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: always show loaded GGUF weight placement (#17449)

oobaboogaa4a46cfb6829
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index fc297d7acc4..23e4cb2ac68 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -479,6 +479,7 @@ import {
} from "@/lib/evicted-toast";
import {
type OffloadCounts,
+ expertSpillNote,
offloadCountsFrom,
offloadWarning,
} from "../lib/partial-offload";
@@ -3482,6 +3483,7 @@ async function autoLoadSmallestModel(options?: AutoLoadOptions): Promise<{
cpuFallbackReason,
mmprojFallbackReason,
offloadWarning(offloadCounts ?? {}),
+ expertSpillNote(offloadCounts ?? {}),
);
const options = {
description: notice.description,
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
index 21f3856b996..2bb7c259fa1 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts
@@ -47,6 +47,7 @@ import {
} from "../lib/server-wide-reload";
import {
type OffloadCounts,
+ expertSpillNote,
offloadCountsFrom,
offloadWarning,
} from "../lib/partial-offload";
@@ -3441,6 +3442,7 @@ export function useChatModelRuntime() {
cpuFallbackReason,
mmprojFallbackReason,
offloadWarning(offloadCounts),
+ expertSpillNote(offloadCounts),
);
const loadedTitle = notice.title;
const loadedDescription = notice.description;
diff --git a/studio/frontend/src/features/chat/lib/partial-offload.ts b/studio/frontend/src/features/chat/lib/partial-offload.ts
index a3cd2aacb82..ddd3e3ee463 100644
--- a/studio/frontend/src/features/chat/lib/partial-offload.ts
+++ b/studio/frontend/src/features/chat/lib/partial-offload.ts
@@ -16,6 +16,10 @@ export interface OffloadCounts {
offloadOverridden?: boolean | null;
cpuFallbackReason?: string | null;
gpuBackendUnavailable?: boolean | null;
+ overflowLayers?: number | null;
+ expertCacheMib?: number | null;
+ gpuWeightsMib?: number | null;
+ hostWeightsMib?: number | null;
}
export interface OffloadCountsSource {
@@ -26,6 +30,10 @@ export interface OffloadCountsSource {
offload_overridden?: boolean | null;
cpu_fallback_reason?: string | null;
gpu_backend_unavailable?: boolean | null;
+ overflow_layers?: number | null;
+ expert_cache_mib?: number | null;
+ gpu_weights_mib?: number | null;
+ host_weights_mib?: number | null;
}
export function offloadCountsFrom(
@@ -39,6 +47,10 @@ export function offloadCountsFrom(
offloadOverridden: response.offload_overridden,
cpuFallbackReason: response.cpu_fallback_reason,
gpuBackendUnavailable: response.gpu_backend_unavailable,
+ overflowLayers: response.overflow_layers,
+ expertCacheMib: response.expert_cache_mib,
+ gpuWeightsMib: response.gpu_weights_mib,
+ hostWeightsMib: response.host_weights_mib,
};
}
@@ -47,57 +59,6 @@ export interface OffloadWarning {
description: string;
}
-/** Why a loaded model is not fully on the GPU, or `null`. A known cpu_fallback_reason
- * wins over the counts: a Vulkan CPU fallback also logs 0/M, and is not a size problem.
- * Manual pins only count with gpuLayers >= 0 (Manual + Auto layers is llama.cpp's fit). */
-export function offloadWarning(counts: OffloadCounts): OffloadWarning | null {
- const {
- offloaded,
- total,
- gpuMemoryMode,
- gpuLayers,
- offloadOverridden,
- cpuFallbackReason,
- gpuBackendUnavailable,
- } = counts;
- const cpuFallbackDescription = cpuFallbackMessage(cpuFallbackReason);
- if (cpuFallbackDescription) {
- return {
- titleSuffix: " on CPU",
- description: cpuFallbackDescription,
- };
- }
- if (cpuFallbackReason) return null;
- if (typeof offloaded !== "number" || typeof total !== "number") return null;
- if (total <= 0 || offloaded >= total) return null;
- // Before the pin check: a failed GPU backend is never what was asked for.
- if (offloaded <= 0 && gpuBackendUnavailable && gpuLayers !== 0) {
- return {
- titleSuffix: ", on CPU",
- description:
- "The GPU was found but llama.cpp could not use it, so the model runs " +
- "entirely on CPU and generation will be slow. This is a backend " +
- "problem rather than a size one, so a smaller quantization will not " +
- "help. The llama-server log in Settings > Logs has the reason.",
- };
- }
- const manualPin =
- gpuMemoryMode === "manual" &&
- typeof gpuLayers === "number" &&
- gpuLayers >= 0;
- if (manualPin || offloadOverridden) return null;
- if (offloaded <= 0) {
- return {
- titleSuffix: ", on CPU",
- description: `None of the ${total} layers fit on the GPU, so the model runs entirely on CPU and generation will be slow. A smaller quantization or a shorter context may leave room on the GPU.`,
- };
- }
- return {
- titleSuffix: ", partly on CPU",
- description: `${offloaded} of ${total} layers are on the GPU. The rest run on CPU, so generation will be slower. A smaller quantization or a shorter context may let more of it fit on the GPU.`,
- };
-}
-
function count(value: unknown): number | null {
return typeof value === "number" && Number.isFinite(value) && value >= 0
? value
@@ -111,6 +72,95 @@ function weightAmount(mib: number): string {
: formatGiB(mib / 1024);
}
+/** Manual pins only count with gpuLayers >= 0 (Manual + Auto layers is llama.cpp's fit). */
+function splitChosenByUser(counts: OffloadCounts): boolean {
+ return (
+ (counts.gpuMemoryMode === "manual" &&
+ typeof counts.gpuLayers === "number" &&
+ counts.gpuLayers >= 0) ||
+ counts.offloadOverridden === true
+ );
+}
+
+type ShortfallCause = "fallback" | "backend" | "noneFit" | "partial";
+
+/** Why a load is not fully on the GPU although nothing asked for that, or `null`. A cpu_fallback_reason
+ * wins over the counts: a Vulkan CPU fallback also logs 0/M, and is not a size problem. */
+function shortfallCause(counts: OffloadCounts): ShortfallCause | null {
+ const { offloaded, total, gpuLayers, gpuBackendUnavailable } = counts;
+ if (counts.cpuFallbackReason) return "fallback";
+ if (typeof offloaded !== "number" || typeof total !== "number") return null;
+ if (total <= 0) return null;
+ // A GPU build whose backend failed to start still logs "offloaded N/N"; its buffers are all host ones.
+ const noGpuBuffers =
+ count(counts.gpuWeightsMib) === 0 && (count(counts.hostWeightsMib) ?? 0) > 0;
+ const onGpu = gpuBackendUnavailable && noGpuBuffers ? 0 : offloaded;
+ if (onGpu >= total) return null;
+ // Before the pin check: a failed GPU backend is never what was asked for.
+ if (onGpu <= 0 && gpuBackendUnavailable && gpuLayers !== 0) return "backend";
+ if (splitChosenByUser(counts)) return null;
+ return onGpu <= 0 ? "noneFit" : "partial";
+}
+
+/** Weights only, no KV cache. Names the device, not VRAM/RAM: those are one pool on some machines. */
+function weightsSentence(counts: OffloadCounts): string {
+ const gpu = count(counts.gpuWeightsMib) ?? 0;
+ const host = count(counts.hostWeightsMib) ?? 0;
+ const parts: string[] = [];
+ if (gpu > 0) parts.push(`${weightAmount(gpu)} on the GPU`);
+ if (host > 0) parts.push(`${weightAmount(host)} on the CPU`);
+ return parts.length > 0 ? ` Model weights: ${parts.join(", ")}.` : "";
+}
+
+export function offloadWarning(counts: OffloadCounts): OffloadWarning | null {
+ const { offloaded, total } = counts;
+ switch (shortfallCause(counts)) {
+ case "fallback": {
+ // A reason this build does not know is left to the caller's own fallback wording.
+ const description = cpuFallbackMessage(counts.cpuFallbackReason);
+ return description ? { titleSuffix: " on CPU", description } : null;
+ }
+ case "backend":
+ return {
+ titleSuffix: ", on CPU",
+ description:
+ "The GPU was found but llama.cpp could not use it, so the model runs " +
+ "entirely on CPU and generation will be slow. This is a backend " +
+ "problem rather than a size one, so a smaller quantization will not " +
+ "help. The llama-server log in Settings > Logs has the reason.",
+ };
+ case "noneFit":
+ return {
+ titleSuffix: ", on CPU",
+ description: `None of the ${total} layers fit on the GPU, so the model runs entirely on CPU and generation will be slow. A smaller quantization or a shorter context may leave room on the GPU.${weightsSentence(counts)}`,
+ };
+ case "partial":
+ return {
+ titleSuffix: ", partly on CPU",
+ description: `${offloaded} of ${total} layers are on the GPU. The rest run on CPU, so generation will be slower. A smaller quantization or a shorter context may let more of it fit on the GPU.${weightsSentence(counts)}`,
+ };
+ default:
+ return null;
+ }
+}
+
+/** Not a warning: --fit keeping MoE expert weights on the CPU is how a large MoE fits, far cheaper than layers. */
+export function expertSpillNote(counts: OffloadCounts): string | null {
+ const { offloaded, total } = counts;
+ if (typeof offloaded !== "number" || typeof total !== "number") return null;
+ if (total <= 0 || offloaded < total) return null;
+ if (shortfallCause(counts) !== null || splitChosenByUser(counts)) return null;
+ const cache = count(counts.expertCacheMib) ?? 0;
+ const overflow = count(counts.overflowLayers) ?? 0;
+ if (cache > 0) {
+ return `All ${total} layers are on the GPU. Every expert is on the CPU, with a ${weightAmount(cache)} cache of the most used ones on the GPU.${weightsSentence(counts)}`;
+ }
+ if (overflow > 0) {
+ return `All ${total} layers are on the GPU. The expert weights of ${overflow} ${overflow === 1 ? "layer" : "layers"} are on the CPU.${weightsSentence(counts)}`;
+ }
+ return null;
+}
+
/** --fit reports its own expert spill; a Manual load says it through n_cpu_moe instead. */
export function loadedPlacementFrom(response: {
is_gguf?: boolean | null;
diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts
index 54f47d42634..ff8d9e8c6db 100644
--- a/studio/frontend/src/features/chat/types/api.ts
+++ b/studio/frontend/src/features/chat/types/api.ts
@@ -291,6 +291,10 @@ export interface LoadModelResponse {
gpu_layers?: number;
offloaded_layers?: number | null;
offload_total_layers?: number | null;
+ overflow_layers?: number | null;
+ gpu_weights_mib?: number | null;
+ host_weights_mib?: number | null;
+ expert_cache_mib?: number | null;
offload_overridden?: boolean | null;
gpu_backend_unavailable?: boolean | null;
cpu_fallback_reason?: CpuFallbackReason | null;
diff --git a/studio/frontend/src/features/chat/utils/mmproj-fallback.ts b/studio/frontend/src/features/chat/utils/mmproj-fallback.ts
index 38dbdf43a14..41baa3dd6e6 100644
--- a/studio/frontend/src/features/chat/utils/mmproj-fallback.ts
+++ b/studio/frontend/src/features/chat/utils/mmproj-fallback.ts
@@ -60,12 +60,14 @@ export function mmprojLoadNotice(
};
}
-/** Covers both fallbacks: a CPU-fallback replay can set both, and neither may be dropped. */
+/** Covers both fallbacks: a CPU-fallback replay can set both, and neither may be dropped.
+ * `placementNote` is information for a load that is otherwise fine, so it never makes the notice a warning. */
export function loadFallbackNotice(
baseTitle: string,
cpuFallbackReason: CpuFallbackReason | null | undefined,
mmprojFallbackReason: MmprojFallbackReason | null | undefined,
offloadNotice?: { titleSuffix: string; description: string } | null,
+ placementNote?: string | null,
): { title: string; description: string | undefined; degraded: boolean } {
const textOnly = isTextOnlyMmprojFallback(mmprojFallbackReason);
const offload = cpuFallbackReason ? null : (offloadNotice ?? null);
@@ -91,6 +93,8 @@ export function loadFallbackNotice(
}
if (offload) {
parts.push(offload.description);
+ } else if (!cpuFallbackReason && placementNote) {
+ parts.push(placementNote);
}
if (mmprojFallbackReason) {
parts.push(mmprojFallbackMessage(mmprojFallbackReason));
diff --git a/studio/frontend/src/features/model-picker/components/loaded-placement-row.tsx b/studio/frontend/src/features/model-picker/components/loaded-placement-row.tsx
index d975ce34a4a..90b9895e902 100644
--- a/studio/frontend/src/features/model-picker/components/loaded-placement-row.tsx
+++ b/studio/frontend/src/features/model-picker/components/loaded-placement-row.tsx
@@ -4,18 +4,17 @@
import {
describePlacement,
type PlacementPool,
- useChatPreferencesStore,
useChatRuntimeStore,
} from "@/features/chat";
import { useT } from "@/i18n";
/** Where the resident GGUF's weights landed. Under Auto, --fit decides, and the layer count alone hides a
- * MoE spill, so the experts and the RAM side are shown too. `pool` is the panel's single memory pool, if any. */
+ * MoE spill, so the experts and the RAM side are shown too. `pool` is the panel's single memory pool, if any.
+ * Always shown: the estimate above it is a forecast, and this is the only lasting account of the load. */
export function LoadedPlacementRow({ pool }: { pool: PlacementPool | null }) {
const t = useT();
- const enabled = useChatPreferencesStore((s) => s.showGenerationDetails);
const placement = useChatRuntimeStore((s) => s.loadedPlacement);
- if (!enabled || placement === null) return null;
+ if (placement === null) return null;
const { layers, memory } = describePlacement(placement, t, pool);
if (layers === null && memory === null) return null;
return (
diff --git a/studio/frontend/src/i18n/locales/ar.ts b/studio/frontend/src/i18n/locales/ar.ts
index a621a228e99..1851b3fd951 100644
--- a/studio/frontend/src/i18n/locales/ar.ts
+++ b/studio/frontend/src/i18n/locales/ar.ts
@@ -3043,11 +3043,11 @@ export const ar = {
showReplyTimesDescription: "إظهار وقت انتهاء كل رد أسفله.",
showReplyTimesKeywords:
"طابع زمني وقت ساعة تاريخ متى انتهى رد رسالة timestamp time clock date reply",
- showGenerationDetails: "معلومات مفصّلة عن السرعة والتحميل",
+ showGenerationDetails: "معلومات مفصّلة عن السرعة",
showGenerationDetailsDescription:
- "يضيف سرعة قراءة الموجّه وعدد الرموز وقبول المسودات إلى القراءات المباشرة، ويعرض مكان تحميل طبقات النموذج.",
+ "يضيف سرعة قراءة الموجّه وعدد الرموز وقبول المسودات إلى القراءات المباشرة.",
showGenerationDetailsKeywords:
- "تفاصيل سرعة رموز إحصاءات طبقات ذاكرة verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "تفاصيل سرعة رموز إحصاءات verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "إحصاءات السرعة والسياق",
generationStatsUpdatesDescription:
"يحسب الوضع الفوري الرموز في الثانية واستخدام السياق أثناء بث الرد. أما عند الانتهاء فيحدّثها عند اكتمال كل رد.",
diff --git a/studio/frontend/src/i18n/locales/cs.ts b/studio/frontend/src/i18n/locales/cs.ts
index 8cf2323d0f0..dfd14f2b308 100644
--- a/studio/frontend/src/i18n/locales/cs.ts
+++ b/studio/frontend/src/i18n/locales/cs.ts
@@ -3308,11 +3308,11 @@ export const cs = {
"Pod každou odpovědí zobrazovat, kdy byla dokončena.",
showReplyTimesKeywords:
"časové razítko čas hodiny datum kdy dokončeno odpověď zpráva timestamp time clock date hour when finished arrived reply response message",
- showGenerationDetails: "Podrobné informace o rychlosti a načítání",
+ showGenerationDetails: "Podrobné informace o rychlosti",
showGenerationDetailsDescription:
- "Přidá do živých údajů rychlost čtení výzvy, počty tokenů a přijetí návrhů a ukáže, kam se načetly vrstvy modelu.",
+ "Přidá do živých údajů rychlost čtení výzvy, počty tokenů a přijetí návrhů.",
showGenerationDetailsKeywords:
- "podrobné vývojář statistiky rychlost vrstvy verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "podrobné vývojář statistiky rychlost verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "Statistiky rychlosti a kontextu",
generationStatsUpdatesDescription:
"Režim V reálném čase počítá tokeny za sekundu a využití kontextu během streamování odpovědi. Po dokončení je aktualizuje, až každá odpověď skončí.",
diff --git a/studio/frontend/src/i18n/locales/de.ts b/studio/frontend/src/i18n/locales/de.ts
index 05c54d5cb06..0ba69c0d068 100644
--- a/studio/frontend/src/i18n/locales/de.ts
+++ b/studio/frontend/src/i18n/locales/de.ts
@@ -3084,11 +3084,11 @@ export const de = {
showReplyTimesDescription: "Unter jeder Antwort anzeigen, wann sie fertig war.",
showReplyTimesKeywords:
"Zeitstempel Uhrzeit Datum wann fertig Antwort Nachricht timestamp time clock date reply",
- showGenerationDetails: "Detaillierte Geschwindigkeits- und Ladeinfos",
+ showGenerationDetails: "Detaillierte Geschwindigkeitsinfos",
showGenerationDetailsDescription:
- "Ergänzt die Live-Anzeigen um Prompt-Lesegeschwindigkeit, Tokenzahlen und Entwurfsannahme und zeigt, wo die Schichten eines Modells geladen wurden.",
+ "Ergänzt die Live-Anzeigen um Prompt-Lesegeschwindigkeit, Tokenzahlen und Entwurfsannahme.",
showGenerationDetailsKeywords:
- "Details Geschwindigkeit Token Statistik Schichten Auslagerung verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "Details Geschwindigkeit Token Statistik verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "Geschwindigkeits- und Kontextstatistik",
generationStatsUpdatesDescription:
"Echtzeit zählt Token pro Sekunde und Kontextnutzung, während eine Antwort gestreamt wird. Nach Abschluss aktualisiert sie, sobald jede Antwort fertig ist.",
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index cb985aa8cae..efa6d8cdce4 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -3062,11 +3062,11 @@ export const en = {
showReplyTimesDescription: "Show when each reply finished, below the reply.",
showReplyTimesKeywords:
"timestamp time clock date hour when finished arrived reply response message",
- showGenerationDetails: "Detailed speed and loading info",
+ showGenerationDetails: "Detailed speed info",
showGenerationDetailsDescription:
- "Adds prompt-reading speed, token counts and draft acceptance to the live readouts, and shows where a model's layers loaded.",
+ "Adds prompt-reading speed, token counts and draft acceptance to the live readouts.",
showGenerationDetailsKeywords:
- "generation details verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "generation details verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "Speed and context stats",
generationStatsUpdatesDescription:
"Realtime counts tokens per second and context usage while a reply streams. When done updates them as each reply finishes.",
diff --git a/studio/frontend/src/i18n/locales/es-419.ts b/studio/frontend/src/i18n/locales/es-419.ts
index b9c924baec4..8ec201457da 100644
--- a/studio/frontend/src/i18n/locales/es-419.ts
+++ b/studio/frontend/src/i18n/locales/es-419.ts
@@ -3337,11 +3337,11 @@ export const es419 = {
"Muestra debajo de cada respuesta cuándo terminó.",
showReplyTimesKeywords:
"marca de tiempo hora reloj fecha cuándo terminó llegó respuesta mensaje timestamp time clock date hour when finished arrived reply response message",
- showGenerationDetails: "Información detallada de velocidad y carga",
+ showGenerationDetails: "Información detallada de velocidad",
showGenerationDetailsDescription:
- "Agrega a las lecturas en vivo la velocidad de lectura del prompt, el conteo de tokens y la aceptación de borradores, y muestra dónde se cargaron las capas del modelo.",
+ "Agrega a las lecturas en vivo la velocidad de lectura del prompt, el conteo de tokens y la aceptación de borradores.",
showGenerationDetailsKeywords:
- "detallado desarrollador estadísticas velocidad capas trasladar expertos verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "detallado desarrollador estadísticas velocidad verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "Estadísticas de velocidad y contexto",
generationStatsUpdatesDescription:
"En tiempo real cuenta los tokens por segundo y el uso del contexto mientras se transmite una respuesta. Al terminar los actualiza cuando termina cada respuesta.",
diff --git a/studio/frontend/src/i18n/locales/es.ts b/studio/frontend/src/i18n/locales/es.ts
index d1cec5f1558..b977fc83565 100644
--- a/studio/frontend/src/i18n/locales/es.ts
+++ b/studio/frontend/src/i18n/locales/es.ts
@@ -3075,11 +3075,11 @@ export const es = {
showReplyTimesDescription: "Mostrar debajo de cada respuesta cuándo terminó.",
showReplyTimesKeywords:
"marca de tiempo hora fecha cuándo terminó respuesta mensaje timestamp time clock date reply",
- showGenerationDetails: "Información detallada de velocidad y carga",
+ showGenerationDetails: "Información detallada de velocidad",
showGenerationDetailsDescription:
- "Añade a las lecturas en directo la velocidad de lectura del prompt, el recuento de tokens y la aceptación de borradores, y muestra dónde se cargaron las capas del modelo.",
+ "Añade a las lecturas en directo la velocidad de lectura del prompt, el recuento de tokens y la aceptación de borradores.",
showGenerationDetailsKeywords:
- "detalles velocidad tokens estadísticas capas descarga verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "detalles velocidad tokens estadísticas verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "Estadísticas de velocidad y contexto",
generationStatsUpdatesDescription:
"En tiempo real cuenta los tokens por segundo y el uso del contexto mientras se transmite una respuesta. Al terminar los actualiza cuando acaba cada respuesta.",
diff --git a/studio/frontend/src/i18n/locales/fa.ts b/studio/frontend/src/i18n/locales/fa.ts
index 1e470a9a3e6..2e653774cb1 100644
--- a/studio/frontend/src/i18n/locales/fa.ts
+++ b/studio/frontend/src/i18n/locales/fa.ts
@@ -3278,11 +3278,11 @@ export const fa = {
"زمان پایان هر پاسخ، زیر همان پاسخ نمایش داده شود.",
showReplyTimesKeywords:
"مهر زمانی زمان ساعت تاریخ پایان رسیدن پاسخ پیام timestamp",
- showGenerationDetails: "اطلاعات دقیق سرعت و بارگذاری",
+ showGenerationDetails: "اطلاعات دقیق سرعت",
showGenerationDetailsDescription:
- "سرعت خواندن پرامپت، شمار توکن‌ها و پذیرش پیش‌نویس را به نمایش زنده می‌افزاید و نشان می‌دهد لایه‌های مدل کجا بارگذاری شده‌اند.",
+ "سرعت خواندن پرامپت، شمار توکن‌ها و پذیرش پیش‌نویس را به نمایش زنده می‌افزاید.",
showGenerationDetailsKeywords:
- "مفصل توسعه‌دهنده آمار سرعت tok/s prefill decode لایه‌ها offload vram experts",
+ "مفصل توسعه‌دهنده آمار سرعت tok/s prefill decode",
generationStatsUpdates: "آمار سرعت و زمینه",
generationStatsUpdatesDescription:
"حالت بی‌درنگ توکن در ثانیه و مصرف زمینه را هنگام پخش پاسخ می‌شمارد. حالت پس از پایان آن‌ها را با تمام شدن هر پاسخ به‌روز می‌کند.",
diff --git a/studio/frontend/src/i18n/locales/fr.ts b/studio/frontend/src/i18n/locales/fr.ts
index 0c9a832fafa..3e4d65b040e 100644
--- a/studio/frontend/src/i18n/locales/fr.ts
+++ b/studio/frontend/src/i18n/locales/fr.ts
@@ -3080,11 +3080,11 @@ export const fr = {
showReplyTimesDescription: "Afficher sous chaque réponse l'heure à laquelle elle s'est terminée.",
showReplyTimesKeywords:
"horodatage heure date quand terminé réponse message timestamp time clock date reply",
- showGenerationDetails: "Infos détaillées de vitesse et de chargement",
+ showGenerationDetails: "Infos détaillées de vitesse",
showGenerationDetailsDescription:
- "Ajoute aux indicateurs en direct la vitesse de lecture du prompt, le nombre de tokens et l'acceptation des brouillons, et indique où les couches du modèle ont été chargées.",
+ "Ajoute aux indicateurs en direct la vitesse de lecture du prompt, le nombre de tokens et l'acceptation des brouillons.",
showGenerationDetailsKeywords:
- "détails vitesse jetons statistiques couches déchargement verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "détails vitesse jetons statistiques verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "Statistiques de vitesse et de contexte",
generationStatsUpdatesDescription:
"En temps réel compte les tokens par seconde et l'utilisation du contexte pendant la diffusion d'une réponse. À la fin les met à jour quand chaque réponse se termine.",
diff --git a/studio/frontend/src/i18n/locales/he.ts b/studio/frontend/src/i18n/locales/he.ts
index 56f18ded81b..0535192fe99 100644
--- a/studio/frontend/src/i18n/locales/he.ts
+++ b/studio/frontend/src/i18n/locales/he.ts
@@ -3123,11 +3123,11 @@ export const he = {
showReplyTimesDescription: "הצג מתחת לכל תשובה מתי היא הסתיימה.",
showReplyTimesKeywords:
"חותמת זמן שעה תאריך מתי הסתיים תשובה הודעה timestamp time clock date reply",
- showGenerationDetails: "מידע מפורט על מהירות וטעינה",
+ showGenerationDetails: "מידע מפורט על מהירות",
showGenerationDetailsDescription:
- "מוסיף למדדים החיים את מהירות קריאת ההנחיה, ספירת טוקנים וקבלת טיוטות, ומציג היכן נטענו שכבות המודל.",
+ "מוסיף למדדים החיים את מהירות קריאת ההנחיה, ספירת טוקנים וקבלת טיוטות.",
showGenerationDetailsKeywords:
- "פרטים מהירות אסימונים סטטיסטיקה שכבות verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "פרטים מהירות אסימונים סטטיסטיקה verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "נתוני מהירות והקשר",
generationStatsUpdatesDescription:
"מצב זמן אמת סופר טוקנים לשנייה ושימוש בהקשר בזמן שתשובה מוזרמת. מצב בסיום מעדכן אותם כשכל תשובה מסתיימת.",
diff --git a/studio/frontend/src/i18n/locales/hi.ts b/studio/frontend/src/i18n/locales/hi.ts
index e524036abeb..db74dd7a284 100644
--- a/studio/frontend/src/i18n/locales/hi.ts
+++ b/studio/frontend/src/i18n/locales/hi.ts
@@ -3052,11 +3052,11 @@ export const hi = {
showReplyTimesDescription: "हर जवाब के नीचे दिखाएँ कि वह कब पूरा हुआ।",
showReplyTimesKeywords:
"टाइमस्टैम्प समय घड़ी तारीख कब पूरा जवाब संदेश timestamp time clock date reply",
- showGenerationDetails: "गति और लोडिंग की विस्तृत जानकारी",
+ showGenerationDetails: "गति की विस्तृत जानकारी",
showGenerationDetailsDescription:
- "लाइव रीडआउट में प्रॉम्प्ट पढ़ने की गति, टोकन गिनती और ड्राफ़्ट स्वीकृति जोड़ता है, और दिखाता है कि मॉडल की लेयर कहाँ लोड हुईं।",
+ "लाइव रीडआउट में प्रॉम्प्ट पढ़ने की गति, टोकन गिनती और ड्राफ़्ट स्वीकृति जोड़ता है।",
showGenerationDetailsKeywords:
- "विवरण गति टोकन आँकड़े परतें verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "विवरण गति टोकन आँकड़े verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "गति और संदर्भ आँकड़े",
generationStatsUpdatesDescription:
"रीयल-टाइम उत्तर स्ट्रीम होते समय प्रति सेकंड टोकन और संदर्भ उपयोग गिनता है। पूरा होने पर हर उत्तर खत्म होने पर इन्हें अपडेट करता है।",
diff --git a/studio/frontend/src/i18n/locales/id.ts b/studio/frontend/src/i18n/locales/id.ts
index 5e04a1c2964..06adccdfd7f 100644
--- a/studio/frontend/src/i18n/locales/id.ts
+++ b/studio/frontend/src/i18n/locales/id.ts
@@ -3292,11 +3292,11 @@ export const id = {
"Tampilkan kapan setiap balasan selesai, di bawah balasan.",
showReplyTimesKeywords:
"timestamp time clock date hour when finished arrived reply response message waktu jam tanggal balasan pesan",
- showGenerationDetails: "Info kecepatan dan pemuatan terperinci",
+ showGenerationDetails: "Info kecepatan terperinci",
showGenerationDetailsDescription:
- "Menambahkan kecepatan membaca prompt, jumlah token, dan penerimaan draf ke tampilan langsung, serta menunjukkan tempat lapisan model dimuat.",
+ "Menambahkan kecepatan membaca prompt, jumlah token, dan penerimaan draf ke tampilan langsung.",
showGenerationDetailsKeywords:
- "verbose developer stats speed tok/s prefill decode layers offload vram experts statistik kecepatan lapisan",
+ "verbose developer stats speed tok/s prefill decode statistik kecepatan",
generationStatsUpdates: "Statistik kecepatan dan konteks",
generationStatsUpdatesDescription:
"Waktu nyata menghitung token per detik dan penggunaan konteks saat balasan sedang dialirkan. Setelah selesai memperbaruinya saat setiap balasan selesai.",
diff --git a/studio/frontend/src/i18n/locales/it.ts b/studio/frontend/src/i18n/locales/it.ts
index 417f81a02ab..a5bf011a16e 100644
--- a/studio/frontend/src/i18n/locales/it.ts
+++ b/studio/frontend/src/i18n/locales/it.ts
@@ -2978,11 +2978,11 @@ export const it = {
showReplyTimesDescription: "Mostra sotto ogni risposta quando è terminata.",
showReplyTimesKeywords:
"orario ora data quando terminata risposta messaggio timestamp time clock date reply",
- showGenerationDetails: "Info dettagliate su velocità e caricamento",
+ showGenerationDetails: "Info dettagliate sulla velocità",
showGenerationDetailsDescription:
- "Aggiunge alle letture dal vivo la velocità di lettura del prompt, il conteggio dei token e l'accettazione delle bozze, e mostra dove sono stati caricati i livelli del modello.",
+ "Aggiunge alle letture dal vivo la velocità di lettura del prompt, il conteggio dei token e l'accettazione delle bozze.",
showGenerationDetailsKeywords:
- "dettagli velocità token statistiche livelli scaricamento verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "dettagli velocità token statistiche verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "Statistiche di velocità e contesto",
generationStatsUpdatesDescription:
"In tempo reale conta i token al secondo e l'uso del contesto mentre una risposta è in streaming. Al termine li aggiorna quando ogni risposta finisce.",
diff --git a/studio/frontend/src/i18n/locales/ja.ts b/studio/frontend/src/i18n/locales/ja.ts
index 408fcd14fed..b943d1f3b3e 100644
--- a/studio/frontend/src/i18n/locales/ja.ts
+++ b/studio/frontend/src/i18n/locales/ja.ts
@@ -3029,11 +3029,11 @@ export const ja = {
showReplyTimesDescription: "各応答の下に、完了した時刻を表示します。",
showReplyTimesKeywords:
"タイムスタンプ 時刻 時間 日付 完了 応答 メッセージ timestamp time clock date reply",
- showGenerationDetails: "速度と読み込みの詳細情報",
+ showGenerationDetails: "速度の詳細情報",
showGenerationDetailsDescription:
- "ライブ表示にプロンプトの読み取り速度、トークン数、ドラフトの採用率を追加し、モデルのレイヤーの読み込み先を表示します。",
+ "ライブ表示にプロンプトの読み取り速度、トークン数、ドラフトの採用率を追加します。",
showGenerationDetailsKeywords:
- "詳細 速度 トークン 統計 レイヤー オフロード verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "詳細 速度 トークン 統計 verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "速度とコンテキストの統計",
generationStatsUpdatesDescription:
"リアルタイムでは、応答のストリーミング中に毎秒トークン数とコンテキスト使用量を数えます。完了時は各応答の終了時に更新します。",
diff --git a/studio/frontend/src/i18n/locales/ko.ts b/studio/frontend/src/i18n/locales/ko.ts
index 555e0d9dfd4..86a3c41c3fc 100644
--- a/studio/frontend/src/i18n/locales/ko.ts
+++ b/studio/frontend/src/i18n/locales/ko.ts
@@ -3036,11 +3036,11 @@ export const ko = {
showReplyTimesDescription: "각 응답 아래에 완료된 시간을 표시합니다.",
showReplyTimesKeywords:
"타임스탬프 시간 시각 날짜 완료 응답 메시지 timestamp time clock date reply",
- showGenerationDetails: "자세한 속도 및 로딩 정보",
+ showGenerationDetails: "자세한 속도 정보",
showGenerationDetailsDescription:
- "실시간 표시에 프롬프트 읽기 속도, 토큰 수, 초안 채택률을 더하고 모델 레이어가 로드된 위치를 보여 줍니다.",
+ "실시간 표시에 프롬프트 읽기 속도, 토큰 수, 초안 채택률을 더합니다.",
showGenerationDetailsKeywords:
- "세부 속도 토큰 통계 레이어 오프로드 verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "세부 속도 토큰 통계 verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "속도 및 컨텍스트 통계",
generationStatsUpdatesDescription:
"실시간은 응답이 스트리밍되는 동안 초당 토큰과 컨텍스트 사용량을 셉니다. 완료 시는 각 응답이 끝날 때 업데이트합니다.",
diff --git a/studio/frontend/src/i18n/locales/nl.ts b/studio/frontend/src/i18n/locales/nl.ts
index 402853442b9..73ffcd04c72 100644
--- a/studio/frontend/src/i18n/locales/nl.ts
+++ b/studio/frontend/src/i18n/locales/nl.ts
@@ -3333,11 +3333,11 @@ export const nl = {
"Onder elk antwoord tonen wanneer het klaar was.",
showReplyTimesKeywords:
"tijdstempel tijd klok datum uur wanneer voltooid aangekomen antwoord reactie bericht timestamp time",
- showGenerationDetails: "Gedetailleerde snelheids- en laadinfo",
+ showGenerationDetails: "Gedetailleerde snelheidsinfo",
showGenerationDetailsDescription:
- "Voegt promptleessnelheid, tokenaantallen en conceptacceptatie toe aan de live weergaven en toont waar de lagen van een model zijn geladen.",
+ "Voegt promptleessnelheid, tokenaantallen en conceptacceptatie toe aan de live weergaven.",
showGenerationDetailsKeywords:
- "uitgebreid ontwikkelaar statistieken snelheid tok/s prefill decode lagen offload vram experts verbose developer stats speed layers",
+ "uitgebreid ontwikkelaar statistieken snelheid tok/s prefill decode verbose developer stats speed",
generationStatsUpdates: "Snelheids- en contextstatistieken",
generationStatsUpdatesDescription:
"Realtime telt tokens per seconde en contextgebruik terwijl een antwoord wordt gestreamd. Na afloop werkt ze bij zodra elk antwoord klaar is.",
diff --git a/studio/frontend/src/i18n/locales/pl.ts b/studio/frontend/src/i18n/locales/pl.ts
index 927aea3ef8e..d6dbad63303 100644
--- a/studio/frontend/src/i18n/locales/pl.ts
+++ b/studio/frontend/src/i18n/locales/pl.ts
@@ -3328,11 +3328,11 @@ export const pl = {
"Pokazuj pod odpowiedzią, kiedy się zakończyła.",
showReplyTimesKeywords:
"timestamp time clock date hour when finished arrived reply response message znacznik czasu godzina data odpowiedź wiadomość",
- showGenerationDetails: "Szczegółowe informacje o szybkości i ładowaniu",
+ showGenerationDetails: "Szczegółowe informacje o szybkości",
showGenerationDetailsDescription:
- "Dodaje do odczytów na żywo szybkość czytania promptu, liczbę tokenów i akceptację szkiców oraz pokazuje, gdzie załadowano warstwy modelu.",
+ "Dodaje do odczytów na żywo szybkość czytania promptu, liczbę tokenów i akceptację szkiców.",
showGenerationDetailsKeywords:
- "verbose developer stats speed tok/s prefill decode layers offload vram experts szczegóły statystyki szybkość warstwy eksperci",
+ "verbose developer stats speed tok/s prefill decode szczegóły statystyki szybkość",
generationStatsUpdates: "Statystyki szybkości i kontekstu",
generationStatsUpdatesDescription:
"Tryb Na bieżąco liczy tokeny na sekundę i użycie kontekstu podczas strumieniowania odpowiedzi. Po zakończeniu aktualizuje je, gdy każda odpowiedź się skończy.",
diff --git a/studio/frontend/src/i18n/locales/pt-br.ts b/studio/frontend/src/i18n/locales/pt-br.ts
index 972893d3d71..008af6e90e5 100644
--- a/studio/frontend/src/i18n/locales/pt-br.ts
+++ b/studio/frontend/src/i18n/locales/pt-br.ts
@@ -3057,11 +3057,11 @@ export const ptBR = {
showReplyTimesDescription: "Mostrar abaixo de cada resposta quando ela terminou.",
showReplyTimesKeywords:
"carimbo de data hora data quando terminou resposta mensagem timestamp time clock date reply",
- showGenerationDetails: "Informações detalhadas de velocidade e carregamento",
+ showGenerationDetails: "Informações detalhadas de velocidade",
showGenerationDetailsDescription:
- "Adiciona às leituras ao vivo a velocidade de leitura do prompt, a contagem de tokens e a aceitação de rascunhos, e mostra onde as camadas do modelo foram carregadas.",
+ "Adiciona às leituras ao vivo a velocidade de leitura do prompt, a contagem de tokens e a aceitação de rascunhos.",
showGenerationDetailsKeywords:
- "detalhes geração velocidade tokens estatísticas camadas especialistas desenvolvedor verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "detalhes geração velocidade tokens estatísticas desenvolvedor verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "Estatísticas de velocidade e contexto",
generationStatsUpdatesDescription:
"Em tempo real conta tokens por segundo e uso do contexto enquanto uma resposta é transmitida. Ao terminar atualiza quando cada resposta termina.",
diff --git a/studio/frontend/src/i18n/locales/ru.ts b/studio/frontend/src/i18n/locales/ru.ts
index 47bb1f2b9f2..d2682e52502 100644
--- a/studio/frontend/src/i18n/locales/ru.ts
+++ b/studio/frontend/src/i18n/locales/ru.ts
@@ -3056,11 +3056,11 @@ export const ru = {
showReplyTimesDescription: "Показывать под каждым ответом, когда он завершился.",
showReplyTimesKeywords:
"время метка дата когда завершено ответ сообщение timestamp time clock date reply",
- showGenerationDetails: "Подробности о скорости и загрузке",
+ showGenerationDetails: "Подробности о скорости",
showGenerationDetailsDescription:
- "Добавляет к живым показателям скорость чтения запроса, число токенов и принятие черновиков, а также показывает, куда загружены слои модели.",
+ "Добавляет к живым показателям скорость чтения запроса, число токенов и принятие черновиков.",
showGenerationDetailsKeywords:
- "подробности скорость токены статистика слои выгрузка verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "подробности скорость токены статистика verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "Статистика скорости и контекста",
generationStatsUpdatesDescription:
"В реальном времени считает токены в секунду и использование контекста, пока ответ передаётся. По завершении обновляет их, когда каждый ответ закончен.",
diff --git a/studio/frontend/src/i18n/locales/sv.ts b/studio/frontend/src/i18n/locales/sv.ts
index a174dc9ca2e..f0f456b7260 100644
--- a/studio/frontend/src/i18n/locales/sv.ts
+++ b/studio/frontend/src/i18n/locales/sv.ts
@@ -3205,11 +3205,11 @@ export const sv = {
showReplyTimesDescription: "Visa under varje svar när det blev klart.",
showReplyTimesKeywords:
"tidsstämpel tid klocka datum när klar svar meddelande timestamp time clock date reply",
- showGenerationDetails: "Detaljerad hastighets- och laddningsinfo",
+ showGenerationDetails: "Detaljerad hastighetsinfo",
showGenerationDetailsDescription:
- "Lägger till läshastighet för prompten, tokenantal och utkastgodkännande i livevisningen och visar var modellens lager laddades.",
+ "Lägger till läshastighet för prompten, tokenantal och utkastgodkännande i livevisningen.",
showGenerationDetailsKeywords:
- "detaljer hastighet token statistik lager avlastning verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "detaljer hastighet token statistik verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "Hastighets- och kontextstatistik",
generationStatsUpdatesDescription:
"Realtid räknar token per sekund och kontextanvändning medan ett svar strömmas. När klart uppdaterar dem när varje svar är färdigt.",
diff --git a/studio/frontend/src/i18n/locales/th.ts b/studio/frontend/src/i18n/locales/th.ts
index 29b4f24331d..5f6fa7ba800 100644
--- a/studio/frontend/src/i18n/locales/th.ts
+++ b/studio/frontend/src/i18n/locales/th.ts
@@ -3156,11 +3156,11 @@ export const th = {
showReplyTimesDescription: "แสดงเวลาที่การตอบกลับแต่ละครั้งเสร็จสิ้นไว้ใต้การตอบกลับ",
showReplyTimesKeywords:
"เวลา นาฬิกา วันที่ ชั่วโมง เมื่อใด เสร็จ มาถึง ตอบกลับ คำตอบ ข้อความ timestamp time clock date reply response message",
- showGenerationDetails: "ข้อมูลความเร็วและการโหลดแบบละเอียด",
+ showGenerationDetails: "ข้อมูลความเร็วแบบละเอียด",
showGenerationDetailsDescription:
- "เพิ่มความเร็วการอ่านพรอมต์ จำนวนโทเค็น และอัตราการยอมรับร่างลงในการแสดงผลสด และแสดงว่าเลเยอร์ของโมเดลโหลดไว้ที่ใด",
+ "เพิ่มความเร็วการอ่านพรอมต์ จำนวนโทเค็น และอัตราการยอมรับร่างลงในการแสดงผลสด",
showGenerationDetailsKeywords:
- "ละเอียด นักพัฒนา สถิติ ความเร็ว เลเยอร์ verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "ละเอียด นักพัฒนา สถิติ ความเร็ว verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "สถิติความเร็วและบริบท",
generationStatsUpdatesDescription:
"แบบเรียลไทม์จะนับโทเค็นต่อวินาทีและการใช้บริบทระหว่างที่คำตอบกำลังสตรีม แบบเมื่อเสร็จจะอัปเดตเมื่อแต่ละคำตอบจบ",
diff --git a/studio/frontend/src/i18n/locales/tr.ts b/studio/frontend/src/i18n/locales/tr.ts
index 1b13e33bc10..f49980ee3fe 100644
--- a/studio/frontend/src/i18n/locales/tr.ts
+++ b/studio/frontend/src/i18n/locales/tr.ts
@@ -3290,11 +3290,11 @@ export const tr = {
"Her yanıtın ne zaman tamamlandığını yanıtın altında gösterin.",
showReplyTimesKeywords:
"zaman damgası zaman saat tarih ne zaman bitti geldi yanıt cevap mesaj timestamp",
- showGenerationDetails: "Ayrıntılı hız ve yükleme bilgisi",
+ showGenerationDetails: "Ayrıntılı hız bilgisi",
showGenerationDetailsDescription:
- "Canlı göstergelere istem okuma hızını, token sayılarını ve taslak kabulünü ekler ve modelin katmanlarının nereye yüklendiğini gösterir.",
+ "Canlı göstergelere istem okuma hızını, token sayılarını ve taslak kabulünü ekler.",
showGenerationDetailsKeywords:
- "ayrıntılı geliştirici istatistik hız tok/s prefill decode katmanlar aktarım offload vram uzmanlar experts",
+ "ayrıntılı geliştirici istatistik hız tok/s prefill decode",
generationStatsUpdates: "Hız ve bağlam istatistikleri",
generationStatsUpdatesDescription:
"Gerçek zamanlı, bir yanıt akarken saniye başına token ve bağlam kullanımını sayar. Bitince, her yanıt tamamlandığında bunları günceller.",
diff --git a/studio/frontend/src/i18n/locales/uk.ts b/studio/frontend/src/i18n/locales/uk.ts
index 2052c0e7ada..c0b08a07799 100644
--- a/studio/frontend/src/i18n/locales/uk.ts
+++ b/studio/frontend/src/i18n/locales/uk.ts
@@ -3318,11 +3318,11 @@ export const uk = {
"Показувати під кожною відповіддю, коли її завершено.",
showReplyTimesKeywords:
"позначка часу час годинник дата година коли завершено надійшло відповідь повідомлення timestamp",
- showGenerationDetails: "Докладна інформація про швидкість і завантаження",
+ showGenerationDetails: "Докладна інформація про швидкість",
showGenerationDetailsDescription:
- "Додає до живих показників швидкість читання запиту, кількість токенів і прийняття чернеток, а також показує, куди завантажено шари моделі.",
+ "Додає до живих показників швидкість читання запиту, кількість токенів і прийняття чернеток.",
showGenerationDetailsKeywords:
- "докладно розробник статистика швидкість tok/s prefill decode шари вивантаження offload vram експерти",
+ "докладно розробник статистика швидкість tok/s prefill decode",
generationStatsUpdates: "Статистика швидкості й контексту",
generationStatsUpdatesDescription:
"У реальному часі рахує токени за секунду й використання контексту, поки відповідь передається. Після завершення оновлює їх, коли кожна відповідь закінчена.",
diff --git a/studio/frontend/src/i18n/locales/vi.ts b/studio/frontend/src/i18n/locales/vi.ts
index d8a06f983ca..860f7136a22 100644
--- a/studio/frontend/src/i18n/locales/vi.ts
+++ b/studio/frontend/src/i18n/locales/vi.ts
@@ -3291,11 +3291,11 @@ export const vi = {
"Hiển thị thời điểm mỗi câu trả lời hoàn tất, bên dưới câu trả lời.",
showReplyTimesKeywords:
"timestamp time clock date hour when finished arrived reply response message dấu thời gian giờ ngày hoàn tất trả lời tin nhắn",
- showGenerationDetails: "Thông tin chi tiết về tốc độ và tải",
+ showGenerationDetails: "Thông tin chi tiết về tốc độ",
showGenerationDetailsDescription:
- "Thêm tốc độ đọc lời nhắc, số token và tỷ lệ chấp nhận bản nháp vào chỉ số trực tiếp, và cho biết các lớp của mô hình được tải ở đâu.",
+ "Thêm tốc độ đọc lời nhắc, số token và tỷ lệ chấp nhận bản nháp vào chỉ số trực tiếp.",
showGenerationDetailsKeywords:
- "verbose developer stats speed tok/s prefill decode layers offload vram experts chi tiết nhà phát triển thống kê tốc độ lớp",
+ "verbose developer stats speed tok/s prefill decode chi tiết nhà phát triển thống kê tốc độ",
generationStatsUpdates: "Thống kê tốc độ và ngữ cảnh",
generationStatsUpdatesDescription:
"Thời gian thực đếm token mỗi giây và mức dùng ngữ cảnh trong khi phản hồi đang truyền. Khi xong cập nhật chúng khi mỗi phản hồi kết thúc.",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index 5271da54524..e3e379f7345 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -3000,11 +3000,11 @@ export const zhCN = {
showReplyTimesDescription: "在每条回复下方显示其完成的时间。",
showReplyTimesKeywords:
"时间戳 时间 时刻 日期 完成 回复 消息 timestamp time clock date reply",
- showGenerationDetails: "详细速度与加载信息",
+ showGenerationDetails: "详细速度信息",
showGenerationDetailsDescription:
- "在实时读数中加入提示词读取速度、token 数和草稿采纳率,并显示模型各层加载在哪里。",
+ "在实时读数中加入提示词读取速度、token 数和草稿采纳率。",
showGenerationDetailsKeywords:
- "详情 速度 统计 层 卸载 verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "详情 速度 统计 verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "速度与上下文统计",
generationStatsUpdatesDescription:
"实时模式在回复流式输出时统计每秒 token 数和上下文用量。完成后模式在每条回复结束时更新。",
diff --git a/studio/frontend/src/i18n/locales/zh-TW.ts b/studio/frontend/src/i18n/locales/zh-TW.ts
index 7be3fb50adb..73cc4a237dd 100644
--- a/studio/frontend/src/i18n/locales/zh-TW.ts
+++ b/studio/frontend/src/i18n/locales/zh-TW.ts
@@ -3138,11 +3138,11 @@ export const zhTW = {
showReplyTimesDescription: "在每則回覆下方顯示其完成時間。",
showReplyTimesKeywords:
"時間戳記 時間 時鐘 日期 小時 完成 抵達 回覆 回應 訊息 timestamp time clock date hour when finished arrived reply response message",
- showGenerationDetails: "詳細速度與載入資訊",
+ showGenerationDetails: "詳細速度資訊",
showGenerationDetailsDescription:
- "在即時讀數中加入提示詞讀取速度、token 數和草稿採用率,並顯示模型各層載入在哪裡。",
+ "在即時讀數中加入提示詞讀取速度、token 數和草稿採用率。",
showGenerationDetailsKeywords:
- "詳細 開發人員 統計 速度 預填 解碼 層 卸載 專家 verbose developer stats speed tok/s prefill decode layers offload vram experts",
+ "詳細 開發人員 統計 速度 預填 解碼 verbose developer stats speed tok/s prefill decode",
generationStatsUpdates: "速度與上下文統計",
generationStatsUpdatesDescription:
"即時模式在回覆串流輸出時統計每秒 token 數和上下文用量。完成後模式在每則回覆結束時更新。",
diff --git a/studio/frontend/tests/generation-details.test.ts b/studio/frontend/tests/generation-details.test.ts
index 79713b7ff26..e81482e35a8 100644
--- a/studio/frontend/tests/generation-details.test.ts
+++ b/studio/frontend/tests/generation-details.test.ts
@@ -475,6 +475,15 @@ test("the placement readout names expert weights and gives both memory figures",
assert.equal(loadedPlacementFrom({ is_gguf: true }), null);
});
+test("the placement row needs no setting", () => {
+ // Source text: the row cannot be mounted here, since it pulls in the chat feature's React tree.
+ const row = readSrc("features/model-picker/components/loaded-placement-row.tsx");
+ assert.doesNotMatch(row, /showGenerationDetails|useChatPreferencesStore/);
+ assert.match(row, /if \(placement === null\) return null;/);
+ assert.equal(en.settings.chat.showGenerationDetails, "Detailed speed info");
+ assert.doesNotMatch(en.settings.chat.showGenerationDetailsDescription, /layers/);
+});
+
test("a GPU build that fell back to the CPU is not reported as GPU layers", () => {
// CUDA build, no usable device: llama.cpp still logs "offloaded 29/29 layers to GPU", buffers AMX + CPU_Mapped.
assert.deepEqual(
diff --git a/studio/frontend/tests/mmproj-fallback.test.ts b/studio/frontend/tests/mmproj-fallback.test.ts
index 02ec5b2c62f..d18fe4bf5ec 100644
--- a/studio/frontend/tests/mmproj-fallback.test.ts
+++ b/studio/frontend/tests/mmproj-fallback.test.ts
@@ -201,3 +201,27 @@ test("a clean load is not degraded and carries no description", () => {
assert.equal(notice.description, undefined);
assert.equal(notice.degraded, false);
});
+
+test("a placement note is said on a success toast and gives way to any warning", () => {
+ const note = "All 31 layers are on the GPU. The expert weights of 7 layers are on the CPU.";
+ const informed = loadFallbackNotice("X loaded", null, null, null, note);
+ assert.equal(informed.title, "X loaded");
+ assert.equal(informed.description, note);
+ assert.equal(informed.degraded, false);
+ // With a projector on the CPU the note still leads, and the load is a warning for the projector's sake.
+ const withProjector = loadFallbackNotice("X loaded", null, "cpu_offload", null, note);
+ assert.equal(
+ withProjector.description,
+ `${note} ${mmprojFallbackMessage("cpu_offload")}`,
+ );
+ assert.equal(withProjector.degraded, true);
+ const split = { titleSuffix: ", partly on CPU", description: "38 of 60 layers are on the GPU." };
+ assert.equal(
+ loadFallbackNotice("X loaded", null, null, split, note).description,
+ split.description,
+ );
+ assert.doesNotMatch(
+ loadFallbackNotice("X loaded", "vulkan_startup_crash", null, null, note).description ?? "",
+ /expert weights/,
+ );
+});
diff --git a/studio/frontend/tests/partial-offload.test.ts b/studio/frontend/tests/partial-offload.test.ts
index 914a48aa09f..0381b9b068f 100644
--- a/studio/frontend/tests/partial-offload.test.ts
+++ b/studio/frontend/tests/partial-offload.test.ts
@@ -8,7 +8,7 @@ import { registerBundlerResolver } from "./helpers/kit.ts";
registerBundlerResolver();
-const { offloadCountsFrom, offloadWarning } = await import(
+const { expertSpillNote, offloadCountsFrom, offloadWarning } = await import(
"../src/features/chat/lib/partial-offload.ts"
);
@@ -97,6 +97,14 @@ test("every load path reads the response the same way", () => {
cpu_fallback_reason: null,
// biome-ignore lint/style/useNamingConvention: api schema
gpu_backend_unavailable: false,
+ // biome-ignore lint/style/useNamingConvention: api schema
+ overflow_layers: 7,
+ // biome-ignore lint/style/useNamingConvention: api schema
+ expert_cache_mib: null,
+ // biome-ignore lint/style/useNamingConvention: api schema
+ gpu_weights_mib: 1228.8,
+ // biome-ignore lint/style/useNamingConvention: api schema
+ host_weights_mib: 2048,
}),
{
offloaded: 38,
@@ -106,6 +114,10 @@ test("every load path reads the response the same way", () => {
offloadOverridden: false,
cpuFallbackReason: null,
gpuBackendUnavailable: false,
+ overflowLayers: 7,
+ expertCacheMib: null,
+ gpuWeightsMib: 1228.8,
+ hostWeightsMib: 2048,
},
);
});
@@ -200,3 +212,85 @@ test("the advice does not promise that a smaller quantization fits", () => {
assert.doesNotMatch(text, /would fit entirely|would leave room/);
}
});
+
+test("the notice says how much of the weights is on the GPU and on the CPU", () => {
+ // A T4 load of a 27B: CUDA0 12677.16 MiB, CPU_Mapped 5833.19 MiB.
+ assert.match(
+ offloadWarning({ offloaded: 45, total: 66, gpuWeightsMib: 12677.16, hostWeightsMib: 5833.19 })
+ ?.description ?? "",
+ /may let more of it fit on the GPU\. Model weights: 12 GiB on the GPU, 5\.7 GiB on the CPU\.$/,
+ );
+ assert.match(
+ offloadWarning({ offloaded: 0, total: 60, gpuWeightsMib: 0, hostWeightsMib: 22400 })
+ ?.description ?? "",
+ /may leave room on the GPU\. Model weights: 22 GiB on the CPU\.$/,
+ );
+ assert.match(
+ offloadWarning({ offloaded: 9, total: 25, gpuWeightsMib: 40 })?.description ?? "",
+ /Model weights: 40 MiB on the GPU\.$/,
+ );
+ assert.doesNotMatch(
+ offloadWarning({ offloaded: 54, total: 66, gpuWeightsMib: 19456, hostWeightsMib: 4813 })
+ ?.description ?? "",
+ /VRAM|RAM/,
+ );
+});
+
+test("missing or nonsense weight figures leave the notice as it was", () => {
+ for (const weights of [
+ {},
+ { gpuWeightsMib: null, hostWeightsMib: null },
+ { gpuWeightsMib: 0, hostWeightsMib: 0 },
+ { gpuWeightsMib: -5, hostWeightsMib: Number.NaN },
+ ]) {
+ assert.match(
+ offloadWarning({ offloaded: 38, total: 60, ...weights })?.description ?? "",
+ /may let more of it fit on the GPU\.$/,
+ );
+ }
+});
+
+test("expert weights the fit kept on the CPU are stated, not warned about", () => {
+ // gemma-4-26B-A4B Q8_0 on a 3090: llama.cpp logs "offloaded 31/31" with seven layers' experts in CUDA_Host.
+ const spill = {
+ offloaded: 31,
+ total: 31,
+ gpuMemoryMode: "auto",
+ gpuLayers: -1,
+ overflowLayers: 7,
+ gpuWeightsMib: 20709.15,
+ hostWeightsMib: 5639.4,
+ };
+ assert.equal(offloadWarning(spill), null);
+ assert.equal(
+ expertSpillNote(spill),
+ "All 31 layers are on the GPU. The expert weights of 7 layers are on the CPU. Model weights: 20 GiB on the GPU, 5.5 GiB on the CPU.",
+ );
+ assert.match(expertSpillNote({ ...spill, overflowLayers: 1 }) ?? "", /of 1 layer are on the CPU/);
+ assert.equal(
+ expertSpillNote({ offloaded: 31, total: 31, overflowLayers: 30, expertCacheMib: 6144 }),
+ "All 31 layers are on the GPU. Every expert is on the CPU, with a 6.0 GiB cache of the most used ones on the GPU.",
+ );
+});
+
+test("only a spill nobody chose, on an otherwise full load, gets the expert note", () => {
+ const spill = { offloaded: 31, total: 31, overflowLayers: 7 };
+ assert.equal(expertSpillNote({ offloaded: 31, total: 31, overflowLayers: 0 }), null);
+ assert.equal(expertSpillNote({ offloaded: 31, total: 31 }), null);
+ assert.equal(expertSpillNote({ ...spill, gpuMemoryMode: "manual", gpuLayers: 31 }), null);
+ assert.equal(expertSpillNote({ ...spill, offloadOverridden: true }), null);
+ // Manual with GPU Layers on Auto is still the fit's doing.
+ assert.notEqual(expertSpillNote({ ...spill, gpuMemoryMode: "manual", gpuLayers: -1 }), null);
+ assert.equal(expertSpillNote({ ...spill, offloaded: 20 }), null);
+ assert.equal(expertSpillNote({ ...spill, cpuFallbackReason: "vulkan_startup_crash" }), null);
+ assert.equal(expertSpillNote({}), null);
+});
+
+test("a full count with every buffer in system memory is not a success", () => {
+ const hostOnly = { offloaded: 29, total: 29, gpuWeightsMib: 0, hostWeightsMib: 499.8 };
+ const broken = offloadWarning({ ...hostOnly, gpuBackendUnavailable: true });
+ assert.equal(broken?.titleSuffix, ", on CPU");
+ assert.match(broken?.description ?? "", /could not use it/);
+ // The same buffers on a machine with no GPU to miss are the expected placement.
+ assert.equal(offloadWarning(hostOnly), null);
+});
diff --git a/tests/studio/test_chat_autoload_failure_gate.py b/tests/studio/test_chat_autoload_failure_gate.py
index be251947fe7..0f253e45de4 100644
--- a/tests/studio/test_chat_autoload_failure_gate.py
+++ b/tests/studio/test_chat_autoload_failure_gate.py
@@ -435,6 +435,9 @@ function offloadWarning(counts: any) {
if (total <= 0 || offloaded >= total) return null;
return { titleSuffix: ", partly on CPU", description: `${offloaded} of ${total}` };
}
+function expertSpillNote(_counts: any) {
+ return null;
+}
async function prepareHfTokenForUse(token: any) {
EVENTS.push({ kind: "prepareHfToken" });
if (SCENARIO.tokenDecision === "decline") return { proceed: false, token };