fix(cpu): simplify device mapping to prevent meta tensor issues

Refactored model initialization by using explicit device_map instead of complex meta tensor handling. Added HF_HOME environment configuration and removed to_empty() method calls, resulting in cleaner and more reliable CPU device placement.

app.py

@@ -13,41 +13,38 @@ model_name = 'Marqo/marqo-fashionSigLIP'
 # Force CPU usage to avoid device mapping issues
 device = torch.device('cpu')
 
-# Handle meta tensor initialization properly
+# Set environment variables to prevent meta tensor issues
+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
 try:
-    # Load model with empty weights initialization to avoid meta tensor issues
+    # Load model with specific configuration to prevent meta tensor creation
     model = AutoModel.from_pretrained(
         model_name,
         trust_remote_code=True,
-        torch_dtype=torch.float32
+        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
     )
     
-    # 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
+    # Fallback method - load with explicit CPU device mapping
     try:
         model = AutoModel.from_pretrained(
             model_name,
-            trust_remote_code=True
+            trust_remote_code=True,
+            torch_dtype=torch.float32,
+            device_map="cpu"  # Force CPU mapping
         )
-        # 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
+        # Last resort - load with basic configuration and manual device placement
         model = AutoModel.from_pretrained(
             model_name,
             trust_remote_code=True,
-            low_cpu_mem_usage=False  # Disable to avoid accelerate issues
+            torch_dtype=torch.float32
         )
         model = model.to(device)