Training, Optimizers, Device Handling & Z-Image Improvements#713

Draft

daanforever wants to merge 17 commits intoostris:mainfrom

daanforever:pr-main

Commits on Feb 13, 2026

Commits on Feb 17, 2026

fix(optimizers): adjust max_lr and improve learning rate clamping
daanforever
committed
refactor(z_image): streamline embedding conversion and device management
daanforever
committed
fix(train): free memory before checkpoint save to reduce OOM risk
daanforever
committed
fix(peft): apply alpha key fix for all peft types when loading LoRA
daanforever
committed
refactor(dataloader): remove redundant token shuffling logic
daanforever
committed
refactor(dataloader): remove redundant token shuffling logic
daanforever
committed
fix(adafactor): apply new min_lr/max_lr on restart after loading checkpoint
daanforever
committed
refactor(paths): rename normalize_model_path to normalize_path and strip whitespace
daanforever
committed
feat(debug): add memory_debug util and use for CUDA load/unload logging
daanforever
committed
fix(sampling), feat(lora): sampling VRAM, single LoRA, path and debug wiring
daanforever
committed
Merge branch 'pr-main' of github.com:daanforever/ai-toolkit into pr-main
daanforever
committed