#!/bin/bash
# PostTrainBench one-time setup on this Slurm cluster: container images, then the full upstream
# HuggingFace cache (14 models + 382 dataset splits, hundreds of GB, hours).
#   sbatch scripts/ptb_setup.sbatch                  # images + full cache
#   FULL_CACHE=0 sbatch scripts/ptb_setup.sbatch     # images only (no HF_TOKEN needed)
#
# Apptainer only exists on the compute nodes, and the compute nodes reach the internet directly --
# the inherited X2P proxy on the login node is not routable from them.
# Needs a GPU node: upstream's defs install vllm with `uv pip install --torch-backend=auto`, which
# resolves to the CPU wheel index (and then fails on xformers) when no NVIDIA driver is visible.
#SBATCH --job-name=ptb-setup
#SBATCH --partition=h100
#SBATCH --account=midpri
#SBATCH --qos=midpri
#SBATCH --gres=gpu:1
#SBATCH --time=04:00:00
#SBATCH --cpus-per-task=16
#SBATCH --mem=256G
#SBATCH --output=logs/ptb-setup-%j.out

set -euo pipefail
unset http_proxy https_proxy HTTP_PROXY HTTPS_PROXY no_proxy NO_PROXY X2P_PROXY_URL

cd /home/jiaruiliu/local/git-repos/auto-impl
export PYTHONPATH=src

echo "node=$(hostname) apptainer=$(apptainer --version)"
[[ "${FULL_CACHE:-1}" == "1" ]] && CACHE_FLAG=--full-cache || CACHE_FLAG=
python3 -m posttrainbench setup \
    --containers ${CONTAINERS-opus_5_local vllm_debug} \
    ${CACHE_FLAG} \
    --models ${MODELS-Qwen/Qwen3-1.7B-Base} \
    --tasks ${TASKS-gsm8k}