Update app.py

app.py

@@ -7,9 +7,18 @@ import spaces
 import re
 from typing import List, Tuple
 
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+# Force CPU usage and patch torch.load to handle CUDA tensors on CPU
+DEVICE = "cpu"  # Force CPU since you don't have GPU access
 print(f"🚀 Running on device: {DEVICE}")
 
+# Patch torch.load to automatically map CUDA tensors to CPU
+original_load = torch.load
+def patched_load(f, map_location=None, **kwargs):
+    if map_location is None:
+        map_location = 'cpu'  # Always map to CPU
+    return original_load(f, map_location=map_location, **kwargs)
+torch.load = patched_load
+
 # --- Global Model Initialization ---
 MODEL = None
 
@@ -20,12 +29,43 @@ def get_or_load_model():
     if MODEL is None:
         print("Model not loaded, initializing...")
         try:
-            MODEL = ChatterboxTTS.from_pretrained(DEVICE)
-            if hasattr(MODEL, 'to') and str(MODEL.device) != DEVICE:
-                MODEL.to(DEVICE)
-            print(f"Model loaded successfully. Internal device: {getattr(MODEL, 'device', 'N/A')}")
+            # Try multiple loading strategies for CPU
+            print("Attempting to load model on CPU...")
+            
+            # Strategy 1: Direct CPU loading
+            try:
+                MODEL = ChatterboxTTS.from_pretrained("cpu")
+                print("✅ Model loaded successfully with direct CPU method")
+            except Exception as e1:
+                print(f"Direct CPU loading failed: {e1}")
+                
+                # Strategy 2: Try with explicit map_location if supported
+                try:
+                    MODEL = ChatterboxTTS.from_pretrained(DEVICE, map_location='cpu')
+                    print("✅ Model loaded successfully with map_location method")
+                except Exception as e2:
+                    print(f"map_location method failed: {e2}")
+                    
+                    # Strategy 3: Load with default then move to CPU
+                    try:
+                        MODEL = ChatterboxTTS.from_pretrained()
+                        if hasattr(MODEL, 'to'):
+                            MODEL = MODEL.to('cpu')
+                        print("✅ Model loaded successfully with default then CPU move")
+                    except Exception as e3:
+                        print(f"All loading strategies failed. Last error: {e3}")
+                        raise e3
+            
+            # Ensure model is on CPU
+            if hasattr(MODEL, 'to'):
+                MODEL = MODEL.to('cpu')
+            if hasattr(MODEL, 'device'):
+                print(f"Model device: {MODEL.device}")
+            
+            print(f"Model loaded successfully on CPU")
+            
         except Exception as e:
-            print(f"Error loading model: {e}")
+            print(f"CRITICAL: All model loading attempts failed: {e}")
             raise
     return MODEL
 
@@ -38,9 +78,7 @@ except Exception as e:
 def set_seed(seed: int):
     """Sets the random seed for reproducibility across torch, numpy, and random."""
     torch.manual_seed(seed)
-    if DEVICE == "cuda":
-        torch.cuda.manual_seed(seed)
-        torch.cuda.manual_seed_all(seed)
+    # Remove CUDA seed setting since we're on CPU only
     random.seed(seed)
     np.random.seed(seed)
 
@@ -155,7 +193,7 @@ def concatenate_audio_chunks(audio_chunks: List[Tuple[int, np.ndarray]],
     
     return sample_rate, np.concatenate(combined_audio)
 
-@spaces.GPU
+@spaces.GPU  # This decorator might not work on CPU, but keeping it for compatibility
 def generate_tts_audio_chunked(
     text_input: str,
     audio_prompt_path_input: str = None,
@@ -307,4 +345,4 @@ with gr.Blocks() as demo:
         outputs=[audio_output],
     )
 
-demo.launch(mcp_server=True)
+demo.launch(share=True, mcp_server=True)