Qwen2-vl lora train？

[kike-0304]

Reminder

• I have read the README and searched the existing issues.

System Info

None

Reproduction

lora qwen2-vl

Expected behavior

Using Lora to train qwen2-vl, will the visual encoder freeze? How can Lora train the language model while training all the parameters of the visual model?

Others

No response

[CaiJichang212]

请问这里的--lora_target all all包含ViT模块的参数吗？

[hiyouga]

We freeze the vision encoder by default

LLaMA-Factory/src/llamafactory/hparams/finetuning_args.py

Lines 333 to 336 in 3eebae8

	freeze_vision_tower: bool = field(
	default=True,
	metadata={"help": "Whether ot not to freeze vision tower in MLLM training."},
	)