Update vision_analyzer.py

vision_analyzer.py

@@ -81,22 +81,36 @@ import sys
 device = "cpu" # Explicitly ensure CPU usage
 
 try:
-    # Load the model, forcing the 'eager' (CPU-compatible) attention implementation
+try:
+    # Modified model loading code
     vision_language_model_large = Blip2ForConditionalGeneration.from_pretrained(
         "Salesforce/blip2-opt-2.7b",
+        torch_dtype=torch.float32,  # Explicitly set dtype
+        device_map="cpu",          # Force CPU
+        low_cpu_mem_usage=True,    # Optimize for low memory
         trust_remote_code=True
-    ).to(device).eval()
+    ).eval()  # Don't call .to(device) since device_map handles it
+    
     vision_language_processor_large = AutoProcessor.from_pretrained(
         "Salesforce/blip2-opt-2.7b",
         trust_remote_code=True
     )
-    print("BLIP-2 base model and processor loaded successfully on CPU.")
+    print("BLIP-2 model loaded successfully on CPU.")
 except Exception as e:
-    print(f"Error loading BLIP-2 model on CPU: {e}")
-    print("Please ensure you have enough RAM and a compatible PyTorch version.")
-    vision_language_model_large = None
-    vision_language_processor_large = None
-
+    print(f"Error loading BLIP-2 model: {e}")
+    # Fallback to smaller model if available
+    try:
+        vision_language_model_large = Blip2ForConditionalGeneration.from_pretrained(
+            "Salesforce/blip2-opt-1.3b",  # Smaller variant
+            torch_dtype=torch.float32,
+            device_map="cpu",
+            low_cpu_mem_usage=True,
+            trust_remote_code=True
+        ).eval()
+        print("Loaded smaller 1.3B parameter model instead.")
+    except:
+        vision_language_model_large = None
+        vision_language_processor_large = None
     
 def log_message(message: str):
     """Log messages with timestamp"""