Refactor model loading: replace open_clip monkey patch with transformers device mapping for CPU and fp32, add robust fallbacks to prevent meta tensor issues

app.py

@@ -17,38 +17,39 @@ device = torch.device('cpu')
 import os
 os.environ['HF_HOME'] = '/tmp/hf_cache'  # Use temporary cache directory
 
-# Monkey patch open_clip to prevent meta tensor issues
+# Prevent meta tensor creation by controlling model initialization
 try:
-    import open_clip
-    original_create_model = open_clip.factory.create_model
+    # Import transformers and patch the model loading if needed
+    import transformers
     
-    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 tensors
     model = AutoModel.from_pretrained(
         model_name,
         trust_remote_code=True,
-        torch_dtype=torch.float32
+        torch_dtype=torch.float32,
+        low_cpu_mem_usage=False,  # Disable to avoid accelerate issues
+        device_map={"": "cpu"}     # Explicitly map to CPU
     )
-    model = model.to(device)
     
 except Exception as e:
-    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)
+    print(f"Primary loading method failed: {e}")
+    try:
+        # Fallback method - load with explicit device control
+        model = AutoModel.from_pretrained(
+            model_name,
+            trust_remote_code=True,
+            torch_dtype=torch.float32,
+            device_map="cpu"
+        )
+    except Exception as e2:
+        print(f"Fallback method also failed: {e2}")
+        # Last resort - load and manually move to device
+        model = AutoModel.from_pretrained(
+            model_name,
+            trust_remote_code=True,
+            torch_dtype=torch.float32
+        )
+        model = model.to(device)
 
 processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)