Fix: list_repo_files returns strings directly, skip Phase A correctly

Persist phase state via model repo files — skip completed phases on restart

Add 8 workers for parallel augmentation + prefetch_factor=4

Fix OOM: preload as fp16 (13GB) + pre-allocate tensors

Preload all data into RAM: zero I/O wait, num_workers=0

Boost: batch=32, lr=1e-3, num_workers=8, cudnn.benchmark, resume from prev best

Fix: total_mem -> total_memory attribute

Fix Dockerfile: use nvidia/cuda base + pip install torch

Fix: use mounted volume + add progress web server on 7860

Upload training pipeline (Dockerfile + train_hf.py)

initial commit