0.4.2rc1 is released on 25th Feb 2025

What's Changed

• [Ray] Add Support for Disaggregating VAE and DiT by @lihuahua123 in #422
• [Fix] Add method to prepare extra step kwargs for scheduler in xFuserCogVideoXPipeline by @LazyBusyYang in #426
• [Bugfix] avoid yunchang undefined & format by @DefTruth in #434
• [Tests]: Context and Sequence Parallel by @antferdom in #438
• [Ray] Add more Ray examples by @lihuahua123 in #427
• [doc] ray launch parallel inferene by @feifeibear in #442
• [Bugfix] Fix for only one DIT worker and a separate VAE worker. by @lihuahua123 in #443
• [Fix] Make some dependencies optional by @xibosun in #445
• [Feat] Add implementations for USP by @xibosun in #446
• Add TeaCache and FBCache by @Binary2355 in #451
• refactory cache directory and update readme by @feifeibear in #452
• Cache applied with torch.compile by @Binary2355 in #453
• feat: Add support for Tensor Parallelism to the Step-Video-T2V model by @LiaoYuanF in #454
• bump version to 0.4.2rc1 by @feifeibear in #455

New Contributors

• @LazyBusyYang made their first contribution in #426
• @antferdom made their first contribution in #438
• @Binary2355 made their first contribution in #451
• @LiaoYuanF made their first contribution in #454

Full Changelog: 0.4.1...0.4.2rc1

0.4.1 version is released on Jan 1st 2025

What's Changed

• Host script fix by @fy1214 in #386
• add hunyuanvideo performance by @feifeibear in #387
• fix readme typos for hunyuanvideo by @feifeibear in #390
• closes #334; prefer newer version of flash-attn by @chengzeyi in #394
• fix bugs in cogvideo patch_embed by @xibosun in #399
• add warning for old flash attn by @chengzeyi in #397
• Add docs for adding new models by @xibosun in #403
• update readme with new adding model tutorial by @feifeibear in #404
• [ray] launch multiple GPU with ray by @lihuahua123 in #396
• add hunyuan_video_usp_example.py by @chengzeyi in #401
• ConsisID for xdit by @pkuhxy in #405
• Fix README related by @SHYuanBest in #408
• remove dependency on flash_attn by @feifeibear in #410
• hotfix pipefusion using flash_attn by @feifeibear in #411
• update readme with flash_attn and diffusers optional. by @feifeibear in #412
• fix flux on diffusers version 0.32 by @xibosun in #413
• correct perf numbers and add details by @chengzeyi in #414
• bump to version 0.4.1 by @feifeibear in #418

New Contributors

• @lihuahua123 made their first contribution in #396
• @pkuhxy made their first contribution in #405
• @SHYuanBest made their first contribution in #408

Full Changelog: 0.4.0...0.4.1

version 0.4.0 is released on 5th Dec 2024

What's Changed

The version provide lightweight USP API for HunyuanVideo

• bump to version 0.4.0 by @feifeibear in #384

Full Changelog: 0.3.6...0.4.0

v0.3.6 released on 5th December 2024!

What's Changed

• Add CogVideoX1.5 support by @xibosun in #357
• concurrent access to text-to-image http service. by @feifeibear in #359
• [FEAT] Slice text embedding in MM-DiT by @xibosun in #361
• Update README.md with paper links by @feifeibear in #362
• ring_flash_attn forward compatible with FA>=2.7.0 by @DefTruth in #364
• Flux.1 performance optimizations on H100 by @xibosun in #365
• support fp8 t5 encoder in examples by @Lay2000 in #366
• Support optimized USP in Flux by @xibosun in #368
• add Flux.1 4xH100 performance by @feifeibear in #369
• Update citation format in README.md by @feifeibear in #371
• fix host.py in single gpu case by @fy1214 in #375
• Update history parts in README.md by @feifeibear in #377
• Optimize USP interface for Flux and CogVideo by @xibosun in #376
• bump to version 0.3.6 by @feifeibear in #379
• FIX: Decouple runtime_state from USP by @xibosun in #382
• diffusers as optional during install by @feifeibear in #383

New Contributors

• @DefTruth made their first contribution in #364
• @fy1214 made their first contribution in #375

Full Changelog: 0.3.5...0.3.6

v0.3.5 released on 20th September 2024!

What's Changed

• update readme with contact info and mochi-xdit by @feifeibear in #341
• readme: correct email address by @feifeibear in #342
• doc: update ComfyUI in README by @Lay2000 in #349
• [hotfix] yunchang 0.4.0 import error by @feifeibear in #351
• Doc/update readme by @Lay2000 in #352
• fix: flux dp bug by @ictzyqq in #353
• fix torch_attn import by @xibosun in #355
• bump to version 0.3.5 by @feifeibear in #356

New Contributors

• @ictzyqq made their first contribution in #353

Full Changelog: 0.3.4...0.3.5

v0.3.4 released on 7th November 2024!

This version offers a set of stable Unified Sequence Parallel (USP) interfaces for MM-DiT.

What's Changed

• [doc] update the docker tags and flux.1 performance by @feifeibear in #317
• add imageio in setup.py by @feifeibear in #320
• support FID compute by @Lay2000 in #327
• [doc] updates hunyuandit performance by @feifeibear in #328
• [unitest] test ring_flash_attn_func by @feifeibear in #329
• add unitest for xfuser attn layer by @feifeibear in #335
• Refactor http service by @feifeibear in #336
• [Refactor] unified interface for long ctx attn by @Eigensystem in #337
• version to 0.3.4 by @feifeibear in #338
• Update README.md by @feifeibear in #339

Full Changelog: 0.3.3...0.3.4

v0.3.3 released on 25th Octorber 2024!

What's Changed

• Fix the CI/CD routine by @xibosun in #279
• fix: use xFuserJointLongContextAttention as the underlying SP method for CogVideoX by @xibosun in #280
• feat: complete sequence parallelism for CogVideo by @xibosun in #285
• move some pictures to other repo by @feifeibear in #286
• update docs and remove figure by @xibosun in #287
• fix: typo by @LRY89757 in #289
• Fix Latte Pipeline support by @xwhzz in #288
• add cogvideo entry by @feifeibear in #290
• fix: diffusers version by @Eigensystem in #292
• Fix parallel vae by @gty111 in #281
• Refactor: move legacy dir to branch legacy & Doc: README.md by @Eigensystem in #294
• Add updates to readme and fix typos by @xibosun in #293
• feat: DiTFastAttn for PixArt by @ZDJeffrey in #297
• add ditfastattn in readme, and seperate cogvideo and ditfastattn in run.sh by @feifeibear in #298
• add scripts for cogvideo and ditfastattn by @feifeibear in #299
• Enable warm up for VAE by @gty111 in #300
• [Feat] support pipefusion in flux model by @Eigensystem in #301
• [doc] polish readme by @feifeibear in #302
• add blog link on the banner by @feifeibear in #304
• Add CogVideoX results on A100 by @xibosun in #305
• Optimize GPU Memory Usage in CogVideo by @xibosun in #306
• [doc] add medium link in comfyui part of readme by @feifeibear in #307
• [example] print model parameter memory usage by @feifeibear in #308
• fix xFuserArgs members by @xibosun in #309
• Add scripts for inference on multiple nodes by @xibosun in #312
• fix bugs in allgather by @xibosun in #314
• Record CogVideo performance on up to 12 GPUs by @xibosun in #313
• upgrade diffusers to 0.31 by @feifeibear in #316

New Contributors

• @LRY89757 made their first contribution in #289
• @xwhzz made their first contribution in #288
• @gty111 made their first contribution in #281
• @ZDJeffrey made their first contribution in #297

Full Changelog: 0.3.2...0.3.3

v0.3.2 released on 14th September 2024!

What's Changed

• [doc] add hunyuandit performance by @feifeibear in #235
• [doc] update flux performance results with onediff and torch.compile by @feifeibear in #236
• [refactor] move cache_manager out of DiTRuntimeState by @feifeibear in #237
• add unitest for hunyuandit attn and refactor cache manager import by @feifeibear in #239
• attention processor unitest by @feifeibear in #240
• v100 disable torch.compile for attention processor by @feifeibear in #241
• Adding CogVideoX into xDiT by @dannyxiaocn in #211
• [doc] add latte, sd3 perf results by @feifeibear in #243
• Cogvideox SP Dev Minor Fix by @dannyxiaocn in #244
• [Feat] support xdit_parallel interface for huggingface by @Eigensystem in #245
• polish the run.sh in example by @feifeibear in #246
• [Fix] dp issue by @Eigensystem in #249
• dockerfile by @feifeibear in #250
• feat: support simple xdit comfyui-demo by @Lay2000 in #238
• [Fix]: image saving bugs by @Eigensystem in #251
• add a http service demo. by @feifeibear in #257
• [docker] service dockerfile by @feifeibear in #258
• Refactor setup.py and add instruction for cu124 by @lixiang007666 in #259
• Fix CogVideoX support by @chengzeyi in #261
• Revert "Fix CogVideoX support" by @Eigensystem in #263
• Make flux work with diffusers 0.30.2 by @chengzeyi in #264
• polish cogvideo example by @feifeibear in #266
• upgrade diffusers to 0.31.dev to support flux and cogvideox at the same time by @feifeibear in #268
• diffusers version check by @feifeibear in #269
• Added the CI/CD feature based on GitHub Actions by @xibosun in #270
• upgrade diffusers to 0.30.3 by @feifeibear in #273
• version 0.3.2 by @feifeibear in #274
• fix setup.py extras_require format error by @feifeibear in #275

New Contributors

• @lixiang007666 made their first contribution in #259
• @chengzeyi made their first contribution in #261
• @xibosun made their first contribution in #270

Full Changelog: 0.3.1...0.3.2

v0.3.1 released at 2024.08.27

1. upgrade flash_attn to 2.6.0+ and yunchang to 0.3.0
2. add torch.compile and onediff for compilation acceleration.

v0.3 is released on 2024.08.21

1. Reconfigure the module names. Therefore, it is not compatible with v0.2.0
2. Optimize the Flux.1 Performance
3. Support Ulysses and Ring Hybrid Attention for SD3 and Flux.1
4. xDiT can be applied on T4, V100.