Update app.py

app.py

@@ -35,33 +35,42 @@ def load_model_from_pickle(pickle_path="best_model.pkl"):
         if not os.path.exists(pickle_path):
             return f"❌ Model file not found: {pickle_path}\n\nPlease ensure best_model.pkl is uploaded to the HuggingFace Space."
         
-        # ULTIMATE FIX: Monkey-patch torch.cuda.is_available during unpickling
-        # This tricks PyTorch into thinking CUDA is available, then we immediately move to CPU
-        import torch.cuda
-        import torch.serialization
+        # NUCLEAR OPTION: Mock the entire CUDA module to prevent driver checks
+        import sys
+        import types
         
-        original_is_available = torch.cuda.is_available
-        original_device_count = torch.cuda.device_count
+        # Create a fake CUDA module that claims everything is available but does nothing
+        fake_cuda = types.ModuleType('cuda')
+        fake_cuda.is_available = lambda: True
+        fake_cuda.device_count = lambda: 1
+        fake_cuda.current_device = lambda: 0
+        fake_cuda.get_device_name = lambda x: "CPU (mocked as CUDA)"
+        fake_cuda.set_device = lambda x: None
+        fake_cuda.device = lambda x: types.SimpleNamespace(__enter__=lambda: None, __exit__=lambda *args: None)
+        fake_cuda.init = lambda: None
+        fake_cuda.is_initialized = lambda: True
+        fake_cuda._initialization_lock = types.SimpleNamespace(__enter__=lambda: None, __exit__=lambda *args: None)
         
-        def fake_is_available():
-            return True
-        
-        def fake_device_count():
-            return 1
+        # Save original
+        original_cuda = torch.cuda
         
         try:
-            # Temporarily make PyTorch think CUDA is available
-            torch.cuda.is_available = fake_is_available
-            torch.cuda.device_count = fake_device_count
+            # Replace torch.cuda with our mock during loading
+            torch.cuda = fake_cuda
+            sys.modules['torch'].cuda = fake_cuda
             
-            # Load the pickle (PyTorch will accept CUDA device references now)
+            # Load with aggressive CPU mapping
             with open(pickle_path, 'rb') as f:
-                model_package = torch.load(f, map_location=lambda storage, loc: storage.cpu())
+                model_package = torch.load(
+                    f, 
+                    map_location='cpu',
+                    weights_only=False
+                )
         
         finally:
-            # Always restore original functions
-            torch.cuda.is_available = original_is_available
-            torch.cuda.device_count = original_device_count
+            # Always restore original CUDA module
+            torch.cuda = original_cuda
+            sys.modules['torch'].cuda = original_cuda
         
         # Success! Model loaded with one of the strategies above
         # Handle a few common package shapes.