This is amazing work—thank you for your contribution and for making it open source! Can you provide some suggestions on how to continue fine-tuning the jinaclip-v2 model with local data? After I comment out @torch.inference_mode(), can I directly load and fine-tune this model?