Skip HiFi-GAN training entirely, use pre-trained model + FAISS index

train_model crashes in ZeroGPU's sandboxed worker (runpy/mp.Process
patterns are incompatible). The retrieval-based approach using the
pre-trained generator + user's FAISS index is the practical solution.
GPU is still used for fast feature extraction (F0 + HuBERT embeddings).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

pipeline/training.py

@@ -426,18 +426,14 @@ def full_training_pipeline(
     # Build FAISS index (fast, CPU-friendly)
     index_path = build_index(model_name)
 
-    if has_gpu:
-        # With GPU: do full training for best quality
-        if progress_callback:
-            progress_callback(0.65, "GPU détecté. Entraînement du modèle...")
-        train_model(model_name, sample_rate, epochs, batch_size)
-        pth_path = find_trained_model(model_name)
-    else:
-        # CPU only: use pre-trained model (skip hours-long training)
-        if progress_callback:
-            progress_callback(0.75, "Mode CPU : utilisation du modèle pré-entraîné...")
-        logger.info("CPU mode: skipping HiFi-GAN training, using pre-trained model.")
-        pth_path = find_pretrained_model(sample_rate)
+    # Use pre-trained RVC generator model + user's FAISS index for voice conversion.
+    # Full HiFi-GAN training is skipped because:
+    # - On CPU: takes hours (impractical)
+    # - On ZeroGPU: worker sandbox doesn't support runpy/multiprocessing patterns
+    # The FAISS index captures the user's voice characteristics for retrieval-based conversion.
+    if progress_callback:
+        progress_callback(0.75, "Finalisation du modèle vocal...")
+    pth_path = find_pretrained_model(sample_rate)
 
     if not pth_path:
         raise RuntimeError("Aucun modèle trouvé. Vérifiez que les modèles pré-entraînés sont téléchargés.")