Upload train.py with huggingface_hub

Bulletproof probe: simulate grad_accum, expandable_segments, 70% safety

Fix: probe with return_logits=False (was wasting 3GB on logits tensor), show peak VRAM

Fix probe: use actual model (no duplicate), add GPU cleanup at start

Maximize GPU usage: probe up to 4096 batch, 92% safety, only catch OOM

Disable torch.compile entirely - causes OOM/crash at 100M scale

Fix: use compile mode=default to avoid CUDA graph + tied embedding crash

Fix torch.compile CUDA graph training on 5090

LUNA 100M: cloud-ready training pipeline