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: pick the quant of a GGUF dictation model in Voice settings (#12900)

Etherl1cd08cb9a0ea
diff --git a/studio/backend/core/inference/stt_audiocpp_sidecar.py b/studio/backend/core/inference/stt_audiocpp_sidecar.py
index 0ebdb566363..a727f51bec1 100644
--- a/studio/backend/core/inference/stt_audiocpp_sidecar.py
+++ b/studio/backend/core/inference/stt_audiocpp_sidecar.py
@@ -181,13 +181,20 @@ def resolve_audio_cpp_stt_model(
try:
require_runnable(found, "asr")
except AudioCppModelError as exc:
+ # cached rows omit absent quants, so callers treat a selected missing quant as downloadable.
+ if not network and (variant or ref_variant):
+ row = resolve(base, None, hf_token, network = False)
+ if row is not None and row.task == "asr" and row.unsupported is None:
+ raise SttModelNotDownloadedError(
+ f"STT model '{base}' ({variant or ref_variant}) is not downloaded. Download "
+ "it in Settings, then Voice, before loading it."
+ ) from exc
raise SttModelIdError(str(exc)) from exc
return found
def resolve_audio_cpp_stt_model_id(model: Optional[str]) -> str:
- """The name dictation reports for ``model``: a legacy key stays that key (Settings compares
- against it), anything else becomes its row id."""
+ """keeps legacy keys for Settings comparisons; all other names become row ids."""
if model is None or not str(model).strip():
return DEFAULT_AUDIO_CPP_STT_MODEL
base, variant = split_variant_ref(str(model).strip())
diff --git a/studio/backend/tests/test_audio_cpp_models.py b/studio/backend/tests/test_audio_cpp_models.py
index f52be09cd0f..2c03e79ecff 100644
--- a/studio/backend/tests/test_audio_cpp_models.py
+++ b/studio/backend/tests/test_audio_cpp_models.py
@@ -2084,10 +2084,22 @@ def test_stt_takes_the_variant_keys_gguf_variants_lists_even_from_a_partial_cach
assert s.resolve_audio_cpp_stt_model(f"{folder}:small/Q8_0").variant.key == "small/Q8_0"
assert s.resolve_audio_cpp_stt_model(f"{folder}:small").variant.key == "small/Q8_0"
assert s.is_model_downloaded(f"{folder}:small/Q8_0")
- # A variant that is not on disk is still refused rather than swapped for another.
+ # variants missing from disk are refused instead of replaced by another variant.
tiny = acm.resolve(folder, "tiny/Q8_0", network = False)
assert "not found" in tiny.unsupported
- # A file stem is a spelling of its own row, not a new scope.
+ # dictation treats this as a download opportunity, not a bad model id.
+ from core.inference.stt_sidecar import SttModelIdError, SttModelNotDownloadedError
+
+ with pytest.raises(SttModelNotDownloadedError, match = r"\(tiny/Q8_0\) is not downloaded"):
+ s.resolve_audio_cpp_stt_model(f"{folder}:tiny/Q8_0")
+ with pytest.raises(SttModelNotDownloadedError):
+ s.resolve_audio_cpp_stt_model(folder, "tiny/Q8_0")
+ assert not s.is_model_downloaded(f"{folder}:tiny/Q8_0")
+ # another task's folder remains invalid regardless of quant.
+ _kokoro(hub)
+ with pytest.raises(SttModelIdError):
+ s.resolve_audio_cpp_stt_model(f"{AUDIO_CPP_REPO}/Kokoro-82M-GGUF:Q8_0")
+ # a file stem names its existing row rather than a new scope.
stem = acm.resolve(folder, "moonshine-streaming-small-q8_0", network = False)
assert stem.variant.key == "Q8_0" and stem.unsupported is None
@@ -2096,7 +2108,7 @@ def test_yue2_length_follows_the_requested_duration():
b, srv = _backend_with(_speech_model("yue2"), MUSIC_REPLY)
b.generate_audio_response("[chorus] oh", instructions = "rock", max_new_tokens = 125)
options = srv.calls[0][1]["request"]["options"]
- # 5 s at 25 frames per second, with the default 200-frame floor lowered to fit.
+ # 5 s at 25 frames per second, with the 200-frame default lowered to fit.
assert options["semantic_max_tokens"] == 125 and options["semantic_min_tokens"] == 125
b.generate_audio_response("[chorus] oh", instructions = "rock", max_new_tokens = 1500)
options = srv.calls[1][1]["request"]["options"]
diff --git a/studio/frontend/src/features/audio/hooks/use-reference-transcribe.ts b/studio/frontend/src/features/audio/hooks/use-reference-transcribe.ts
index 3b6313fc4e5..91eb71474c6 100644
--- a/studio/frontend/src/features/audio/hooks/use-reference-transcribe.ts
+++ b/studio/frontend/src/features/audio/hooks/use-reference-transcribe.ts
@@ -2,14 +2,13 @@
// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
import { sttEngineFor } from "@/features/chat/adapters/studio-model-dictation-adapter";
-import { useVoiceSettingsStore } from "@/features/settings";
+import { useVoiceSettingsStore, withSttVariant } from "@/features/settings";
import { useCallback, useEffect, useRef, useState } from "react";
import { transcribeAudioInput } from "../api";
import { type AudioSourceSelection, sourceRefOf } from "../audio-run-request";
import { sttEngineForRepoId, sttSidecarKeyFor } from "../catalog";
-/** Fills "What's said in the clip" without adding it to the transcript list. No language hint:
- * the page's language is the output's, and a cross-lingual reference is not in it. */
+/** fills the prompt without a transcript entry or language hint for cross-lingual references. */
export function useReferenceTranscribe({
sttRepo,
onText,
@@ -29,7 +28,7 @@ export function useReferenceTranscribe({
const transcribe = useCallback(
async (reference: AudioSourceSelection | null) => {
if (!reference) return;
- // A newer clip wins, so its own transcription is never skipped.
+ // a newer clip wins so its transcription is never skipped.
abort.current?.abort();
const controller = new AbortController();
abort.current = controller;
@@ -41,7 +40,10 @@ export function useReferenceTranscribe({
const result = await transcribeAudioInput(
sourceRefOf(reference),
{
- model,
+ // the Settings model uses its saved quant; a Transcribe selection is already resident.
+ model: sttRepo
+ ? model
+ : withSttVariant(model, voice.sttGgufVariant),
engine: sttRepo ? sttEngineForRepoId(sttRepo) : sttEngineFor(model),
device: voice.sttDevice,
purpose,
diff --git a/studio/frontend/src/features/chat/adapters/studio-dictation-adapter.tsx b/studio/frontend/src/features/chat/adapters/studio-dictation-adapter.tsx
index b3ca0f62f6a..00de2a576c3 100644
--- a/studio/frontend/src/features/chat/adapters/studio-dictation-adapter.tsx
+++ b/studio/frontend/src/features/chat/adapters/studio-dictation-adapter.tsx
@@ -4,6 +4,7 @@
import { requestSttDownload } from "@/features/settings/stores/stt-download-prompt-store";
import {
type DictationEngine,
+ sttModelVariant,
useVoiceSettingsStore,
} from "@/features/settings/stores/voice-settings-store";
import { toast } from "@/lib/toast";
@@ -17,18 +18,17 @@ import {
StudioModelDictationAdapter,
fetchSttStatus,
sttEngineStatusFor,
+ sttQuantDownloaded,
} from "./studio-model-dictation-adapter";
import {
type StudioDictationSession,
StudioWebSpeechDictationAdapter,
} from "./studio-web-speech-dictation-adapter";
-// The one live dictation session, so Escape can discard it without going through assistant-ui,
-// which only exposes stop. Cancelling emits no transcript, so composer text is untouched.
-// assistant-ui exposes only stop, i.e. transcribe, which is why Escape needs its own path.
+// one live session lets Escape cancel because assistant-ui only exposes stop, which transcribes.
let activeSession: StudioDictationSession | null = null;
-/** Discard the current dictation without transcribing. Safe to call when idle. */
+/** discards the active dictation without transcribing; safe when idle. */
export function cancelActiveStudioDictation(): void {
const session = activeSession;
activeSession = null;
@@ -148,25 +148,22 @@ export function notifyStudioDictationUnavailable(
return;
}
if (usesRecordedAudio(dictationEngine)) {
- // defensive: media recording is effectively always present here.
toast.error("Voice recording isn't available in this browser.");
return;
}
- // Browser Web Speech is missing (e.g. Firefox). Local dictation is the only way to type by
- // voice here, so offer it rather than describing it.
+ // Firefox lacks Web Speech, so offer local dictation.
void offerLocalDictation();
}
-/** Move a browser with no speech service onto local dictation. Already downloaded means one
- * switch; otherwise the same confirmation the mic raises, which flips the engine only if
- * accepted. */
+/** offers local dictation when Web Speech is missing, prompting before a needed download. */
async function offerLocalDictation(): Promise<void> {
- const { sttModel, setDictationEngine } = useVoiceSettingsStore.getState();
+ const { sttModel, sttGgufVariant, setDictationEngine } =
+ useVoiceSettingsStore.getState();
+ const ggufVariant = sttModelVariant(sttModel, sttGgufVariant);
try {
const status = await fetchSttStatus(undefined, sttModel);
const engine = sttEngineStatusFor(status, sttModel);
- // An engine with no runtime installed cannot load what it downloads, so say what is missing
- // rather than asking for gigabytes first.
+ // avoid offering a large download when the runtime cannot load it.
if (engine && !engine.available) {
toast.error("Local transcription isn't installed on this server.", {
description:
@@ -174,7 +171,10 @@ async function offerLocalDictation(): Promise<void> {
});
return;
}
- if (engine?.downloaded_models.includes(sttModel)) {
+ if (
+ engine?.downloaded_models.includes(sttModel) &&
+ (await sttQuantDownloaded(sttModel, ggufVariant))
+ ) {
setDictationEngine("model");
toast.success("Switched to local transcription.", {
description:
@@ -183,7 +183,7 @@ async function offerLocalDictation(): Promise<void> {
return;
}
} catch {
- // Status is unreachable; the download path reports its own failure.
+ // the download path reports status or quant lookup failures.
}
- requestSttDownload(sttModel, { selectLocalEngine: true });
+ requestSttDownload(sttModel, { selectLocalEngine: true, ggufVariant });
}
diff --git a/studio/frontend/src/features/chat/adapters/studio-model-dictation-adapter.ts b/studio/frontend/src/features/chat/adapters/studio-model-dictation-adapter.ts
index e8ebb954036..e530f5b6e05 100644
--- a/studio/frontend/src/features/chat/adapters/studio-model-dictation-adapter.ts
+++ b/studio/frontend/src/features/chat/adapters/studio-model-dictation-adapter.ts
@@ -7,7 +7,9 @@ import {
isAudioCppFolderId,
} from "../../audio/audio-cpp-catalog";
import { authFetch } from "@/features/auth";
+import { listGgufVariants } from "@/features/hub/inventory/api";
import { hubTokenHeader } from "@/features/hub/lib/hub-token-header";
+import { getHfToken, hfApiToken } from "@/features/hub/stores/hf-token-store";
import { useSettingsDialogStore } from "@/features/settings/stores/settings-dialog-store";
import { requestSttDownload } from "@/features/settings/stores/stt-download-prompt-store";
import {
@@ -18,7 +20,10 @@ import {
isCuratedSttModel,
recordRecentDictation,
resolveModelDictationLanguage,
+ sttListedQuantDownloaded,
+ sttModelVariant,
useVoiceSettingsStore,
+ withSttVariant,
} from "@/features/settings/stores/voice-settings-store";
import type { DictationAdapter } from "@assistant-ui/react";
import { toast } from "sonner";
@@ -29,7 +34,7 @@ import { useExternalProvidersStore } from "../stores/external-providers-store";
import { startDictationLevelMeter } from "./dictation-level";
import { type SegmentRecorder, createAudioRecorder } from "./pcm-recorder";
import { SttModelNotDownloadedError, sttRequestError } from "./stt-errors";
-// Re-exported so the one public entry point for dictation is unchanged.
+// re-export preserves the public dictation entry point
export { SttModelNotDownloadedError } from "./stt-errors";
import {
beginDictationSession,
@@ -122,7 +127,6 @@ async function sttErrorDetail(response: Response): Promise<string> {
return body?.detail ?? body?.error?.message ?? `HTTP ${response.status}`;
}
-/** post recorded audio to the selected STT backend and return its transcript. */
export async function transcribeAudioBlob(
blob: Blob,
options: {
@@ -130,6 +134,8 @@ export async function transcribeAudioBlob(
language?: string;
engine?: SttEngine;
device?: SttDevice;
+ /** package-folder quant; defaults to the saved quant when `model` is also omitted */
+ ggufVariant?: string | null;
providerId?: string;
signal?: AbortSignal;
} = {},
@@ -193,7 +199,19 @@ export async function transcribeAudioBlob(
const language = resolveModelDictationLanguage(model, languageSetting);
const engine = options.engine ?? sttEngineFor(model);
- const params = new URLSearchParams({ model, fast: "true", engine });
+ // cold sidecars resolve bare rows to their default quant, so the pick travels as `row:variant`
+ const variant =
+ options.ggufVariant !== undefined
+ ? options.ggufVariant
+ : options.model === undefined
+ ? sttModelVariant(model, settings.sttGgufVariant)
+ : null;
+ const params = new URLSearchParams({
+ model:
+ engine === "audiocpp" && variant ? withSttVariant(model, variant) : model,
+ fast: "true",
+ engine,
+ });
if (language) params.set("language", language);
params.set("device", options.device ?? settings.sttDevice);
const response = await authFetch(
@@ -222,10 +240,9 @@ export interface SttDownloadStatus {
downloading: boolean;
model: string | null;
error: string | null;
- /** The last download was stopped by the user rather than failing. */
+ /** whether the user stopped the last download */
cancelled?: boolean;
- /** Which model that cancellation applies to. `model` goes null once the worker thread stops,
- * so this is the only way to tell a settled cancellation from an unrelated one. */
+ /** cancellation target retained after the worker clears `model` */
cancelled_model?: string | null;
bytes_total: number | null;
bytes_done: number | null;
@@ -234,6 +251,8 @@ export interface SttDownloadStatus {
export interface SttEngineStatus {
available: boolean;
loaded_model: string | null;
+ /** resident audiocpp quant; absent for other engines */
+ loaded_variant?: string | null;
loading: boolean;
device: string | null;
keep_alive_seconds: number;
@@ -325,8 +344,20 @@ export async function validateSttModel(
}
}
-/** The quant an audiocpp pick names, as the request field; every other engine takes none. A saved
- * dictation key already implies its package, so callers pass nothing for one. */
+/** checks the quant because row status is true for any cached quant; unreadable listings defer to it */
+export async function sttQuantDownloaded(
+ model: string,
+ ggufVariant: string | null | undefined,
+ signal?: AbortSignal,
+): Promise<boolean> {
+ if (!ggufVariant) return true;
+ const listing = await listGgufVariants(model, hfApiToken(getHfToken()), {
+ signal,
+ }).catch(() => null);
+ return !listing || sttListedQuantDownloaded(listing, ggufVariant);
+}
+
+/** sends audiocpp quant picks; saved keys imply a package, and other engines take none */
function sttVariantBody(
engine: SttEngine,
ggufVariant: string | null | undefined,
@@ -337,7 +368,6 @@ function sttVariantBody(
: {};
}
-/** Load a selected model that is already downloaded. */
export function loadSttModel(
model: string,
engine?: SttEngine,
@@ -494,6 +524,9 @@ export class StudioModelDictationAdapter implements DictationAdapter {
const sessionEngine = usesExternalEndpoint
? undefined
: sttEngineFor(sessionModel);
+ const sessionVariant = usesExternalEndpoint
+ ? null
+ : sttModelVariant(sessionModel, settings.sttGgufVariant);
const sessionChatId = resolveDictationChatId(this.chatId);
const speechStartCallbacks = new Set<() => void>();
@@ -550,13 +583,12 @@ export class StudioModelDictationAdapter implements DictationAdapter {
if (reportedTranscriptionError || cancelled || ended) return;
reportedTranscriptionError = true;
console.error("STT transcription error:", error);
- // An undownloaded model is the ordinary first-run state, not a failure. Point at the
- // download; never start it here.
+ // undownloaded models require confirmation; never download from dictation
if (
!usesExternalEndpoint &&
error instanceof SttModelNotDownloadedError
) {
- requestSttDownload(sessionModel);
+ requestSttDownload(sessionModel, { ggufVariant: sessionVariant });
finishSession("cancelled");
return;
}
@@ -570,9 +602,7 @@ export class StudioModelDictationAdapter implements DictationAdapter {
onClick: () => useSettingsDialogStore.getState().openDialog("voice"),
},
});
- // The preload runs cache-only, so nothing it reports is transient: a missing runtime or a
- // load refused for training means no segment of this session can be transcribed. End it
- // rather than let the user keep speaking into a recorder whose audio is already lost.
+ // cache-only preload failures make the session unusable, so stop before more audio is lost
if (stage === "preload") finishSession("cancelled");
};
@@ -622,7 +652,7 @@ export class StudioModelDictationAdapter implements DictationAdapter {
}
};
- // Transcribe queued segments one at a time so the backend is never flooded.
+ // serialize segment transcription to avoid flooding the backend
const processQueue = () => {
if (worker || cancelled || ended) return;
const item = queue.shift();
@@ -637,14 +667,14 @@ export class StudioModelDictationAdapter implements DictationAdapter {
model: sessionModel,
language: sessionLanguage,
engine: sessionEngine,
+ ggufVariant: sessionVariant,
providerId: sessionProviderId,
signal: abortController.signal,
});
if (!cancelled) results[item.index] = text;
} catch (error) {
if (!cancelled && !abortController.signal.aborted) {
- // Keep transcribed segments, but never hide that part was lost. Only a lost segment is
- // partial: the model preload shares this reporter and can fail without costing any audio.
+ // only a lost segment is partial; preload failures cost no recorded audio
markDictationFailed();
reportTranscriptionError(error);
}
@@ -850,9 +880,15 @@ export class StudioModelDictationAdapter implements DictationAdapter {
return;
}
if (!usesExternalEndpoint && sessionEngine) {
- // warm the model only after mic access; the backend never downloads here.
- void loadSttModel(sessionModel, sessionEngine).catch(
- (error: unknown) => reportTranscriptionError(error, "preload"),
+ // warm the model only after mic access; the backend never downloads here
+ void loadSttModel(
+ sessionModel,
+ sessionEngine,
+ undefined,
+ undefined,
+ sessionVariant,
+ ).catch((error: unknown) =>
+ reportTranscriptionError(error, "preload"),
);
}
stopLevelMeter = startDictationLevelMeter(stream, (rawRms, now) => {
diff --git a/studio/frontend/src/features/chat/hooks/use-chat-audio-upload.ts b/studio/frontend/src/features/chat/hooks/use-chat-audio-upload.ts
index e95d151d9d9..93ab2c5b5f5 100644
--- a/studio/frontend/src/features/chat/hooks/use-chat-audio-upload.ts
+++ b/studio/frontend/src/features/chat/hooks/use-chat-audio-upload.ts
@@ -10,6 +10,7 @@ import {
applyDictationDictionary,
recordRecentDictation,
requestSttDownload,
+ sttModelVariant,
useSettingsDialogStore,
useVoiceSettingsStore,
} from "@/features/settings";
@@ -30,6 +31,7 @@ import {
fetchSttStatus,
sttEngineFor,
sttEngineStatusFor,
+ sttQuantDownloaded,
transcribeAudioBlob,
} from "../adapters/studio-model-dictation-adapter";
import { resolveDictationChatId } from "../adapters/studio-web-speech-dictation-adapter";
@@ -60,6 +62,8 @@ export type ChatAudioUploadReadiness =
interface ChatAudioUploadSnapshot extends ChatAudioUploadFence {
model: string;
engine: SttEngine;
+ /** quant of a package folder model, pinned with it */
+ ggufVariant: string | null;
language: string;
device: "auto" | "cpu";
chatId: string | undefined;
@@ -89,6 +93,7 @@ export function useChatAudioUpload({
}: UseChatAudioUploadOptions) {
const t = useT();
const model = useVoiceSettingsStore((state) => state.sttModel);
+ const ggufVariant = useVoiceSettingsStore((state) => state.sttGgufVariant);
const language = useVoiceSettingsStore((state) => state.dictationLanguage);
const device = useVoiceSettingsStore((state) => state.sttDevice);
const [dialogOpen, setDialogOpen] = useState(false);
@@ -165,17 +170,19 @@ export function useChatAudioUpload({
return retry
? {
model: retry.snapshot.model,
+ ggufVariant: retry.snapshot.ggufVariant,
language: retry.snapshot.language,
device: retry.snapshot.device,
engine: retry.snapshot.engine,
}
: {
model: model.trim(),
+ ggufVariant: sttModelVariant(model.trim(), ggufVariant),
language,
device,
engine: sttEngineFor(model.trim()),
};
- }, [device, language, model]);
+ }, [device, ggufVariant, language, model]);
const refreshReadiness = useCallback(
async (silent = false) => {
@@ -197,6 +204,12 @@ export function useChatAudioUpload({
}
try {
const status = await fetchSttStatus(undefined, targetModel, signal);
+ // status tracks rows; a pinned quant is ready only when its own files are
+ const quantDownloaded = await sttQuantDownloaded(
+ targetModel,
+ target.ggufVariant,
+ signal,
+ );
if (
!queue.isCurrent(attempt) ||
ownerRef.current !== ownerAtStart ||
@@ -229,9 +242,11 @@ export function useChatAudioUpload({
return;
}
setReadiness({
- state: engineStatus.downloaded_models.includes(targetModel)
- ? "ready"
- : "missing",
+ state:
+ engineStatus.downloaded_models.includes(targetModel) &&
+ quantDownloaded
+ ? "ready"
+ : "missing",
model: targetModel,
});
} catch {
@@ -285,7 +300,9 @@ export function useChatAudioUpload({
}
if (readiness.model !== target.model || readiness.state !== "ready") {
if (readiness.model === target.model && readiness.state === "missing") {
- requestSttDownload(target.model);
+ requestSttDownload(target.model, {
+ ggufVariant: target.ggufVariant,
+ });
} else if (readiness.state === "error") {
void refreshReadiness();
}
@@ -299,6 +316,7 @@ export function useChatAudioUpload({
authSessionEpoch: getAuthSessionEpoch(),
model: target.model,
engine: target.engine,
+ ggufVariant: target.ggufVariant,
language: target.language,
device: target.device,
chatId: resolveDictationChatId(chatId),
@@ -345,6 +363,7 @@ export function useChatAudioUpload({
transcribeAudioBlob(file, {
model: source.model,
engine: source.engine,
+ ggufVariant: source.ggufVariant,
language: source.language,
device: source.device,
signal: controller.signal,
@@ -371,7 +390,9 @@ export function useChatAudioUpload({
if (controller.signal.aborted) return;
if (error instanceof SttModelNotDownloadedError) {
setReadiness({ state: "missing", model: source.model });
- requestSttDownload(source.model);
+ requestSttDownload(source.model, {
+ ggufVariant: source.ggufVariant,
+ });
}
failure =
error instanceof Error && error.message
@@ -450,7 +471,9 @@ export function useChatAudioUpload({
readiness.model === failed.snapshot.model &&
readiness.state === "missing"
) {
- requestSttDownload(failed.snapshot.model);
+ requestSttDownload(failed.snapshot.model, {
+ ggufVariant: failed.snapshot.ggufVariant,
+ });
} else {
void refreshReadiness();
}
diff --git a/studio/frontend/src/features/settings/components/stt-download-prompt.tsx b/studio/frontend/src/features/settings/components/stt-download-prompt.tsx
index 4a6b9019b88..b13aab555ac 100644
--- a/studio/frontend/src/features/settings/components/stt-download-prompt.tsx
+++ b/studio/frontend/src/features/settings/components/stt-download-prompt.tsx
@@ -12,12 +12,13 @@ import {
AlertDialogMedia,
AlertDialogTitle,
} from "@/components/ui/alert-dialog";
+import { audioCppSizeLabel } from "@/features/audio/audio-cpp-catalog";
import { startSttDownload } from "@/features/chat";
-import { hfApiToken, useHfTokenStore } from "@/features/hub";
+import { hfApiToken, listGgufVariants, useHfTokenStore } from "@/features/hub";
import { useT } from "@/i18n";
import { MicIcon } from "@/lib/mic-icon";
import { toast } from "@/lib/toast";
-import type { ReactNode } from "react";
+import { type ReactNode, useEffect, useState } from "react";
import { trackSttDownload } from "../lib/stt-download-mirror";
import {
type SttDownloadRequest,
@@ -29,7 +30,7 @@ import {
useVoiceSettingsStore,
} from "../stores/voice-settings-store";
-/** Emphasise the model name in translated copy; plain text if absent. */
+/** emphasise the model name in translated copy; use plain text if absent */
function highlightModel(text: string, model: string): ReactNode {
const at = model ? text.indexOf(model) : -1;
if (at === -1) return text;
@@ -42,25 +43,58 @@ function highlightModel(text: string, model: string): ReactNode {
);
}
-/**
- * App-level confirmation for a dictation model download. Mounted once so the
- * mic can raise it before Voice settings is ever opened.
- */
+/** app-level confirmation mounted once so the mic can raise it before Voice settings opens */
export function SttDownloadPrompt() {
const t = useT();
const pending = useSttDownloadPromptStore((s) => s.pending);
const dismiss = useSttDownloadPromptStore((s) => s.dismiss);
const hfToken = useHfTokenStore((state) => state.token);
+ const pendingModel = pending?.model ?? null;
+ const variant = pending?.ggufVariant ?? null;
+ // a package folder has no curated size, so read its quant's from the cached listing
+ const [listedSize, setListedSize] = useState<{
+ model: string;
+ variant: string | null;
+ size: string;
+ } | null>(null);
+ useEffect(() => {
+ if (!pendingModel || sttModelSize(pendingModel)) return;
+ let cancelled = false;
+ listGgufVariants(pendingModel, hfApiToken(hfToken))
+ .then((listing) => {
+ const row = listing.variants.find(
+ (candidate) =>
+ candidate.quant === (variant ?? listing.default_variant),
+ );
+ if (cancelled || !row) return;
+ setListedSize({
+ model: pendingModel,
+ variant,
+ size: audioCppSizeLabel(row.download_size_bytes ?? row.size_bytes),
+ });
+ })
+ .catch(() => {});
+ return () => {
+ cancelled = true;
+ };
+ }, [pendingModel, variant, hfToken]);
const confirm = async (request: SttDownloadRequest) => {
- // Only on accept: a cancel must not leave the engine changed.
+ // change the engine only on accept because cancel must leave it unchanged
if (request.selectLocalEngine) {
useVoiceSettingsStore.getState().setDictationEngine("model");
}
try {
- await startSttDownload(request.model, hfApiToken(hfToken));
- // Progress goes to the shared download panel; the model loads itself when it lands.
- trackSttDownload(request.model);
+ await startSttDownload(
+ request.model,
+ hfApiToken(hfToken),
+ undefined,
+ request.ggufVariant,
+ );
+ // progress uses the shared download panel; the model loads when the download finishes
+ trackSttDownload(request.model, {
+ ggufVariant: request.ggufVariant ?? null,
+ });
} catch (error) {
toast.error(t("settings.voice.dictation.sttDownloadFailed"), {
description: error instanceof Error ? error.message : undefined,
@@ -68,8 +102,12 @@ export function SttDownloadPrompt() {
}
};
- const pendingModel = pending?.model ?? null;
- const size = pendingModel ? sttModelSize(pendingModel) : "";
+ const size = pendingModel
+ ? sttModelSize(pendingModel) ||
+ (listedSize?.model === pendingModel && listedSize.variant === variant
+ ? listedSize.size
+ : "")
+ : "";
return (
<AlertDialog
open={pending !== null}
@@ -79,8 +117,7 @@ export function SttDownloadPrompt() {
>
<AlertDialogContent>
<AlertDialogHeader>
- {/* The glyph fills its viewBox, unlike the padded hugeicons the
- default circle is sized for, so both come down together. */}
+ {/* the glyph fills its viewBox, unlike the padded hugeicons, so size both together */}
<AlertDialogMedia className="size-12">
<MicIcon className="text-muted-foreground size-5" />
</AlertDialogMedia>
diff --git a/studio/frontend/src/features/settings/index.ts b/studio/frontend/src/features/settings/index.ts
index c135eea2e6f..267a2830d70 100644
--- a/studio/frontend/src/features/settings/index.ts
+++ b/studio/frontend/src/features/settings/index.ts
@@ -117,7 +117,9 @@ export {
applyDictationDictionary,
recordRecentDictation,
sttModelName,
+ sttModelVariant,
useVoiceSettingsStore,
+ withSttVariant,
} from "./stores/voice-settings-store";
export type { Palette, ResolvedTheme, Theme } from "./stores/theme-store";
diff --git a/studio/frontend/src/features/settings/lib/stt-download-mirror.ts b/studio/frontend/src/features/settings/lib/stt-download-mirror.ts
index 41d5750c5e5..290ed5eb6c8 100644
--- a/studio/frontend/src/features/settings/lib/stt-download-mirror.ts
+++ b/studio/frontend/src/features/settings/lib/stt-download-mirror.ts
@@ -1,6 +1,7 @@
// 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 { isAudioCppFolderId } from "@/features/audio/audio-cpp-catalog";
import {
type SttEngine,
cancelSttDownload,
@@ -11,6 +12,7 @@ import {
} from "@/features/chat";
import {
finishExternalJob,
+ invalidateGgufVariantsCache,
startExternalJob,
updateExternalJob,
} from "@/features/hub";
@@ -20,6 +22,7 @@ import {
type SttModel,
getSttModelRepo,
sttModelName,
+ sttModelVariant,
useVoiceSettingsStore,
} from "../stores/voice-settings-store";
import { SttDownloadTrackers } from "./stt-download-trackers";
@@ -37,6 +40,8 @@ const START_GRACE_MS = 8_000;
const trackers = new SttDownloadTrackers();
const warmSelectedVoiceModelOnComplete = new Map<string, boolean>();
+// The quant each tracked download fetches, when its starter knew it.
+const trackedVariants = new Map<string, string | null>();
function trackerKey(model: SttModel, engine?: SttEngine): string {
return engine && engine !== "transformers" ? `${engine}:${model}` : model;
@@ -51,9 +56,10 @@ function jobKey(model: SttModel, engine?: SttEngine): string {
async function loadAndAnnounce(
model: SttModel,
engine?: SttEngine,
+ ggufVariant?: string | null,
): Promise<void> {
try {
- await loadSttModel(model, engine);
+ await loadSttModel(model, engine, undefined, undefined, ggufVariant);
toast.success(
translate("settings.voice.dictation.sttModelReady", {
model: sttModelName(model),
@@ -74,21 +80,31 @@ function settle(
): void {
const key = trackerKey(model, engine);
finishExternalJob(jobKey(model, engine), outcome, error);
+ // A cached listing would still call the new quant not downloaded.
+ if (outcome === "complete" && isAudioCppFolderId(model)) {
+ invalidateGgufVariantsCache(model);
+ }
trackers.stop(key);
const shouldWarmVoiceModel =
warmSelectedVoiceModelOnComplete.get(key) ?? true;
warmSelectedVoiceModelOnComplete.delete(key);
- // Only warm what the user is still pointed at. Selecting another model, or
- // leaving local dictation, during the download means this one is not wanted
- // and loading it would undo the unload that switch performed.
- const { sttModel, dictationEngine } = useVoiceSettingsStore.getState();
+ const tracked = trackedVariants.get(key);
+ trackedVariants.delete(key);
+ // Only warm what the user is still pointed at. Selecting another model or
+ // quant, or leaving local dictation, during the download means this one is
+ // not wanted and loading it would undo the unload that switch performed.
+ const { sttModel, sttGgufVariant, dictationEngine } =
+ useVoiceSettingsStore.getState();
+ const variant = sttModelVariant(model, sttGgufVariant);
if (
shouldWarmVoiceModel &&
outcome === "complete" &&
dictationEngine === "model" &&
- sttModel === model
+ sttModel === model &&
+ // An adopted download's quant is unknown; it can only be what an unpinned row runs.
+ (tracked === undefined ? variant === null : tracked === variant)
) {
- void loadAndAnnounce(model, engine);
+ void loadAndAnnounce(model, engine, variant);
}
}
@@ -121,11 +137,9 @@ async function poll(
return;
}
- if (engineStatus?.downloaded_models.includes(model)) {
- settle(model, "complete", undefined, engine);
- return;
- }
- if (download?.cancelled) {
+ // Before the downloaded check: a row lists as downloaded once any quant of it is cached, so a
+ // stopped download of another quant would read as complete. start() clears both flags.
+ if (download?.cancelled && (download.cancelled_model ?? model) === model) {
settle(model, "cancelled", undefined, engine);
return;
}
@@ -133,6 +147,10 @@ async function poll(
settle(model, "error", download.error, engine);
return;
}
+ if (engineStatus?.downloaded_models.includes(model)) {
+ settle(model, "complete", undefined, engine);
+ return;
+ }
if (Date.now() - startedAt > START_GRACE_MS) {
settle(
model,
@@ -162,6 +180,8 @@ export function trackSttDownload(
warmSelectedVoiceModelOnComplete?: boolean;
engine?: SttEngine;
repoId?: string;
+ /** The quant this download fetches, so a quant picked meanwhile is not warmed. */
+ ggufVariant?: string | null;
} = {},
): void {
const resolvedEngine = options.engine ?? sttEngineFor(model);
@@ -177,6 +197,9 @@ export function trackSttDownload(
key,
options.warmSelectedVoiceModelOnComplete ?? true,
);
+ if (options.ggufVariant !== undefined) {
+ trackedVariants.set(key, options.ggufVariant);
+ }
startExternalJob({
key: jobKey(model, resolvedEngine),
repoId: options.repoId ?? getSttModelRepo(model),
diff --git a/studio/frontend/src/features/settings/stores/stt-download-prompt-store.ts b/studio/frontend/src/features/settings/stores/stt-download-prompt-store.ts
index 8196d1b396c..670ddc68c7e 100644
--- a/studio/frontend/src/features/settings/stores/stt-download-prompt-store.ts
+++ b/studio/frontend/src/features/settings/stores/stt-download-prompt-store.ts
@@ -6,19 +6,15 @@ import type { SttModel } from "./voice-settings-store";
export interface SttDownloadRequest {
model: SttModel;
- /** Also switch dictation to local on confirm, for a browser whose speech
- * service cannot work at all. Left alone if the user cancels. */
+ /** pinned quant for a package folder model. */
+ ggufVariant?: string | null;
+ /** switches dictation to local only after confirmation when browser speech is unavailable. */
selectLocalEngine?: boolean;
}
-/**
- * The one pending "download this dictation model?" confirmation.
- *
- * A store, not local state, so the mic can raise the dialog without Voice
- * settings being open.
- */
+/** pending download confirmation shared so the mic can open it without Voice settings mounted. */
interface SttDownloadPromptState {
- /** Request awaiting a yes/no, or null when nothing is asked. */
+ /** request awaiting confirmation, or null when idle. */
pending: SttDownloadRequest | null;
requestDownload: (request: SttDownloadRequest) => void;
dismiss: () => void;
diff --git a/studio/frontend/src/features/settings/stores/stt-model-catalog.ts b/studio/frontend/src/features/settings/stores/stt-model-catalog.ts
index 4c2061aa6f6..3c4363ea5da 100644
--- a/studio/frontend/src/features/settings/stores/stt-model-catalog.ts
+++ b/studio/frontend/src/features/settings/stores/stt-model-catalog.ts
@@ -166,6 +166,59 @@ export function sttModelName(model: SttModel): string {
);
}
+/** package folder quant; "" uses the resident or default, while saved keys encode their own. */
+export function sttModelVariant(
+ model: SttModel,
+ variant: string,
+): string | null {
+ return variant && isAudioCppFolderId(model) ? variant : null;
+}
+
+/** folds a quant into the audio runtime model id as `row:variant`. */
+export function withSttVariant(model: SttModel, variant: string): string {
+ const quant = sttModelVariant(model, variant);
+ return quant ? `${model}:${quant}` : model;
+}
+
+/** picks the pinned, loaded, cached, or default quant; resolves loose cache keys like `Q8_0`. */
+export function sttShownVariant(
+ pinned: string | null,
+ loaded: string | null,
+ listing: {
+ default_variant: string | null;
+ variants: readonly { quant: string; downloaded?: boolean }[];
+ } | null,
+): string | null {
+ if (pinned) return pinned;
+ if (!listing) return loaded;
+ if (loaded) {
+ if (listing.variants.some((variant) => variant.quant === loaded)) {
+ return loaded;
+ }
+ const scoped = listing.variants.filter(
+ (variant) => variant.downloaded && variant.quant.endsWith(`/${loaded}`),
+ );
+ if (scoped.length === 1) return scoped[0].quant;
+ }
+ return (
+ listing.variants.find((variant) => variant.downloaded)?.quant ??
+ listing.default_variant
+ );
+}
+
+/** Whether a listing leaves the pinned quant possibly on disk. Only the row with that exact key can
+ * say no: a cache-only (offline) listing keys cached files by what tells them apart ("ctc/F16" for
+ * "v3-ctc/F16"), so a key it leaves out may still be cached; the backend matches it on load. */
+export function sttListedQuantDownloaded(
+ listing: { variants: readonly { quant: string; downloaded?: boolean }[] },
+ pinned: string,
+): boolean {
+ return (
+ listing.variants.find((variant) => variant.quant === pinned)?.downloaded !==
+ false
+ );
+}
+
export function sttModelSize(model: SttModel): string {
return STT_MODEL_SIZES[model as DefaultSttModel] ?? "";
}
diff --git a/studio/frontend/src/features/settings/stores/voice-settings-store.ts b/studio/frontend/src/features/settings/stores/voice-settings-store.ts
index 4a3921ce499..8e3a8bd3ce0 100644
--- a/studio/frontend/src/features/settings/stores/voice-settings-store.ts
+++ b/studio/frontend/src/features/settings/stores/voice-settings-store.ts
@@ -103,23 +103,23 @@ export function isCuratedSttModel(model: SttModel): boolean {
}
export interface VoiceSettingsState {
- /** Input device for dictation. "default" = system default microphone. */
+ /** input device; "default" selects the system microphone. */
micDeviceId: string;
setMicDeviceId: (value: string) => void;
- /**
- * "browser": Web Speech API. "model": local transcription; the model decides
- * the backend (whisper.cpp for curated GGML, Transformers for custom repos).
- */
+ /** "browser" uses Web Speech API; "model" chooses the local backend from the model. */
dictationEngine: DictationEngine;
setDictationEngine: (value: DictationEngine) => void;
- /** STT model to use when dictationEngine is "model". */
+ /** STT model used when dictationEngine is "model". */
sttModel: SttModel;
setSttModel: (value: SttModel) => void;
- /** "cpu" holds the dictation model in system RAM instead of the GPU. Sent
- * with every load and transcribe, so a change applies on the next load. */
+ /** quant for a package folder `sttModel`; "" uses its resident or default quant. */
+ sttGgufVariant: string;
+ setSttGgufVariant: (value: string) => void;
+
+ /** "cpu" keeps the model in system RAM; loads and transcriptions apply the current value. */
sttDevice: SttDevice;
setSttDevice: (value: SttDevice) => void;
@@ -128,11 +128,11 @@ export interface VoiceSettingsState {
sttProviderModel: string;
setSttProviderModel: (value: string) => void;
- /** bcp 47 tag for speech recognition, or "auto" for engine-specific detection. */
+ /** BCP 47 tag, or "auto" for engine-specific language detection. */
dictationLanguage: string;
setDictationLanguage: (value: string) => void;
- /** Exact spellings applied to matching transcript words and phrases. */
+ /** exact spellings applied to matching transcript words and phrases. */
dictionary: string[];
addDictionaryEntry: (value: string) => void;
updateDictionaryEntry: (index: number, value: string) => void;
@@ -234,17 +234,22 @@ export const useVoiceSettingsStore = create<VoiceSettingsState>()(
sttModel: DEFAULT_STT_MODEL,
setSttModel: (value) =>
set((state) => {
- const sttModel = normalizeSttModel(value);
- return {
- sttModel: isSttModelLanguageCompatible(
- sttModel,
- state.dictationLanguage,
- )
- ? sttModel
- : DEFAULT_STT_MODEL,
- };
+ const normalized = normalizeSttModel(value);
+ const sttModel = isSttModelLanguageCompatible(
+ normalized,
+ state.dictationLanguage,
+ )
+ ? normalized
+ : DEFAULT_STT_MODEL;
+ // a quant belongs to the model it was selected for.
+ return sttModel === state.sttModel
+ ? { sttModel }
+ : { sttModel, sttGgufVariant: "" };
}),
+ sttGgufVariant: "",
+ setSttGgufVariant: (sttGgufVariant) => set({ sttGgufVariant }),
+
sttDevice: DEFAULT_STT_DEVICE,
setSttDevice: (value) => set({ sttDevice: normalizeSttDevice(value) }),
@@ -255,15 +260,15 @@ export const useVoiceSettingsStore = create<VoiceSettingsState>()(
dictationLanguage: "auto",
setDictationLanguage: (dictationLanguage) =>
- set((state) => ({
- dictationLanguage,
- sttModel: isSttModelLanguageCompatible(
- state.sttModel,
- dictationLanguage,
- )
- ? state.sttModel
- : DEFAULT_STT_MODEL,
- })),
+ set((state) =>
+ isSttModelLanguageCompatible(state.sttModel, dictationLanguage)
+ ? { dictationLanguage }
+ : {
+ dictationLanguage,
+ sttModel: DEFAULT_STT_MODEL,
+ sttGgufVariant: "",
+ },
+ ),
dictionary: [],
addDictionaryEntry: (value) =>
@@ -388,6 +393,10 @@ export const useVoiceSettingsStore = create<VoiceSettingsState>()(
micDeviceId: asString(saved?.micDeviceId, "default"),
dictationEngine,
sttModel,
+ sttGgufVariant:
+ sttModel === savedSttModel
+ ? asString(saved?.sttGgufVariant, "")
+ : "",
sttDevice: normalizeSttDevice(saved?.sttDevice),
sttProviderId: asString(saved?.sttProviderId, ""),
sttProviderModel: asString(saved?.sttProviderModel, ""),
diff --git a/studio/frontend/src/features/settings/tabs/voice-tab.tsx b/studio/frontend/src/features/settings/tabs/voice-tab.tsx
index 0a1a617baa7..8375c9ad50b 100644
--- a/studio/frontend/src/features/settings/tabs/voice-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/voice-tab.tsx
@@ -44,7 +44,10 @@ import {
validateSttModel,
} from "@/features/chat";
import {
+ type GgufVariantsResponse,
+ ggufVariantDisplayLabel,
hfApiToken,
+ invalidateGgufVariantsCache,
listGgufVariants,
useHfTokenStore,
useHubModelSearch,
@@ -87,6 +90,9 @@ import {
isSttModelLanguageCompatible,
sttModelName,
sttModelSize,
+ sttListedQuantDownloaded,
+ sttModelVariant,
+ sttShownVariant,
type TtsEngine,
useVoiceSettingsStore,
} from "../stores/voice-settings-store";
@@ -478,6 +484,10 @@ export function VoiceTab() {
const setDictationEngine = useVoiceSettingsStore((s) => s.setDictationEngine);
const sttModel = useVoiceSettingsStore((s) => s.sttModel);
const setSttModel = useVoiceSettingsStore((s) => s.setSttModel);
+ const sttGgufVariant = useVoiceSettingsStore((s) => s.sttGgufVariant);
+ const setSttGgufVariant = useVoiceSettingsStore((s) => s.setSttGgufVariant);
+ // The quant a pinned package folder row runs; null leaves the row's resident or default one.
+ const sttVariant = sttModelVariant(sttModel, sttGgufVariant);
// Named apart from the `sttDevice` state below, which the sidecar reports back.
const sttDevicePreference = useVoiceSettingsStore((s) => s.sttDevice);
const setSttDevicePreference = useVoiceSettingsStore((s) => s.setSttDevice);
@@ -629,31 +639,85 @@ export function VoiceTab() {
// so one throttled timer or bursty poll set the displayed speed outright.
// Model whose download this tab watched; completion auto-loads it.
const watchedDownloadRef = useRef<string | null>(null);
+ // The quant this tab started downloading; an adopted transfer's quant is unknown.
+ const startedDownloadRef = useRef<{
+ model: string;
+ variant: string | null;
+ } | null>(null);
+ // Quant resident for the selected row, so an unpinned Select shows what dictation would run.
+ const [sttLoadedVariant, setSttLoadedVariant] = useState<string | null>(null);
+ // The selected package folder's quants, from the same cached listing the Transcribe picker reads.
+ const [sttVariantListing, setSttVariantListing] = useState<{
+ model: string;
+ listing: GgufVariantsResponse;
+ } | null>(null);
+ const [sttListingNonce, setSttListingNonce] = useState(0);
// Selecting a model (or finishing its download) loads it without a Load
// click. A model that is not downloaded fails quietly and stays on demand.
- const autoLoadSttModel = useCallback(async (model: string) => {
- setSttPhase("loading");
- try {
- await loadSttModel(model);
- } catch {
- // Not downloaded (or the engine is busy): the status poll resets the phase
- // and the user still sees the Download button.
- } finally {
- setStatusNonce((nonce) => nonce + 1);
- }
- }, []);
+ const autoLoadSttModel = useCallback(
+ async (model: string, variant: string | null = null) => {
+ setSttPhase("loading");
+ try {
+ await loadSttModel(model, undefined, undefined, undefined, variant);
+ } catch {
+ // Not downloaded (or the engine is busy): the status poll resets the phase
+ // and the user still sees the Download button.
+ } finally {
+ setStatusNonce((nonce) => nonce + 1);
+ }
+ },
+ [],
+ );
const sttRepoId = getSttModelRepo(sttModel);
const hfToken = useHfTokenStore((state) => state.token);
+ // biome-ignore lint/correctness/useExhaustiveDependencies: sttListingNonce is the refetch trigger once a quant lands.
+ useEffect(() => {
+ if (!isLocalEngine || !isAudioCppFolderId(sttModel)) return;
+ let cancelled = false;
+ listGgufVariants(sttModel, hfApiToken(hfToken))
+ .then((listing) => {
+ if (!cancelled) setSttVariantListing({ model: sttModel, listing });
+ })
+ .catch(() => {});
+ return () => {
+ cancelled = true;
+ };
+ }, [isLocalEngine, sttModel, hfToken, sttListingNonce]);
+ const sttVariants =
+ sttVariantListing?.model === sttModel
+ ? sttVariantListing.listing.variants
+ : [];
+ const shownSttVariant = sttShownVariant(
+ sttVariant,
+ sttLoadedVariant,
+ sttVariantListing?.model === sttModel ? sttVariantListing.listing : null,
+ );
+ const shownSttLabel = ggufVariantDisplayLabel(
+ sttVariants.find((variant) => variant.quant === shownSttVariant) ?? {
+ quant: shownSttVariant ?? "",
+ },
+ );
const [sttDownloadStarting, setSttDownloadStarting] = useState(false);
const [sttDownloadAvailability, setSttDownloadAvailability] = useState<{
repoId: string;
state: SttDownloadAvailability;
}>({ repoId: "", state: "checking" });
- const effectiveSttDownloadAvailability =
+ const rowSttDownloadAvailability =
sttDownloadAvailability.repoId === sttRepoId
? sttDownloadAvailability.state
: "checking";
+ // Status knows rows, not quants: a pinned quant of a row with another one cached is missing, or
+ // failed when its download did, as a row's would.
+ const effectiveSttDownloadAvailability =
+ rowSttDownloadAvailability === "downloaded" &&
+ sttVariant &&
+ sttVariantListing?.model === sttModel &&
+ !sttListedQuantDownloaded(sttVariantListing.listing, sttVariant)
+ ? sttDownload?.error
+ ? "error"
+ : "missing"
+ : rowSttDownloadAvailability;
useEffect(() => {
if (!isLocalEngine || !modelSttSupported) {
return;
@@ -685,6 +749,14 @@ export function VoiceTab() {
setSttPhase("unavailable");
return;
}
+ const rowLoaded = engineStatus.loaded_model === sttModel;
+ setSttLoadedVariant(
+ rowLoaded ? (engineStatus.loaded_variant ?? null) : null,
+ );
+ // A pinned quant is ready only when that quant is the resident one.
+ const loaded =
+ rowLoaded &&
+ (!sttVariant || engineStatus.loaded_variant === sttVariant);
const download = engineStatus.download;
setSttDownload(download);
setSttDownloadAvailability({
@@ -709,13 +781,26 @@ export function VoiceTab() {
} else {
const finished = watchedDownloadRef.current;
watchedDownloadRef.current = null;
+ const started = startedDownloadRef.current;
+ startedDownloadRef.current = null;
+ // Only the quant that landed may load: a quant picked meanwhile is not on disk.
+ const landedPinned =
+ started?.model === finished
+ ? started.variant === sttVariant
+ : sttVariant === null;
+ if (finished === sttModel && isAudioCppFolderId(sttModel)) {
+ // The cached listing still has this quant as not downloaded.
+ invalidateGgufVariantsCache(sttModel);
+ setSttListingNonce((nonce) => nonce + 1);
+ }
if (
finished === sttModel &&
engineStatus.downloaded_models.includes(sttModel) &&
- engineStatus.loaded_model !== sttModel
+ !loaded &&
+ landedPinned
) {
// The download this tab watched just finished; load the model.
- void autoLoadSttModel(sttModel);
+ void autoLoadSttModel(sttModel, sttVariant);
return;
}
}
@@ -726,7 +811,7 @@ export function VoiceTab() {
}, 600);
return;
}
- if (engineStatus.loaded_model === sttModel && !engineStatus.loading) {
+ if (loaded && !engineStatus.loading) {
setSttDevice(engineStatus.device);
setSttPhase("ready");
window.setTimeout(
@@ -754,6 +839,7 @@ export function VoiceTab() {
}, [
isLocalEngine,
sttModel,
+ sttVariant,
sttRepoId,
modelSttSupported,
statusNonce,
@@ -824,8 +910,14 @@ export function VoiceTab() {
const beginSttDownload = async () => {
setSttDownloadStarting(true);
try {
- await startSttDownload(sttModel, hfApiToken(hfToken));
- trackSttDownload(sttModel);
+ await startSttDownload(
+ sttModel,
+ hfApiToken(hfToken),
+ undefined,
+ sttVariant,
+ );
+ trackSttDownload(sttModel, { ggufVariant: sttVariant });
+ startedDownloadRef.current = { model: sttModel, variant: sttVariant };
// The status effect only re-polls while it can see a download. Its last read was before this
// one existed, and the on-demand branch schedules nothing, so without a nudge the tab shows
// Download for the whole transfer.
@@ -842,7 +934,7 @@ export function VoiceTab() {
const warmSttModel = async () => {
setSttPhase("loading");
try {
- await loadSttModel(sttModel);
+ await loadSttModel(sttModel, undefined, undefined, undefined, sttVariant);
setStatusNonce((nonce) => nonce + 1);
} catch (error) {
setSttPhase("error");
@@ -1126,7 +1218,11 @@ export function VoiceTab() {
modelSttSupported ? (
<SettingsRow
label={t("settings.voice.dictation.sttModelLabel")}
- description={t("settings.voice.dictation.sttModelDescription")}
+ description={
+ sttVariants.length > 1
+ ? t("settings.voice.dictation.sttQuantDescription")
+ : t("settings.voice.dictation.sttModelDescription")
+ }
// Progress lives in the shared downloads panel; a second bar here
// said the same thing twice.
below={
@@ -1224,7 +1320,7 @@ export function VoiceTab() {
</div>
}
>
- <div className="w-56">
+ <div className="flex w-56 flex-col gap-1.5">
<SttModelPicker
value={sttModel}
language={dictationLanguage}
@@ -1237,6 +1333,61 @@ export function VoiceTab() {
setSttModel(next);
}}
/>
+ {/* A saved key names its own quant, so only package folder rows offer one. */}
+ {sttVariants.length > 1 ? (
+ // Bound to the pin, not the shown quant: picking the quant that runs now
+ // must still save it, or an idle unload drops back to another one.
+ <Select
+ value={sttVariant ?? ""}
+ onValueChange={(next) => {
+ setSttGgufVariant(next);
+ if (next === shownSttVariant) return;
+ void unloadSttModel().catch(() => {});
+ void autoLoadSttModel(sttModel, next);
+ }}
+ >
+ <SelectTrigger
+ data-testid="stt-quant-trigger"
+ aria-label={t("settings.voice.dictation.sttQuantLabel")}
+ className="w-full font-mono text-xs"
+ size="sm"
+ >
+ {/* Unpinned, the value is "" and Radix renders the placeholder. */}
+ <SelectValue placeholder={shownSttLabel}>
+ {shownSttLabel}
+ </SelectValue>
+ </SelectTrigger>
+ <SelectContent align="end">
+ {sttVariants.map((variant) => (
+ <SelectItem
+ key={variant.quant}
+ value={variant.quant}
+ // Size flush right, as in the Agents quant list.
+ className="[&>span:last-child]:w-full [&>span:last-child]:justify-between"
+ >
+ <span className="flex items-center gap-1.5 font-mono text-xs whitespace-nowrap">
+ {/* The slot stays when empty, so the labels share a column. */}
+ {variant.downloaded ? (
+ <span
+ role="img"
+ aria-label={t("picker.onDevice")}
+ className="size-[calc(5px*var(--ui-space-scale,1))] shrink-0 rounded-full bg-status-success"
+ />
+ ) : (
+ <span className="size-[calc(5px*var(--ui-space-scale,1))] shrink-0" />
+ )}
+ {ggufVariantDisplayLabel(variant)}
+ </span>
+ <span className="text-ui-10 tabular-nums whitespace-nowrap text-muted-foreground">
+ {audioCppSizeLabel(
+ variant.download_size_bytes ?? variant.size_bytes,
+ )}
+ </span>
+ </SelectItem>
+ ))}
+ </SelectContent>
+ </Select>
+ ) : null}
</div>
</SettingsRow>
) : (
diff --git a/studio/frontend/src/i18n/locales/ar.ts b/studio/frontend/src/i18n/locales/ar.ts
index 415a22db20c..3e06ef6877d 100644
--- a/studio/frontend/src/i18n/locales/ar.ts
+++ b/studio/frontend/src/i18n/locales/ar.ts
@@ -1290,6 +1290,9 @@ export const ar = {
customModelDescription: "اسم النموذج المرسل إلى /v1/audio/transcriptions.",
sttModelLabel: "نموذج التعرّف على الكلام",
sttModelDescription: "اختر نموذج STT أو ابحث عنه لتشغيله محليًا.",
+ sttQuantLabel: "التكميم",
+ sttQuantDescription:
+ "اختر نموذج STT أو ابحث عنه لتشغيله محليًا، ثم اختر التكميم.",
sttDeviceLabel: "التحميل إلى",
sttDeviceAuto: "وحدة معالجة الرسومات عند توفرها",
sttDeviceCpu: "ذاكرة المعالج",
diff --git a/studio/frontend/src/i18n/locales/de.ts b/studio/frontend/src/i18n/locales/de.ts
index 5700f95beb2..17fc34b41ba 100644
--- a/studio/frontend/src/i18n/locales/de.ts
+++ b/studio/frontend/src/i18n/locales/de.ts
@@ -1296,6 +1296,9 @@ export const de = {
sttModelLabel: "Spracherkennungsmodell",
sttModelDescription:
"Wählen oder suchen Sie ein STT-Modell für die lokale Ausführung.",
+ sttQuantLabel: "Quantisierung",
+ sttQuantDescription:
+ "Wählen oder suchen Sie ein STT-Modell für die lokale Ausführung, dann seine Quantisierung.",
sttDeviceLabel: "Laden in",
sttDeviceAuto: "GPU, wenn verfügbar",
sttDeviceCpu: "CPU-RAM",
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index ef364165212..e00753f17a0 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -1337,6 +1337,9 @@ export const en = {
customModelDescription: "Model name sent to /v1/audio/transcriptions.",
sttModelLabel: "Speech recognition model",
sttModelDescription: "Choose or search a STT model to run locally.",
+ sttQuantLabel: "Quantization",
+ sttQuantDescription:
+ "Choose or search a STT model to run locally, then its quantization.",
sttDeviceLabel: "Load into",
sttDeviceAuto: "GPU when available",
sttDeviceCpu: "CPU RAM",
diff --git a/studio/frontend/src/i18n/locales/es.ts b/studio/frontend/src/i18n/locales/es.ts
index 116fd361031..956accd21a4 100644
--- a/studio/frontend/src/i18n/locales/es.ts
+++ b/studio/frontend/src/i18n/locales/es.ts
@@ -1295,6 +1295,9 @@ export const es = {
sttModelLabel: "Modelo de reconocimiento de voz",
sttModelDescription:
"Elige o busca un modelo STT para ejecutarlo en local.",
+ sttQuantLabel: "Cuantización",
+ sttQuantDescription:
+ "Elige o busca un modelo STT para ejecutarlo en local y luego su cuantización.",
sttDeviceLabel: "Cargar en",
sttDeviceAuto: "GPU cuando esté disponible",
sttDeviceCpu: "RAM de la CPU",
diff --git a/studio/frontend/src/i18n/locales/fr.ts b/studio/frontend/src/i18n/locales/fr.ts
index 7e7792b894f..4fcd556c2c1 100644
--- a/studio/frontend/src/i18n/locales/fr.ts
+++ b/studio/frontend/src/i18n/locales/fr.ts
@@ -1296,6 +1296,9 @@ export const fr = {
sttModelLabel: "Modèle de reconnaissance vocale",
sttModelDescription:
"Choisissez ou recherchez un modèle STT à exécuter en local.",
+ sttQuantLabel: "Quantification",
+ sttQuantDescription:
+ "Choisissez ou recherchez un modèle STT à exécuter en local, puis sa quantification.",
sttDeviceLabel: "Charger dans",
sttDeviceAuto: "GPU si disponible",
sttDeviceCpu: "RAM du CPU",
diff --git a/studio/frontend/src/i18n/locales/he.ts b/studio/frontend/src/i18n/locales/he.ts
index a4b40b0bd2b..96c979d6f58 100644
--- a/studio/frontend/src/i18n/locales/he.ts
+++ b/studio/frontend/src/i18n/locales/he.ts
@@ -1389,6 +1389,9 @@ export const he = {
customModelDescription: "שם המודל שנשלח אל /v1/audio/transcriptions.",
sttModelLabel: "מודל זיהוי דיבור",
sttModelDescription: "בחר או חפש מודל STT להרצה מקומית.",
+ sttQuantLabel: "קוונטיזציה",
+ sttQuantDescription:
+ "בחר או חפש מודל STT להרצה מקומית, ואז את הקוונטיזציה שלו.",
sttDeviceLabel: "טען אל",
sttDeviceAuto: "GPU כאשר זמין",
sttDeviceCpu: "זיכרון RAM של ה-CPU",
diff --git a/studio/frontend/src/i18n/locales/hi.ts b/studio/frontend/src/i18n/locales/hi.ts
index 2492b61532e..9382a0be379 100644
--- a/studio/frontend/src/i18n/locales/hi.ts
+++ b/studio/frontend/src/i18n/locales/hi.ts
@@ -1294,6 +1294,9 @@ export const hi = {
"/v1/audio/transcriptions को भेजा जाने वाला मॉडल नाम।",
sttModelLabel: "स्पीच रिकग्निशन मॉडल",
sttModelDescription: "लोकली चलाने के लिए STT मॉडल चुनें या खोजें।",
+ sttQuantLabel: "क्वांटाइज़ेशन",
+ sttQuantDescription:
+ "लोकली चलाने के लिए STT मॉडल चुनें या खोजें, फिर उसका क्वांटाइज़ेशन चुनें।",
sttDeviceLabel: "इसमें लोड करें",
sttDeviceAuto: "उपलब्ध होने पर GPU",
sttDeviceCpu: "CPU RAM",
diff --git a/studio/frontend/src/i18n/locales/it.ts b/studio/frontend/src/i18n/locales/it.ts
index 9f07f487c99..f5ab5d88979 100644
--- a/studio/frontend/src/i18n/locales/it.ts
+++ b/studio/frontend/src/i18n/locales/it.ts
@@ -1253,6 +1253,9 @@ export const it = {
sttModelLabel: "Modello di riconoscimento vocale",
sttModelDescription:
"Scegli o cerca un modello STT da eseguire in locale.",
+ sttQuantLabel: "Quantizzazione",
+ sttQuantDescription:
+ "Scegli o cerca un modello STT da eseguire in locale, poi la sua quantizzazione.",
sttDeviceLabel: "Carica in",
sttDeviceAuto: "GPU quando disponibile",
sttDeviceCpu: "RAM della CPU",
diff --git a/studio/frontend/src/i18n/locales/ja.ts b/studio/frontend/src/i18n/locales/ja.ts
index be53e905c01..afad5ee5ddc 100644
--- a/studio/frontend/src/i18n/locales/ja.ts
+++ b/studio/frontend/src/i18n/locales/ja.ts
@@ -1294,6 +1294,9 @@ export const ja = {
sttModelLabel: "音声認識モデル",
sttModelDescription:
"ローカルで実行する STT モデルを選択または検索します。",
+ sttQuantLabel: "量子化",
+ sttQuantDescription:
+ "ローカルで実行する STT モデルを選択または検索し、量子化を選びます。",
sttDeviceLabel: "読み込み先",
sttDeviceAuto: "GPU(利用できる場合)",
sttDeviceCpu: "CPU メモリ",
diff --git a/studio/frontend/src/i18n/locales/ko.ts b/studio/frontend/src/i18n/locales/ko.ts
index c191187f654..38debcbfbab 100644
--- a/studio/frontend/src/i18n/locales/ko.ts
+++ b/studio/frontend/src/i18n/locales/ko.ts
@@ -1291,6 +1291,9 @@ export const ko = {
"/v1/audio/transcriptions에 전송할 모델 이름입니다.",
sttModelLabel: "음성 인식 모델",
sttModelDescription: "로컬에서 실행할 STT 모델을 선택하거나 검색하세요.",
+ sttQuantLabel: "양자화",
+ sttQuantDescription:
+ "로컬에서 실행할 STT 모델을 선택하거나 검색한 다음 양자화를 고르세요.",
sttDeviceLabel: "로드 위치",
sttDeviceAuto: "GPU(사용 가능할 때)",
sttDeviceCpu: "CPU RAM",
diff --git a/studio/frontend/src/i18n/locales/pt-br.ts b/studio/frontend/src/i18n/locales/pt-br.ts
index 5a93e96e91e..4e0196b05fc 100644
--- a/studio/frontend/src/i18n/locales/pt-br.ts
+++ b/studio/frontend/src/i18n/locales/pt-br.ts
@@ -1294,6 +1294,9 @@ export const ptBR = {
sttModelLabel: "Modelo de reconhecimento de fala",
sttModelDescription:
"Escolha ou busque um modelo STT para executar localmente.",
+ sttQuantLabel: "Quantização",
+ sttQuantDescription:
+ "Escolha ou busque um modelo STT para executar localmente e depois sua quantização.",
sttDeviceLabel: "Carregar em",
sttDeviceAuto: "GPU quando disponível",
sttDeviceCpu: "RAM da CPU",
diff --git a/studio/frontend/src/i18n/locales/ru.ts b/studio/frontend/src/i18n/locales/ru.ts
index ff799ff289b..a5c1fce8c58 100644
--- a/studio/frontend/src/i18n/locales/ru.ts
+++ b/studio/frontend/src/i18n/locales/ru.ts
@@ -1293,6 +1293,9 @@ export const ru = {
sttModelLabel: "Модель распознавания речи",
sttModelDescription:
"Выберите или найдите модель STT для локального запуска.",
+ sttQuantLabel: "Квантизация",
+ sttQuantDescription:
+ "Выберите или найдите модель STT для локального запуска, затем её квантизацию.",
sttDeviceLabel: "Загружать в",
sttDeviceAuto: "GPU, если доступен",
sttDeviceCpu: "ОЗУ процессора",
diff --git a/studio/frontend/src/i18n/locales/sv.ts b/studio/frontend/src/i18n/locales/sv.ts
index 0194ae1e483..66ba105da1a 100644
--- a/studio/frontend/src/i18n/locales/sv.ts
+++ b/studio/frontend/src/i18n/locales/sv.ts
@@ -1427,6 +1427,9 @@ export const sv = {
sttModelLabel: "Taligenkänningsmodell",
sttModelDescription:
"Välj eller sök efter en STT-modell som ska köras lokalt.",
+ sttQuantLabel: "Kvantisering",
+ sttQuantDescription:
+ "Välj eller sök efter en STT-modell som ska köras lokalt och sedan dess kvantisering.",
sttDeviceLabel: "Läs in i",
sttDeviceAuto: "GPU när tillgänglig",
sttDeviceCpu: "CPU-RAM",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index e35a014a4d4..933c9057b50 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -1288,6 +1288,9 @@ export const zhCN = {
"发送到 /v1/audio/transcriptions 的模型名称。",
sttModelLabel: "语音识别模型",
sttModelDescription: "选择或搜索要在本地运行的 STT 模型。",
+ sttQuantLabel: "量化",
+ sttQuantDescription:
+ "选择或搜索要在本地运行的 STT 模型,再选择其量化。",
sttDeviceLabel: "加载到",
sttDeviceAuto: "有 GPU 时使用 GPU",
sttDeviceCpu: "CPU 内存",
diff --git a/studio/frontend/tests/audio-cpp-catalog.test.ts b/studio/frontend/tests/audio-cpp-catalog.test.ts
index 525eb4f3407..e78760ed4a2 100644
--- a/studio/frontend/tests/audio-cpp-catalog.test.ts
+++ b/studio/frontend/tests/audio-cpp-catalog.test.ts
@@ -73,6 +73,10 @@ import {
STT_PICKER_MODELS,
sttModelName,
sttModelSize,
+ sttModelVariant,
+ sttListedQuantDownloaded,
+ sttShownVariant,
+ withSttVariant,
} from "../src/features/settings/stores/stt-model-catalog.ts";
import {
@@ -523,9 +527,132 @@ test("an Audio-page ASR pick sends its quant; other engines and saved keys send
page.match(/controller\.signal,\s*undefined,\s*ggufVariant,/g)?.length,
2,
);
- // Settings > Voice passes a saved key and no variant.
+ // Settings > Voice sends its saved quant only for package folders.
const voiceTab = readSrc("features/settings/tabs/voice-tab.tsx");
- assert.match(voiceTab, /await startSttDownload\(sttModel, hfApiToken\(hfToken\)\);/);
+ assert.match(
+ voiceTab,
+ /await startSttDownload\(\s*sttModel,\s*hfApiToken\(hfToken\),\s*undefined,\s*sttVariant,\s*\);/,
+ );
+ assert.match(voiceTab, /const sttVariant = sttModelVariant\(sttModel, sttGgufVariant\);/);
+});
+
+const GIGAAM = `${AUDIO_CPP_REPO}/GigaAM-ASR-GGUF`;
+
+test("only a package folder row runs a saved quant; saved keys and other engines ignore it", () => {
+ assert.equal(sttModelVariant(GIGAAM, "v3-ctc/F16"), "v3-ctc/F16");
+ assert.equal(withSttVariant(GIGAAM, "v3-ctc/F16"), `${GIGAAM}:v3-ctc/F16`);
+ // an empty pin keeps the resident quant, or the default when none is resident.
+ assert.equal(sttModelVariant(GIGAAM, ""), null);
+ assert.equal(withSttVariant(GIGAAM, ""), GIGAAM);
+ for (const model of ["audiocpp-moonshine-small", "qwen3-asr-0.6b", "small", "openai/whisper-small"]) {
+ assert.equal(sttModelVariant(model, "Q8_0"), null, model);
+ assert.equal(withSttVariant(model, "Q8_0"), model, model);
+ }
+});
+
+test("the saved quant is cleared with its model and rehydrated only beside it", () => {
+ const store = readSrc("features/settings/stores/voice-settings-store.ts");
+ assert.match(
+ store,
+ /return sttModel === state\.sttModel\s*\?\s*\{ sttModel \}\s*:\s*\{ sttModel, sttGgufVariant: "" \};/,
+ );
+ assert.match(store, /sttModel: DEFAULT_STT_MODEL,\s*sttGgufVariant: "",/);
+ assert.match(
+ store,
+ /sttGgufVariant:\s*sttModel === savedSttModel\s*\?\s*asString\(saved\?\.sttGgufVariant, ""\)\s*:\s*"",/,
+ );
+});
+
+test("every Settings dictation path carries the saved quant", () => {
+ const adapter = readSrc("features/chat/adapters/studio-model-dictation-adapter.ts");
+ // a raw transcription folds the quant into the id so a cold sidecar does not load the default.
+ assert.match(
+ adapter,
+ /model:\s*engine === "audiocpp" && variant \? withSttVariant\(model, variant\) : model,/,
+ );
+ assert.match(adapter, /sttModelVariant\(model, settings\.sttGgufVariant\)/);
+ // the session pins the quant for preload and every segment.
+ assert.match(adapter, /const sessionVariant = usesExternalEndpoint\s*\?\s*null\s*:\s*sttModelVariant\(sessionModel, settings\.sttGgufVariant\);/);
+ assert.match(adapter, /ggufVariant: sessionVariant,/);
+ assert.match(adapter, /sessionEngine,\s*undefined,\s*undefined,\s*sessionVariant,/);
+
+ const voiceTab = readSrc("features/settings/tabs/voice-tab.tsx");
+ assert.match(voiceTab, /await loadSttModel\(sttModel, undefined, undefined, undefined, sttVariant\);/);
+ assert.match(voiceTab, /void autoLoadSttModel\(sttModel, sttVariant\);/);
+ // ready requires the pinned quant to be resident, not merely its row.
+ assert.match(voiceTab, /\(!sttVariant \|\| engineStatus\.loaded_variant === sttVariant\)/);
+ // status is row-level; the listing shows whether the pinned quant is cached.
+ assert.match(
+ voiceTab,
+ /!sttListedQuantDownloaded\(sttVariantListing\.listing, sttVariant\)\s*\?\s*sttDownload\?\.error\s*\?\s*"error"\s*:\s*"missing"/,
+ );
+ // binding Select to the pin preserves an active quant selection.
+ assert.match(voiceTab, /value=\{sttVariant \?\? ""\}/);
+ assert.match(voiceTab, /setSttGgufVariant\(next\);\s*if \(next === shownSttVariant\) return;/);
+ assert.match(voiceTab, /<SelectValue placeholder=\{shownSttLabel\}>/);
+
+ // the prompt fetches the requester's pin, not the current Voice setting.
+ const prompt = readSrc("features/settings/components/stt-download-prompt.tsx");
+ assert.match(prompt, /hfApiToken\(hfToken\),\s*undefined,\s*request\.ggufVariant,/);
+ assert.match(prompt, /const variant = pending\?\.ggufVariant \?\? null;/);
+ assert.doesNotMatch(prompt, /sttGgufVariant/);
+ assert.match(prompt, /candidate\.quant === \(variant \?\? listing\.default_variant\)/);
+ assert.match(adapter, /requestSttDownload\(sessionModel, \{ ggufVariant: sessionVariant \}\)/);
+ const offer = readSrc("features/chat/adapters/studio-dictation-adapter.tsx");
+ assert.match(offer, /const ggufVariant = sttModelVariant\(sttModel, sttGgufVariant\);/);
+ assert.match(offer, /requestSttDownload\(sttModel, \{ selectLocalEngine: true, ggufVariant \}\)/);
+
+ const mirror = readSrc("features/settings/lib/stt-download-mirror.ts");
+ // a stopped second-quant download must not look complete because its row is already listed.
+ const cancelledAt = mirror.indexOf("download?.cancelled && (download.cancelled_model ?? model) === model");
+ const erroredAt = mirror.indexOf("if (download?.error)");
+ const downloadedAt = mirror.indexOf("if (engineStatus?.downloaded_models.includes(model))");
+ assert.ok(cancelledAt > 0 && cancelledAt < erroredAt && erroredAt < downloadedAt);
+ // do not warm a new pin after another quant lands; adopted downloads warm only unpinned rows.
+ assert.match(
+ mirror,
+ /\(tracked === undefined \? variant === null : tracked === variant\)/,
+ );
+ assert.match(voiceTab, /trackSttDownload\(sttModel, \{ ggufVariant: sttVariant \}\)/);
+ // The tab's own watcher loads a landed download only when it is still the pinned quant.
+ assert.match(voiceTab, /startedDownloadRef\.current = \{ model: sttModel, variant: sttVariant \};/);
+ assert.match(
+ voiceTab,
+ /started\?\.model === finished\s*\?\s*started\.variant === sttVariant\s*:\s*sttVariant === null;/,
+ );
+ assert.match(voiceTab, /!loaded &&\s*landedPinned\s*\)/);
+ assert.match(prompt, /trackSttDownload\(request\.model, \{\s*ggufVariant: request\.ggufVariant \?\? null,\s*\}\)/);
+ assert.match(mirror, /sttModelVariant\(model, sttGgufVariant\)/);
+ assert.match(mirror, /outcome === "complete" && isAudioCppFolderId\(model\)[\s\S]*invalidateGgufVariantsCache\(model\)/);
+
+ const upload = readSrc("features/chat/hooks/use-chat-audio-upload.ts");
+ assert.match(upload, /ggufVariant: sttModelVariant\(model\.trim\(\), ggufVariant\),/);
+ assert.match(upload, /ggufVariant: source\.ggufVariant,/);
+ for (const holder of ["target", "source", "failed\\.snapshot"]) {
+ assert.match(upload, new RegExp(`requestSttDownload\\([^)]*\\{\\s*ggufVariant: ${holder}\\.ggufVariant,`));
+ }
+ // status lists a row after any quant is cached; a pin needs its own listing row.
+ assert.match(adapter, /return !listing \|\| sttListedQuantDownloaded\(listing, ggufVariant\);/);
+ // upload readiness and the no-service fallback require the pinned quant on disk.
+ assert.match(upload, /sttQuantDownloaded\(\s*targetModel,\s*target\.ggufVariant,\s*signal,\s*\)/);
+ assert.match(
+ upload,
+ /engineStatus\.downloaded_models\.includes\(targetModel\) &&\s*quantDownloaded\s*\?\s*"ready"/,
+ );
+ assert.match(
+ offer,
+ /engine\?\.downloaded_models\.includes\(sttModel\) &&\s*\(await sttQuantDownloaded\(sttModel, ggufVariant\)\)/,
+ );
+
+ const reference = readSrc("features/audio/hooks/use-reference-transcribe.ts");
+ assert.match(reference, /withSttVariant\(model, voice\.sttGgufVariant\)/);
+});
+
+test("the quant Select appears for package folders with more than one quant", () => {
+ const voiceTab = readSrc("features/settings/tabs/voice-tab.tsx");
+ assert.match(voiceTab, /\{sttVariants\.length > 1 \? \([\s\S]{0,200}?<Select\s/);
+ assert.match(voiceTab, /if \(!isLocalEngine \|\| !isAudioCppFolderId\(sttModel\)\) return;/);
+ assert.match(voiceTab, /aria-label=\{t\("settings\.voice\.dictation\.sttQuantLabel"\)\}/);
});
test("sttEngineFor sends saved keys, folders and GGUF repos to the audiocpp engine", () => {
@@ -761,3 +888,57 @@ test("the picked STT quant is remembered with the repo across a restart", () =>
assert.match(page, /lastSttRepo && lastSttVariant \? \[\[lastSttRepo\.toLowerCase\(\), lastSttVariant\]\]/);
assert.match(page, /setLastSttVariant\(sttGgufVariants\.current\.get\(repo\.toLowerCase\(\)\) \?\? ""\)/);
});
+
+test("the quant Select shows the quant a dictation would actually run", () => {
+ const listing = {
+ default_variant: "multilingual-ctc/F16",
+ variants: [
+ { quant: "multilingual-ctc/F16", downloaded: false },
+ { quant: "v3-ctc/F16", downloaded: true },
+ { quant: "v3-rnnt/F16", downloaded: true },
+ ],
+ };
+ // a pin overrides the resident quant.
+ assert.equal(sttShownVariant("v3-rnnt/F16", "v3-ctc/F16", listing), "v3-rnnt/F16");
+ // without a pin, the resident quant runs.
+ assert.equal(sttShownVariant(null, "v3-rnnt/F16", listing), "v3-rnnt/F16");
+ // without a resident quant, an offline bare load uses the first cached quant.
+ assert.equal(sttShownVariant(null, null, listing), "v3-ctc/F16");
+ // without a cache, use the default.
+ const cold = { ...listing, variants: listing.variants.map((v) => ({ ...v, downloaded: false })) };
+ assert.equal(sttShownVariant(null, null, cold), "multilingual-ctc/F16");
+ // a cache-only load reports a loose key that maps to its sole cached quant.
+ const moonshine = {
+ default_variant: "tiny/Q8_0",
+ variants: [
+ { quant: "tiny/Q8_0", downloaded: false },
+ { quant: "small/Q8_0", downloaded: true },
+ ],
+ };
+ assert.equal(sttShownVariant(null, "Q8_0", moonshine), "small/Q8_0");
+ // without a listing, use the resident quant.
+ assert.equal(sttShownVariant(null, "Q8_0", null), "Q8_0");
+ const voiceTab = readSrc("features/settings/tabs/voice-tab.tsx");
+ assert.match(voiceTab, /const shownSttVariant = sttShownVariant\(\s*sttVariant,\s*sttLoadedVariant,/);
+});
+
+test("only the pinned quant's own listing row can say it is not on disk", () => {
+ const online = {
+ variants: [
+ { quant: "multilingual-ctc/F16", downloaded: true },
+ { quant: "v3-ctc/F16", downloaded: false },
+ ],
+ };
+ assert.equal(sttListedQuantDownloaded(online, "multilingual-ctc/F16"), true);
+ assert.equal(sttListedQuantDownloaded(online, "v3-ctc/F16"), false);
+ // A cache-only listing keys cached files by what tells them apart, so a pinned key it leaves out
+ // may still be cached (the backend matches it on load): defer instead of calling it missing.
+ const offline = {
+ variants: [
+ { quant: "ctc/F16", downloaded: true },
+ { quant: "rnnt/F16", downloaded: true },
+ ],
+ };
+ assert.equal(sttListedQuantDownloaded(offline, "v3-ctc/F16"), true);
+ assert.equal(sttListedQuantDownloaded({ variants: [] }, "small/Q8_0"), true);
+});