[training] fixes to the quantization training script and add AdEMAMix optimizer as an option by sayakpaul · Pull Request #9806 · huggingface/diffusers (original) (raw)
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service andprivacy statement. We’ll occasionally send you account related emails.
Already on GitHub?Sign in to your account
Conversation13 Commits4 Checks11 Files changed
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
[ Show hidden characters]({{ revealButtonHref }})
| @@ -1059,7 +1076,7 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32): |
|---|
| ) |
| # handle guidance |
| if transformer.config.guidance_embeds: |
| if unwrap_model(transformer).config.guidance_embeds: |
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
So that things are compatible with DeepSpeed.
Comment on lines +1040 to 1047
| vae_scale_factor = 2 ** (len(vae_config_block_out_channels) - 1) |
|---|
| latent_image_ids = FluxPipeline._prepare_latent_image_ids( |
| model_input.shape[0], |
| model_input.shape[2], |
| model_input.shape[3], |
| model_input.shape[2] // 2, |
| model_input.shape[3] // 2, |
| accelerator.device, |
| weight_dtype, |
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Follows what we do in the Flux LoRA scripts.
Comment on lines +1102 to +1103
| height=model_input.shape[2] * vae_scale_factor, |
|---|
| width=model_input.shape[3] * vae_scale_factor, |
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Same as above.
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Cool! did you get a chance to play with AdEMAMix? should we consider adding it to the other scripts as well?
Cool! did you get a chance to play with AdEMAMix? should we consider adding it to the other scripts as well?
Testing the memory requirement as we speak. Will report back.
| else: |
|---|
| optimizer_class = bnb.optim.AdEMAMix |
| optimizer = optimizer_class(params_to_optimize) |
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
should we support betas and weight_decay here?
We could use the existing args like we did for prodigy, i.e.
optimizer = optimizer_class(params_to_optimize,
betas=(args.adam_beta1, args.adam_beta2),
weight_decay=args.adam_weight_decay)
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Umm I didn't want to actually to keep the separations of concern very clear. We could maybe revisit if the community finds the optimizer worth the go?
a-r-r-o-w pushed a commit that referenced this pull request
… optimizer as an option (#9806)
fixes
more fixes.
sayakpaul added a commit that referenced this pull request
… optimizer as an option (#9806)
fixes
more fixes.