guard against exploded loss (loss_cap/pred_cap) + adamw eps 1e-6

set Attention.upcast_attention=True per-module (method removed in 0.31)

fix NaN root cause: upcast attention to fp32 (AttnProcessor); weight-integrity guards before any upload

drop GradScaler (fp16 grads unsupported); manual finite-grad skip

train_continue.py: NaN-hardening (fp32 VAE, autocast UNet, GradScaler skip-step)

train_continue.py: DATA_DIR env for images (keep Kaggle output small)

train_continue.py: DATA_DIR env for images (keep Kaggle output small)

train_continue.py: DATA_DIR env for images (keep Kaggle output small)

train_continue.py: download images.tar.gz archive (avoid HF 429 rate-limit on 20k files); resolve Indoor/Outdoor casing

train_continue.py portable (RUN_DIR, kaggle secret token, RUN_DIR output dirs)

upload train script train_continue.py