Skip to content

Conversation

@kohya-ss
Copy link
Owner

@kohya-ss kohya-ss commented Oct 23, 2025

When block swap is enabled only for the forward pass, extra blocks are mistakenly swapped. This caused an error after sample image generation (which is the forward pass only).

This also significantly improves inference speed when using block swap.

Additional tests are needed for the inference scripts for each architecture.

closes #654

@FurkanGozukara
Copy link
Contributor

wow this important

i had some block swap implementation based on musubi tuner maybe i have same error as well :D

@FurkanGozukara
Copy link
Contributor

@kohya-ss i tried to implement this into my app but got errors after several times testing

maybe i am doing wrong but here some findings

image

@kohya-ss
Copy link
Owner Author

@kohya-ss i tried to implement this into my app but got errors after several times testing

maybe i am doing wrong but here some findings

Thanks for reporting it, indeed the current code doesn't work when "block_to_swap >= num_blocks // 2". I'll fix it.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[BUG]Qwen-edit-plus training with sample

2 participants