support of model preloading for multiple inferences

[CA-TT-AC]

I've been working with your ChartVLM model and noticed that the current implementation loads the model before each inference call. This appears to be inefficient for my use case where I need to perform multiple inferences.

Looking at the code in ChartVLM.py, each time infer_ChartVLM() is called, it seems to reload the adapter and decoders. I'm wondering if you have an alternative implementation to load the model components once and reuse them for multiple inferences?

This would significantly improve performance for batch processing scenarios.

Thank you for your time and for sharing this excellent work.