fix: actually quantize model with quantize_4bit()

test_model_loading.py

@@ -248,6 +248,12 @@ def test_strategy_6():
             model,
             use_gradient_checkpointing=False,
         )
+        print("   ✓ Model prepared for k-bit training")
+        
+        # Actually quantize the model
+        from bitsandbytes.nn.modules import Params4bit
+        print("   Quantizing weights to 4-bit...")
+        model.quantize_4bit()
         print("   ✓ Model quantized to 4-bit (~17.5GB)")
         
         print("\n   Step 3: Move to GPU 0...")