1.1.0( DiT model Acceleration and nexfort compiler backend Beta Release)

Hightlight

• Acceleration of DiT model (with Pixart alpha as example): https://github.com/siliconflow/onediff/tree/main/onediff_diffusers_extensions/examples/pixart_alpha
• nexfort( a lightweight compiler backend for DiT models): https://github.com/siliconflow/onediff/tree/main/src/onediff/infer_compiler/backends/nexfort

What's Changed

• add user enterprise key guide by @lijunliangTG in #819
• Refining the Deversus Quantization Documentation by @lijunliangTG in #802
• Fixed minor error in README.md by @iamrohitanshu in #823
• update sd-webui doc and api_example by @marigoold in #712
• ComfyUI online quantitative usage documentation in wiki by @fmk345 in #809
• onediffx add get_active_adapters, and related unittest by @marigoold in #833
• fix bug in webui when switching from quantized to non-quantized by @marigoold in #830
• onediffx supports lycoris by @marigoold in #817
• Refine SVD diffusers example for graph save/load by @lixiang007666 in #836
• Add Apache 2.0 LICENSE by @strint in #842
• Update README.md file by @lijunliangTG in #846
• Support 8x dimensions by hijacking cat in webui by @nono-Sang in #847
• Update onediff/onediff_comfy_nodes/README.md by @ccssu in #853
• fix removeprefix for python 3.8 by @nono-Sang in #856
• fix bug of always recompile in enterprise by @marigoold in #858
• [Fix] diffusers enterprise example seed by @lixiang007666 in #860
• Fix grammar in README by @jackalcooper in #803
• Refine README details by @jackalcooper in #862
• Dev ipa+isntantid by @ccssu in #814
• fix error caused by interpolate by @nono-Sang in #863
• add ONEFLOW_CONV2D_KERNEL_ENABLE_TUNING_WARMUP Environment variable by @lijunliangTG in #871
• Fix IPAdapter save load graph by @ccssu in #872
• define removesuffix in webui for python38 compatible by @marigoold in #874
• webui support save/load graph, refine UI by @marigoold in #825
• api for new backend by @hjchen2 in #794
• Version 1.1.0.dev1 by @jackalcooper in #876
• Fix Ci Error by @ccssu in #875
• Refine manual seed by @lixiang007666 in #877
• Fix bug for torch2oflow demo test by @nono-Sang in #878
• Upgrade ipa support by @ccssu in #879
• fix bug of onediff.infer_compiler.register by @marigoold in #885
• Fix ipa weight update by @ccssu in #891
• add nexfort and pixart alpha by @strint in #882
• Add throughput for benchmark by @lixiang007666 in #894
• Pixart performance data by @strint in #902
• Modify after Cancelling Quantitative Model by @ccssu in #897
• Fix Ci Test by @ccssu in #905
• Separate compile backend config by @strint in #909
• Add CI for Stable Diffusion web UI by @marigoold in #826
• webui extension refine and add close button for hint message by @marigoold in #908
• rename cache by @strint in #898
• renew readme in comfy by @lijunliangTG in #883
• Add nexfort cache docstr by @lixiang007666 in #917
• fix unet graph file cache by @ccssu in #887
• refactor test_lora by @marigoold in #821
• [fix bug] OneDiffX load_and_fuse_lora when lora_scale == 0 makes w_up to allzeros by @marigoold in #926
• add sd-webui CI Test SSIM by @lijunliangTG in #915
• remove extra generate lines by @lijunliangTG in #932
• Fix IPAdapter Style & Composition SDXL Support by @ccssu in #927
• Avoid installing both oneflow and nextforce by @ccssu in #938
• Add nexfort backend quant and docstr by @lixiang007666 in #919
• Add nexfort dynamic shape for pixart by @lixiang007666 in #931
• Update README.md for nexfort pixart by @lixiang007666 in #939
• Update text_to_image.py args default value by @lixiang007666 in #940

New Contributors

• @iamrohitanshu made their first contribution in #823
• @fmk345 made their first contribution in #809
• @nono-Sang made their first contribution in #847

Full Changelog: 1.0.0...1.1.0

1.0.0

OneDiff 1.0

OneDiff 1.0 is for Stable Diffusion and Stable Video Diffusion models(UNet/VAE/CLIP based) acceleration. We have got a lot of support/feedback from the community: https://github.com/siliconflow/onediff/wiki , big thanks!

The later version 2.0 will focus on DiT/Sora-like models.

OneDiff 1.0 's updates are mainly the issues in milestone 0.13

• New Features
  • Quantize tools for enterprise edition
    • https://github.com/siliconflow/onediff/tree/main/src/onediff/quantization
    • https://github.com/siliconflow/onediff/blob/main/README_ENTERPRISE.md#onediff-enterprise
    • SD-WebUI supports offline quantized model
  • OneDiff Quality Evaluation
    • https://github.com/siliconflow/OneDiffGenMetrics
  • Reuse compiled graph
    • #581
    • https://github.com/siliconflow/onediff/tree/main/onediff_sd_webui_extensions#fast-model-switching
    • reuse graph with constant folding enabled #782
  • Refine support for Playground v2.5
    • #703
  • support ComfyUI-AnimateDiff-Evolved
    • #575
    • https://github.com/siliconflow/onediff/tree/main/onediff_comfy_nodes/modules/oneflow/hijack_animatediff
  • support ComfyUI_IPAdapter_plus
    • https://github.com/siliconflow/onediff/tree/main/onediff_comfy_nodes/modules/oneflow/hijack_ipadapter_plus
  • support stable cascade
    • #659
• Improvements
  • Improve performance of VAE
    • #667
• Bug fixes
  • Fix graph save/load
  • Fix cuda memory usage on cuda 0 when run on other device
  • Fix the quality problem
  • many others bugs/issues from the community are fixed: https://github.com/siliconflow/onediff/milestone/2?closed=1
• Documentation
  • Add more intro about onediff: https://github.com/siliconflow/onediff?tab=readme-ov-file#about-onediff

What's Changed

• Add -U in code blocks of installation using aliyun index link by @jackalcooper in #699
• Fix black images issues with v2.1 by @doombeaker in #701
• enterprise doc add sdwebui by @marigoold in #705
• Dev quantize pipeline by @hjchen2 in #704
• fix quantize pipeline by @hjchen2 in #706
• update script of webui in enterprise doc by @marigoold in #709
• add quantize tools for diffusers by @hjchen2 in #710
• QuantPipeline support from_single_file by @marigoold in #711
• Fix graph save by @doombeaker in #702
• support stable cascade by @chengzeyi in #659
• Allow image_only_indicator to be None (comfy) by @lixiang007666 in #720
• open_conv_bias_add_act_fusion_default by @clackhan in #721
• fix ci error by @ccssu in #729
• [Fix] black images issues with diffusers SD2.1 by @lixiang007666 in #725
• [feat] Add SVD1.1 example by @lixiang007666 in #726
• doc for rsd to by @strint in #713
• SD-WebUI supports offline quantized model by @marigoold in #718
• Refactor onediffx deepcache by @clackhan in #733
• fix pipeline save load by @strint in #734
• fix_sdxl_light_test_scrip by @clackhan in #735
• use diffusers' builtin instant id by @chengzeyi in #738
• fix upcast_vae by @hjchen2 in #741
• fix import error by @ccssu in #746
• onediffx.lora supports diffusers>=0.19.3 by @marigoold in #740
• sd-webui supports reuse compiled graph by @marigoold in #742
• import torch first by @strint in #751
• support ComfyUI-AnimateDiff-Evolved by @ccssu in #575
• Reuse pipe by @strint in #647
• fix_multi_resulotion_error by @clackhan in #754
• refine_spatio_temporal_oflow by @clackhan in #755
• sd-webui override send_model_to_cpu by @marigoold in #758
• remove reuse pipe ci test by @marigoold in #759
• [add] deep-cache test by @ccssu in #739
• refine readme by @doombeaker in #768
• fix_cuda0_mem_usage_error by @clackhan in #770
• fix bug of WebUI which quantization is always True by @marigoold in #775
• Add SD 1.5 quant example for diffusers by @lixiang007666 in #757
• Fix: Change ComfyUI-AnimateDiff-Evolved load to warning. by @ccssu in #777
• Fix AttributeError in OneFlowDeepCacheSpeedUpModelPatcher by @ccssu in #774
• fix ComfyUI_IPAdapter_plus use by @ccssu in #732
• Fix Torch-to-OneFlow Transformation Time by @ccssu in #786
• reuse graph with constant folding for sd webui by @marigoold in #782
• refactor oneflow_compile to support multiple backends by @hjchen2 in #787
• Fix redundant warning printing by @ccssu in #788
• develop compile options and rm use_graph by @hjchen2 in #791
• test_quantize_custom_model by @ccssu in #756
• Imporve ENTERPRISE doc by @lijunliangTG in #784
• Update quant ci examples by @lijunliangTG in #792
• refactor graph reuse by @marigoold in #790
• Dev refactor onediff comfy nodes by @ccssu in #796
• fix bug in generate_constant_folding_info by @marigoold in #811

New Contributors

• @lijunliangTG made their first contribution in #784

Full Changelog: 0.12.1...1.0.0

0.12.1(Stable acceleration of SD and SVD for production environment)

SOTA performance update for SDXL and SVD

https://github.com/siliconflow/onediff/tree/main?tab=readme-ov-file#state-of-the-art-performance

Fully support dynamic resolution run of SD and SVD

https://github.com/siliconflow/onediff/blob/55627d50157d4a0c4b484ba76b088c90f39179ff/onediff_diffusers_extensions/examples/text_to_image_sdxl.py#L96

Compile/Save/Load pipeline for HF diffusers

https://github.com/siliconflow/onediff/tree/main/onediff_diffusers_extensions#compile-save-and-load-pipeline

Fast LoRA loading and switching for HF diffusers

https://github.com/siliconflow/onediff/tree/main/onediff_diffusers_extensions#fast-lora-loading-and-switching

Accelerate InstantID

https://github.com/siliconflow/onediff/blob/main/benchmarks/instant_id.py
https://github.com/siliconflow/onediff/tree/main/onediff_comfy_nodes#instantid

Accelerate SDXL lightning

https://github.com/siliconflow/onediff/blob/main/onediff_diffusers_extensions/examples/text_to_image_sdxl_light.py

Reuse compiled components between pipelines

https://github.com/siliconflow/onediff/blob/main/onediff_diffusers_extensions/examples/reuse_compiled_pipeline_components.py

Compatible with HF diffisers 0.26

SD Webui support img2img

Change log

• Official LoRA compatible with OneDiff by @marigoold in #507
• use dict.setdefault in oneflow_compile by @marigoold in #515
• fix compilation failure of resnet50 and faster_rcnn by @haoyang9804 in #491
• Add SILICON_ONEDIFF_LICENSE_KEY by @jackalcooper in #519
• Dev quant tools and fix graph file management by @ccssu in #495
• refine document of setup by @doombeaker in #522
• replace reshape by flatten and unflatten to speedup svd by @hjchen2 in #516
• graph load fix by @strint in #525
• compile vae decoder to avoid oneflow module to while upcasting vae by @hjchen2 in #526
• Debug Comfy in CI by @jackalcooper in #521
• simplify test by @strint in #529
• Enable docker hub cache by @jackalcooper in #532
• Add links for NA/EU by @jackalcooper in #542
• Reorder links by @jackalcooper in #544
• support_deepcache_svd_pipeline by @clackhan in #514
• view comfyui log by @doombeaker in #534
• Documents for OneDiff Enterprise by @doombeaker in #545
• fix adpater id argument by @doombeaker in #478
• Dev optimize svd by @chengzeyi in #533
• fix bug of _torch_module_to_with_check using oneflow args by @marigoold in #548
• transform diffusers AttnProcessor by @hjchen2 in #550
• update readme by @doombeaker in #553
• Update hf sd1.5 model url by @doombeaker in #554
• rm dynamo by @hjchen2 in #558
• Fix Quantitative Execution in sd_webui by @ccssu in #517
• Publish onediff on pypi automatically by @jackalcooper in #560
• add sd-webui api doc and script by @marigoold in #559
• Fix ControlNet Usage by @ccssu in #561
• align dynamic param with torch by @strint in #537
• update readme by @doombeaker in #562
• Add community version controlnet performance testing by @ccssu in #557
• [diffusers] support dynamic batch size by @hjchen2 in #552
• Enable dynamic for comfy by @strint in #535
• fix ControlLora usage by @ccssu in #565
• add ci for deepcache and int8 by @doombeaker in #556
• Enable MLIR logging flags in CI by @jackalcooper in #568
• speedup node create mixed modelpatcher by @marigoold in #472
• Support YGemm by @chengzeyi in #563
• Support SVD dynamic shape[feat] by @lixiang007666 in #564
• add cached load_lora_weight by @marigoold in #524
• support custom svd by @strint in #536
• make diffusers_extensions.utils a package by @isidentical in #579
• [Fix] dynamic switch for uncommon resolutions by @lixiang007666 in #573
• Fix typo in README.md by @jackalcooper in #583
• fix_multi_resolutions_bug by @clackhan in #578
• add graph cache key by @ccssu in #576
• [perf] Remove redundant reshape op by @lixiang007666 in #586
• Modify readme ee for svd by @lixiang007666 in #588
• implement compile_pipe and oneflow_compiler_config by @chengzeyi in #584
• add quant_unet by @ccssu in #592
• Fix Check Failed Error by @ccssu in #585
• rename_diffusers_extension_to_onediffx by @clackhan in #593
• Update onediffx README.md by @strint in #596
• fix env var by @chengzeyi in #598
• Dev support instant id by @chengzeyi in #597
• Update setup.py by @lixiang007666 in #606
• Refactor dualmodule._modules by @marigoold in #604
• add repr of dualmodule by @marigoold in #610
• Update enterprise readme for comfy svd by @lixiang007666 in #608
• [feat] SVD dynamic shape for comfyui by @lixiang007666 in #600
• Fix svd for comfy by @lixiang007666 in #611
• add attention_allow_quantization option by @chengzeyi in #612
• Diffusers ext LoRA support text encoder by @marigoold in #587
• reduce transpose operations to speedup by @hjchen2 in #589
• use EulerAncestralDiscreteScheduler for instant_id and text_to_image by @chengzeyi in #623
• Fix duplicate compilation by @ccssu in #622
• point DualModule buffers and params to torch_module by @marigoold in #617
• fix_readme by @clackhan in #625
• Supporting obj is not an object. by @ccssu in #621
• fix bug of dual module (setattr, and compatible with DualModule input) by @marigoold in #613
• Black config by @strint in #627
• Fix typo installing onediff_diffusers_extensions by @jackalcooper in #632
• fix compatibility with peft and diffusers 0.26.1 by @chengzeyi in #626
• diffusers tests run with vm by @strint in #634
• update readme with news by @strint in #638
• Fix svd deepcache dynamic shape by @lixiang007666 in #635
• support save graph which has dataclass output and add reverse_proxy_class by @strint in #633
• Refine SVD test example by @lixiang007666 in #639
• remove sdxl turbo custom scheduler to refine quality by @doombeaker in #614
• Run turbo in examples by @jackalcooper in #637
• Revert "Run turbo in examples" by @jackalcooper in #640
• fix no module named 'onediffx.utils' error by @chengzeyi in #624
• fix diffusers>=0.26 compatibility for deepcache/onediffx by @isidentical in #641
• [Fix] deepcache uncommon resolution base by @lixiang007666 in #620
• mv examples to diffusers extension by @strint in #643
• Enhance debug of transform error by @strint in #645
• Reuse compiled components betwent pipelines by @strint in https://github.com/silicon...

0.12.0(Enable vm runtime as default)

onediff 0.12 (change image size will have no cost for most of the models now)

• switch image size will not trigger re-compilation(i.e. no time cost);
• faster to save and load a graph;
• smaller static memory needed;
• much easier to switch graphs;

Note

1. After updating onediff to 0.12, oneflow also needs to be updated to the latest with https://github.com/siliconflow/onediff?tab=readme-ov-file#1-install-oneflow
2. If you are using save/load graph. Please use this new version to save the graph again.
3. For SVD(Stable Video Diffusion) onediff users, please use os.environ["VM_REBUILD_DYNAMIC_SHAPE"] = "1" to support DYNAMIC SHAPE with SVD. There has some re-compile cost, when enabling VM_REBUILD_DYNAMIC_SHAPE, so loading graph is needed. We will do more work on this.

OneDiff SDXL inference has been integrated into fal.ai playground

SDXL inference at the speed of thought: https://www.fal.ai/models/onediff-sdxl

And OneDiff ComfyUI nodes are in ComfyUI-Manager Now

0.11.4(Support SD web UI and update performance)

• Release new performance data on SDXL.
• Support SD web UI SD and SDXL acceleration.
• Add SVD/DeepCache diffusers pipeline and optimize performance.

What's Changed

• Fix int8 lora by @doombeaker in #431
• Fix deepcache for svd by @lixiang007666 in #438
• Update onediff_comfy_nodes README by @ccssu in #439
• Update README.md by @eltociear in #440
• Fix diffusers 0 25 dev by @ccssu in #444
• add_deep_cache_with_lora_workflow by @clackhan in #442
• fix reamde typos by @doombeaker in #455
• support mode parameter in PyTorch 2.1 by @haoyang9804 in #458
• use fused_multi_head_attention_inference_v2 to implement CrossAttention by @fpzh2011 in #459
• Add OneDiffCheckpointLoader by @ccssu in #457
• Add onediffcheckpointloadersimple refine by @ccssu in #460
• Fix SVD VAE compile for diffuser by @lixiang007666 in #461
• Feature/svd support by @chengzeyi in #454
• Update README.md for SVD by @lixiang007666 in #462
• Revert "Fix SVD VAE compile for diffuser" by @hjchen2 in #464
• Integrate deep cache diffusers pipeline by @clackhan in #450
• support v1-5-pruned model in sd-webui by @fpzh2011 in #463
• use object.setattr in DualModule and DeployableModule to avoid unexpected key in self._modules by @marigoold in #468
• fix an op convertion error by @haoyang9804 in #469
• Quick fix vae load by @strint in #470
• add perf chart by @strint in #474
• Revert "use object.setattr in DualModule and DeployableModule to avoid unexpected key in self._modules" by @marigoold in #476
• update_onediff_comfy_nodes_README by @ccssu in #477
• Sd Web Unet Model Quantization and Black Formatting by @ccssu in #445
• Add a notebook showing how to save graph locally and load from it by @kopyl in #475
• prefer cudnn convolution by @chengzeyi in #480
• Fix lora parameter update by @ccssu in #479
• Migrate to new CI deployments by @jackalcooper in #482
• enable dual gemm glu by @chengzeyi in #481
• Opt deep cache sdxl pipeline by @clackhan in #485
• Change default env val by @clackhan in #487
• add vae compile to sd-webui by @fpzh2011 in #473
• fix read me of webui by @fpzh2011 in #489
• Update README.md for connection by @strint in #492

New Contributors

• @eltociear made their first contribution in #440
• @haoyang9804 made their first contribution in #458
• @chengzeyi made their first contribution in #454
• @kopyl made their first contribution in #475

Full Changelog: 0.11.3...0.11.4

0.11.3(DeepCache, SVD supported)

What's Changed

• explain quantization feature by @doombeaker in #402
• Update_onediff_comfy_nodes_readme by @ccssu in #404
• Update README.md by @strint in #405
• Add SVD compile node by @lixiang007666 in #409
• Dev support controlnet in onediff comfy nodes by @ccssu in #403
• remove rewrite_self_attention by @hjchen2 in #412
• Fuse LoRA in examples/text_to_image_sdxl_lora.py by @marigoold in #386
• all dynamic shape by @strint in #411
• add image to image vae compile demo by @strint in #406
• add comfyui quantization tool node by @doombeaker in #413
• Run tests regularly by @jackalcooper in #417
• add sd-webui extension by @fpzh2011 in #401
• Compatibility with the latest ComfyUI update (commit b454a67) by @ccssu in #418
• Update README.md by @strint in #419
• add context manager TensorInplaceAssign by @marigoold in #396
• add sd-webui performance data; fix refiner model bug by @fpzh2011 in #420
• Add badge for regular tests by @jackalcooper in #421
• add mock class for specific workflow by @doombeaker in #416
• update benchmark scripts by @hjchen2 in #425
• add oss model url by @doombeaker in #429
• update benchmark readme by @hjchen2 in #430
• dev_optimize_mock_torch by @ccssu in #410
• Integrate deep cache by @clackhan in #426

Full Changelog: 0.11.2...0.11.3

0.11.2 (ComfyUI, ControlNet and SDXL Turbo)

Speedup for ComfyUI, ControlNet, and SDXL Turbo.

0.11.1 is released

• Add more examples.
• Some code refactoring and bug fixes.

0.11.0 is released

• New performance data.
• Added support for SD2.1 and SDXL.
• New plug-and-play compilation interface oneflow_compile.