Decrease accuracy when running llava model

[nghoaithuong]

I successfully converted the model to TensorRT-LLM by following the tutorials. However, when I evaluated the accuracy of this model on both the A100 device and Jetson Orin AGX, the accuracy decreased significantly (from 97% to 93%) when using FP16/FP8 precision and (from 97% to 89%) when using INT4.
Do you have any suggestions for resolving this issue?

[TriDefender]

Use higher precison