Running nanoGPT Fineweb Training in tmux | Generated by AI

Home 2026.05

cd /mnt/data/nanoGPT
tmux new -s nano        # or: tmux attach -t nano

# do NOT activate the venv β€” use system python3.11
PYTORCH_ALLOC_CONF=expandable_segments:True \
python3.11 -u train.py config/train_fineweb_gpt3.py --compile=False 2>&1 \
  | tee -a out-fineweb-gpt3/train.log

Detach with Ctrl-b d, reattach with tmux attach -t nano.

Resume after kill/restart: append --init_from=resume.

Expected: ~11s/iter eager β†’ ~59 hrs for the full 19073-iter run, checkpoint every 500 iters.


Back Donate