feat: monkey-patch open_clip to avoid meta tensors and load model on CPU

- Patch open_clip.factory.create_model to force CPU device and fp32, preventing meta tensor creation
- Simplify model loading: trust_remote_code + torch_dtype=float32 then .to(device); streamline fallback
- Set HF_HOME to /tmp/hf_cache and update exception messages

app.py

@@ -17,36 +17,38 @@ device = torch.device('cpu')
 import os
 os.environ['HF_HOME'] = '/tmp/hf_cache'  # Use temporary cache directory
 
-# Handle meta tensor initialization properly by controlling device mapping at the source
+# Monkey patch open_clip to prevent meta tensor issues
+try:
+    import open_clip
+    original_create_model = open_clip.factory.create_model
+    
+    def patched_create_model(*args, **kwargs):
+        # Force device to CPU to prevent meta tensor creation
+        kwargs['device'] = 'cpu'
+        kwargs['precision'] = 'fp32'  # Force float32 precision
+        return original_create_model(*args, **kwargs)
+    
+    open_clip.factory.create_model = patched_create_model
+except Exception as e:
+    print(f"Could not patch open_clip: {e}")
+
+# Load model with patched open_clip to prevent meta tensor issues
 try:
-    # Load model with specific configuration to prevent meta tensor creation
     model = AutoModel.from_pretrained(
         model_name,
         trust_remote_code=True,
-        torch_dtype=torch.float32,
-        device_map={"": "cpu"},  # Explicitly map all modules to CPU to avoid meta tensors
-        low_cpu_mem_usage=False   # Disable low CPU mem usage to avoid accelerate issues
+        torch_dtype=torch.float32
     )
+    model = model.to(device)
     
 except Exception as e:
-    print(f"Primary loading method failed: {e}")
-    # Fallback method - load with explicit CPU device mapping
-    try:
-        model = AutoModel.from_pretrained(
-            model_name,
-            trust_remote_code=True,
-            torch_dtype=torch.float32,
-            device_map="cpu"  # Force CPU mapping
-        )
-    except Exception as e2:
-        print(f"Fallback method also failed: {e2}")
-        # Last resort - load with basic configuration and manual device placement
-        model = AutoModel.from_pretrained(
-            model_name,
-            trust_remote_code=True,
-            torch_dtype=torch.float32
-        )
-        model = model.to(device)
+    print(f"Model loading failed: {e}")
+    # Fallback - try loading with different configuration
+    model = AutoModel.from_pretrained(
+        model_name,
+        trust_remote_code=True
+    )
+    model = model.to(device)
 
 processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)