ai-toolkit · Issues· 130 open

Dataset caches thumbnail previews even when replacing them

MiniMax-H3 ref2va training: clean hang on first step with triton; OOM at load without triton (RTX 5090 32GB, Windows, torch 2.14+cu130)

`split_model_over_gpus` + LoRA training on multi-GPU: five independent device/signature bugs break FLUX LoRA runs (16 GB cards)

One stuck `running` job row permanently clogs the queue — worker trusts exit code 0 and never reaps dead/hung jobs

ModuleNotFoundError: No module named 'extensions_built_in.llm_models' when initializing job

Auto-captioning audio fails with 'NoneType' object has no attribute 'device' in low-VRAM mode

Bug Report: Ideogram 4 training fails during transformer dequantization

H3 unable to train LORA either OOM or error

Krea 2: q/k/v dtypes can diverge under gradient checkpointing, breaking backward

Krea 2 stopped training and gives noise since 10 September

ai-toolkit 0.3.6 starts ok - New job start - Lora creation - Krea-2 (raw) - but it is VERY slow. Sometimes it works. No changes.

"Bug" - memory leak

EMA shadow weights diverge — `sub_` changed to `add_` in commit 4a41f14

ever since the update z-image lora training stopped working

Windows: stale cron worker survives UI shutdown, breaks next startup with EPERM on prisma generate