| #SBATCH --job-name=hf_ckpt_sweeper | |
| #SBATCH --mail-type=FAIL,END | |
| #SBATCH --mail-user=dingqy@umich.edu | |
| #SBATCH --nodes=1 | |
| #SBATCH --ntasks-per-node=1 | |
| # SA ckpts are ~18 GB and huggingface_hub buffers during upload — need | |
| # enough RAM. Memory rule: always 11 GB/cpu; scale CPU to reach the target | |
| # memory. 6 CPU × 11 GB = 66 GB total. | |
| #SBATCH --cpus-per-task=6 | |
| #SBATCH --mem-per-cpu=11GB | |
| #SBATCH --time=48:00:00 | |
| # CPU-only sweeper — uses lsa3 (not ahowens1) so it doesn't consume the AI | |
| # lab's GPU allocation budget on a job that needs zero GPUs. | |
| #SBATCH --account=lsa3 | |
| #SBATCH --partition=standard | |
| #SBATCH --output=/nfs/turbo/coe-ahowens-nobackup/dingqy/hf_ckpt_sweeper-%j.log | |
| # | |
| # CPU-only background sweeper: uploads training ckpts to HF dataset | |
| # AE-W/ckpt, then deletes the local copy. Keeps last.ckpt local so SLURM | |
| # preemption / requeue can still resume. Run alongside the 4 training jobs. | |
| # | |
| set -eu | |
| source /home/dingqy/.hf_token | |
| PY=/nfs/turbo/coe-ahowens-nobackup/dingqy/miniforge3/envs/V2A/bin/python | |
| "$PY" /nfs/turbo/coe-ahowens-nobackup/dingqy/hf_ckpt_sweeper.py | |