fix: revert to eager model loading — L4 GPU loads models fast, deferred loading caused inference errors

app.py

@@ -120,31 +120,15 @@ from musetalk.utils.preprocessing import get_landmark_and_bbox, read_imgs, coord
 from musetalk.utils.blending import get_image
 from musetalk.utils.utils import load_all_model
 
-audio_processor = None
-vae = None
-unet = None
-pe = None
-device = None
-timesteps = None
-
-
-def ensure_models_loaded():
-    global audio_processor, vae, unet, pe, device, timesteps
-
-    if audio_processor is not None:
-        return
-
-    print("Loading models (first inference)...")
-    audio_processor, vae, unet, pe = load_all_model()
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    timesteps = torch.tensor([0], device=device)
-    print("Models loaded.")
+print("Loading models...")
+audio_processor, vae, unet, pe = load_all_model()
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+timesteps = torch.tensor([0], device=device)
+print(f"Models loaded on {device}.")
 
 
 @torch.no_grad()
 def inference(audio_path, video_path, bbox_shift, progress=gr.Progress(track_tqdm=True)):
-    ensure_models_loaded()
-
     args_dict = {
         "result_dir": "./results",
         "fps": 25,