LoRA for Conv2d layer, script to convert kohya_ss LoRA to PEFT #461
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Wow, I was literally just looking for something like this. Nice! Hopefully it gets pulled soon |
|
The documentation is not available anymore as the PR was closed or merged. |
…o peft conversion script
|
Hi! |
|
Hey, @pacman100, would highly appreciate it if you could have a look at this PR. Thanks in advance! |
pacman100
approved these changes
Jun 15, 2023
There was a problem hiding this comment.
Thank you @kovalexal for adding support for Conv2D layers with LoRA method and converting the Kohya ckpts to PEFT format, LGTM! 🤗🚀🔥
|
Hi @kovalexal , Thanks for your nice code to make LoRA work better! For those who want to use PEFT LoRAs in Automatic1111 WebUI, do you have any suggestion on how to convert LoRA in PEFT format back to the Kohya ckpts or safetensors? I know the question is not quite related to this PR (and there is #212) but I can't find any answer. Thank you very much! |
Guy-Bilitski
pushed a commit
to Guy-Bilitski/peft
that referenced
this pull request
May 13, 2025
…ngface#461) * Added LoRA for Conv2d layer, script to convert kohya_ss linear lora to PEFT * Fixed code style, added missing safetensors dependency for kohya_ss to peft conversion script
cyyever
pushed a commit
to cyyever/peft
that referenced
this pull request
Sep 4, 2025
* WIP skeleton * minimal working poc * cleanup * rename variables * quick typo fix * add v1 masking (huggingface#429) * add v1 masking * working v1 * adapt from suggestion * avoid warning `Setting `pad_token_id` to `eos_token_id`:50256 for open-end generation.` * fix masking - mask the responses from API call only * quality * address comments * Update trl/environment/base.py Co-authored-by: Leandro von Werra <[email protected]> * adapt a bit * wip on tokenization/masking in textenv * small fixes * update viz * add example * print debug text and pass masks * style * format and move tensor to device * update example * update example * This seems to work * fix masking * fix rich output to console --------- Co-authored-by: Costa Huang <[email protected]> Co-authored-by: Leandro von Werra <[email protected]> Co-authored-by: leandro <[email protected]> * Add masking (huggingface#461) * add v1 masking * working v1 * adapt from suggestion * avoid warning `Setting `pad_token_id` to `eos_token_id`:50256 for open-end generation.` * fix masking - mask the responses from API call only * quality * address comments * Update trl/environment/base.py Co-authored-by: Leandro von Werra <[email protected]> * adapt a bit * wip on tokenization/masking in textenv * small fixes * update viz * add example * print debug text and pass masks * style * format and move tensor to device * update example * update example * This seems to work * fix masking * fix rich output to console * fix batched generation * improve stopping criteria * improve error handling in tool call --------- Co-authored-by: younesbelkada <[email protected]> Co-authored-by: Younes Belkada <[email protected]> Co-authored-by: Costa Huang <[email protected]> * fix uknown tool * fix rewards and increase bs * remove unused script * ugly WIP fix * do not return modified obj for in-place operations * do not return modified obj for in-place operations * clean up stopping criterium * push updates * push update * format, add docs * rename file * add kwargs to reward fn * simplify example * simplify example * bug fix * add a trivia example * pre-commit * max tool response length * fix regex for multi-line * refactor tool exceptions * fix exceptions in tool * add docs * fix style * make rich optional * add docstrings * add tests * add TextEnv tests (WIP) * update triviaqa code * update docs * refactor text env * update tests (WIP) * add end2end test * update docs * upload tool demo * refactor * customizable system prompt * add text env docs * update index and toc * fix `TextHistory` show methods * add max length * fix style * fix typo * refactor to kwargs in init and tasks to queries * kwargs for reward docs * Update examples/triviaqa.py Co-authored-by: Younes Belkada <[email protected]> * Update examples/tool_demo.py Co-authored-by: Younes Belkada <[email protected]> * Update docs/source/learning_tools.mdx Co-authored-by: Younes Belkada <[email protected]> * Update docs/source/learning_tools.mdx Co-authored-by: Younes Belkada <[email protected]> * Update docs/source/learning_tools.mdx Co-authored-by: Younes Belkada <[email protected]> * Update docs/source/text_environments.md Co-authored-by: Younes Belkada <[email protected]> * Update examples/triviaqa.py Co-authored-by: Younes Belkada <[email protected]> * Update examples/triviaqa.py Co-authored-by: Younes Belkada <[email protected]> * move to tool folder * remove assets * remove tool demo * move rich import test to import utils * add copyright * fixes for masks in ppo trainer * add text env api docs * make precommit + add ppo test with mask * move examples and add python * fix style * update triviaqa example * add more docs * update docs * Update docs/source/learning_tools.mdx * Apply suggestions from code review * precommit --------- Co-authored-by: Costa Huang <[email protected]> Co-authored-by: Younes Belkada <[email protected]> Co-authored-by: younesbelkada <[email protected]> Co-authored-by: leandro von werra <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Hi!
Thank you for your awesome library, it pushes the limits of tuning large-scale models on consumer-like hardware even further 🤟
Some researches have shown, that allowing Conv2d layers to be also trained in LoRA helps to get even better results for SD models.
I've made a small modification, which allows to train Conv2d layers, and also made a script to convert basic checkpoints from kohya_ss to be used in PEFT.
It would be beneficial if it will get eventually merged and will allow anyone to use some of pretrained LoRAs in diffusers framework.