Directly load model on CUDA + display progress

app.py

@@ -8,9 +8,9 @@ from chatterbox.tts_turbo import ChatterboxTurboTTS
 
 # --- 1. FORCE CPU FOR GLOBAL LOADING ---
 # ZeroGPU forbids CUDA during startup. We only move to CUDA inside the decorated function.
-DEVICE = "cpu" 
+DEVICE = "cuda" 
 
-MODEL = None
+MODEL = ChatterboxTurboTTS.from_pretrained(DEVICE)
 
 EVENT_TAGS = [
     "[clear throat]", "[sigh]", "[shush]", "[cough]", "[groan]",
@@ -75,13 +75,6 @@ def set_seed(seed: int):
     random.seed(seed)
     np.random.seed(seed)
 
-
-def load_model():
-    global MODEL
-    print(f"Loading Chatterbox-Turbo on {DEVICE}...")
-    MODEL = ChatterboxTurboTTS.from_pretrained(DEVICE)
-    return MODEL
-
 @spaces.GPU
 def generate(
         text,
@@ -92,16 +85,9 @@ def generate(
         top_p,
         top_k,
         repetition_penalty,
-        norm_loudness
+        norm_loudness,
+        progress=gr.Progress(track_tqdm=True),
 ):
-    global MODEL
-    # Reload if the worker lost the global state
-    if MODEL is None:
-        MODEL = ChatterboxTurboTTS.from_pretrained("cpu")
-
-    # --- MOVE TO GPU HERE ---
-    MODEL.to("cuda")
-
     if seed_num != 0:
         set_seed(int(seed_num))
 
@@ -162,9 +148,6 @@ with gr.Blocks(title="Chatterbox Turbo") as demo:
                 min_p = gr.Slider(0.00, 1.00, step=0.01, label="Min P (Set to 0 to disable)", value=0.00)
                 norm_loudness = gr.Checkbox(value=True, label="Normalize Loudness (-27 LUFS)")
 
-    # Load on startup (CPU)
-    demo.load(fn=load_model, inputs=[], outputs=[])
-
     run_btn.click(
         fn=generate,
         inputs=[