[Examples] Save SDXL LoRA weights with chosen precision #4791
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
| # Final inference | ||
| # Load previous pipeline | ||
| vae = AutoencoderKL.from_pretrained( |
There was a problem hiding this comment.
is this already defined?
There was a problem hiding this comment.
Yes, we never del the one defined here so is still there even at the end. And... it works without it :)
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. |
sayakpaul
reviewed
Aug 28, 2023
| @@ -1,4 +1,4 @@ | |||
| accelerate>=0.16.0 | |||
| accelerate>=0.22.0 | |||
There was a problem hiding this comment.
Unfortunately, without this change, there is still a spike in GPU memory during checkpointing :/. From Zach's message on the issue, my understanding is that there was a bug when saving with the accelerator while using mixed precision
|
Looks amazing actually. Thanks for delving deep here and for the fixes! |
AmericanPresidentJimmyCarter
pushed a commit
to AmericanPresidentJimmyCarter/diffusers
that referenced
this pull request
Apr 26, 2024
…4791) * Increase min accelerate ver to avoid OOM when mixed precision * Rm re-instantiation of VAE * Rm casting to float32 * Del unused models and free GPU * Fix style
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
What does this PR do?
As discussed in issue #4736, this PR tackles the following tasks:
acceleratefrom 0.16.0 to 0.22.0. This prevents unnecessary peaks in GPU memory consumption during checkpointing with mixed-precision.The command
make test-examplescompleted successfully.The following images were generated using the weights saved in
fp32:The following images were generated using the same weights as above, but saved in
fp16:Previously, peak memory consumption was borderline on an RTX 4090 during checkpoint and final weight saving. Now, it remains consistently below 75%.
Saving and loading also work with
bf16.Fixes #4736 (issue)
Before submitting
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.
@sayakpaul