Does LoRA fine-tuning freeze the Vision Encoder?

[zcyqyq]

Hi, thanks for your great work, which is really impressive! I noticed that when using LoRA fine-tuning configurations (e.g., pi0_fast_libero_low_mem_finetune), the Vision Encoder (SigLIP, PaliGemma/img/*) appears to be trainable rather than frozen. Is this true or did I misunderstand something? Looking at get_freeze_filter(), it only matches .llm. but not .img., which suggests the vision encoder is being trained. Is this intentional?