Add LoRA support for Cosmos Predict 2.5 and several fixes to match the official repo#1

Open

terarachang wants to merge 215 commits into

mainfrom

cosmos_predict_2.5_lora

Commits on Feb 27, 2026

Fix LTX-2 image-to-video generation failure in two stages generation (#13187)

Songrui625
and
github-actions[bot]
authored
Fixing Kohya loras loading: Flux.1-dev loras with TE ("lora_te1_" prefix) (#13188)
christopher5106
authored
[Modular] update the auto pipeline blocks doc (#13148)

authored

Commits on Feb 28, 2026

Commits on Mar 2, 2026

Commits on Mar 3, 2026

Commits on Mar 6, 2026

Commits on Mar 7, 2026

post release 0.37.0 (#13215)
sayakpaul
authored

Commits on Mar 8, 2026

Commits on Mar 10, 2026

[CI] Potential fix for code scanning alert no. 2150: Workflow does not contain permissions (#13230)

authored

Commits on Mar 12, 2026

klein 9b kv (#13262)

huemin-art
and
github-actions[bot]
authored

Commits on Mar 13, 2026

Commits on Mar 14, 2026

Add AGENTS.md (#13259)

authored

Commits on Mar 16, 2026

[docs] updates (#13248)
stevhliu
authored

Commits on Mar 18, 2026

Commits on Mar 19, 2026

Commits on Mar 21, 2026

[tests] test load_components in modular (#13245)
sayakpaul
authored

Commits on Mar 23, 2026

Commits on Mar 24, 2026

Commits on Mar 26, 2026

Commits on Mar 27, 2026

Commits on Mar 28, 2026

change minimum version guard for torchao to 0.15.0 (#13355)
howardzhang-cv
authored

Commits on Mar 31, 2026

Commits on Apr 1, 2026

Commits on Apr 2, 2026

Commits on Apr 6, 2026

Commits on Apr 7, 2026

Commits on Apr 8, 2026

[CI] Use finegrained token for Issue Labeler (#13433)
DN6
authored

Commits on Apr 10, 2026

Commits on Apr 11, 2026

Commits on Apr 13, 2026

Commits on Apr 16, 2026

Commits on Apr 18, 2026

Commits on Apr 21, 2026

Commits on Apr 23, 2026

Improve `trust_remote_code` (#13448)

authored

Commits on Apr 24, 2026

Commits on Apr 27, 2026

Commits on Apr 28, 2026

support lora for inference
Ting-Yun Chang
committed
support lora for training
Ting-Yun Chang
committed
fix args to match the offical cosmos repo: (1) use deterministic vae instead of sampling (2) use flash attn implementation for the text encoder
Ting-Yun Chang
committed
Add scripts to fix the differences between diffusers and cosmos's video processors. torchvision.transforms.functional.resize uses bilinear interpolation by default
Ting-Yun Chang
committed
refactor training code
Ting-Yun Chang
committed
fix numeric precision of sigma_t in pipeline
Ting-Yun Chang
committed
fix text_encoder device when attn_implementation specified
Ting-Yun Chang
committed
refactor inference code. To match the cosmos github, some changes (1) Strip the new line of the prompt in .json; (2) Support invariant seeds with numpy noise sampling
Ting-Yun Chang
committed
refeactor Cosmos2_5_PredictBasePipeline
Ting-Yun Chang
committed
add scripts to extract inputs/outputs from i4 models to compare with diffusers implementation
Ting-Yun Chang
committed
the default conditional_frame_timestep should be scaled by timestep_scale=0.001
Ting-Yun Chang
committed
auto cast adaln and final layer of DiT to fp32, following cosmos implementation
Ting-Yun Chang
committed
Apply the same fix for lora training; the default conditional_frame_timestep should be scaled by timestep_scale=0.001
Ting-Yun Chang
committed
use the same optimizer as the cosmos codebase for lora training
Ting-Yun Chang
committed
add cfg during lora training
Ting-Yun Chang
committed
reverse sigma_t back to the diffusers implementations; support no image and no video inputs
Ting-Yun Chang
committed
(1) set fa2 as the default attn implementation for text encoder (2) add autocast_fp32 flag to allow users to trade precisions off inference speed
Ting-Yun Chang
committed
support taking pre-processed videos as inputs (e.g. inputs from train dataloader)
Ting-Yun Chang
committed
save lora layers + final generation after lora training
Ting-Yun Chang
committed
support multi-gpu training
Ting-Yun Chang
committed
support conditional_frames_probs is training
Ting-Yun Chang
committed
hardcode to args
Ting-Yun Chang
committed
remove redundant code in lora_pipeline
Ting-Yun Chang
committed
checkpoint lora weights during training
Ting-Yun Chang
committed
remove redundant args in train and add eval scripts on gr1_object dataset
Ting-Yun Chang
committed
refactor
Ting-Yun Chang
committed
fuse lora/dora at inference and enable multi-seed evaluation
Ting-Yun Chang
committed
add comos-reason2 as judge
Ting-Yun Chang
committed
refactor and add assets
Ting-Yun Chang
committed
add generated videos
Ting-Yun Chang
committed
clean up scripts
Ting-Yun Chang
committed
set revision to fullname diffusers/base/post-trained for consistency
Ting-Yun Chang
committed
fix download script and safety checker
Ting-Yun Chang
committed
remove misc
Ting-Yun Chang
committed
autocast_fp32 default to False and only set to True at training
Ting-Yun Chang
committed
edit data preprocess script to avoid downloading cosmos2.5 repo
Ting-Yun Chang
committed
add readme, requirements, and make flash attn optional
Ting-Yun Chang
committed
use built-in optimizer
Ting-Yun Chang
committed
make style
Ting-Yun Chang
committed
make style
Ting-Yun Chang
committed
add CosmosLoraLoaderMixin to lora.md docs
Ting-Yun Chang
committed
add CosmosLoraLoaderMixin to TYPE_CHECKING import block
Ting-Yun Chang
committed
fix(scheduler): device mismatch from upstream b114620d8 - move rk and b to device before torch.stack
Ting-Yun Chang
committed

Commits on Apr 30, 2026

upload YAML prompt files for LLM judges evaluating videos
Ting-Yun Chang
committed