v4.10.0 - high-quality MOSS-Image & Flux 3 Klein training

Features

• MOSS-Image with VRAM-optimised configuration examples
• Flux 3 Klein Image with a distillation training adapter (assistant LoRA)
• KohakuFA attention is now set by default when attention-cute extra is installed

Bugfixes

• MultiGPU (FSDP2 and DDP) training correctness fixes
• WebUI metrics display and collection fixes
• VAE caching should be a little faster now with large datasets

What's Changed

• Fix Qwen-Image 2.1 FSDP metadata after model selection by @hjinnkim in #3258
• Fix rank-dependent VAE bucket geometry by @hjinnkim in #3254
• Fix redundant image resizing during VAE caching by @hjinnkim in #3255
• Document the LyCORIS 4 LoKR kernel-size workaround by @bghira in #3259
• Fix training reports with large metric and timestep histories by @bghira in #3260
• Keep Qwen Image VAE attention independent of denoiser kernel selection by @bghira in #3263
• Normalize text-file captions before training cache lookups by @bghira in #3262
• Preserve sparse evaluation metrics when downsampling training records by @bghira in #3261
• Add ComfyUI export for fused Krea2 full-matrix LoKR and DoRA by @bghira in #3264
• Support value gradient clipping under FSDP2 by @hjinnkim in #3265
• Fix LoRA and LyCORIS training under DDP and FSDP2 by @hjinnkim in #3266
• Fix LR scheduler position being overwritten on resume by @hjinnkim in #3267
• Require CaptionFlow 0.5.7 and PyTorch 2.13 by @bghira in #3269
• Keep container device-list paths out of CUDA_VISIBLE_DEVICES by @bghira in #3270
• Make automatic Kohaku attention the training default by @bghira in #3256
• Add FLUX.3 Image Klein training and image editing support by @bghira in #3271
• Preserve model-card base identity for local checkpoints by @bghira in #3272
• Fix Flux.3 Image Klein segmented checkpointing and add GPU measurements by @bghira in #3274
• Use the released FLUX.3 Klein training assistant by default by @bghira in #3275
• Add MOSS-Image Base and SFT training support by @bghira in #3273
• merge by @bghira in #3276

Full Changelog: v4.9.3...v4.10.0