Freeze LLM set to True

[AdityaKulshrestha]

In the code, when enabling the LoRA the freeze_llm should be set to False. Shouldn't it be the opposite? Like in your other repository including QwenVL-Finetune.

assert not (training_args.lora_enable and training_args.freeze_llm), 'When using LoRA, the LLM should not be frozen. If you want to freeze the LLM, please disable LoRA.'

Constrastingly, in the QwenVL repo

raise ValueError("If `lora_enable` is True, `freeze_llm` must also be True.")

[2U1]

Yes—the LLM is frozen during training. The PEFT library automatically freezes the base model, so only the LoRA adapters are updated. I originally phrased it as “training the LLM with LoRA” to help first-time users think of it as training, but that wording caused confusion. I’ve switched the wording to match the Qwen2 repo’s style, and I’ll refactor this repo as well (same folder structure, docs, and code layout).