Qwen TTS : serveur GPU distant (Unraid) réglé dans les paramètres

- qwen-tts : GPU NVIDIA détecté automatiquement (bfloat16), image CUDA 12.8
  (RTX 50xx comprises) via TORCH_VARIANT=cu128, docker-compose.gpu.yml
  dédié (clé obligatoire, choix de la carte) et guide d'installation Unraid.
- Paramètres → « Qwen TTS (voix locale) » : adresse + clé du service, test
  de connexion (modèle et carte graphique affichés), stockés en base
  (app_config.qwen_tts_url / qwen_tts_api_key) et transmis à ffmpeg-api à
  chaque voix ; QWEN_TTS_URL reste un défaut.
- Le service qwen-tts CPU du docker-compose principal devient optionnel
  (profil qwen-local).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Upu97wMmsRkBoj6iVM4rH6

server/services/ffmpeg.ts

@@ -30,6 +30,7 @@ export interface ReelRenderOptions {
     ttsEngine?: TtsEngine;
     ttsStyle?: TtsStyle;
     geminiApiKey?: string;
+    qwen?: QwenTarget;
     fontSize?: number;
     musicVolume?: number;
     drawText?: boolean;
@@ -71,9 +72,17 @@ export interface TimedWord {
     end: number;
 }
 
+/** Service Qwen TTS (local ou serveur GPU) réglé dans les paramètres. */
+export interface QwenTarget {
+    url: string;
+    apiKey?: string;
+}
+
 export interface VoiceCatalog {
     qwen: VoiceOption[];
     qwen_available: boolean;
+    /** Modèle et carte graphique du service Qwen, null s'il ne répond pas. */
+    qwen_status: { model?: string; device?: string; busy?: boolean } | null;
     gemini: VoiceOption[];
     edge: VoiceOption[];
 }
@@ -149,6 +158,8 @@ export class FFmpegService {
             tts_engine: options.ttsEngine,
             tts_style: options.ttsStyle,
             gemini_api_key: options.geminiApiKey,
+            qwen_tts_url: options.qwen?.url,
+            qwen_tts_api_key: options.qwen?.apiKey,
             font_size: options.fontSize ?? 64,
             music_volume: options.musicVolume ?? 0.25,
             draw_text: options.drawText ?? true,
@@ -221,6 +232,8 @@ export class FFmpegService {
                 tts_engine: options.ttsEngine,
                 tts_style: options.ttsStyle,
                 gemini_api_key: options.geminiApiKey,
+                qwen_tts_url: options.qwen?.url,
+                qwen_tts_api_key: options.qwen?.apiKey,
                 music_volume: options.musicVolume ?? 0.25,
                 draw_text: options.drawText ?? true,
                 stabilize: options.stabilize ?? false,
@@ -294,15 +307,18 @@ export class FFmpegService {
     }
 
     /** Voix proposées par moteur ; `qwen_available` indique si le service local répond. */
-    async listVoices(): Promise<VoiceCatalog> {
-        const response = await this.call('/voices', { method: 'GET', timeoutMs: 15_000 });
+    async listVoices(qwen?: QwenTarget): Promise<VoiceCatalog> {
+        const headers: Record<string, string> = {};
+        if (qwen?.url) headers['X-Qwen-Url'] = qwen.url;
+        if (qwen?.apiKey) headers['X-Qwen-Key'] = qwen.apiKey;
+        const response = await this.call('/voices', { method: 'GET', headers, timeoutMs: 30_000 });
         return await response.json() as VoiceCatalog;
     }
 
     /** Génère la voix seule (aperçu), avec le minutage de chaque mot. */
     async previewVoice(
         text: string,
-        options: { voice?: string; engine?: TtsEngine; style?: TtsStyle; geminiApiKey?: string } = {},
+        options: { voice?: string; engine?: TtsEngine; style?: TtsStyle; geminiApiKey?: string; qwen?: QwenTarget } = {},
     ): Promise<VoicePreview> {
         const response = await this.call('/preview-tts', {
             method: 'POST',
@@ -313,6 +329,8 @@ export class FFmpegService {
                 tts_engine: options.engine,
                 tts_style: options.style,
                 gemini_api_key: options.geminiApiKey,
+                qwen_tts_url: options.qwen?.url,
+                qwen_tts_api_key: options.qwen?.apiKey,
             }),
             timeoutMs: TTS_TIMEOUT_MS,
         });

server/services/reels/assets.ts

@@ -3,6 +3,7 @@
  * magasin), jusqu'ici recopiée dans chaque route.
  */
 
+import type { QwenTarget } from "../ffmpeg";
 import { storage } from "../../storage";
 import { buildMinioUrl, resolveInternalUrl } from "../minio";
 
@@ -49,6 +50,31 @@ export async function resolveGeminiApiKey(ttsEngine?: string): Promise<string |
   return appConfig?.geminiApiKey ?? process.env.GEMINI_API_KEY ?? undefined;
 }
 
+/** Service Qwen TTS réglé dans les paramètres (sinon ffmpeg-api prend QWEN_TTS_URL). */
+export async function resolveQwenTarget(ttsEngine?: string): Promise<QwenTarget | undefined> {
+  if (ttsEngine !== "qwen") return undefined;
+  return qwenTargetFromConfig(await storage.getAppConfig());
+}
+
+/** Adresse http(s) du service Qwen saisie dans les paramètres, sans « / » final ; null si invalide. */
+export function normalizeQwenUrl(value: unknown): string | null {
+  if (typeof value !== "string" || !value.trim()) return null;
+  try {
+    const url = new URL(value.trim());
+    if (url.protocol !== "http:" && url.protocol !== "https:") return null;
+    return url.toString().replace(/\/+$/, "");
+  } catch {
+    return null;
+  }
+}
+
+export function qwenTargetFromConfig(
+  appConfig: { qwenTtsUrl?: string | null; qwenTtsApiKey?: string | null } | undefined,
+): QwenTarget | undefined {
+  if (!appConfig?.qwenTtsUrl) return undefined;
+  return { url: appConfig.qwenTtsUrl, apiKey: appConfig.qwenTtsApiKey ?? undefined };
+}
+
 /** Nom affiché dans l'outro : celui de la page choisie, sinon de la première page de l'utilisateur. */
 export async function resolveStoreName(
   userId: string,

server/services/reels/imagesPipeline.ts

@@ -11,7 +11,7 @@ import { imagesReelParamsSchema, type ImagesReelResult } from "@shared/reel";
 import { ffmpegService } from "../ffmpeg";
 import { generateVideoThumbnail } from "../thumbnail";
 import type { JobContext } from "./queue";
-import { resolveGeminiApiKey, resolveLogoPath } from "./assets";
+import { resolveGeminiApiKey, resolveLogoPath, resolveQwenTarget } from "./assets";
 import { REMOTION_TEMP_DIR, localHttpUrl, tempFileUrl, toDataUrl } from "./remotionAssets";
 import { renderReelComposition } from "./remotionRenderer";
 
@@ -42,6 +42,7 @@ export async function runImagesReelJob({ job, progress }: JobContext): Promise<I
         engine: params.ttsEngine,
         style: params.ttsStyle,
         geminiApiKey: await resolveGeminiApiKey(params.ttsEngine),
+        qwen: await resolveQwenTarget(params.ttsEngine),
       });
       for (const warning of voice.warnings) console.warn(`⚠️ [Reels] ${warning}`);
 

server/services/reels/videoPipeline.ts

@@ -17,7 +17,7 @@ import { storage } from "../../storage";
 import { ffmpegService, type ReelRenderOptions } from "../ffmpeg";
 import { resolveInternalUrl } from "../minio";
 import type { JobContext } from "./queue";
-import { resolveGeminiApiKey, resolveLogoPath, resolveMusicUrl } from "./assets";
+import { resolveGeminiApiKey, resolveLogoPath, resolveMusicUrl, resolveQwenTarget } from "./assets";
 import { describePublishFailure, publishReelToPages, storeRenderedVideo } from "./publish";
 import { REMOTION_TEMP_DIR, tempFileUrl, toDataUrl } from "./remotionAssets";
 import { renderReelComposition } from "./remotionRenderer";
@@ -39,10 +39,11 @@ export async function runVideoReelJob({ job, progress }: JobContext) {
     throw new Error("Vidéo source introuvable ou invalide");
   }
 
-  const [musicUrl, logoPath, geminiApiKey] = await Promise.all([
+  const [musicUrl, logoPath, geminiApiKey, qwen] = await Promise.all([
     resolveMusicUrl(params.musicTrackId, params.musicUrl),
     resolveLogoPath(),
     resolveGeminiApiKey(params.ttsEngine),
+    resolveQwenTarget(params.ttsEngine),
   ]);
 
   const options: ReelRenderOptions = {
@@ -54,6 +55,7 @@ export async function runVideoReelJob({ job, progress }: JobContext) {
     ttsEngine: params.ttsEngine,
     ttsStyle: params.ttsStyle,
     geminiApiKey,
+    qwen,
     fontSize: params.fontSize,
     musicVolume: params.musicVolume,
     drawText: params.drawText,