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: optional detailed Live monitor with GPU load, temperature and power (#13351)

Daniel Han2e387701402c
diff --git a/studio/backend/main.py b/studio/backend/main.py
index 28693d8949a..a1d7c0bd77a 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -2308,6 +2308,7 @@ def _get_cached_system_gpu_info(
"""Return training and inference GPU info with bounded live-probe churn."""
import time
from utils.hardware import (
+ LIVE_GPU_TELEMETRY_KEYS,
amd_driver_warning_report,
get_backend_visible_gpu_info,
get_cross_vendor_inference_gpu_info,
@@ -2382,6 +2383,7 @@ def _get_cached_system_gpu_info(
enriched_dev["vram_utilization_pct"] = util.get(
"vram_utilization_pct", dev.get("vram_utilization_pct")
)
+ enriched_dev.update({k: util[k] for k in LIVE_GPU_TELEMETRY_KEYS if k in util})
enriched_devices.append(enriched_dev)
# The tile divides the aggregate by the SUMMED per-device totals, so both must describe the same cards. The
diff --git a/studio/backend/tests/test_system_gpu_live_telemetry.py b/studio/backend/tests/test_system_gpu_live_telemetry.py
new file mode 100644
index 00000000000..e4f35ba34d4
--- /dev/null
+++ b/studio/backend/tests/test_system_gpu_live_telemetry.py
@@ -0,0 +1,133 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from types import SimpleNamespace
+
+import main
+
+_LOGGER = SimpleNamespace(debug = lambda *args: None)
+
+
+def _patch(
+ monkeypatch,
+ util_device,
+ util_backend = "cuda",
+):
+ import utils.hardware as hardware
+ from core.inference.llama_cpp import LlamaCppBackend
+
+ monkeypatch.setattr(main, "_system_gpu_cache", None)
+ monkeypatch.setattr(LlamaCppBackend, "_is_vulkan_backend", staticmethod(lambda: False))
+ monkeypatch.setattr(LlamaCppBackend, "_backend_lacks_gpu_lib", staticmethod(lambda: False))
+ monkeypatch.setattr(hardware, "get_vulkan_inference_gpu_info", lambda: None)
+ monkeypatch.setattr(hardware, "get_cross_vendor_inference_gpu_info", lambda: None)
+ monkeypatch.setattr(
+ hardware,
+ "get_backend_visible_gpu_info",
+ lambda: {
+ "available": True,
+ "backend": "cuda",
+ "index_kind": "physical",
+ "devices": [{"index": 0, "name": "NVIDIA B200", "memory_total_gb": 179.0}],
+ },
+ )
+ calls = []
+
+ def utilization():
+ calls.append(1)
+ return {"available": True, "backend": util_backend, "devices": [util_device]}
+
+ monkeypatch.setattr(hardware, "get_visible_gpu_utilization", utilization)
+ return calls
+
+
+def test_system_gpu_devices_carry_the_live_telemetry_the_probe_already_read(monkeypatch):
+ calls = _patch(
+ monkeypatch,
+ {
+ "index": 0,
+ "vram_used_gb": 10.0,
+ "vram_total_gb": 179.0,
+ "vram_utilization_pct": 5.6,
+ "gpu_utilization_pct": 42.0,
+ "temperature_c": 61.0,
+ "power_draw_w": 312.5,
+ "power_limit_w": 1000.0,
+ "power_utilization_pct": 31.2,
+ },
+ )
+
+ gpu, _ = main._get_cached_system_gpu_info(_LOGGER)
+
+ (device,) = gpu["devices"]
+ assert device["gpu_utilization_pct"] == 42.0
+ assert device["temperature_c"] == 61.0
+ assert device["power_draw_w"] == 312.5
+ assert device["power_limit_w"] == 1000.0
+ assert device["vram_used_gb"] == 10.0
+ # No extra probe: one utilization read per cache window, as before.
+ main._get_cached_system_gpu_info(_LOGGER)
+ assert len(calls) == 1
+
+
+def test_system_gpu_devices_omit_telemetry_the_probe_did_not_report(monkeypatch):
+ _patch(monkeypatch, {"index": 0, "vram_used_gb": 1.0, "vram_total_gb": 179.0})
+
+ gpu, _ = main._get_cached_system_gpu_info(_LOGGER)
+
+ (device,) = gpu["devices"]
+ for key in ("gpu_utilization_pct", "temperature_c", "power_draw_w", "power_limit_w"):
+ assert key not in device
+
+
+def test_system_gpu_devices_never_take_telemetry_from_another_backend(monkeypatch):
+ _patch(
+ monkeypatch,
+ {"index": 0, "temperature_c": 70.0, "power_draw_w": 100.0},
+ util_backend = "vulkan",
+ )
+
+ gpu, _ = main._get_cached_system_gpu_info(_LOGGER)
+
+ (device,) = gpu["devices"]
+ assert "temperature_c" not in device
+ assert "power_draw_w" not in device
+
+
+def test_cuda_inference_inventory_carries_the_same_telemetry(monkeypatch):
+ """No-torch / CPU-torch host with a CUDA llama.cpp: the monitor shows inference_gpu."""
+ import utils.hardware.hardware as hw
+ import utils.hardware.nvidia as nvidia
+ from core.inference.llama_cpp import LlamaCppBackend
+
+ monkeypatch.delenv("CUDA_VISIBLE_DEVICES", raising = False)
+ monkeypatch.setattr(LlamaCppBackend, "_visible_devices_mask", staticmethod(lambda _name: None))
+ monkeypatch.setattr(
+ nvidia,
+ "get_physical_gpu_inventory",
+ lambda: {"devices": [{"index": 0, "name": "RTX 4090", "memory_total_gb": 24.0}]},
+ )
+ monkeypatch.setattr(
+ nvidia,
+ "get_visible_gpu_utilization",
+ lambda _ids: {
+ "devices": [
+ {
+ "index": 0,
+ "vram_used_gb": 2.0,
+ "gpu_utilization_pct": 7.0,
+ "temperature_c": 44.0,
+ "power_draw_w": 35.0,
+ "power_limit_w": 450.0,
+ }
+ ]
+ },
+ )
+
+ (device,) = hw._nvidia_inference_devices()
+
+ assert device["vram_used_gb"] == 2.0
+ assert device["gpu_utilization_pct"] == 7.0
+ assert device["temperature_c"] == 44.0
+ assert device["power_draw_w"] == 35.0
+ assert device["power_limit_w"] == 450.0
diff --git a/studio/backend/utils/hardware/__init__.py b/studio/backend/utils/hardware/__init__.py
index a8b179eca9c..e31a71960a4 100644
--- a/studio/backend/utils/hardware/__init__.py
+++ b/studio/backend/utils/hardware/__init__.py
@@ -7,6 +7,7 @@ from typing import Optional
from . import hardware as _hardware
from .hardware import (
+ LIVE_GPU_TELEMETRY_KEYS,
DeviceType,
DEVICE,
CHAT_ONLY,
@@ -89,6 +90,7 @@ def get_torch_device_str() -> str:
__all__ = [
+ "LIVE_GPU_TELEMETRY_KEYS",
"DeviceType",
"DEVICE",
"CHAT_ONLY",
diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py
index b1958397b7b..4b3de1f21ed 100644
--- a/studio/backend/utils/hardware/hardware.py
+++ b/studio/backend/utils/hardware/hardware.py
@@ -6287,6 +6287,10 @@ def _installed_llama_backend() -> Optional[str]:
return marker_backend(read_install_marker(LlamaCppBackend._find_llama_server_binary()))
+# Readings the utilization probes already return, passed through for the Live monitor's detailed view.
+LIVE_GPU_TELEMETRY_KEYS = ("gpu_utilization_pct", "temperature_c", "power_draw_w", "power_limit_w")
+
+
def _smi_inference_device(
index: int,
ordinal: int,
@@ -6346,15 +6350,15 @@ def _nvidia_inference_devices() -> list[Dict[str, Any]]:
devices = []
for ordinal, row in enumerate(rows):
util = usage_by_index.get(row["index"], {})
- devices.append(
- _smi_inference_device(
- row["index"],
- ordinal,
- row.get("name"),
- row.get("memory_total_gb") or util.get("vram_total_gb"),
- util.get("vram_used_gb"),
- )
+ device = _smi_inference_device(
+ row["index"],
+ ordinal,
+ row.get("name"),
+ row.get("memory_total_gb") or util.get("vram_total_gb"),
+ util.get("vram_used_gb"),
)
+ device.update({k: util[k] for k in LIVE_GPU_TELEMETRY_KEYS if k in util})
+ devices.append(device)
return devices
diff --git a/studio/frontend/src/components/floating-monitor.tsx b/studio/frontend/src/components/floating-monitor.tsx
index 00171993674..406e7ab9e3e 100644
--- a/studio/frontend/src/components/floating-monitor.tsx
+++ b/studio/frontend/src/components/floating-monitor.tsx
@@ -26,6 +26,13 @@ import { CpuIcon, GripVerticalIcon, XIcon } from "lucide-react";
import { AnimatePresence, motion } from "motion/react";
import { useEffect, useRef, useState, useSyncExternalStore } from "react";
+import {
+ formatPower,
+ formatTemperature,
+ gpuTelemetryRows,
+ selectedDetailMetrics,
+ temperatureLevel,
+} from "./monitor-details";
import {
FLOATING_MONITOR_WIDTH,
floatingMonitorConstraintStyle,
@@ -57,6 +64,48 @@ function usageTextClass(percent: number): string {
return "text-primary";
}
+const LEVEL_INDICATOR_CLASS = {
+ ok: "bg-control-accent",
+ warm: "bg-amber-500",
+ hot: "bg-destructive",
+} as const;
+
+const LEVEL_TEXT_CLASS = {
+ ok: "text-primary",
+ warm: "text-amber-600 dark:text-amber-400",
+ hot: "text-destructive",
+} as const;
+
+function DetailBar({
+ label,
+ value,
+ percent,
+ indicatorClassName,
+ textClassName,
+}: {
+ label: string;
+ value: string;
+ percent: number;
+ indicatorClassName: string;
+ textClassName: string;
+}) {
+ return (
+ <div className="space-y-0.5">
+ <div className="flex justify-between gap-2 text-xs font-mono text-muted-foreground">
+ <span className="truncate">{label}</span>
+ <span className={cn("shrink-0 tabular-nums", textClassName)}>
+ {value}
+ </span>
+ </div>
+ <Progress
+ value={clampPercent(percent)}
+ className="h-1 rounded-full bg-muted"
+ indicatorClassName={indicatorClassName}
+ />
+ </div>
+ );
+}
+
function formatGiB(value: number): string {
// RAM/VRAM come from the backend in binary units (bytes / 1024**3), matching
// nvidia-smi and PyTorch, so label the readout GiB rather than GB.
@@ -156,6 +205,20 @@ function FloatingMonitorPanel({
);
const unknownLabel = t("settings.resources.environment.unknown");
+ const showDetails = useMonitorOverlayStore((state) => state.showDetails);
+ const storedDetailMetrics = useMonitorOverlayStore(
+ (state) => state.detailMetrics,
+ );
+ const detailMetrics = selectedDetailMetrics(storedDetailMetrics);
+ const cpuPercent =
+ showDetails && detailMetrics.has("cpuLoad") && systemInfo.status === "ready"
+ ? clampPercent(systemInfo.cpu?.usage_percent ?? 0)
+ : null;
+ const telemetryRows =
+ showDetails && displayedGpu?.available
+ ? gpuTelemetryRows(displayedGpu.devices, detailMetrics)
+ : [];
+
const hasGpu =
(displayedGpu?.available ?? false) &&
(displayedGpu?.devices.length ?? 0) > 0;
@@ -355,6 +418,72 @@ function FloatingMonitorPanel({
</span>
</div>
)}
+ {cpuPercent !== null && (
+ <div
+ className="border-t border-border/60 pt-2"
+ data-testid="floating-monitor-cpu"
+ >
+ <DetailBar
+ label={t("settings.resources.liveMonitor.cpuLoad")}
+ value={`${Math.round(cpuPercent)}%`}
+ percent={cpuPercent}
+ indicatorClassName={usageIndicatorClass(cpuPercent)}
+ textClassName={usageTextClass(cpuPercent)}
+ />
+ </div>
+ )}
+ {telemetryRows.map((row) => (
+ <div
+ key={row.key}
+ className="space-y-1.5 border-t border-border/60 pt-2"
+ data-testid="floating-monitor-gpu-details"
+ >
+ <div className="truncate text-ui-11 font-medium font-mono">
+ {row.name}
+ </div>
+ {row.loadPct !== null && (
+ <DetailBar
+ label={t("settings.resources.liveMonitor.gpuLoad")}
+ value={`${Math.round(row.loadPct)}%`}
+ percent={row.loadPct}
+ indicatorClassName={usageIndicatorClass(row.loadPct)}
+ textClassName={usageTextClass(row.loadPct)}
+ />
+ )}
+ {row.temperatureC !== null && (
+ <DetailBar
+ label={t("settings.resources.liveMonitor.temperature")}
+ value={formatTemperature(row.temperatureC)}
+ percent={row.temperatureC}
+ indicatorClassName={
+ LEVEL_INDICATOR_CLASS[temperatureLevel(row.temperatureC)]
+ }
+ textClassName={
+ LEVEL_TEXT_CLASS[temperatureLevel(row.temperatureC)]
+ }
+ />
+ )}
+ {row.powerDrawW !== null &&
+ (row.powerPct !== null ? (
+ <DetailBar
+ label={t("settings.resources.liveMonitor.power")}
+ value={formatPower(row.powerDrawW, row.powerLimitW)}
+ percent={row.powerPct}
+ indicatorClassName={usageIndicatorClass(row.powerPct)}
+ textClassName={usageTextClass(row.powerPct)}
+ />
+ ) : (
+ <div className="flex justify-between gap-2 text-xs font-mono text-muted-foreground">
+ <span className="truncate">
+ {t("settings.resources.liveMonitor.power")}
+ </span>
+ <span className="shrink-0 tabular-nums text-primary">
+ {formatPower(row.powerDrawW, null)}
+ </span>
+ </div>
+ ))}
+ </div>
+ ))}
</div>
</div>
</motion.div>
diff --git a/studio/frontend/src/components/monitor-details.ts b/studio/frontend/src/components/monitor-details.ts
new file mode 100644
index 00000000000..937c563aaea
--- /dev/null
+++ b/studio/frontend/src/components/monitor-details.ts
@@ -0,0 +1,103 @@
+// 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 } from "../hooks/use-system.ts";
+
+export const MONITOR_DETAIL_METRICS = [
+ "cpuLoad",
+ "gpuLoad",
+ "temperature",
+ "power",
+] as const;
+
+export type MonitorDetailMetric = (typeof MONITOR_DETAIL_METRICS)[number];
+
+/** An empty list is a choice; anything else with no known metric (older state, hand edits) means all. */
+export function selectedDetailMetrics(
+ stored: unknown,
+): ReadonlySet<MonitorDetailMetric> {
+ const known = Array.isArray(stored)
+ ? MONITOR_DETAIL_METRICS.filter((metric) => stored.includes(metric))
+ : [];
+ return new Set(
+ known.length > 0 || (Array.isArray(stored) && stored.length === 0)
+ ? known
+ : MONITOR_DETAIL_METRICS,
+ );
+}
+
+export interface GpuTelemetryRow {
+ key: string;
+ name: string;
+ loadPct: number | null;
+ temperatureC: number | null;
+ powerDrawW: number | null;
+ powerLimitW: number | null;
+ powerPct: number | null;
+}
+
+function finite(value: number | null | undefined): number | null {
+ return typeof value === "number" && Number.isFinite(value) ? value : null;
+}
+
+/** One row per device that reported at least one selected reading; nothing is fabricated. */
+export function gpuTelemetryRows(
+ devices: readonly GpuDevice[],
+ metrics: ReadonlySet<MonitorDetailMetric>,
+): GpuTelemetryRow[] {
+ const rows: GpuTelemetryRow[] = [];
+ devices.forEach((device, position) => {
+ const ordinal = device.visible_ordinal ?? device.index ?? position;
+ const loadPct = metrics.has("gpuLoad")
+ ? finite(device.gpu_utilization_pct)
+ : null;
+ const temperatureC = metrics.has("temperature")
+ ? finite(device.temperature_c)
+ : null;
+ const powerDrawW = metrics.has("power")
+ ? finite(device.power_draw_w)
+ : null;
+ const powerLimitW =
+ powerDrawW !== null ? finite(device.power_limit_w) : null;
+ if (loadPct === null && temperatureC === null && powerDrawW === null) {
+ return;
+ }
+ rows.push({
+ key: `${ordinal}`,
+ name:
+ devices.length > 1
+ ? `GPU ${ordinal}${device.name ? `: ${device.name}` : ""}`
+ : (device.name ?? `GPU ${ordinal}`),
+ loadPct,
+ temperatureC,
+ powerDrawW,
+ powerLimitW,
+ powerPct:
+ powerDrawW !== null && powerLimitW !== null && powerLimitW > 0
+ ? (powerDrawW / powerLimitW) * 100
+ : null,
+ });
+ });
+ return rows;
+}
+
+/** Most cards throttle in the high 80s, so the bar warns before that. */
+export function temperatureLevel(celsius: number): "ok" | "warm" | "hot" {
+ if (celsius >= 90) {
+ return "hot";
+ }
+ if (celsius >= 80) {
+ return "warm";
+ }
+ return "ok";
+}
+
+export function formatTemperature(celsius: number): string {
+ return `${Math.round(celsius)}°C`;
+}
+
+export function formatPower(drawW: number, limitW: number | null): string {
+ return limitW !== null && limitW > 0
+ ? `${Math.round(drawW)} / ${Math.round(limitW)} W`
+ : `${Math.round(drawW)} W`;
+}
diff --git a/studio/frontend/src/features/settings/stores/monitor-overlay-store.ts b/studio/frontend/src/features/settings/stores/monitor-overlay-store.ts
index 1af804a19e8..e89452c5ff0 100644
--- a/studio/frontend/src/features/settings/stores/monitor-overlay-store.ts
+++ b/studio/frontend/src/features/settings/stores/monitor-overlay-store.ts
@@ -4,11 +4,22 @@
import { create } from "zustand";
import { persist } from "zustand/middleware";
+import {
+ MONITOR_DETAIL_METRICS,
+ type MonitorDetailMetric,
+ selectedDetailMetrics,
+} from "@/components/monitor-details";
+
interface MonitorOverlayState {
isOpen: boolean;
isMinimized: boolean;
+ /** Off by default: the monitor looks as it always has until a user opts in. */
+ showDetails: boolean;
+ detailMetrics: MonitorDetailMetric[];
setIsOpen: (open: boolean) => void;
toggleMinimized: () => void;
+ setShowDetails: (show: boolean) => void;
+ setDetailMetric: (metric: MonitorDetailMetric, on: boolean) => void;
}
export const useMonitorOverlayStore = create<MonitorOverlayState>()(
@@ -16,9 +27,21 @@ export const useMonitorOverlayStore = create<MonitorOverlayState>()(
(set) => ({
isOpen: false,
isMinimized: false,
+ showDetails: false,
+ detailMetrics: [...MONITOR_DETAIL_METRICS],
setIsOpen: (isOpen) => set({ isOpen }),
toggleMinimized: () => set((state) => ({ isMinimized: !state.isMinimized })),
+ setShowDetails: (showDetails) => set({ showDetails }),
+ setDetailMetric: (metric, on) =>
+ set((state) => {
+ const current = selectedDetailMetrics(state.detailMetrics);
+ return {
+ detailMetrics: MONITOR_DETAIL_METRICS.filter((m) =>
+ m === metric ? on : current.has(m),
+ ),
+ };
+ }),
}),
{ name: "unsloth_monitor_overlay" }
)
-);
\ No newline at end of file
+);
diff --git a/studio/frontend/src/features/settings/tabs/resources-tab.tsx b/studio/frontend/src/features/settings/tabs/resources-tab.tsx
index ea5c0f86509..a5c9ce3bbbb 100644
--- a/studio/frontend/src/features/settings/tabs/resources-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/resources-tab.tsx
@@ -1,7 +1,15 @@
// 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 {
+ MONITOR_DETAIL_METRICS,
+ formatPower,
+ formatTemperature,
+ gpuTelemetryRows,
+ selectedDetailMetrics,
+} from "@/components/monitor-details";
import { Button } from "@/components/ui/button";
+import { Checkbox } from "@/components/ui/checkbox";
import { Input } from "@/components/ui/input";
import { Progress } from "@/components/ui/progress";
import { Switch } from "@/components/ui/switch";
@@ -264,7 +272,15 @@ export function ResourcesTab() {
const dismissDriverWarning = useSettingsPanelPrefsStore(
(s) => s.dismissDriverWarning,
);
- const { isOpen, setIsOpen } = useMonitorOverlayStore();
+ const {
+ isOpen,
+ setIsOpen,
+ showDetails,
+ setShowDetails,
+ detailMetrics: storedDetailMetrics,
+ setDetailMetric,
+ } = useMonitorOverlayStore();
+ const detailMetrics = selectedDetailMetrics(storedDetailMetrics);
// always fetch once: the switch is persisted now, so gating the hook on it
// would leave a session that opens with it off reading zeros forever.
const systemInfo = useSystemInfo({
@@ -294,6 +310,10 @@ export function ResourcesTab() {
const inferenceVramTotal = separateInferenceGpu
? aggregateGpuMemoryTotalGb(inferenceDisplay.usageDevices)
: 0;
+ const telemetryRows =
+ showDetails && displayedGpu?.available
+ ? gpuTelemetryRows(displayedGpu.devices, detailMetrics)
+ : [];
useEffect(() => {
let cancelled = false;
@@ -575,6 +595,7 @@ export function ResourcesTab() {
: t("settings.resources.liveMonitor.currentLoad")
}
percent={hostUnread ? null : (systemInfo.cpu?.usage_percent ?? 0)}
+ showUsage={!showDetails || detailMetrics.has("cpuLoad")}
/>
<MetricTile
label={t("settings.resources.liveMonitor.ram")}
@@ -668,7 +689,73 @@ export function ResourcesTab() {
}
percent={metrics.vramUsageKnown ? metrics.vramPercent : null}
/>
+ {telemetryRows.map((row) => (
+ <MetricTile
+ key={row.key}
+ label={row.name}
+ value={
+ [
+ row.temperatureC !== null
+ ? formatTemperature(row.temperatureC)
+ : null,
+ row.powerDrawW !== null
+ ? formatPower(row.powerDrawW, row.powerLimitW)
+ : null,
+ ]
+ .filter(Boolean)
+ .join(" · ") ||
+ (row.loadPct !== null ? `${Math.round(row.loadPct)}%` : "--")
+ }
+ detail={[
+ row.temperatureC !== null &&
+ t("settings.resources.liveMonitor.temperature"),
+ row.powerDrawW !== null &&
+ t("settings.resources.liveMonitor.power"),
+ row.loadPct !== null &&
+ t("settings.resources.liveMonitor.gpuLoad"),
+ ]
+ .filter(Boolean)
+ .join(" · ")}
+ showUsage={row.loadPct !== null}
+ percent={hostUnread ? null : row.loadPct}
+ />
+ ))}
</div>
+ <SettingsRow
+ label={t("settings.resources.liveMonitor.detailedView")}
+ description={t(
+ "settings.resources.liveMonitor.detailedViewDescription",
+ )}
+ className="border-t border-border/60"
+ >
+ <Switch
+ aria-label={t("settings.resources.liveMonitor.detailedView")}
+ checked={showDetails}
+ onCheckedChange={setShowDetails}
+ />
+ </SettingsRow>
+ {showDetails && (
+ <div className="flex flex-wrap gap-x-5 gap-y-2 pb-3">
+ {MONITOR_DETAIL_METRICS.map((metric) => {
+ const label = t(`settings.resources.liveMonitor.${metric}`);
+ return (
+ <label
+ key={metric}
+ className="flex cursor-pointer items-center gap-2 text-xs font-medium text-foreground"
+ >
+ <Checkbox
+ checked={detailMetrics.has(metric)}
+ onCheckedChange={(checked) =>
+ setDetailMetric(metric, checked === true)
+ }
+ aria-label={label}
+ />
+ {label}
+ </label>
+ );
+ })}
+ </div>
+ )}
</SettingsSection>
<SettingsSection title={t("settings.resources.gpu.title")}>
diff --git a/studio/frontend/src/hooks/use-system.ts b/studio/frontend/src/hooks/use-system.ts
index 6de9c606112..ce06fc4945a 100644
--- a/studio/frontend/src/hooks/use-system.ts
+++ b/studio/frontend/src/hooks/use-system.ts
@@ -24,6 +24,11 @@ export interface GpuDevice {
unified_memory?: boolean;
/** host-backed portion of the shared pool; the rest is reserved GPU memory. */
shared_memory_host_backed_gb?: number | null;
+ /** Live readings from the same probe as vram_used_gb; absent where the backend has none. */
+ gpu_utilization_pct?: number | null;
+ temperature_c?: number | null;
+ power_draw_w?: number | null;
+ power_limit_w?: number | null;
}
export interface SystemGpuInfo {
diff --git a/studio/frontend/src/i18n/locales/ar.ts b/studio/frontend/src/i18n/locales/ar.ts
index d12517d89f1..4f12d915cb2 100644
--- a/studio/frontend/src/i18n/locales/ar.ts
+++ b/studio/frontend/src/i18n/locales/ar.ts
@@ -2106,6 +2106,13 @@ export const ar = {
noGpu: "لا يوجد GPU مرئي",
gpuUnusable: "GPU غير قابل للاستخدام",
gpuUnusableDetail: "تم اكتشافه، لكن PyTorch لا يستطيع استخدامه",
+ detailedView: "عرض مفصل",
+ detailedViewDescription:
+ "اعرض حمل المعالج وحمل كل وحدة GPU وحرارتها واستهلاكها للطاقة في النافذة العائمة وهنا.",
+ cpuLoad: "حمل المعالج",
+ gpuLoad: "حمل GPU",
+ temperature: "الحرارة",
+ power: "الطاقة",
},
gpu: {
title: "أجهزة GPU",
diff --git a/studio/frontend/src/i18n/locales/de.ts b/studio/frontend/src/i18n/locales/de.ts
index 6719b7eb322..09455e0fbe2 100644
--- a/studio/frontend/src/i18n/locales/de.ts
+++ b/studio/frontend/src/i18n/locales/de.ts
@@ -2134,6 +2134,13 @@ export const de = {
noGpu: "Keine sichtbare GPU",
gpuUnusable: "GPU nicht nutzbar",
gpuUnusableDetail: "Erkannt, aber PyTorch kann sie nicht nutzen",
+ detailedView: "Detailansicht",
+ detailedViewDescription:
+ "Zeigt CPU-Last sowie Last, Temperatur und Leistungsaufnahme jeder GPU im schwebenden Fenster und hier.",
+ cpuLoad: "CPU-Last",
+ gpuLoad: "GPU-Last",
+ temperature: "Temperatur",
+ power: "Leistung",
},
gpu: {
title: "GPU-Geräte",
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 9591f991c11..f009a1b70ac 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -2146,6 +2146,13 @@ export const en = {
noGpu: "No visible GPU",
gpuUnusable: "GPU unusable",
gpuUnusableDetail: "Detected, but PyTorch cannot use it",
+ detailedView: "Detailed view",
+ detailedViewDescription:
+ "Show CPU load and each GPU's load, temperature and power in the floating window and here.",
+ cpuLoad: "CPU load",
+ gpuLoad: "GPU load",
+ temperature: "Temperature",
+ power: "Power",
},
gpu: {
title: "GPU devices",
diff --git a/studio/frontend/src/i18n/locales/es.ts b/studio/frontend/src/i18n/locales/es.ts
index edf42039945..fd59561187d 100644
--- a/studio/frontend/src/i18n/locales/es.ts
+++ b/studio/frontend/src/i18n/locales/es.ts
@@ -2129,6 +2129,13 @@ export const es = {
noGpu: "No hay GPU visible",
gpuUnusable: "GPU no utilizable",
gpuUnusableDetail: "Detectada, pero PyTorch no puede usarla",
+ detailedView: "Vista detallada",
+ detailedViewDescription:
+ "Muestra la carga de la CPU y la carga, temperatura y consumo de cada GPU en la ventana flotante y aquí.",
+ cpuLoad: "Carga de CPU",
+ gpuLoad: "Carga de GPU",
+ temperature: "Temperatura",
+ power: "Consumo",
},
gpu: {
title: "Dispositivos GPU",
diff --git a/studio/frontend/src/i18n/locales/fr.ts b/studio/frontend/src/i18n/locales/fr.ts
index c07db5a6658..e42b6f5a584 100644
--- a/studio/frontend/src/i18n/locales/fr.ts
+++ b/studio/frontend/src/i18n/locales/fr.ts
@@ -2133,6 +2133,13 @@ export const fr = {
noGpu: "Aucun GPU visible",
gpuUnusable: "GPU inutilisable",
gpuUnusableDetail: "Détecté, mais PyTorch ne peut pas l'utiliser",
+ detailedView: "Vue détaillée",
+ detailedViewDescription:
+ "Affiche la charge du CPU ainsi que la charge, la température et la consommation de chaque GPU dans la fenêtre flottante et ici.",
+ cpuLoad: "Charge CPU",
+ gpuLoad: "Charge GPU",
+ temperature: "Température",
+ power: "Puissance",
},
gpu: {
title: "Périphériques GPU",
diff --git a/studio/frontend/src/i18n/locales/he.ts b/studio/frontend/src/i18n/locales/he.ts
index b86a7ffd62f..01feca41746 100644
--- a/studio/frontend/src/i18n/locales/he.ts
+++ b/studio/frontend/src/i18n/locales/he.ts
@@ -2187,6 +2187,13 @@ export const he = {
noGpu: "לא זוהה GPU",
gpuUnusable: "GPU אינו שמיש",
gpuUnusableDetail: "זוהה, אך PyTorch אינו יכול להשתמש בו",
+ detailedView: "תצוגה מפורטת",
+ detailedViewDescription:
+ "הצגת עומס המעבד ועומס, טמפרטורה וצריכת חשמל של כל GPU בחלון הצף וכאן.",
+ cpuLoad: "עומס מעבד",
+ gpuLoad: "עומס GPU",
+ temperature: "טמפרטורה",
+ power: "הספק",
},
gpu: {
title: "התקני GPU",
diff --git a/studio/frontend/src/i18n/locales/hi.ts b/studio/frontend/src/i18n/locales/hi.ts
index a0bbbd42021..478a38faaed 100644
--- a/studio/frontend/src/i18n/locales/hi.ts
+++ b/studio/frontend/src/i18n/locales/hi.ts
@@ -2109,6 +2109,13 @@ export const hi = {
noGpu: "कोई दृश्यमान GPU नहीं",
gpuUnusable: "GPU अनुपयोगी",
gpuUnusableDetail: "पहचाना गया, लेकिन PyTorch इसका उपयोग नहीं कर सकता",
+ detailedView: "विस्तृत दृश्य",
+ detailedViewDescription:
+ "फ़्लोटिंग विंडो में और यहाँ CPU लोड तथा हर GPU का लोड, तापमान और पावर दिखाएँ।",
+ cpuLoad: "CPU लोड",
+ gpuLoad: "GPU लोड",
+ temperature: "तापमान",
+ power: "पावर",
},
gpu: {
title: "GPU डिवाइस",
diff --git a/studio/frontend/src/i18n/locales/it.ts b/studio/frontend/src/i18n/locales/it.ts
index 4dd30e3294c..dca424ba4bc 100644
--- a/studio/frontend/src/i18n/locales/it.ts
+++ b/studio/frontend/src/i18n/locales/it.ts
@@ -2093,6 +2093,13 @@ export const it = {
noGpu: "Nessuna GPU visibile",
gpuUnusable: "GPU non utilizzabile",
gpuUnusableDetail: "Rilevata, ma PyTorch non può usarla",
+ detailedView: "Vista dettagliata",
+ detailedViewDescription:
+ "Mostra il carico della CPU e carico, temperatura e consumo di ogni GPU nella finestra mobile e qui.",
+ cpuLoad: "Carico CPU",
+ gpuLoad: "Carico GPU",
+ temperature: "Temperatura",
+ power: "Consumo",
},
gpu: {
title: "Dispositivi GPU",
diff --git a/studio/frontend/src/i18n/locales/ja.ts b/studio/frontend/src/i18n/locales/ja.ts
index ee345118a7e..65ce1ea44c0 100644
--- a/studio/frontend/src/i18n/locales/ja.ts
+++ b/studio/frontend/src/i18n/locales/ja.ts
@@ -2085,6 +2085,13 @@ export const ja = {
noGpu: "認識できる GPU がありません",
gpuUnusable: "GPU を利用できません",
gpuUnusableDetail: "検出されましたが、PyTorch から利用できません",
+ detailedView: "詳細表示",
+ detailedViewDescription:
+ "フローティングウィンドウとここに、CPU 負荷と各 GPU の負荷、温度、消費電力を表示します。",
+ cpuLoad: "CPU 負荷",
+ gpuLoad: "GPU 負荷",
+ temperature: "温度",
+ power: "消費電力",
},
gpu: {
title: "GPU デバイス",
diff --git a/studio/frontend/src/i18n/locales/ko.ts b/studio/frontend/src/i18n/locales/ko.ts
index 974837cd9b5..1a1fa4d9fcc 100644
--- a/studio/frontend/src/i18n/locales/ko.ts
+++ b/studio/frontend/src/i18n/locales/ko.ts
@@ -2102,6 +2102,13 @@ export const ko = {
noGpu: "인식되는 GPU 없음",
gpuUnusable: "GPU 사용 불가",
gpuUnusableDetail: "감지되었지만 PyTorch에서 사용할 수 없습니다",
+ detailedView: "자세히 보기",
+ detailedViewDescription:
+ "플로팅 창과 여기에 CPU 부하와 각 GPU의 부하, 온도, 전력을 표시합니다.",
+ cpuLoad: "CPU 부하",
+ gpuLoad: "GPU 부하",
+ temperature: "온도",
+ power: "전력",
},
gpu: {
title: "GPU 장치",
diff --git a/studio/frontend/src/i18n/locales/pt-br.ts b/studio/frontend/src/i18n/locales/pt-br.ts
index dba47a7e755..62066f37ff1 100644
--- a/studio/frontend/src/i18n/locales/pt-br.ts
+++ b/studio/frontend/src/i18n/locales/pt-br.ts
@@ -2118,6 +2118,13 @@ export const ptBR = {
noGpu: "Nenhuma GPU visível",
gpuUnusable: "GPU inutilizável",
gpuUnusableDetail: "Detectada, mas o PyTorch não consegue usá-la",
+ detailedView: "Visualização detalhada",
+ detailedViewDescription:
+ "Mostra a carga da CPU e a carga, temperatura e consumo de cada GPU na janela flutuante e aqui.",
+ cpuLoad: "Carga da CPU",
+ gpuLoad: "Carga da GPU",
+ temperature: "Temperatura",
+ power: "Consumo",
},
gpu: {
title: "Dispositivos de GPU",
diff --git a/studio/frontend/src/i18n/locales/ru.ts b/studio/frontend/src/i18n/locales/ru.ts
index a0cb37d6f29..eeb2c263ce0 100644
--- a/studio/frontend/src/i18n/locales/ru.ts
+++ b/studio/frontend/src/i18n/locales/ru.ts
@@ -2112,6 +2112,13 @@ export const ru = {
noGpu: "GPU не обнаружен",
gpuUnusable: "GPU недоступен",
gpuUnusableDetail: "Обнаружен, но PyTorch не может его использовать",
+ detailedView: "Подробный вид",
+ detailedViewDescription:
+ "Показывать загрузку CPU, а также загрузку, температуру и мощность каждой GPU в плавающем окне и здесь.",
+ cpuLoad: "Загрузка CPU",
+ gpuLoad: "Загрузка GPU",
+ temperature: "Температура",
+ power: "Мощность",
},
gpu: {
title: "Устройства GPU",
diff --git a/studio/frontend/src/i18n/locales/sv.ts b/studio/frontend/src/i18n/locales/sv.ts
index 4e52051d430..37bb726cd8f 100644
--- a/studio/frontend/src/i18n/locales/sv.ts
+++ b/studio/frontend/src/i18n/locales/sv.ts
@@ -2257,6 +2257,13 @@ export const sv = {
noGpu: "Ingen synlig GPU",
gpuUnusable: "GPU oanvändbar",
gpuUnusableDetail: "Identifierad, men PyTorch kan inte använda den",
+ detailedView: "Detaljerad vy",
+ detailedViewDescription:
+ "Visa CPU-belastning samt belastning, temperatur och effekt för varje GPU i det flytande fönstret och här.",
+ cpuLoad: "CPU-belastning",
+ gpuLoad: "GPU-belastning",
+ temperature: "Temperatur",
+ power: "Effekt",
},
gpu: {
title: "GPU-enheter",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index c14df4bce5e..db2929b3ebe 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -2079,6 +2079,13 @@ export const zhCN = {
noGpu: "无可见 GPU",
gpuUnusable: "GPU 不可用",
gpuUnusableDetail: "已检测到,但 PyTorch 无法使用",
+ detailedView: "详细视图",
+ detailedViewDescription:
+ "在浮动窗口和此处显示 CPU 负载以及每个 GPU 的负载、温度和功耗。",
+ cpuLoad: "CPU 负载",
+ gpuLoad: "GPU 负载",
+ temperature: "温度",
+ power: "功耗",
},
gpu: {
title: "GPU 设备",
diff --git a/studio/frontend/tests/monitor-details.test.ts b/studio/frontend/tests/monitor-details.test.ts
new file mode 100644
index 00000000000..a90b2e4851b
--- /dev/null
+++ b/studio/frontend/tests/monitor-details.test.ts
@@ -0,0 +1,124 @@
+// 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 {
+ MONITOR_DETAIL_METRICS,
+ formatPower,
+ formatTemperature,
+ gpuTelemetryRows,
+ selectedDetailMetrics,
+ temperatureLevel,
+} from "../src/components/monitor-details.ts";
+
+const ALL = selectedDetailMetrics(undefined);
+
+test("stored metric selections survive and garbage falls back to every metric", () => {
+ assert.deepEqual([...ALL], [...MONITOR_DETAIL_METRICS]);
+ assert.deepEqual(
+ [...selectedDetailMetrics("power")],
+ [...MONITOR_DETAIL_METRICS],
+ );
+ assert.deepEqual(
+ [...selectedDetailMetrics(["power", "bogus", "temperature"])],
+ ["temperature", "power"],
+ );
+ assert.deepEqual([...selectedDetailMetrics([])], []);
+ assert.deepEqual(
+ [...selectedDetailMetrics(["bogus", 1])],
+ [...MONITOR_DETAIL_METRICS],
+ );
+});
+
+test("rows carry only readings the backend reported", () => {
+ const rows = gpuTelemetryRows(
+ [
+ {
+ index: 0,
+ name: "NVIDIA B200",
+ gpu_utilization_pct: 42,
+ temperature_c: 61,
+ power_draw_w: 312.5,
+ power_limit_w: 1000,
+ },
+ { index: 1, name: "NVIDIA B200" },
+ { index: 2, name: "NVIDIA B200", temperature_c: 55, power_draw_w: 80 },
+ ],
+ ALL,
+ );
+ assert.deepEqual(rows, [
+ {
+ key: "0",
+ name: "GPU 0: NVIDIA B200",
+ loadPct: 42,
+ temperatureC: 61,
+ powerDrawW: 312.5,
+ powerLimitW: 1000,
+ powerPct: 31.25,
+ },
+ {
+ key: "2",
+ name: "GPU 2: NVIDIA B200",
+ loadPct: null,
+ temperatureC: 55,
+ powerDrawW: 80,
+ powerLimitW: null,
+ powerPct: null,
+ },
+ ]);
+});
+
+test("a deselected metric is dropped, and a device left with nothing is skipped", () => {
+ const device = {
+ index: 0,
+ name: "RTX 4090",
+ temperature_c: 70,
+ gpu_utilization_pct: 10,
+ };
+ const rows = gpuTelemetryRows(
+ [device],
+ selectedDetailMetrics(["temperature"]),
+ );
+ assert.equal(rows.length, 1);
+ assert.equal(rows[0].name, "RTX 4090");
+ assert.equal(rows[0].loadPct, null);
+ assert.equal(rows[0].temperatureC, 70);
+ assert.deepEqual(
+ gpuTelemetryRows([device], selectedDetailMetrics(["power"])),
+ [],
+ );
+ assert.deepEqual(
+ gpuTelemetryRows([{ index: 0, temperature_c: Number.NaN }], ALL),
+ [],
+ );
+});
+
+test("temperature turns amber at 80 C and red at 90 C", () => {
+ assert.equal(temperatureLevel(79.9), "ok");
+ assert.equal(temperatureLevel(80), "warm");
+ assert.equal(temperatureLevel(90), "hot");
+ assert.equal(formatTemperature(61.4), "61°C");
+ assert.equal(formatPower(312.5, 1000), "313 / 1000 W");
+ assert.equal(formatPower(80, null), "80 W");
+});
+
+test("the detailed view is opt-in and adds no request of its own", () => {
+ const store = readFileSync(
+ new URL(
+ "../src/features/settings/stores/monitor-overlay-store.ts",
+ import.meta.url,
+ ),
+ "utf8",
+ );
+ assert.match(store, /showDetails: false,/);
+ const monitor = readFileSync(
+ new URL("../src/components/floating-monitor.tsx", import.meta.url),
+ "utf8",
+ );
+ // Rows render from the /api/system poll the monitor already runs.
+ assert.doesNotMatch(monitor, /useGpuUtilization|api\/train\/hardware/);
+ assert.match(monitor, /showDetails && displayedGpu\?\.available/);
+});