What's Changed

• ema: offload to cpu, update every n steps by @bghira in https://github.com/bghira/SimpleTuner/pull/517
• ema: move correctly by @bghira in https://github.com/bghira/SimpleTuner/pull/520
• EMA: refactor to support CPU offload, step-skipping, and DiT models
• pixart: reduce max grad norm by default, forcibly by @bghira in https://github.com/bghira/SimpleTuner/pull/521
• remove incorrect log line when using cpu offload for ema by @bghira in https://github.com/bghira/SimpleTuner/pull/523
• add --dataloader_prefetch option for speed-up by @bghira in https://github.com/bghira/SimpleTuner/pull/535
• parquet metadata: retrieve captions near-instantly at startup
• multi-gpu: logging cleanup, performance fixes
• settings: no invisible default minimum_image_size
• text embed cache fix for repeatedly attempting to write files that already exist
• text embed cache fix for AWS prefixes not being referenced correctly
• embed cache (text & vae) compression via --compress_disk_cache by @bghira in https://github.com/bghira/SimpleTuner/pull/540

Full Changelog: https://github.com/bghira/SimpleTuner/compare/v0.9.7.4...v0.9.7.5