Changelog

All notable changes to this project. Format follows Keep a Changelog.

Unreleased — dev branch

SDXL NaN loss (root cause) — _tokenize_tags left overflow/padding chunks with all-False CLIP attention masks. exp(dtype.min) underflows to 0 in bf16/fp32, producing 0/0 = NaN in softmax. Fixed by setting ones_mask_chunk on three overflow paths in strategy_base.py.
min_snr_gamma_enabled missing from 4 presets — Illustrious AdamW8bit, Illustrious CAME Conservative, Pony Prodigy, and pony_template predated the _enabled flag convention.
Custom optimizer crash/corruption bugs — adammini bare param not wrapped in list; dehaze NameError on adaptive_muon path; fmarscrop TypeError; adopt in-place sqrt corruption; laprop alias mutation.
min_snr_gamma injected on every run — default value of 5.0 caused SNR weighting on every run regardless of intent. Now gated on min_snr_gamma_enabled.
disable_cross_attn_mask flag not read — argparse flag existed but was never connected to the training loop.
optimizer_args sent as array instead of string — caused silent arg-parsing failures for custom optimizers.

README restructured — removed conversational copy, expanded documentation links.

CAME NaN (network_train_unet_only default True → False)
Conservative CAME preset fp16+conv_dim:32 → bf16+conv_dim:0
Log cursor wrap-around — deque-relative index replaced with absolute counter
Training log polling — visibility-aware polling with fixed cadence
Log poller race condition at epoch boundary on tab switch
ZIP dataset upload — request body piped instead of buffered (fixes ECONNRESET)
Dataset tags_present flag — tag editor now shows status correctly
LoggingCard Radix fix — log_with uses 'none' sentinel instead of empty string
TOML generation path — config written directly to config/
enable_bucket now respects user config (was hardcoded True)
network_train_unet_only no longer passed in checkpoint training mode
HF upload form — token pre-filled from saved settings
LoRA resize — removed alpha input field (SVD auto-calculates it)
Merge/resize subprocesses now have timeouts
CUDA availability check before --device cuda
CheckpointTrainingConfig TypeScript type includes SD3.5, Chroma, Anima
AbortError console noise on navigation suppressed
Anima presets — clip_skip: 1, keep_tokens: 1 applied across all configs

Initial alpha. Core LoRA training pipeline working end-to-end on VastAI and local NVIDIA GPU.

Web UI (Next.js 15 + FastAPI) for LoRA training on SDXL, SD1.5, Flux, SD3/SD3.5, Lumina
Dataset upload, WD14/BLIP/GIT auto-tagging, caption editor
Training configuration with preset system
Civitai model browser and HuggingFace upload
LoRA merge, resize, and checkpoint merge utilities
VastAI one-click deploy