Update app.py

app.py

@@ -4,6 +4,8 @@ import numpy as np
 import torch
 from src.chatterbox.mtl_tts import ChatterboxMultilingualTTS, SUPPORTED_LANGUAGES
 import gradio as gr
+import os
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
 
 
 
@@ -131,29 +133,27 @@ def get_supported_languages_display() -> str:
 """
 
 
-os.environ["CUDA_VISIBLE_DEVICES"] = ""
-
 DEVICE = "cpu"
 MODEL = None
 
 def get_or_load_model():
-    """Loads ChatterboxMultilingualTTS strictly on CPU."""
+    """Loads ChatterboxMultilingualTTS strictly on CPU (library-safe)."""
     global MODEL
 
     if MODEL is None:
         print("Model not loaded, initializing on CPU only...")
 
         try:
-            MODEL = ChatterboxMultilingualTTS.from_pretrained(
-                DEVICE,
-                map_location="cpu"
-            )
+            # ❌ Do NOT pass map_location (not supported)
+            MODEL = ChatterboxMultilingualTTS.from_pretrained(DEVICE)
 
-            # Absolute safety
+            # ✅ Force CPU after load
             if hasattr(MODEL, "to"):
                 MODEL = MODEL.to("cpu")
 
             MODEL.eval()
+
+            # Disable grads for CPU inference
             for p in MODEL.parameters():
                 p.requires_grad = False
 
@@ -165,7 +165,6 @@ def get_or_load_model():
 
     return MODEL
 
-
 # Load at startup
 try:
     get_or_load_model()
@@ -176,11 +175,11 @@ except Exception as e:
     )
 
 def set_seed(seed: int):
-    """CPU-only reproducibility."""
     torch.manual_seed(seed)
     random.seed(seed)
     np.random.seed(seed)
 
+
     
 def resolve_audio_prompt(language_id: str, provided_path: str | None) -> str | None:
     """