fix: improve model loading with meta tensor handling and CPU fallbacks

- Use torch_dtype to avoid meta tensor issues
- Check for to_empty method for device placement
- Add fallback methods for robust CPU loading

app.py

@@ -7,17 +7,49 @@ from io import BytesIO
 
 fashion_items = ['top', 'trousers', 'jumper']
 
-# Load model and processor with CPU device to avoid meta tensor issues
+# Load model and processor with proper meta tensor handling
 model_name = 'Marqo/marqo-fashionSigLIP'
 
 # Force CPU usage to avoid device mapping issues
 device = torch.device('cpu')
 
-# Simple loading approach - let model handle device placement
-model = AutoModel.from_pretrained(
-    model_name,
-    trust_remote_code=True
-).to(device)
+# Handle meta tensor initialization properly
+try:
+    # Load model with empty weights initialization to avoid meta tensor issues
+    model = AutoModel.from_pretrained(
+        model_name,
+        trust_remote_code=True,
+        torch_dtype=torch.float32
+    )
+    
+    # Check if model has the to_empty method and use it for meta tensor initialization
+    if hasattr(model, 'model') and hasattr(model.model, 'to_empty'):
+        model.model.to_empty(device=device)
+    elif hasattr(model, 'to_empty'):
+        model.to_empty(device=device)
+    else:
+        # Fallback to regular to() method
+        model = model.to(device)
+        
+except Exception as e:
+    print(f"Primary loading method failed: {e}")
+    # Fallback method - load with minimal configuration
+    try:
+        model = AutoModel.from_pretrained(
+            model_name,
+            trust_remote_code=True
+        )
+        # Move to CPU after loading
+        model = model.to(device)
+    except Exception as e2:
+        print(f"Fallback method also failed: {e2}")
+        # Last resort - try loading with low CPU memory usage
+        model = AutoModel.from_pretrained(
+            model_name,
+            trust_remote_code=True,
+            low_cpu_mem_usage=False  # Disable to avoid accelerate issues
+        )
+        model = model.to(device)
 
 processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)