Update app.py

app.py

@@ -27,30 +27,48 @@ MAX_GPU_MEMORY = "40GiB"  # A100 memory allocation
 # --- Model Loading ---
 @GPU(memory=40) # ****** THIS DECORATOR IS ESSENTIAL FOR SPACES STARTUP ******
 def load_model():
-    """Load the LLM model optimized for A100 GPU."""
-    print(f"Attempting to load model: {MODEL_ID}")
+    """Load the LLM model optimized for A100 GPU using 4-bit quantization."""
+    print(f"Attempting to load model: {MODEL_ID} with 4-bit quantization")
     try:
-        # Configure quantization
-        quantization_config = BitsAndBytesConfig(load_in_8bit=True)
+        # Configure quantization for 4-bit
+        quantization_config = BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_quant_type="nf4",            # NF4 is often recommended
+            bnb_4bit_compute_dtype=torch.bfloat16, # Use bfloat16 for compute
+            bnb_4bit_use_double_quant=True,       # Use double quantization to save more memory
+        )
 
+        # Device map will handle placing layers, relying on accelerate
+        # No need to explicitly set max_memory when using device_map="auto" typically
         pipe = pipeline(
             "text-generation",
             model=MODEL_ID,
+            # Note: torch_dtype is sometimes ignored when quantization_config is used,
+            # but specifying compute_dtype in BitsAndBytesConfig is key.
+            # Keep torch_dtype=torch.bfloat16 here for consistency if needed by other parts.
             torch_dtype=torch.bfloat16,
-            device_map="auto", # Relies on accelerate
+            device_map="auto", # Let accelerate handle layer placement
             model_kwargs={
                 "quantization_config": quantization_config,
                 "use_cache": True,
-                # "max_memory": {0: MAX_GPU_MEMORY} # Often handled by device_map="auto"
+                # "trust_remote_code=True" # Add if model requires it (check model card)
             }
         )
-        print(f"Model {MODEL_ID} loaded successfully on device: {pipe.device}")
+        print(f"Model {MODEL_ID} loaded successfully on device: {pipe.device} (using 4-bit quantization)")
         return pipe
     except Exception as e:
-        print(f"FATAL Error loading model '{MODEL_ID}': {e}")
-        # Raise the error to potentially get more detailed logs in Spaces
+        print(f"FATAL Error loading model '{MODEL_ID}' (check memory/config): {e}")
+        # Raise the error to ensure it's visible in Spaces logs
         raise e
-        # return None # Returning None might hide the root cause in Spaces logs
+
+# --- REST OF THE CODE REMAINS THE SAME ---
+# (search_person, create_synthetic_profile, extract_text_from_search_results,
+#  parse_llm_output, generate_enhanced_persona, generate_system_prompt_with_llm,
+#  generate_response, PersonaChat class, create_interface function, __main__ block)
+# ... include the rest of the Python code from the previous correct version here ...
+
+# Make sure the rest of your app.py file follows this modified load_model function.
+# Keep all other functions and the Gradio interface definition as they were.
 
 # --- Web Search ---
 # (Keep search_person, create_synthetic_profile, extract_text_from_search_results as before)