Add LoRA support for Cosmos Predict 2.5 and fix pipeline to match official Cosmos repo#13664
Open
terarachang wants to merge 15 commits intohuggingface:mainfrom
Open
Add LoRA support for Cosmos Predict 2.5 and fix pipeline to match official Cosmos repo#13664terarachang wants to merge 15 commits intohuggingface:mainfrom
terarachang wants to merge 15 commits intohuggingface:mainfrom
Conversation
added 8 commits
April 30, 2026 19:52
…ncoder attention implementation, and timestep scaling
… to device before torch.stack
github-actions
Bot
added
documentation
github-actions
Bot
added
size/L
github-actions
Bot
added
size/L
yiyixuxu
reviewed
May 4, 2026
Collaborator
yiyixuxu
left a comment
yiyixuxu
left a comment
There was a problem hiding this comment.
thanks, i left a question
yiyixuxu
reviewed
May 4, 2026
yiyixuxu
reviewed
May 4, 2026
yiyixuxu
reviewed
May 4, 2026
yiyixuxu
reviewed
May 4, 2026
github-actions
Bot
added
size/L
terarachang
force-pushed
the
cosmos_predict_2.5_lora_clean
branch
from
0cc6351 to
c8513c2
Compare
github-actions
Bot
added
size/L
github-actions
Bot
added
size/L
yiyixuxu
reviewed
May 5, 2026
|
|
||
| device = sample.device | ||
| sigma_t, sigma_s0 = self.sigmas[self.step_index + 1].to(device), self.sigmas[self.step_index].to(device) | ||
| sigma_t, sigma_s0 = self.sigmas[self.step_index + 1], self.sigmas[self.step_index] |
Collaborator
There was a problem hiding this comment.
ohh I think the change here might not be intended, no?
it seeem to have reverted https://github.com/huggingface/diffusers/pull/13489/changes
yiyixuxu
approved these changes
May 5, 2026
Collaborator
yiyixuxu
left a comment
yiyixuxu
left a comment
There was a problem hiding this comment.
thanks, looks good to me once we revert the change in scheculer
github-actions
Bot
added
size/L
sayakpaul
reviewed
May 6, 2026
Member
sayakpaul
left a comment
sayakpaul
left a comment
There was a problem hiding this comment.
I left some questions and suggestions. LMK if anything is unclear.
| [[autodoc]] loaders.lora_pipeline.CosmosLoraLoaderMixin | ||
|
|
||
| ## KandinskyLoraLoaderMixin | ||
| [[autodoc]] loaders.lora_pipeline.KandinskyLoraLoaderMixin |
Member
There was a problem hiding this comment.
Figures should be hosted somewhere else.
|
|
||
| @classmethod | ||
| @validate_hf_hub_args | ||
| def lora_state_dict( |
Member
There was a problem hiding this comment.
We should be able to also use "# Copied from ..." comment here?
| else: | ||
| return state_dict | ||
|
|
||
| def load_lora_weights( |
| safe_serialization=safe_serialization, | ||
| ) | ||
|
|
||
| def fuse_lora( |
Comment on lines
+6096
to
+6118
| network_alphas = {} | ||
| for k in list(state_dict.keys()): | ||
| if "alpha" in k: | ||
| alpha_value = state_dict.get(k) | ||
| if (torch.is_tensor(alpha_value) and torch.is_floating_point(alpha_value)) or isinstance( | ||
| alpha_value, float | ||
| ): | ||
| network_alphas[k] = state_dict.pop(k) | ||
| else: | ||
| raise ValueError( | ||
| f"The alpha key ({k}) seems to be incorrect. If you think this error is unexpected, please open as issue." | ||
| ) | ||
|
|
||
| if return_alphas or return_lora_metadata: | ||
| return cls._prepare_outputs( | ||
| state_dict, | ||
| metadata=metadata, | ||
| alphas=network_alphas, | ||
| return_alphas=return_alphas, | ||
| return_metadata=return_lora_metadata, | ||
| ) | ||
| else: | ||
| return state_dict |
Member
There was a problem hiding this comment.
Do we need this setup in cosmos?
| components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names, **kwargs | ||
| ) | ||
|
|
||
| def unfuse_lora(self, components: list[str] = ["transformer"], **kwargs): |
Author
There was a problem hiding this comment.
Our implementation doesn't need it. But if users trained multiple LoRA adaptors they may want to use this method. Would you suggest I remove unfuse_lora?
Member
There was a problem hiding this comment.
I am not suggesting to remove it. I am suggesting to supplement it with "# Copied from ..." statement.
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
What this PR does
Adds LoRA fine-tuning support for Cosmos Predict 2.5 (nvidia/Cosmos-Predict2.5-2B) and fixes the pipeline to match the official Cosmos reference implementation.
LoRA support
CosmosLoraLoaderMixininsrc/diffusers/loaders/lora_pipeline.pyfor LoRA loading/saving onCosmosTransformer3DModelexamples/cosmos/train_cosmos_predict25_lora.pyusingaccelerate+peftexamples/cosmos/eval_cosmos_predict25_lora.pyCosmosLoraLoaderMixintodocs/source/en/api/loaders/lora.mdFixes to match the official Cosmos repo
conditional_frame_timestepscaling bytimestep_scale=0.001Test plan
check_copies,check_dummies,check_support_list)