join(_read("features/hub/catalog/local-on-device-card.tsx").split())
assert "repoId: modelId, hfToken, preferLocalCache: true, localPath: localGgufPath," in card
# The backend scans a path in the repo_id position before the validation that would 400.
@@ -3204,50 +3036,6 @@ def test_a_local_quant_folder_resolves_its_variants_by_path():
assert "_is_valid_repo_id(repo_id)" in scan[1], "the branch has to come first"
-def test_no_settings_target_is_built_on_an_unread_store():
- """Every path out of both handlers seeds the editor from the store, and Apply
- reloads with what it seeded, so a target built before the status read lands can
- persist and launch the settings of the model an API switch displaced."""
- page = " ".join(_read("features/hub/hub-page.tsx").split())
- # Both handlers read first and drop the open if a newer one started meanwhile, and
- # the variant lookup's await gets the same treatment. Every read carries the guard.
- assert page.count("await refreshResidentModelStatus();") == 3
- assert (
- page.count(
- "await refreshResidentModelStatus(); if (settingsOpenSeq.current !== openSeq) return;"
- )
- == 3
- )
- # Nothing may build a target off a concurrent read instead of an awaited one.
- assert "Promise.all([ listGgufVariants(" not in page
-
-
-def test_an_empty_status_is_read_against_the_idle_unload_setting():
- """/status cannot say whether an empty answer is an idle eviction that reloads
- or a real unload, so the Hub reads the only endpoint that knows and keeps the
- checkpoint only while the loop is armed."""
- hub = " ".join(_read("features/hub/hub-page.tsx").split())
- adopt = " ".join(_read("features/hub/lib/adopt-inference-status.ts").split())
- # Awaited, not raced: the first status read is the one most likely to land on an evicted
- # model, and an unresolved default of false would clear a checkpoint that is coming back.
- assert (
- "Promise.all([getInferenceStatus(), readIdleUnloadArmed()]) "
- ".then(([status, idleUnloadArmed]) => {" in hub
- )
- assert "idleUnloadArmed," in hub
- # Read with every status read, not cached for the life of the page: the idle timeout
- # is editable from Settings while this page stays mounted.
- assert "idleUnloadRead.current ??=" not in hub
- assert "idleUnloadArmed.current = settings.idleUnloadActive;" in hub
- # A failed read keeps the last answer. The default is disarmed, which is the side
- # that clears the checkpoint, so falling back to it would drop a live selection.
- assert ".catch(() => idleUnloadArmed.current)" in hub
- # An Audio load taking the single slot is not an idle eviction: nothing stashes the
- # chat model, so the exemption must not swallow that case.
- assert "if (state.idleUnloadArmed && !status.speechOnly) { return false; }" in adopt
- assert "actions.clearCheckpoint?.();" in adopt
-
-
def test_a_repo_id_ending_in_gguf_keeps_its_quant():
"""Repo ids ending in .gguf are real on the Hub, an iMat repo among them, and
those hold every quant. A suffix-only test reads one as a single file, drops
@@ -3259,24 +3047,6 @@ def test_a_repo_id_ending_in_gguf_keeps_its_quant():
assert "isNativeFileLabel(modelId)" in identity
-def test_cached_repo_settings_key_follows_the_row_not_the_view():
- """A repo in an inactive HF cache loads by snapshot path while its settings are keyed
- by repo id."""
- page = " ".join(_read("features/hub/hub-page.tsx").split())
- assert 'if (kind !== "cache" && resource.source !== "hub_cache") {' in page
-
-
-def test_detail_settings_defers_a_derived_quant_to_a_fresh_status_read():
- """The on-device card resolves the quant it shows from the store's active variant, and
- nothing re-reads status while the window keeps focus, so an API-driven switch leaves
- that quant naming the model it displaced."""
- page = " ".join(_read("features/hub/hub-page.tsx").split())
- card = " ".join(_read("features/hub/catalog/local-on-device-card.tsx").split())
- assert "if (!quantIsUserPicked) { const settled = useChatRuntimeStore.getState();" in page
- assert "variant = settled.activeGgufVariant;" in page
- assert "onOpenSettings(selectedQuant ?? null, quantIsUserPicked)" in card
-
-
def test_a_gpu_pin_is_mirrored_to_the_server_with_its_index_space():
"""The same integers are Vulkan ordinals under Vulkan and device indices elsewhere, so
a pin mirrored without its namespace would, after a backend change, address a different
@@ -3290,9 +3060,8 @@ def test_a_gpu_pin_is_mirrored_to_the_server_with_its_index_space():
assert 'if (gpuIndexKind !== "physical") { payload.gpu_index_kind = gpuIndexKind; }' in mirror
-def test_a_cached_repo_keeps_the_settings_saved_under_its_old_key():
- """A cached repo was keyed by the snapshot path it loads from and is now keyed by its
- repo id; the server backfill only mirrors what is stored, so nothing else moves it."""
+def test_a_cached_repo_moves_legacy_settings_to_the_current_key():
+ """A cached repo's legacy snapshot-path settings move to its current repo-id key."""
config = " ".join(_read("features/model-picker/model-config/per-model-config.ts").split())
assert "export function adoptLegacyConfigKey(" in config
# The key is renamed in one write. Saving the second copy first puts a full map one entry
@@ -3310,9 +3079,6 @@ def test_a_cached_repo_keeps_the_settings_saved_under_its_old_key():
"delete map[legacyKey]; deleteConfigEntriesForModelVariant(map, legacyModelId, "
"ggufVariant);" in config
)
- # Both entry points move it before anything reads the new key.
- page = " ".join(_read("features/hub/hub-page.tsx").split())
- assert page.count("adoptLegacyConfigKey(") == 2
def test_clearing_the_log_keeps_a_request_that_is_still_running():