关于find_unused_parameters的作用

Hi guys, 我想请教一下关于find_unused_parameters的作用

在qwen-image-edit的lora训练中，如果不设置find_unused_parameters的话会出现以下报错：

Parameters which did not receive grad for rank 0:

• pipe.dit.transformer_blocks.59.txt_mlp.net.2.lora_B.default.weight
• pipe.dit.transformer_blocks.59.txt_mlp.net.2.lora_A.default.weight
• pipe.dit.transformer_blocks.59.attn.to_add_out.lora_B.default.weight
• pipe.dit.transformer_blocks.59.attn.to_add_out.lora_A.default.weight

请问一下：
这个txt_mlp.net.2和attn.to_add_out在lora_target_modules中又指定，但是为什么在DDP中又需要设置为unused呢？这个会影响LoRA training的效果吗？以及find_unused_parameters的作用是什么？谢谢！

Start with the qwen-image-edit LoRA training configuration and the DDP setting for find_unused_parameters. Trace why txt_mlp.net.2 and attn.to_add_out are listed in lora_target_modules but do not receive gradients, then explain whether this changes training behavior and what the setting does. No file or test is named in the issue; done means providing a clear, evidence-based explanation.