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.

Studio: change Parallel Slots from the API page's Slots readout (#14225)

Daniel Hancd2595f19d6e
diff --git a/studio/frontend/scripts/check-bundle-budget.ts b/studio/frontend/scripts/check-bundle-budget.ts
index 2315f1b7ca9..f7410578e43 100644
--- a/studio/frontend/scripts/check-bundle-budget.ts
+++ b/studio/frontend/scripts/check-bundle-budget.ts
@@ -281,8 +281,14 @@ export const BUDGET = {
// Naming a chat export after its model, title and day (#12586) adds 615 B raw / 65 B transfer (6,842,183 ->
// 6,842,798 B raw, 2,060,321 -> 2,060,386 B transfer against 46e4f50bea, one machine and build, Node 24,
// 96 chunks): the name builder sits beside the chat exporters in the eager chat chunk. Raised by that growth.
- transferBytes: 2_060_794,
- rawBytes: 6_843_649,
+ // Opening Parallel Slots from the API page's Slots readout (#14225) adds 546 B raw / 212 B transfer (6,842,518 ->
+ // 6,843,064 B raw, 2,060,427 -> 2,060,639 B transfer against bff0a25b4, one machine and build, Node 24, 96 chunks):
+ // the field focus and per-visit Advanced reveal live in the Run settings page and the picker, both in the eager chat
+ // chunk. Raised to the measured build plus the same 160 B raw / 189 B transfer margin.
+ // Reconciled with main at 87c6408455: the combined build measures 6,844,101 B raw / 2,060,934 B transfer
+ // (Node 24, 96 chunks). The API Settings button stays in its lazy page. Keep the same 160 B / 189 B margin.
+ transferBytes: 2_061_123,
+ rawBytes: 6_844_261,
};
// The chunk count is reported but not budgeted. Splitting a page out of the entry raises it while lowering the
diff --git a/studio/frontend/src/features/api-monitor/api-monitor-page.tsx b/studio/frontend/src/features/api-monitor/api-monitor-page.tsx
index b44840b5d45..5f742954637 100644
--- a/studio/frontend/src/features/api-monitor/api-monitor-page.tsx
+++ b/studio/frontend/src/features/api-monitor/api-monitor-page.tsx
@@ -25,6 +25,10 @@ import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
import type { ApiMonitorEntry } from "@/features/chat";
import { isExternalModelId } from "@/features/chat/external-providers";
import { modelIdsMatch } from "@/features/hub/lib/model-identity";
+import {
+ type ModelConfigHandoffRequest,
+ createModelConfigHandoffRequestId,
+} from "@/features/model-picker";
import {
lanApiUrls,
loadLanAccess,
@@ -53,6 +57,13 @@ import { isLifecycleEntry, lifecycleLabel } from "./lifecycle";
import { PAYLOAD_SHOW_ALL_MAX_CHARS, payloadPreview } from "./payload-preview";
import { ToolbarTooltip } from "./components/toolbar-tooltip";
import { unloadResident } from "./unload-resident";
+import {
+ residentHasRunSettings,
+ residentHasSlotsSetting,
+ runSettingsRequest,
+ slotsSettingsRequest,
+} from "./slots-settings-request";
+import { useResidentRunTarget } from "./use-resident-run-target";
import {
type MonitorStatusFilter,
filterEntries,
@@ -604,6 +615,13 @@ export function ApiMonitorPage(): ReactElement {
// block raw unloads while a picker load could finish afterward without a lifecycle guard.
const modelLoading = useChatRuntimeStore((s) => s.modelLoading);
const [unloadError, setUnloadError] = useState<string | null>(null);
+ const resident = useResidentRunTarget(data?.active_model);
+ const [configRequest, setConfigRequest] =
+ useState<ModelConfigHandoffRequest | null>(null);
+ const canEditSettings =
+ !modelLoading && !unloading && residentHasRunSettings(resident);
+ const canEditSlots =
+ !modelLoading && !unloading && residentHasSlotsSetting(resident);
useEffect(() => {
const refreshRemoteBase = () => {
@@ -744,6 +762,15 @@ export function ApiMonitorPage(): ReactElement {
: serverStatus === "ready"
? "Ready"
: "No model loaded";
+ const slotsLabel = data?.queue
+ ? `${data.queue.active}/${data.queue.capacity} busy${
+ data.queue.queued > 0 ? ` · ${data.queue.queued} queued` : ""
+ }`
+ : "";
+ const slotsClassName = cn(
+ "text-ui-12 tabular-nums text-foreground",
+ (data?.queue?.queued ?? 0) > 0 && "text-amber-700 dark:text-amber-500",
+ );
return (
<main className="mx-auto flex w-full max-w-6xl 3xl:max-w-[calc(1440px*var(--ui-space-scale,1))] 4xl:max-w-[calc(1760px*var(--ui-space-scale,1))] flex-col gap-6 px-6 pb-10 pt-12 max-sm:px-4 max-sm:pt-8 font-heading sm:px-10">
@@ -763,6 +790,19 @@ export function ApiMonitorPage(): ReactElement {
onSettled={refresh}
onUnloadActive={unloadActiveModel}
unloading={unloading}
+ configRequest={configRequest}
+ onConfigRequestAdopted={() => setConfigRequest(null)}
+ onOpenSettings={
+ canEditSettings
+ ? () =>
+ setConfigRequest(
+ runSettingsRequest(
+ resident,
+ createModelConfigHandoffRequestId(),
+ ),
+ )
+ : undefined
+ }
/>
<ToolbarTooltip
label={paused ? "Resume the live feed" : "Pause the live feed"}
@@ -910,15 +950,29 @@ export function ApiMonitorPage(): ReactElement {
<span className="text-ui-10 font-medium uppercase tracking-wider text-muted-foreground">
Slots
</span>
- <span
- className={cn(
- "text-ui-12 tabular-nums text-foreground",
- data.queue.queued > 0 && "text-amber-700 dark:text-amber-500",
- )}
- >
- {data.queue.active}/{data.queue.capacity} busy
- {data.queue.queued > 0 ? ` · ${data.queue.queued} queued` : ""}
- </span>
+ {canEditSlots ? (
+ <ToolbarTooltip label="Change parallel slots">
+ <button
+ type="button"
+ onClick={() =>
+ setConfigRequest(
+ slotsSettingsRequest(
+ resident,
+ createModelConfigHandoffRequestId(),
+ ),
+ )
+ }
+ className={cn(
+ slotsClassName,
+ "w-fit cursor-pointer text-left underline decoration-dotted decoration-muted-foreground/60 underline-offset-4 hover:decoration-foreground",
+ )}
+ >
+ {slotsLabel}
+ </button>
+ </ToolbarTooltip>
+ ) : (
+ <span className={slotsClassName}>{slotsLabel}</span>
+ )}
</div>
) : null}
<div className="flex min-w-0 flex-1 flex-col">
diff --git a/studio/frontend/src/features/api-monitor/components/api-model-load-controls.tsx b/studio/frontend/src/features/api-monitor/components/api-model-load-controls.tsx
index c52af249dff..dbf39c45354 100644
--- a/studio/frontend/src/features/api-monitor/components/api-model-load-controls.tsx
+++ b/studio/frontend/src/features/api-monitor/components/api-model-load-controls.tsx
@@ -22,17 +22,17 @@ import {
} from "@/features/hub";
import {
type LoraModelOption,
+ type ModelConfigHandoffRequest,
type ModelOption,
ModelSelector,
type ModelSelectorChangeMeta,
currentRuntimePerModelConfig,
resolveResidentInitialConfig,
- splitQuantSuffix,
useActiveModelConfig,
} from "@/features/model-picker";
import { isNpuModelId } from "@/features/npu";
import { cn } from "@/lib/utils";
-import { ReloadIcon } from "@hugeicons/core-free-icons";
+import { ReloadIcon, Settings02Icon } from "@hugeicons/core-free-icons";
import { HugeiconsIcon } from "@hugeicons/react";
import {
type Dispatch,
@@ -47,6 +47,7 @@ import {
RELOAD_MISSING_HISTORY_MESSAGE,
reloadLastModel,
} from "../reload-last-model";
+import { monitorSelection } from "../use-resident-run-target";
import { ToolbarTooltip } from "./toolbar-tooltip";
// A pick superseded by a newer one rejects too; the newer pick owns the outcome.
@@ -124,22 +125,6 @@ function localSelection(
};
}
-// The monitor reports a llama.cpp model as "<id>:<quant>"; Ollama ids carry their own ":<tag>".
-function monitorSelection(activeModel: string | null | undefined): {
- id: string;
- ggufVariant: string | null;
-} {
- if (!activeModel) {
- return { id: "", ggufVariant: null };
- }
- const split = activeModel.startsWith("ollama/")
- ? null
- : splitQuantSuffix(activeModel);
- return split
- ? { id: split[0], ggufVariant: split[1] }
- : { id: activeModel, ggufVariant: null };
-}
-
type PendingDownload = {
id: string;
meta: ModelSelectorChangeMeta;
@@ -226,12 +211,18 @@ export function ApiModelLoadControls({
onSettled,
onUnloadActive,
unloading,
+ configRequest = null,
+ onConfigRequestAdopted,
+ onOpenSettings,
}: {
activeModel: string | null | undefined;
onSettled: () => void;
onUnloadActive: () => void;
// The page's unload rereads the resident on its second pass, so a load started meanwhile would be unloaded.
unloading: boolean;
+ configRequest?: ModelConfigHandoffRequest | null;
+ onConfigRequestAdopted?: (requestId: string) => void;
+ onOpenSettings?: () => void;
}): ReactElement {
const { selectModel, loadNpuModel, refresh, ejectModel } =
useChatModelRuntime();
@@ -273,6 +264,13 @@ export function ApiModelLoadControls({
refreshLastLoadLabel();
}, [refresh, refreshLocalModels, refreshLastLoadLabel]);
+ // Opened during render, as the picker adopts the request, so the same pass mounts its page.
+ const [openedRequestId, setOpenedRequestId] = useState<string | null>(null);
+ if (configRequest && configRequest.requestId !== openedRequestId) {
+ setOpenedRequestId(configRequest.requestId);
+ setSelectorOpen(true);
+ }
+
// An API auto-switch changes the resident model behind Chat's store; this page's own loads
// reconcile themselves.
useEffect(() => {
@@ -427,7 +425,7 @@ export function ApiModelLoadControls({
publicModelId(runtimeCheckpoint) === effectiveModel;
return (
- <div className="flex flex-wrap items-center gap-2">
+ <div className="flex min-w-0 max-w-full flex-wrap items-center gap-2">
<ModelSelector
models={models}
loraModels={loraModels}
@@ -438,6 +436,8 @@ export function ApiModelLoadControls({
activeLoadedContextLength={
runtimeIsResident ? loadedContextLength : null
}
+ configRequest={configRequest}
+ onConfigRequestAdopted={onConfigRequestAdopted}
onValueChange={(value, meta) => {
handlePick(value, meta);
}}
@@ -456,8 +456,35 @@ export function ApiModelLoadControls({
placeholder="Load model"
open={selectorOpen && !unloading}
onOpenChange={setSelectorOpen}
- className={cn("h-9 rounded-full")}
+ className="h-9 max-w-full rounded-full"
+ contentClassName="max-sm:max-w-[calc(100vw-3rem)]"
/>
+ <ToolbarTooltip
+ label={
+ activeModel
+ ? "Model run settings"
+ : "Load a model to change its run settings"
+ }
+ >
+ <Button
+ type="button"
+ variant="outline"
+ size="sm"
+ aria-label="Model settings"
+ aria-haspopup="dialog"
+ aria-expanded={selectorOpen && !unloading}
+ onClick={onOpenSettings}
+ disabled={busy || reloading || !onOpenSettings}
+ className="h-9 gap-1.5 rounded-full"
+ >
+ <HugeiconsIcon
+ icon={Settings02Icon}
+ strokeWidth={1.75}
+ className="size-4"
+ />
+ Settings
+ </Button>
+ </ToolbarTooltip>
<ToolbarTooltip label={reloadTitle(lastLoadLabel, Boolean(activeModel))}>
<Button
type="button"
diff --git a/studio/frontend/src/features/api-monitor/slots-settings-request.ts b/studio/frontend/src/features/api-monitor/slots-settings-request.ts
new file mode 100644
index 00000000000..cf0ab0eea2c
--- /dev/null
+++ b/studio/frontend/src/features/api-monitor/slots-settings-request.ts
@@ -0,0 +1,106 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// eslint-disable-next-line no-restricted-imports -- The barrel pulls in React components, which node tests cannot load.
+import {
+ type ModelConfigHandoffRequest,
+ modelConfigTarget,
+ modelConfigTargetIsResident,
+} from "@/features/model-picker/model-config/model-config-handoff";
+
+/** The resident model as Chat's runtime holds it: the same fields Chat's own reload sends. */
+export type ResidentRunTarget = {
+ checkpoint: string;
+ ggufVariant: string | null;
+ isGguf: boolean;
+ isLora: boolean;
+ isDiffusion: boolean;
+ loadId: string | null;
+ nativePathToken: string | null;
+ nativePathExpiresAtMs: number | null;
+};
+
+/** Run can only apply settings while access to the resident model is still valid. */
+export function residentHasRunSettings(
+ resident: ResidentRunTarget | null,
+ nowMs: number = Date.now(),
+): resident is ResidentRunTarget {
+ if (!resident) {
+ return false;
+ }
+ // An expired file token cannot reload the model, so there is nothing to apply the change to.
+ return !(
+ resident.nativePathToken &&
+ resident.nativePathExpiresAtMs != null &&
+ nowMs >= resident.nativePathExpiresAtMs
+ );
+}
+
+/** Only a llama-server load reports slots to the monitor; the diffusion runner gets no --parallel. */
+export function residentHasSlotsSetting(
+ resident: ResidentRunTarget | null,
+ nowMs: number = Date.now(),
+): resident is ResidentRunTarget {
+ return (
+ residentHasRunSettings(resident, nowMs) &&
+ resident.isGguf &&
+ !resident.isDiffusion
+ );
+}
+
+/** Open the loaded model's settings with Advanced visible for this visit. */
+export function runSettingsRequest(
+ resident: ResidentRunTarget | null,
+ requestId: string,
+ nowMs: number = Date.now(),
+): ModelConfigHandoffRequest | null {
+ if (!residentHasRunSettings(resident, nowMs)) {
+ return null;
+ }
+ return {
+ requestId,
+ id: resident.checkpoint,
+ focus: "advanced",
+ meta: {
+ source: "local",
+ isLora: resident.isLora,
+ loadId: resident.loadId,
+ ggufVariant: resident.ggufVariant ?? undefined,
+ nativePathToken: resident.nativePathToken ?? undefined,
+ nativePathExpiresAtMs: resident.nativePathExpiresAtMs,
+ isGguf: resident.isGguf,
+ isDiffusion: resident.isDiffusion,
+ isDownloaded: true,
+ },
+ };
+}
+
+/** Opens the resident's Run settings on Parallel Slots, so Run saves and reloads as it does there. */
+export function slotsSettingsRequest(
+ resident: ResidentRunTarget | null,
+ requestId: string,
+ nowMs: number = Date.now(),
+): ModelConfigHandoffRequest | null {
+ const request = runSettingsRequest(resident, requestId, nowMs);
+ return request && residentHasSlotsSetting(resident, nowMs)
+ ? { ...request, focus: "parallelSlots" }
+ : null;
+}
+
+/** The picker's residency test on the target the request opens. A file-picked GGUF (path vs the
+ * monitor's stem) or a quant Chat has not caught up with fails it and would open on stored settings. */
+export function residentMatchesMonitor(
+ resident: ResidentRunTarget | null,
+ monitor: { id: string; ggufVariant: string | null },
+ nowMs: number = Date.now(),
+): boolean {
+ const request = runSettingsRequest(resident, "", nowMs);
+ return (
+ request !== null &&
+ modelConfigTargetIsResident({
+ target: modelConfigTarget(request.id, request.meta),
+ selectedId: monitor.id,
+ activeGgufVariant: monitor.ggufVariant,
+ })
+ );
+}
diff --git a/studio/frontend/src/features/api-monitor/use-resident-run-target.ts b/studio/frontend/src/features/api-monitor/use-resident-run-target.ts
new file mode 100644
index 00000000000..d7c2cd60a35
--- /dev/null
+++ b/studio/frontend/src/features/api-monitor/use-resident-run-target.ts
@@ -0,0 +1,81 @@
+// 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 { useChatRuntimeStore } from "@/features/chat";
+import { publicModelId } from "@/features/hub";
+import {
+ splitQuantSuffix,
+ useActiveModelConfig,
+} from "@/features/model-picker";
+import { useMemo } from "react";
+import {
+ type ResidentRunTarget,
+ residentMatchesMonitor,
+} from "./slots-settings-request";
+
+// The monitor reports a llama.cpp model as "<id>:<quant>"; Ollama ids carry their own ":<tag>".
+export function monitorSelection(activeModel: string | null | undefined): {
+ id: string;
+ ggufVariant: string | null;
+} {
+ if (!activeModel) {
+ return { id: "", ggufVariant: null };
+ }
+ const split = activeModel.startsWith("ollama/")
+ ? null
+ : splitQuantSuffix(activeModel);
+ return split
+ ? { id: split[0], ggufVariant: split[1] }
+ : { id: activeModel, ggufVariant: null };
+}
+
+/** The resident as Chat's runtime holds it, or null when that is not the model the monitor reports. */
+export function useResidentRunTarget(
+ activeModel: string | null | undefined,
+): ResidentRunTarget | null {
+ const { checkpoint, isGguf } = useActiveModelConfig();
+ const ggufVariant = useChatRuntimeStore((s) => s.activeGgufVariant);
+ const isDiffusion = useChatRuntimeStore((s) => s.loadedIsDiffusion);
+ const loadId = useChatRuntimeStore((s) => s.activeLoadId);
+ const nativePathToken = useChatRuntimeStore((s) => s.activeNativePathToken);
+ const nativePathExpiresAtMs = useChatRuntimeStore(
+ (s) => s.activeNativePathExpiresAtMs,
+ );
+ const models = useChatRuntimeStore((s) => s.models);
+ const loras = useChatRuntimeStore((s) => s.loras);
+ return useMemo(() => {
+ const { id, ggufVariant: monitorVariant } = monitorSelection(activeModel);
+ // The picker's own test for handing Run settings the live config (runtimeIsResident).
+ if (!id || checkpoint == null || publicModelId(checkpoint) !== id) {
+ return null;
+ }
+ const model = models.find((entry) => entry.id === checkpoint);
+ const isLora = model
+ ? model.isLora
+ : loras.find((entry) => entry.id === checkpoint)?.exportType === "lora";
+ const resident = {
+ checkpoint,
+ ggufVariant,
+ isGguf,
+ isLora,
+ isDiffusion,
+ loadId,
+ nativePathToken,
+ nativePathExpiresAtMs,
+ };
+ return residentMatchesMonitor(resident, { id, ggufVariant: monitorVariant })
+ ? resident
+ : null;
+ }, [
+ activeModel,
+ checkpoint,
+ isGguf,
+ ggufVariant,
+ isDiffusion,
+ loadId,
+ nativePathToken,
+ nativePathExpiresAtMs,
+ models,
+ loras,
+ ]);
+}
diff --git a/studio/frontend/src/features/model-picker/components/model-config-page.tsx b/studio/frontend/src/features/model-picker/components/model-config-page.tsx
index c8f063c26bd..1558397e824 100644
--- a/studio/frontend/src/features/model-picker/components/model-config-page.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-config-page.tsx
@@ -220,6 +220,7 @@ import {
import { ChatTemplateEditorDialog } from "./chat-template-editor-dialog";
import { MemoryEstimateRow } from "./memory-estimate-row";
import type { ModelPickTarget } from "./model-selector/types";
+import type { ModelConfigFocus } from "../model-config/model-config-handoff";
import { reconcileTensorSplit } from "@/hooks/gpu-tensor-split";
import {
NumericValueInput,
@@ -1260,11 +1261,25 @@ function ParallelSlotsRow({
config,
update,
hint,
+ focusOnMount = false,
}: {
config: PerModelConfig;
update: (patch: Partial<PerModelConfig>) => void;
hint: string;
+ focusOnMount?: boolean;
}) {
+ const inputRef = useRef<HTMLInputElement>(null);
+ useEffect(() => {
+ if (!focusOnMount) {
+ return;
+ }
+ // Next frame: the popover moves focus to itself after its children mount.
+ const frame = window.requestAnimationFrame(() => {
+ inputRef.current?.scrollIntoView({ block: "center" });
+ inputRef.current?.focus({ preventScroll: true });
+ });
+ return () => window.cancelAnimationFrame(frame);
+ }, [focusOnMount]);
return (
<div className={ROW_CLASS}>
<div className="flex min-w-0 items-center gap-1.5">
@@ -1272,6 +1287,7 @@ function ParallelSlotsRow({
<InfoHint>{hint}</InfoHint>
</div>
<input
+ ref={inputRef}
type="number"
min={N_PARALLEL_MIN}
max={N_PARALLEL_MAX}
@@ -1658,6 +1674,7 @@ function GgufAdvancedSettings({
moeLayersInputRef,
onExtraArgsLoadableChange,
draftKey,
+ focusParallelSlots,
}: {
config: PerModelConfig;
update: (patch: Partial<PerModelConfig>) => void;
@@ -1674,6 +1691,7 @@ function GgufAdvancedSettings({
/** Which stored entries the extra-arguments row reads, most specific first. */
onExtraArgsLoadableChange: (loadable: boolean) => void;
draftKey: string;
+ focusParallelSlots: boolean;
}) {
const batchAdviceId = useId();
const ubatchAdviceId = useId();
@@ -1853,7 +1871,12 @@ function GgufAdvancedSettings({
</div>
)}
- <ParallelSlotsRow config={config} update={update} hint={GGUF_PARALLEL_HINT} />
+ <ParallelSlotsRow
+ config={config}
+ update={update}
+ hint={GGUF_PARALLEL_HINT}
+ focusOnMount={focusParallelSlots}
+ />
{!isDiffusion && (
<div className="space-y-1">
@@ -2342,6 +2365,7 @@ interface ModelConfigPageProps {
/** Page variant only: render the built-in "Run settings" title block. A host that already shows
* the model name as its page heading turns this off. */
showHeader?: boolean;
+ focusField?: ModelConfigFocus;
}
export function ModelConfigPage({
@@ -2354,6 +2378,7 @@ export function ModelConfigPage({
isDiffusion = false,
variant = "page",
showHeader = true,
+ focusField,
}: ModelConfigPageProps) {
const rememberId = useId();
const platformDeviceType = usePlatformStore((s) => s.deviceType);
@@ -2568,10 +2593,16 @@ export function ModelConfigPage({
// and a width that starts applying then has to surface.
const autoOpenForMlxKvQuant = servedByMlx && initialMlxKvQuant != null;
const autoOpenForMlxInt8Prefill = servedByMlx && initialMlxInt8Prefill;
+ // Opened to change one field: show its section for this visit, leaving the saved switch alone.
+ const [revealForFocus, setRevealForFocus] = useState(focusField != null);
const showAdvanced =
- advancedPreference ??
- (autoOpenAdvanced || autoOpenForMlxKvQuant || autoOpenForMlxInt8Prefill);
- const toggleAdvanced = saveAdvancedSettingsOpen;
+ revealForFocus ||
+ (advancedPreference ??
+ (autoOpenAdvanced || autoOpenForMlxKvQuant || autoOpenForMlxInt8Prefill));
+ const toggleAdvanced = (open: boolean) => {
+ setRevealForFocus(false);
+ saveAdvancedSettingsOpen(open);
+ };
const contextInputRef = useRef<NumericValueInputHandle>(null);
const maxSeqLengthInputRef = useRef<NumericValueInputHandle>(null);
const gpuLayersInputRef = useRef<NumericValueInputHandle>(null);
@@ -3813,6 +3844,7 @@ export function ModelConfigPage({
moeLayersInputRef={moeLayersInputRef}
draftKey={draftKey}
onExtraArgsLoadableChange={setExtraArgsLoadable}
+ focusParallelSlots={focusField === "parallelSlots"}
/>
)}
</>
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 24ea2d12fcf..7e903ceaac9 100644
--- a/studio/frontend/src/features/model-picker/components/model-selector.tsx
+++ b/studio/frontend/src/features/model-picker/components/model-selector.tsx
@@ -37,6 +37,7 @@ import {
useState,
} from "react";
import {
+ type ModelConfigFocus,
type ModelConfigHandoffRequest,
modelConfigTarget,
modelConfigTargetIsResident,
@@ -480,6 +481,7 @@ function ModelSelectorContent({
const [adoptedConfigRequestId, setAdoptedConfigRequestId] = useState<
string | null
>(null);
+ const [configFocus, setConfigFocus] = useState<ModelConfigFocus | null>(null);
// The picker remounts on each open but this section state does not, so re-derive the default
// section on the open edge.
@@ -494,6 +496,7 @@ function ModelSelectorContent({
}
if (!open && wasOpen.current) {
setConfigTarget(null);
+ setConfigFocus(null);
}
wasOpen.current = open;
}, [open, wantsConnectedDefault, hasAdditionalOnDeviceModels]);
@@ -545,6 +548,7 @@ function ModelSelectorContent({
? id.slice(NPU_MODEL_PREFIX.length)
: undefined;
setConfigTarget(modelConfigTarget(id, meta, displayName));
+ setConfigFocus(null);
};
const requestedConfigTarget = useMemo(
() =>
@@ -581,6 +585,7 @@ function ModelSelectorContent({
) {
setAdoptedConfigRequestId(configRequest.requestId);
setConfigTarget(requestedConfigTarget);
+ setConfigFocus(configRequest.focus ?? null);
}
useEffect(() => {
if (!configRequest || adoptedConfigRequestId !== configRequest.requestId) {
@@ -652,7 +657,11 @@ function ModelSelectorContent({
visibleLoadedConfig,
)}
target={visibleConfigTarget}
- onBack={() => setConfigTarget(null)}
+ onBack={() => {
+ setConfigTarget(null);
+ setConfigFocus(null);
+ }}
+ focusField={configFocus ?? undefined}
onRun={(config, isDiffusion) =>
onSelect(
visibleConfigTarget.configId ?? visibleConfigTarget.id,
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
index e1639d41126..6164b69701b 100644
--- 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
@@ -24,8 +24,12 @@ export interface ModelConfigHandoffRequest {
id: string;
displayName?: string;
meta: ModelSelectorChangeMeta;
+ /** Section or field the Run settings page opens on. */
+ focus?: ModelConfigFocus;
}
+export type ModelConfigFocus = "advanced" | "parallelSlots";
+
interface ModelConfigHandoffState {
request: ModelConfigHandoffRequest | null;
submit: (request: ModelConfigHandoffRequest) => void;
diff --git a/studio/frontend/tests/api-monitor-slots-settings-request.test.ts b/studio/frontend/tests/api-monitor-slots-settings-request.test.ts
new file mode 100644
index 00000000000..deb52bf0c76
--- /dev/null
+++ b/studio/frontend/tests/api-monitor-slots-settings-request.test.ts
@@ -0,0 +1,137 @@
+// 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 { ResidentRunTarget } from "../src/features/api-monitor/slots-settings-request.ts";
+import { registerBundlerResolver } from "./helpers/kit.ts";
+
+registerBundlerResolver();
+
+const {
+ residentHasRunSettings,
+ residentHasSlotsSetting,
+ residentMatchesMonitor,
+ runSettingsRequest,
+ slotsSettingsRequest,
+} = await import("../src/features/api-monitor/slots-settings-request.ts");
+
+const GGUF: ResidentRunTarget = {
+ checkpoint: "unsloth/Qwen3-4B-GGUF",
+ ggufVariant: "UD-Q4_K_XL",
+ isGguf: true,
+ isLora: false,
+ isDiffusion: false,
+ loadId: null,
+ nativePathToken: null,
+ nativePathExpiresAtMs: null,
+};
+
+test("Settings opens Advanced without focusing Parallel Slots or changing the load identity", () => {
+ const resident = {
+ ...GGUF,
+ loadId: "/cache/qwen",
+ nativePathToken: "token",
+ nativePathExpiresAtMs: 1_000,
+ };
+ const request = runSettingsRequest(resident, "settings", 999);
+ const slots = slotsSettingsRequest(resident, "slots", 999);
+ assert.equal(request?.focus, "advanced");
+ assert.equal(request?.id, resident.checkpoint);
+ assert.equal(request?.meta.loadId, resident.loadId);
+ assert.equal(request?.meta.nativePathToken, "token");
+ assert.deepEqual(request?.meta, slots?.meta);
+ assert.equal(runSettingsRequest(resident, "expired", 1_000), null);
+ assert.equal(residentHasRunSettings(null), false);
+});
+
+test("Settings also supports non-GGUF residents without offering a Slots shortcut", () => {
+ const resident = {
+ ...GGUF,
+ checkpoint: "unsloth/Qwen3-0.6B",
+ isGguf: false,
+ ggufVariant: null,
+ isLora: true,
+ };
+ assert.equal(residentHasRunSettings(resident), true);
+ assert.equal(
+ residentMatchesMonitor(resident, {
+ id: resident.checkpoint,
+ ggufVariant: null,
+ }),
+ true,
+ );
+ assert.equal(runSettingsRequest(resident, "settings")?.meta.isLora, true);
+ assert.equal(slotsSettingsRequest(resident, "slots"), null);
+});
+
+test("a GGUF resident opens its Run settings on Parallel Slots, as Chat's reload names it", () => {
+ assert.deepEqual(slotsSettingsRequest(GGUF, "req-1"), {
+ requestId: "req-1",
+ id: "unsloth/Qwen3-4B-GGUF",
+ focus: "parallelSlots",
+ meta: {
+ source: "local",
+ isLora: false,
+ loadId: null,
+ ggufVariant: "UD-Q4_K_XL",
+ nativePathToken: undefined,
+ nativePathExpiresAtMs: null,
+ isGguf: true,
+ isDiffusion: false,
+ isDownloaded: true,
+ },
+ });
+});
+
+test("models without a Parallel Slots field stay a plain readout", () => {
+ assert.equal(residentHasSlotsSetting(null), false);
+ assert.equal(slotsSettingsRequest(null, "req"), null);
+ // Transformers, MLX and NPU loads: the monitor reports no llama-server slots for them.
+ assert.equal(residentHasSlotsSetting({ ...GGUF, isGguf: false }), false);
+ // The diffusion runner never receives --parallel.
+ assert.equal(residentHasSlotsSetting({ ...GGUF, isDiffusion: true }), false);
+});
+
+test("a file-picked GGUF whose access expired cannot be reloaded, so it is not offered", () => {
+ const picked = {
+ ...GGUF,
+ nativePathToken: "tok",
+ nativePathExpiresAtMs: 1_000,
+ };
+ assert.equal(residentHasSlotsSetting(picked, 999), true);
+ assert.equal(
+ slotsSettingsRequest(picked, "req", 999)?.meta.nativePathToken,
+ "tok",
+ );
+ assert.equal(residentHasSlotsSetting(picked, 1_000), false);
+});
+
+test("the readout is offered only when the picker takes the target for the monitor's model", () => {
+ const monitor = { id: "unsloth/Qwen3-4B-GGUF", ggufVariant: "UD-Q4_K_XL" };
+ assert.equal(residentMatchesMonitor(GGUF, monitor), true);
+ // An API auto-switch to another quant of the same repo, before Chat's runtime catches up:
+ // Run would reload the old quant.
+ assert.equal(
+ residentMatchesMonitor(GGUF, { ...monitor, ggufVariant: "Q8_0" }),
+ false,
+ );
+ // A file-picked GGUF: Chat holds the path, the monitor its bare stem. The picker cannot tie
+ // them, so Run settings would open on stored settings rather than the loaded ones.
+ const picked = {
+ ...GGUF,
+ checkpoint: "/home/user/models/qwen3-4b-Q4_K_M.gguf",
+ ggufVariant: "Q4_K_M",
+ nativePathToken: "tok",
+ nativePathExpiresAtMs: Number.MAX_SAFE_INTEGER,
+ };
+ assert.equal(
+ residentMatchesMonitor(picked, {
+ id: "qwen3-4b-Q4_K_M",
+ ggufVariant: "Q4_K_M",
+ }),
+ false,
+ );
+ assert.equal(residentMatchesMonitor(null, monitor), false);
+});