#!/usr/bin/env bash
# Run the MSQ battery over a list of models, one at a time, waiting for the
# GPU to be free between runs.
set -u
cd /home/adeerkhan/github/fatima
source .venv/bin/activate
export HF_TOKEN="$(cat "$HOME/.config/fatima/hf_token")"
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

wait_for_gpu () {
  while pgrep -f 'run_model.py' > /dev/null 2>&1; do sleep 15; done
  sleep 10
}

for m in "$@"; do
  wait_for_gpu
  echo "[$(date +%T)] running $m"
  python code/run_model.py --model "$m" --out "results/raw/$m.jsonl" \
      --load-in-4bit --batch-size 8 --max-new-tokens 280 \
      > "logs/run_$m.log" 2>&1
  echo "[$(date +%T)] finished $m rc=$?"
done
echo "[$(date +%T)] QUEUE COMPLETE"