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: show Vulkan iGPU memory as shared availability (#11305)

oobabooga857f68d1e375
diff --git a/studio/frontend/src/components/floating-monitor.tsx b/studio/frontend/src/components/floating-monitor.tsx
index ae1db35d8db..9577967bb43 100644
--- a/studio/frontend/src/components/floating-monitor.tsx
+++ b/studio/frontend/src/components/floating-monitor.tsx
@@ -9,6 +9,7 @@ import {
useMonitorFrameStore,
useMonitorOverlayStore,
} from "@/features/settings";
+import { gpuMemoryDisplay } from "@/hooks/gpu-memory-display";
import { gpuMemoryTotalsGb, resolveGpuVramUsedGb } from "@/hooks/gpu-vram";
import { aggregateGpuMemoryTotalGb, useSystemInfo } from "@/hooks/use-system";
import { useT } from "@/i18n";
@@ -460,16 +461,18 @@ function FloatingMonitorPanel({
systemInfo.inference_gpu.backend !== systemInfo.gpu.backend
? systemInfo.inference_gpu
: null;
+ const memoryDisplay = gpuMemoryDisplay(displayedGpu);
+ const inferenceDisplay = gpuMemoryDisplay(separateInferenceGpu);
const inferenceVramTotal = separateInferenceGpu
- ? aggregateGpuMemoryTotalGb(separateInferenceGpu.devices)
+ ? aggregateGpuMemoryTotalGb(inferenceDisplay.usageDevices)
: 0;
- const devices = displayedGpu?.devices ?? [];
+ const devices = memoryDisplay.usageDevices;
const memoryTotals = gpuMemoryTotalsGb(devices);
const vramTotal = memoryTotals.total;
const hasSharedPool = memoryTotals.shared > 0;
// null usage = unknown (e.g. Windows ROCm perf counter); 0 would fabricate a
// readout. The host figure can still be known when no device's is (#7452).
- const resolvedVramUsed = resolveGpuVramUsedGb(displayedGpu);
+ const resolvedVramUsed = resolveGpuVramUsedGb(memoryDisplay.usageGpu);
const vramUsageKnown = resolvedVramUsed !== null;
const vramUsed = resolvedVramUsed ?? 0;
const vramPercent = clampPercent(
@@ -477,7 +480,9 @@ function FloatingMonitorPanel({
);
const unknownLabel = t("settings.resources.environment.unknown");
- const hasGpu = (displayedGpu?.available ?? false) && devices.length > 0;
+ const hasGpu =
+ (displayedGpu?.available ?? false) &&
+ (displayedGpu?.devices.length ?? 0) > 0;
// The container sits on the floating panel layer, above the bottom-right overlay stack. The stack
// is anchored to that same corner and does not move for this monitor, so the two can overlap. The
@@ -574,7 +579,7 @@ function FloatingMonitorPanel({
/>
</div>
- {hasGpu && (
+ {hasGpu && devices.length > 0 && (
<div className="space-y-1">
<div className="flex justify-between text-ui-11 font-medium font-mono">
<span className="truncate flex-1 pr-2">
@@ -610,18 +615,51 @@ function FloatingMonitorPanel({
/>
</div>
)}
+ {hasGpu && memoryDisplay.sharedDevices.length > 0 && (
+ <div className="space-y-1 text-xs">
+ <div className="font-medium">
+ {t("settings.resources.gpu.sharedWithSystemRam")}
+ </div>
+ <div className="font-mono text-muted-foreground">
+ {t("settings.resources.gpu.estimatedAvailable", {
+ value:
+ memoryDisplay.sharedAvailableGb === null
+ ? unknownLabel
+ : formatGiB(memoryDisplay.sharedAvailableGb),
+ })}
+ </div>
+ </div>
+ )}
{separateInferenceGpu && (
- <div className="flex justify-between gap-2 text-ui-11 font-mono">
- <span className="text-muted-foreground">
+ <div className="space-y-1 border-t border-border/60 pt-2 text-ui-11">
+ <span className="block font-medium text-muted-foreground">
{t("settings.resources.gpu.ggufInference")}
</span>
- <span className="uppercase text-foreground">
- {separateInferenceGpu.backend ?? "GPU"}
- {separateInferenceGpu.available
- ? inferenceVramTotal
- ? ` · ${formatGiB(inferenceVramTotal)}`
- : ""
- : ` · ${t("settings.resources.gpu.unavailable")}`}
+ <span className="block font-mono text-foreground">
+ {separateInferenceGpu.backend?.toUpperCase() ?? "GPU"}
+ {separateInferenceGpu.available ? (
+ inferenceVramTotal ? (
+ <span className="block">
+ {t("settings.resources.gpu.vramUtilization")}:{" "}
+ {formatGiB(inferenceVramTotal)}
+ </span>
+ ) : (
+ ""
+ )
+ ) : (
+ ` · ${t("settings.resources.gpu.unavailable")}`
+ )}
+ {separateInferenceGpu.available &&
+ inferenceDisplay.sharedDevices.length > 0 && (
+ <span className="block normal-case">
+ {t("settings.resources.gpu.sharedEstimatedAvailable", {
+ value:
+ inferenceDisplay.sharedAvailableGb === null
+ ? unknownLabel
+ : formatGiB(inferenceDisplay.sharedAvailableGb),
+ })}
+ </span>
+ )}
</span>
</div>
)}
diff --git a/studio/frontend/src/features/settings/tabs/resources-tab.tsx b/studio/frontend/src/features/settings/tabs/resources-tab.tsx
index e2fa569ec54..eafafa6037d 100644
--- a/studio/frontend/src/features/settings/tabs/resources-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/resources-tab.tsx
@@ -11,6 +11,11 @@ import {
openModelsDir,
pickHuggingFaceCacheDir,
} from "@/features/native-intents";
+import {
+ gpuMemoryDisplay,
+ hasSharedMemoryHeadroom,
+ sharedMemoryAvailableGb,
+} from "@/hooks/gpu-memory-display";
import {
gpuMemoryTotalsGb,
gpuVramUsedIsPerDevice,
@@ -112,6 +117,8 @@ function MetricTile({
value,
detail,
percent,
+ showUsage = true,
+ extraDetail,
}: {
label: string;
value: string;
@@ -119,6 +126,8 @@ function MetricTile({
// null = usage unknown (e.g. Windows ROCm perf counter): show a dash and
// empty bar rather than a fabricated 0%.
percent: number | null;
+ showUsage?: boolean;
+ extraDetail?: string;
}) {
const percentKnown = isFiniteNumber(percent);
const safePercent = clampPercent(percent);
@@ -128,16 +137,18 @@ function MetricTile({
<span className="truncate text-ui-11 font-semibold uppercase tracking-[0.08em] text-muted-foreground">
{label}
</span>
- <span
- className={cn(
- "shrink-0 font-mono text-xs tabular-nums",
- percentKnown
- ? usageTextClass(safePercent)
- : "text-muted-foreground",
- )}
- >
- {percentKnown ? formatPercent(safePercent) : "--"}
- </span>
+ {showUsage && (
+ <span
+ className={cn(
+ "shrink-0 font-mono text-xs tabular-nums",
+ percentKnown
+ ? usageTextClass(safePercent)
+ : "text-muted-foreground",
+ )}
+ >
+ {percentKnown ? formatPercent(safePercent) : "--"}
+ </span>
+ )}
</div>
{/* Both lines truncate, so carry the full text: the GPU states are sentences. */}
<div className="min-w-0">
@@ -147,16 +158,26 @@ function MetricTile({
>
{value}
</div>
- <div title={detail} className="mt-0.5 truncate text-xs text-muted-foreground">
+ <div
+ title={detail}
+ className="mt-0.5 truncate text-xs text-muted-foreground"
+ >
{detail}
</div>
</div>
- <Progress
- value={percentKnown ? safePercent : 0}
- aria-label={label}
- className="h-1.5 rounded-full bg-muted dark:bg-black/40"
- indicatorClassName={usageIndicatorClass(safePercent)}
- />
+ {showUsage && (
+ <Progress
+ value={percentKnown ? safePercent : 0}
+ aria-label={label}
+ className="h-1.5 rounded-full bg-muted dark:bg-black/40"
+ indicatorClassName={usageIndicatorClass(safePercent)}
+ />
+ )}
+ {extraDetail && (
+ <div className="whitespace-pre-line text-xs text-muted-foreground">
+ {extraDetail}
+ </div>
+ )}
</div>
);
}
@@ -241,8 +262,10 @@ export function ResourcesTab() {
systemInfo.inference_gpu.backend !== systemInfo.gpu.backend
? systemInfo.inference_gpu
: null;
+ const memoryDisplay = gpuMemoryDisplay(displayedGpu);
+ const inferenceDisplay = gpuMemoryDisplay(separateInferenceGpu);
const inferenceVramTotal = separateInferenceGpu
- ? aggregateGpuMemoryTotalGb(separateInferenceGpu.devices)
+ ? aggregateGpuMemoryTotalGb(inferenceDisplay.usageDevices)
: 0;
useEffect(() => {
@@ -271,16 +294,18 @@ export function ResourcesTab() {
const diskFree = systemInfo.disk?.free_gb ?? 0;
const diskUsed = Math.max(0, diskTotal - diskFree);
const diskPercent = diskTotal > 0 ? (diskUsed / diskTotal) * 100 : 0;
- const gpuMemoryTotals = gpuMemoryTotalsGb(devices);
+ const display = gpuMemoryDisplay(displayedGpu);
+ const usageDevices = display.usageDevices;
+ const gpuMemoryTotals = gpuMemoryTotalsGb(usageDevices);
const vramTotal = gpuMemoryTotals.total;
// null usage = unknown (e.g. Windows ROCm perf counter); 0 would fabricate a
// total, so the device's own row stays unknown. The host figure can still be
// known when no device's is (#7452).
- const perDeviceKnown = gpuVramUsedIsPerDevice(devices);
- const vramUsed = resolveGpuVramUsedGb(displayedGpu);
+ const perDeviceKnown = gpuVramUsedIsPerDevice(usageDevices);
+ const vramUsed = resolveGpuVramUsedGb(display.usageGpu);
const vramUsageKnown = vramUsed !== null;
const vramFree = perDeviceKnown
- ? devices.reduce(
+ ? usageDevices.reduce(
(sum, device) =>
sum +
(device.vram_free_gb ??
@@ -523,14 +548,25 @@ export function ResourcesTab() {
percent={hostUnread ? null : metrics.diskPercent}
/>
<MetricTile
- label={t("settings.resources.liveMonitor.vram")}
+ label={t(
+ memoryDisplay.sharedOnly
+ ? "settings.resources.gpu.memory"
+ : "settings.resources.liveMonitor.vram",
+ )}
value={
gpuUnknown
? unknownLabel
: hasGpu
- ? metrics.vramUsageKnown
- ? `${formatGiB(metrics.vramUsed)} / ${vramCapacityLabel}`
- : `${unknownLabel} / ${vramCapacityLabel}`
+ ? memoryDisplay.sharedOnly
+ ? t("settings.resources.gpu.estimatedAvailable", {
+ value:
+ memoryDisplay.sharedAvailableGb === null
+ ? unknownLabel
+ : formatGiB(memoryDisplay.sharedAvailableGb),
+ })
+ : metrics.vramUsageKnown
+ ? `${formatGiB(metrics.vramUsed)} / ${vramCapacityLabel}`
+ : `${unknownLabel} / ${vramCapacityLabel}`
: gpuMismatch
? t("settings.resources.liveMonitor.gpuUnusable")
: t("settings.resources.liveMonitor.noGpu")
@@ -539,15 +575,24 @@ export function ResourcesTab() {
gpuUnknown
? gpuUnknownLabel
: hasGpu
- ? metrics.vramUsageKnown
- ? t("settings.resources.liveMonitor.free", {
- value: formatGiB(metrics.vramFree),
- })
- : unknownLabel
+ ? memoryDisplay.sharedOnly
+ ? t("settings.resources.gpu.sharedWithSystemRam")
+ : metrics.vramUsageKnown
+ ? t("settings.resources.liveMonitor.free", {
+ value: formatGiB(metrics.vramFree),
+ })
+ : unknownLabel
: gpuMismatch
? t("settings.resources.liveMonitor.gpuUnusableDetail")
: backendLabel
}
+ showUsage={!memoryDisplay.sharedOnly}
+ extraDetail={
+ !memoryDisplay.sharedOnly &&
+ memoryDisplay.sharedDevices.length > 0
+ ? `${t("settings.resources.gpu.sharedWithSystemRam")}\n${t("settings.resources.gpu.estimatedAvailable", { value: memoryDisplay.sharedAvailableGb === null ? unknownLabel : formatGiB(memoryDisplay.sharedAvailableGb) })}`
+ : undefined
+ }
percent={metrics.vramUsageKnown ? metrics.vramPercent : null}
/>
</div>
@@ -586,13 +631,31 @@ export function ResourcesTab() {
<span className="text-muted-foreground">
{t("settings.resources.gpu.ggufInference")}
</span>
- <span className="text-right font-mono text-xs uppercase text-foreground">
- {separateInferenceGpu.backend ?? "GPU"}
- {separateInferenceGpu.available
- ? inferenceVramTotal
- ? ` · ${formatGiB(inferenceVramTotal)}`
- : ""
- : ` · ${t("settings.resources.gpu.unavailable")}`}
+ <span className="text-right font-mono text-xs text-foreground">
+ {separateInferenceGpu.backend?.toUpperCase() ?? "GPU"}
+ {separateInferenceGpu.available ? (
+ inferenceVramTotal ? (
+ <span className="block">
+ {t("settings.resources.gpu.vramUtilization")}:{" "}
+ {formatGiB(inferenceVramTotal)}
+ </span>
+ ) : (
+ ""
+ )
+ ) : (
+ ` · ${t("settings.resources.gpu.unavailable")}`
+ )}
+ {separateInferenceGpu.available &&
+ inferenceDisplay.sharedDevices.length > 0 && (
+ <span className="block normal-case">
+ {t("settings.resources.gpu.sharedEstimatedAvailable", {
+ value:
+ inferenceDisplay.sharedAvailableGb === null
+ ? unknownLabel
+ : formatGiB(inferenceDisplay.sharedAvailableGb),
+ })}
+ </span>
+ )}
</span>
</div>
)}
@@ -648,10 +711,16 @@ export function ResourcesTab() {
</span>
{/* Same accent pill as the New tags, which stays legible
on the light background. */}
- <span className="shrink-0 rounded-full bg-control-accent/10 px-2 py-1 text-ui-10 leading-none font-semibold tabular-nums text-control-accent">
- {percentText}{" "}
- {t("settings.resources.gpu.vramUtilization")}
- </span>
+ {hasSharedMemoryHeadroom(device) ? (
+ <span className="text-xs text-muted-foreground">
+ {t("settings.resources.gpu.sharedWithSystemRam")}
+ </span>
+ ) : (
+ <span className="shrink-0 rounded-full bg-control-accent/10 px-2 py-1 text-ui-10 leading-none font-semibold tabular-nums text-control-accent">
+ {percentText}{" "}
+ {t("settings.resources.gpu.vramUtilization")}
+ </span>
+ )}
</div>
</div>
{/* Meter under the figures, so it spans their width instead of
@@ -659,34 +728,51 @@ export function ResourcesTab() {
Model downloads control, so both blocks start on one edge.
Below 992px the dialog stops filling its 960px cap and the
device name would truncate, so the row stacks instead. */}
- <div className="flex w-[392px] shrink-0 flex-col items-stretch gap-2.5 max-[992px]:w-full">
- {/* Ruled between the three readings: run together they are
+ {hasSharedMemoryHeadroom(device) ? (
+ <div className="text-xs font-mono text-muted-foreground">
+ {t("settings.resources.gpu.estimatedAvailable", {
+ value:
+ sharedMemoryAvailableGb([device]) === null
+ ? unknownLabel
+ : formatGiB(device.vram_free_gb),
+ })}
+ </div>
+ ) : (
+ <div className="flex w-[392px] shrink-0 flex-col items-stretch gap-2.5 max-[992px]:w-full">
+ {/* Ruled between the three readings: run together they are
easy to misread as one number. */}
- {/* min-w-0 on each reading, or truncate cannot fire: a flex
+ {/* min-w-0 on each reading, or truncate cannot fire: a flex
item defaults to min-width:auto and the longest locales
would push past the block instead of ellipsizing. */}
- <div className="flex items-center justify-between gap-3 font-mono text-ui-11 tabular-nums text-muted-foreground">
- <span className="min-w-0 truncate">
- {t("settings.resources.gpu.used", { value: usedText })}
- </span>
- <span aria-hidden className="h-3 w-px shrink-0 bg-border" />
- <span className="min-w-0 truncate">
- {t("settings.resources.gpu.free", { value: freeText })}
- </span>
- <span aria-hidden className="h-3 w-px shrink-0 bg-border" />
- <span className="min-w-0 truncate">
- {t("settings.resources.gpu.total", {
- value: totalText,
- })}
- </span>
+ <div className="flex items-center justify-between gap-3 font-mono text-ui-11 tabular-nums text-muted-foreground">
+ <span className="min-w-0 truncate">
+ {t("settings.resources.gpu.used", { value: usedText })}
+ </span>
+ <span
+ aria-hidden
+ className="h-3 w-px shrink-0 bg-border"
+ />
+ <span className="min-w-0 truncate">
+ {t("settings.resources.gpu.free", { value: freeText })}
+ </span>
+ <span
+ aria-hidden
+ className="h-3 w-px shrink-0 bg-border"
+ />
+ <span className="min-w-0 truncate">
+ {t("settings.resources.gpu.total", {
+ value: totalText,
+ })}
+ </span>
+ </div>
+ <Progress
+ value={safePercent}
+ aria-label={device.name ?? "GPU"}
+ className="h-1.5 w-full rounded-full bg-muted dark:bg-black/40"
+ indicatorClassName={usageIndicatorClass(safePercent)}
+ />
</div>
- <Progress
- value={safePercent}
- aria-label={device.name ?? "GPU"}
- className="h-1.5 w-full rounded-full bg-muted dark:bg-black/40"
- indicatorClassName={usageIndicatorClass(safePercent)}
- />
- </div>
+ )}
</div>
);
})
diff --git a/studio/frontend/src/hooks/gpu-memory-display.ts b/studio/frontend/src/hooks/gpu-memory-display.ts
new file mode 100644
index 00000000000..5d7d6323ba4
--- /dev/null
+++ b/studio/frontend/src/hooks/gpu-memory-display.ts
@@ -0,0 +1,48 @@
+// 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 { GpuDevice, SystemGpuInfo } from "./use-system";
+
+/** Vulkan shared devices report allocation headroom, not a fixed VRAM pool. */
+export function hasSharedMemoryHeadroom(device: GpuDevice): boolean {
+ return device.index_kind === "vulkan" && device.shared_memory === true;
+}
+
+export function sharedMemoryAvailableGb(devices: GpuDevice[]): number | null {
+ if (
+ !devices.length ||
+ devices.some(
+ (device) =>
+ !Number.isFinite(device.vram_free_gb) ||
+ (device.vram_free_gb as number) < 0,
+ )
+ )
+ return null;
+ // Multiple iGPUs are views of the same host pool, not additive capacity.
+ return Math.max(...devices.map((device) => device.vram_free_gb as number));
+}
+
+/** Presentation only. Keep allocation budgets and model-fit capacity unchanged. */
+export function gpuMemoryDisplay(gpu: SystemGpuInfo | null | undefined) {
+ const devices = gpu?.devices ?? [];
+ const sharedDevices = devices.filter(hasSharedMemoryHeadroom);
+ const usageDevices = devices.filter(
+ (device) => !hasSharedMemoryHeadroom(device),
+ );
+ return {
+ sharedDevices,
+ usageDevices,
+ sharedOnly: sharedDevices.length > 0 && usageDevices.length === 0,
+ sharedAvailableGb: sharedMemoryAvailableGb(sharedDevices),
+ usageGpu: gpu
+ ? {
+ ...gpu,
+ devices: usageDevices,
+ // A host aggregate cannot be attributed to a filtered subset of its devices.
+ vram_used_gb_aggregate: sharedDevices.length
+ ? null
+ : gpu.vram_used_gb_aggregate,
+ }
+ : gpu,
+ };
+}
diff --git a/studio/frontend/src/i18n/locales/ar.ts b/studio/frontend/src/i18n/locales/ar.ts
index 02d2f8c1008..c09da9b132f 100644
--- a/studio/frontend/src/i18n/locales/ar.ts
+++ b/studio/frontend/src/i18n/locales/ar.ts
@@ -1259,7 +1259,11 @@ export const ar = {
},
gpu: {
title: "أجهزة GPU",
- ggufInference: "استدلال GGUF",
+ memory: "ذاكرة GPU",
+ sharedWithSystemRam: "مشتركة مع ذاكرة النظام",
+ estimatedAvailable: "المتاح تقديريًا: {value}",
+ sharedEstimatedAvailable: "ذاكرة النظام المشتركة: المتاح تقديريًا {value}",
+ ggufInference: "ذاكرة نماذج GGUF",
unavailable: "غير متاح",
detecting: "جارٍ البحث عن وحدات GPU...",
unreadable: "تعذّرت قراءة عتاد هذا الخادم.",
diff --git a/studio/frontend/src/i18n/locales/de.ts b/studio/frontend/src/i18n/locales/de.ts
index 2ba3d9243fa..5b9f7830245 100644
--- a/studio/frontend/src/i18n/locales/de.ts
+++ b/studio/frontend/src/i18n/locales/de.ts
@@ -1285,7 +1285,11 @@ export const de = {
},
gpu: {
title: "GPU-Geräte",
- ggufInference: "GGUF-Inferenz",
+ memory: "GPU-Speicher",
+ sharedWithSystemRam: "Mit dem System-RAM geteilt",
+ estimatedAvailable: "Geschätzt verfügbar: {value}",
+ sharedEstimatedAvailable: "Gemeinsamer System-RAM: geschätzt {value} verfügbar",
+ ggufInference: "Speicher für GGUF-Modelle",
unavailable: "nicht verfügbar",
detecting: "Suche nach GPUs...",
unreadable: "Die Hardware dieses Servers konnte nicht gelesen werden.",
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index af60d2943cc..831cc81ae73 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -1258,7 +1258,11 @@ export const en = {
},
gpu: {
title: "GPU devices",
- ggufInference: "GGUF inference",
+ memory: "GPU memory",
+ sharedWithSystemRam: "Shared with system RAM",
+ estimatedAvailable: "Estimated available: {value}",
+ sharedEstimatedAvailable: "Shared system RAM: estimated {value} available",
+ ggufInference: "GGUF model memory",
unavailable: "unavailable",
detecting: "Checking for GPUs...",
unreadable: "Could not read this server's hardware.",
diff --git a/studio/frontend/src/i18n/locales/es.ts b/studio/frontend/src/i18n/locales/es.ts
index 735d52db464..a0b5cbcfd4e 100644
--- a/studio/frontend/src/i18n/locales/es.ts
+++ b/studio/frontend/src/i18n/locales/es.ts
@@ -1280,7 +1280,11 @@ export const es = {
},
gpu: {
title: "Dispositivos GPU",
- ggufInference: "Inferencia GGUF",
+ memory: "Memoria GPU",
+ sharedWithSystemRam: "Compartida con la RAM del sistema",
+ estimatedAvailable: "Disponible estimado: {value}",
+ sharedEstimatedAvailable: "RAM compartida del sistema: disponibilidad estimada de {value}",
+ ggufInference: "Memoria para modelos GGUF",
unavailable: "No disponible",
detecting: "Buscando GPU...",
unreadable: "No se pudo leer el hardware de este servidor.",
diff --git a/studio/frontend/src/i18n/locales/fr.ts b/studio/frontend/src/i18n/locales/fr.ts
index 85882feeba4..8375013c692 100644
--- a/studio/frontend/src/i18n/locales/fr.ts
+++ b/studio/frontend/src/i18n/locales/fr.ts
@@ -1284,7 +1284,11 @@ export const fr = {
},
gpu: {
title: "Périphériques GPU",
- ggufInference: "Inférence GGUF",
+ memory: "Mémoire GPU",
+ sharedWithSystemRam: "Partagée avec la RAM système",
+ estimatedAvailable: "Disponibilité estimée : {value}",
+ sharedEstimatedAvailable: "RAM système partagée : disponibilité estimée de {value}",
+ ggufInference: "Mémoire pour les modèles GGUF",
unavailable: "indisponible",
detecting: "Recherche de GPU...",
unreadable: "Impossible de lire le matériel de ce serveur.",
diff --git a/studio/frontend/src/i18n/locales/hi.ts b/studio/frontend/src/i18n/locales/hi.ts
index ac40e2c47a2..6ac3a496b15 100644
--- a/studio/frontend/src/i18n/locales/hi.ts
+++ b/studio/frontend/src/i18n/locales/hi.ts
@@ -1262,7 +1262,11 @@ export const hi = {
},
gpu: {
title: "GPU डिवाइस",
- ggufInference: "GGUF इन्फ़रेंस",
+ memory: "GPU मेमोरी",
+ sharedWithSystemRam: "सिस्टम RAM के साथ साझा",
+ estimatedAvailable: "अनुमानित उपलब्ध: {value}",
+ sharedEstimatedAvailable: "साझा सिस्टम RAM: अनुमानित उपलब्ध {value}",
+ ggufInference: "GGUF मॉडल मेमोरी",
unavailable: "उपलब्ध नहीं",
detecting: "GPU खोजे जा रहे हैं...",
unreadable: "इस सर्वर का हार्डवेयर नहीं पढ़ा जा सका।",
diff --git a/studio/frontend/src/i18n/locales/it.ts b/studio/frontend/src/i18n/locales/it.ts
index d0f4a0b024c..7e3b44ca0da 100644
--- a/studio/frontend/src/i18n/locales/it.ts
+++ b/studio/frontend/src/i18n/locales/it.ts
@@ -1251,7 +1251,11 @@ export const it = {
},
gpu: {
title: "Dispositivi GPU",
- ggufInference: "Inferenza GGUF",
+ memory: "Memoria GPU",
+ sharedWithSystemRam: "Condivisa con la RAM di sistema",
+ estimatedAvailable: "Disponibilità stimata: {value}",
+ sharedEstimatedAvailable: "RAM di sistema condivisa: disponibilità stimata di {value}",
+ ggufInference: "Memoria per i modelli GGUF",
unavailable: "non disponibile",
detecting: "Ricerca di GPU...",
unreadable: "Impossibile leggere l'hardware di questo server.",
diff --git a/studio/frontend/src/i18n/locales/ja.ts b/studio/frontend/src/i18n/locales/ja.ts
index 8d9594fbf05..27d94bc7f00 100644
--- a/studio/frontend/src/i18n/locales/ja.ts
+++ b/studio/frontend/src/i18n/locales/ja.ts
@@ -1239,7 +1239,11 @@ export const ja = {
},
gpu: {
title: "GPU デバイス",
- ggufInference: "GGUF 推論",
+ memory: "GPUメモリ",
+ sharedWithSystemRam: "システムRAMと共有",
+ estimatedAvailable: "推定空き容量: {value}",
+ sharedEstimatedAvailable: "共有システムRAM: 推定空き容量 {value}",
+ ggufInference: "GGUFモデル用メモリ",
unavailable: "利用不可",
detecting: "GPU を確認しています...",
unreadable: "このサーバーのハードウェアを読み取れませんでした。",
diff --git a/studio/frontend/src/i18n/locales/ko.ts b/studio/frontend/src/i18n/locales/ko.ts
index 143da8dd525..4bd48d0f9a8 100644
--- a/studio/frontend/src/i18n/locales/ko.ts
+++ b/studio/frontend/src/i18n/locales/ko.ts
@@ -1255,7 +1255,11 @@ export const ko = {
},
gpu: {
title: "GPU 장치",
- ggufInference: "GGUF 추론",
+ memory: "GPU 메모리",
+ sharedWithSystemRam: "시스템 RAM과 공유",
+ estimatedAvailable: "예상 사용 가능: {value}",
+ sharedEstimatedAvailable: "공유 시스템 RAM: 예상 사용 가능 {value}",
+ ggufInference: "GGUF 모델 메모리",
unavailable: "사용할 수 없음",
detecting: "GPU를 확인하는 중...",
unreadable: "이 서버의 하드웨어를 읽을 수 없습니다.",
diff --git a/studio/frontend/src/i18n/locales/pt-br.ts b/studio/frontend/src/i18n/locales/pt-br.ts
index 6a189732735..ed9df4d96cc 100644
--- a/studio/frontend/src/i18n/locales/pt-br.ts
+++ b/studio/frontend/src/i18n/locales/pt-br.ts
@@ -1269,7 +1269,11 @@ export const ptBR = {
},
gpu: {
title: "Dispositivos de GPU",
- ggufInference: "Inferência com GGUF",
+ memory: "Memória da GPU",
+ sharedWithSystemRam: "Compartilhada com a RAM do sistema",
+ estimatedAvailable: "Disponibilidade estimada: {value}",
+ sharedEstimatedAvailable: "RAM compartilhada do sistema: disponibilidade estimada de {value}",
+ ggufInference: "Memória para modelos GGUF",
unavailable: "indisponível",
detecting: "Procurando GPUs...",
unreadable: "Não foi possível ler o hardware deste servidor.",
diff --git a/studio/frontend/src/i18n/locales/ru.ts b/studio/frontend/src/i18n/locales/ru.ts
index 0cee8991bae..a737730399a 100644
--- a/studio/frontend/src/i18n/locales/ru.ts
+++ b/studio/frontend/src/i18n/locales/ru.ts
@@ -1264,7 +1264,11 @@ export const ru = {
},
gpu: {
title: "Устройства GPU",
- ggufInference: "Инференс GGUF",
+ memory: "Память GPU",
+ sharedWithSystemRam: "Общая с оперативной памятью системы",
+ estimatedAvailable: "Примерно доступно: {value}",
+ sharedEstimatedAvailable: "Общая оперативная память системы: примерно {value} доступно",
+ ggufInference: "Память для моделей GGUF",
unavailable: "недоступно",
detecting: "Поиск GPU...",
unreadable: "Не удалось прочитать оборудование этого сервера.",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index 2e169259b80..04c281afb40 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -1233,7 +1233,11 @@ export const zhCN = {
},
gpu: {
title: "GPU 设备",
- ggufInference: "GGUF 推理",
+ memory: "GPU 内存",
+ sharedWithSystemRam: "与系统 RAM 共享",
+ estimatedAvailable: "预计可用:{value}",
+ sharedEstimatedAvailable: "共享系统 RAM:预计可用 {value}",
+ ggufInference: "GGUF 模型内存",
unavailable: "不可用",
detecting: "正在检查 GPU...",
unreadable: "无法读取此服务器的硬件信息。",
diff --git a/studio/frontend/tests/gpu-memory-display.test.ts b/studio/frontend/tests/gpu-memory-display.test.ts
new file mode 100644
index 00000000000..1b8871a2570
--- /dev/null
+++ b/studio/frontend/tests/gpu-memory-display.test.ts
@@ -0,0 +1,107 @@
+// 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 { gpuMemoryDisplay } from "../src/hooks/gpu-memory-display.ts";
+import {
+ gpuMemoryTotalsGb,
+ resolveGpuVramUsedGb,
+} from "../src/hooks/gpu-vram.ts";
+import type { GpuDevice, SystemGpuInfo } from "../src/hooks/use-system.ts";
+
+const integrated: GpuDevice = {
+ index_kind: "vulkan",
+ shared_memory: true,
+ memory_total_gb: 25.69,
+ vram_free_gb: 25.69,
+};
+const discrete: GpuDevice = {
+ index_kind: "vulkan",
+ shared_memory: false,
+ memory_total_gb: 16,
+ vram_used_gb: 1,
+ vram_free_gb: 15,
+};
+const gpu = (devices: GpuDevice[], aggregate?: number): SystemGpuInfo => ({
+ available: true,
+ devices,
+ vram_used_gb_aggregate: aggregate,
+});
+
+test("Vulkan iGPU presents allocation headroom without manufacturing usage", () => {
+ const display = gpuMemoryDisplay(gpu([integrated]));
+ assert.equal(display.sharedOnly, true);
+ assert.equal(display.sharedAvailableGb, 25.69);
+ assert.deepEqual(display.usageDevices, []);
+ assert.equal(resolveGpuVramUsedGb(display.usageGpu), null);
+});
+
+test("mixed inventory measures dedicated usage against dedicated capacity only", () => {
+ const input = gpu([discrete, integrated], 8);
+ const display = gpuMemoryDisplay(input);
+ assert.equal(display.sharedOnly, false);
+ assert.equal(display.sharedAvailableGb, 25.69);
+ const total = gpuMemoryTotalsGb(display.usageDevices).total;
+ const used = resolveGpuVramUsedGb(display.usageGpu);
+ assert.equal(total, 16);
+ assert.equal(used, 1);
+ assert.equal((used! / total) * 100, 6.25);
+ assert.equal(input.devices.length, 2);
+ assert.equal(gpuMemoryTotalsGb(input.devices).total, 41.69);
+});
+
+test("shared views count once and zero is a known reading", () => {
+ assert.equal(
+ gpuMemoryDisplay(gpu([integrated, { ...integrated, vram_free_gb: 12 }]))
+ .sharedAvailableGb,
+ 25.69,
+ );
+ assert.equal(
+ gpuMemoryDisplay(gpu([{ ...integrated, vram_free_gb: 0 }]))
+ .sharedAvailableGb,
+ 0,
+ );
+});
+
+test("missing or invalid shared free readings never fall back to capacity", () => {
+ for (const free of [undefined, null, NaN, Infinity, -1]) {
+ // null is also possible on the wire despite the optional-number API type.
+ const device = { ...integrated, vram_free_gb: free } as GpuDevice;
+ assert.equal(gpuMemoryDisplay(gpu([device])).sharedAvailableGb, null);
+ assert.equal(
+ gpuMemoryDisplay(gpu([integrated, device])).sharedAvailableGb,
+ null,
+ );
+ }
+});
+
+test("ROCm host aggregate survives and unattributable rows remain unknown", () => {
+ const devices = [
+ { index_kind: "torch", memory_total_gb: 16 },
+ { index_kind: "torch", memory_total_gb: 24 },
+ ];
+ const display = gpuMemoryDisplay(gpu(devices, 3));
+ assert.equal(display.sharedOnly, false);
+ assert.equal(resolveGpuVramUsedGb(display.usageGpu), 3);
+ assert.equal(resolveGpuVramUsedGb(gpu(devices)), null);
+ assert.equal(display.sharedDevices.length, 0);
+});
+
+test("host aggregate cannot supply usage for a filtered mixed inventory", () => {
+ const display = gpuMemoryDisplay(
+ gpu([{ ...discrete, vram_used_gb: undefined }, integrated], 3),
+ );
+ assert.equal(resolveGpuVramUsedGb(display.usageGpu), null);
+});
+
+test("other backends retain their shared-pool usage reporting", () => {
+ for (const index_kind of ["torch", "metal", undefined]) {
+ const device = { ...integrated, index_kind, vram_used_gb: 2 };
+ const display = gpuMemoryDisplay(gpu([device]));
+ assert.equal(display.sharedDevices.length, 0);
+ assert.equal(gpuMemoryTotalsGb(display.usageDevices).total, 25.69);
+ assert.equal(resolveGpuVramUsedGb(display.usageGpu), 2);
+ }
+ assert.deepEqual(gpuMemoryDisplay(undefined).usageDevices, []);
+});
diff --git a/studio/frontend/tests/resources-shared-memory-display.test.ts b/studio/frontend/tests/resources-shared-memory-display.test.ts
index ef78573900e..ca090774952 100644
--- a/studio/frontend/tests/resources-shared-memory-display.test.ts
+++ b/studio/frontend/tests/resources-shared-memory-display.test.ts
@@ -9,7 +9,7 @@ import { readSrc } from "./helpers/kit.ts";
const resourcesTab = readSrc("features/settings/tabs/resources-tab.tsx");
test("Resources separates dedicated VRAM from shared GPU memory", () => {
- assert.match(resourcesTab, /gpuMemoryTotalsGb\(devices\)/);
+ assert.match(resourcesTab, /gpuMemoryTotalsGb\(usageDevices\)/);
assert.match(resourcesTab, /metrics\.vramShared > 0/);
assert.match(resourcesTab, /environment\.vramWithShared/);
assert.match(resourcesTab, /vram: formatGiB\(metrics\.vramDedicated\)/);