v3.0.6 - multiGPU bugfixes, make dataloader less spiteful toward users

What's Changed

• hidream: fix EmbedND return value to match what apply_rope expects by @bghira in #1870
• fix HiDream test by @bghira in #1872
• merge by @bghira in #1871
• update detection for conditioning data blocks to look for auto-generator by @bghira in #1874
• add hf_transfer as base dependency for scheduler download issue by @bghira in #1873
• validation: check whether the declared eval_dataset_id exists before trying to use it by @bghira in #1876
• collate: fix issue where conditioning captions are incorrectly used when just a single training set exists by @bghira in #1875
• disable build maximisation step since we now use less space anyway by @bghira in #1877
• detect when a split nukes the dataset contents and restore from snapshot by @bghira in #1869
• merge by @bghira in #1879

Full Changelog: v3.0.5...v3.0.6

v3.0.5 - bugfix release

What's Changed

• integrate lora_alpha fully with diffusers metadata collection by @bghira in #1853
• fix qwen edit conditioning settings not appearing by @bghira in #1854
• add dataset file browser button to instance data directory field, allowing direct selection of the path by @bghira in #1855
• add /api/version endpoint and display it in the UI sidebar-header by @bghira in #1856
• break dataloader template down into subsection templates by @bghira in #1857
• disable moving of text encoders when group offload enabled by @bghira in #1859
• (#1776) handle another edge case where repeats=0 on cond set, repeats>0 on training set. duplicate settings at startup. by @bghira in #1858
• refactor group offload to target transformer only and not call .to() after enabled by @bghira in #1861
• [wip] self-forcing distillation by @bghira in #1847
• wan group offloading fix for backward pass; move temb to the block device before applying scale/shift by @bghira in #1863
• fix bug when loading predefined conditioning_image_embeds sets by @bghira in #1862
• merge by @bghira in #1864
• (#1860) configure Dynamo for BnB models dynamic outputs by @bghira in #1865
• Add non-Python files to SimpleTuner wheel. by @hameerabbasi in #1866
• resolve bug where duplicating conf environment does not copy the whole subfolder by @bghira in #1867
• Release 3.0.5 by @bghira in #1868

New Contributors

• @hameerabbasi made their first contribution in #1866

Full Changelog: v3.0.4...v3.0.5

v3.0.4 - validation adapters, Chroma masking fix, UI fixes

What's Changed

• (#1823) add a Shutdown button and API endpoint that kill the trainer by @bghira in #1837
• (#1811) disable prompt management options when i2v is selected by @bghira in #1841
• (#1808) add options for validation adapters and fix bug preventing validation page from rendering by @bghira in #1843
• import change from Diffusers upstream for Chroma masking fixes by @bghira in #1844
• add note about multi-adapter validations by @bghira in #1845
• wan i2v refactor: auto-create conditioning dataset when is_i2v is true on a dataset with no conditioning data defined by @bghira in #1846
• refactor model_families and model_flavour retrieval for better usage in simpletuner configure TUI by @bghira in #1848
• correct detection of video / i2v configuration on dataset page by @bghira in #1849
• dataset builder JSON mode fix, do not run stringify on dirty marker when in json mode by @bghira in #1850
• revamp dataset view, persistent collapsibles, slightly better layout by @bghira in #1851
• Release 3.0.4 by @bghira in #1852

Full Changelog: v3.0.3...v3.0.4

v3.0.3 - Pony v7 preset, Qwen Edit, web interface improvements

What's Changed

• ui: adjust interface state sync with backend by @bghira in #1820
• update UI state tracking and fix logs for less spamming by @bghira in #1821
• add qwen edit and edit 2509 support via edit-v1 and edit-v2 flavours by @bghira in #1795
• (#1824) reintroduce missing dataloader config examples by @bghira in #1826
• (#1824) keep example dataloader config by default by @bghira in #1827
• ponyflow: add path and override sequence length to 768 by @bghira in #1829
• (#1825) fix dataloader config subsection styling on Environment page by @bghira in #1828
• event list filtration dropdown by @bghira in #1830
• (#1817) add a Copy button to the event log by @bghira in #1831
• (#1801) fix event list events disappearing upon resize by @bghira in #1832
• (#1804) address some UI state desync when errors emit at startup or disconnection by @bghira in #1833
• (#1807) fix ROCm statistics reporting by @bghira in #1834
• tutorial: update installation source to pip by @bghira in #1835

Full Changelog: v3.0.2...v3.0.3

v3.0.2 - Wan I2V, Chroma bugfixes; Revamped TUI configure tool

What's Changed

• Update WAN_STAGE_OVERRIDES subfolder names by @bghira in #1787
• add webhook wrapper for logging that streams all log-level appropriate events to the UI by @bghira in #1788
• chroma: fix deprecated use of img_ids by @bghira in #1789
• (#1786) wan: monkeypatch execution device getter by @bghira in #1790
• update test coverage for webhooks by @bghira in #1791
• add import-export for config next to environment selector by @bghira in #1792
• deprecate --validation_steps, replace with --validation_step_interval and --validation_epoch_interval by @bghira in #1793
• deprecate --checkpointing_steps, replace with --checkpoint_step_interval and --checkpoint_epoch_interval by @bghira in #1794
• do not complain about webhooks being left unconfigured by @bghira in #1797
• configure script: revive and add missing logic by @bghira in #1798
• adjust logging trapper to not stomp on log_format rules by @bghira in #1800
• fix i2v data cache (wan 2.1) by @bghira in #1799
• resolve conditioning dataset multi-gpu split problem (#1776) by @bghira in #1818
• merge 3.0.2 to release branch by @bghira in #1819

Full Changelog: v3.0.1...v3.0.2

v3.0.1 - Chroma HD, Wan 5B, A14B, Block swapping / group offload

This is considered an essential upgrade for bugfixes.

What's Changed

• Chroma HD by @bghira in #1767
• fix chroma controlnet bugs by @bghira in #1768
• harden config loading for v2.x and earlier datasets that used loose typing by @bghira in #1770
• Qwen Image: fix loading bug by @bghira in #1771
• fixes for rocm pytorch dependency by @bghira in #1772
• detect instance dir and raise error all the way to the web interface by @bghira in #1773
• fix qwen unpack bug by @bghira in #1774
• Wan 2.2 A14B, group offload memory optimisations by @bghira in #1769
• add bucket reporting for zero-length datasets to help understand why it had no images by @bghira in #1782
• UI: cannot unset user prompt library path value by @bghira in #1783
• fix wan2.2 model path by @haoxu2000 in #1784

New Contributors

• @haoxu2000 made their first contribution in #1784

Full Changelog: v3.0.0...v3.0.1

v3.0.0 - web interface, FSDP2 + context-parallel training

SimpleTuner v3

TL;DR

• User‑driven workflows, API‑centric design, and a calmer, smarter training experience.
• Big push on reliability, visibility, and integration: wizards, events, webhooks, CLI, and docs.
• Lays the groundwork for SimpleTuner v3 to truly be simple.

Highlights

• Unified, richer webhook/events system with stable interface that will not change much, moving forward.
• Dataset and training configuration wizards that hold your hand and help make fewer mistakes.
• FSDP2 and DeepSpeed fully integrated into the web interface, with more flexible and complete configuration capability.
• Packaging, CLI, Docker, and documentation updates that make things easier.

WebUI & UX

• Dataset Wizard and Training Configuration Wizard with guided steps.
• Event streaming: lifecycle states, event severity, progress bars, lightbox for convenient viewing of validation images.
• Clean interactivity: collapse/expand sections of the interface with persisted state to disk, ensuring the interfaces stays just as you left it the next time it's loaded.
• Support for custom themes to be expanded upon and fully-fleshed out in the next few releases

Data, Checkpoints & Validation

• Checkpoint overview card with validation previews; lightbox for images.
• Fixes for validation image counts, overwrites, and color profiles.
• Checkpoint deletion, filename/path normalization, and selection widgets.
• Hugging Face uploads: bulk/single with improved feedback and error surfacing.

Training Core

• Robust error detection and graceful recovery/cleanup on abort.
• Improved progress tracking (including text/image embed steps).
• Startup status surfaced in API with startup_progress.
• Safer defaults and guards (optimizers, schedulers, warmup, gradient accumulation).

Distributed & Hardware

• FSDP2 backend support with tests and docs; context parallel option.
• DeepSpeed builder UI and safer defaults; single‑GPU fixes for full-rank tuning big models like Flux on just 24G VRAM
• Hardware route for GPU info to allow displaying it in statistics dashboards or collecting for metrics over time.
• Guardrails for macOS/ROCm that disable or remove incompatible features from the interface to avoid confusion or accidental usage.

Webhooks & Events

• Unified event lifecycle; legacy ad-hoc messages are removed.
• Multi‑webhook backend with list config and SSL support.
• Mandatory readable_type for signal names; richer, raw payloads.
• Hardened against timeouts/connection closure; consistent logging.

API, CLI & Service

• Overhauled FastAPI service with API documentation generator.
• CLI entry point and service worker command added to just allow simpletuner server to launch everything.
• Dataset browse endpoint; improved dataloader factory and registry.
• APIState improvements for in‑context subprocess and progress tracking.

Performance & Stability

• Trainingsample Rust ops integrated for batched transforms (goodbye cv2).
• PyTorch upgraded to 2.8.0 or 2.9.0 depending on platform; TorchDynamo config granularity enhanced, with regional compile now possible to enable.

Security & Networking

• Webhook SSL support with auto‑disable verify if server started with SSL - or bring-your-own certificate and launch with a fully verified chain of authority.
• Env var & accelerate config handling refined; safer pipeline defaults.

Packaging, Install & Docs

• Shift to pip‑first workflows; Docker and notebooks updated accordingly.
• CLI launch recommendations unified across docs.
• FSDP2 and DeepSpeed docs; migration guide (2.1 → 2.2).
• API tutorial and current event list added; lots of errata and readability passes.

Removals & Breaking Changes

• Removed legacy webhook message format; use unified events stream.
• Removed Poetry configs
• No CWD fallback for config loading; be explicit about paths.

Quality & Tests

• Expanded unit/integration tests: webhooks (SSL/JSON/AST), events lifecycle, collapsible sections, hardware, backends, and CLI parsing.
• Better mocks for checkpoint tests; formatting/coverage cleanups.

Upgrade Notes

• Install/update via pip (see README for commands and extras).
• Review WEBHOOKS and Events docs for updated signal names and payloads:
  • documentation/api/WEBHOOKS.md
  • documentation/api/TUTORIAL.md

v2.2.1 - QoL fix for command line

What's Changed

• fix parsing of config.env when using cli tool, and allow passthrough of arguments to trainer by @bghira in #1737

Full Changelog: v2.2.0...v2.2.1

v2.2.0 - project layout overhaul

What's Changed

• Add FastWan (DMD) distillation method by @bghira in #1695
• update transformers+diffusers by @bghira in #1699
• VAECache needs to shuffle bucket contents before splitting across GPUs by @bghira in #1701
• Fix Cosmos2 text to image training by using the right objective and latents by @AmericanPresidentJimmyCarter in #1702
• Fix crash in Cosmos2 latent generation by @AmericanPresidentJimmyCarter in #1703
• add vLLM based captioner script for Qwen-VL 2.5 3B and 7B Instruct by @bghira in #1704
• Fix: DeepSpeed ZeRO Stage 3 Compatibility for Flux by @alexsenden in #1712
• fixes for vllm script by @bghira in #1716
• Expand aspect_bucket_alignment options; allow override with --i_know_what_i_am_doing by @twri in #1713
• Bugfix: wan vae normalisation + deterministic posterior sampling by @squewel in #1722
• wan vae normalisation by @bghira in #1723
• (#1732) sd3 and hidream fix for lora function location by @bghira in #1733
• Rust-based TrainingSample; accelerated, vectorised operations by @bghira in #1734
• add latest trainingsample v0.2.0 for better performance by @bghira in #1735
• simpletuner 2.2: package-friendly layout by @bghira in #1736

Full Changelog: v2.1.3...v2.2.0

v2.1.3 - TREAD softly

Features

• Qwen Image is now supported for finetuning
• TREAD: Uses this CompVis research for a great speed-up and VRAM reduction in FLUX and Wan 2.x models, thanks to American President Jimmy Carter for the great contribution
• Added krea, fluxbooru and libreflux flavour options to Flux, updating the default from dev to krea

Fixes

• Distributed int8 training for torchao is now functional
• Cleaner output on console when training with PNGs, no more iCCP chunk spam
• Cosmos2Image quality improvement by fixing VAE encode
• Aspect ratio errors for square crop configurations are resolved (#1692)
• Fixed end-to-end test runner on MacOS

What's Changed

• pixart: more spots where transformer access needs fixing by @bghira in #1667
• fix tests by @bghira in #1669
• add fluxbooru v0.3 with sensible default overrides by @bghira in #1671
• add neta lumina flavour for anime based lumina2 finetune by @bghira in #1674
• add libreflux flavour defaults by @bghira in #1673
• add TREAD training by @bghira in #1675
• add support for distributed int8 via torchao (experimental) by @bghira in #1676
• disable CLIP eval in default examples, add TREAD example for Flux by @bghira in #1677
• support huggingface datasets for video by @bghira in #1678
• add torchao v0.12 by @bghira in #1679
• add Wan TREAD support and docs by @bghira in #1680
• Add FLUX.1-Krea-dev and make it the default Flux flavour by @bghira in #1686
• when saving model card, cast to str for check by @bghira in #1687
• merge by @bghira in #1690
• remove bad iCCP chunk from PNG data before passing bytes to CV2 (silence libpng console spam) by @bghira in #1691
• include cosmos2image in the posterior sampling classes by @bghira in #1693
• aspect ratio incorrect on square crops by @bghira in #1692
• merge by @bghira in #1694
• fix macos end2end test runner crash by @bghira in #1697
• add qwen image by @bghira in #1696
• merge by @bghira in #1698

Full Changelog: v2.1.2...v2.1.3