82713e9 update using more vram in lora training?

I can train lora model in [b32abdd](https://github.com/kohya-ss/sd-scripts/commit/b32abdd327f314fabc76fbbb627218090afb7b69) version using rtx3050 4g laptop with `--xformers --shuffle_caption --use_8bit_adam --network_train_unet_only --mixed_precision="fp16"`

but when I update to [82713e9](https://github.com/kohya-ss/sd-scripts/commit/82713e9aa62d9a28b17b3920b2b9bccd74051f9b) version(which is lastest) I just out of memory before _`Replace CrossAttention.forward to use xformers`_ hint

I can still training on [b32abdd](https://github.com/kohya-ss/sd-scripts/commit/b32abdd327f314fabc76fbbb627218090afb7b69) version, but I want to know which part of changes cause that?

Compare commits b32abdd and 82713e9, then reproduce the reported LoRA training command on the RTX 3050 4GB setup and observe where memory usage changes before the “Replace CrossAttention.forward to use xformers” hint. Done means identifying the change that causes the out-of-memory failure and documenting the relevant evidence.

Written by the indexing model from the issue text.