Fix ZeroGPU CUDA support for V3 demo

app.py

@@ -6,6 +6,8 @@ import gradio as gr
 import spaces
 
 MODEL = None
+# ZeroGPU supports CUDA placement at module load time via CUDA emulation.
+TARGET_DEVICE = "cuda"
 
 DEFAULT_CONFIG = {
     "audio": 'https://storage.googleapis.com/chatterbox-demo-samples/mtl_prompts/en_f1.flac',
@@ -40,9 +42,15 @@ def default_text_for_ui():
 def get_or_load_model():
     global MODEL
     if MODEL is None:
-        print("Model not loaded, initializing on CPU...")
-        MODEL = ChatterboxTTS.from_pretrained("cpu")
-        print("Model loaded.")
+        print(f"Model not loaded, initializing on {TARGET_DEVICE}...")
+        try:
+            MODEL = ChatterboxTTS.from_pretrained(TARGET_DEVICE)
+        except Exception as exc:
+            if TARGET_DEVICE != "cuda":
+                raise
+            print(f"CUDA model initialization failed, falling back to CPU: {exc}")
+            MODEL = ChatterboxTTS.from_pretrained("cpu")
+        print(f"Model loaded on {MODEL.device}.")
     return MODEL
 
 
@@ -66,9 +74,8 @@ def generate_tts_audio(
     cfgw_input: float = 0.5,
 ):
     """Generate speech from text with optional reference audio styling."""
-    device = "cuda" if torch.cuda.is_available() else "cpu"
     current_model = get_or_load_model()
-    current_model.to(device)
+    device = current_model.device
     if seed_num_input != 0:
         set_seed(int(seed_num_input), device)
     chosen_prompt = audio_prompt_path_input or default_audio_for_ui()
@@ -86,6 +93,9 @@ def generate_tts_audio(
     return (current_model.sr, wav.squeeze(0).cpu().numpy())
 
 
+get_or_load_model()
+
+
 with gr.Blocks() as demo:
     gr.Markdown(
         """

requirements.txt

@@ -1,9 +1,12 @@
+--extra-index-url https://download.pytorch.org/whl/cu128
+torch==2.8.0
+torchaudio==2.8.0
+
 gradio
 numpy==1.26.0
 resampy==0.4.3
 librosa==0.10.0
 s3tokenizer
-torchaudio<2.8
 
 transformers==4.46.3
 diffusers==0.29.0
@@ -11,4 +14,4 @@ omegaconf==2.3.0
 resemble-perth==1.0.1
 silero-vad==5.1.2
 conformer==0.3.2
-safetensors
+safetensors