| #SBATCH --job-name=tm_diffgate | |
| #SBATCH --partition=athena-genai,athena,athena-mini,athena-small | |
| #SBATCH --nodes=1 | |
| #SBATCH --ntasks=1 | |
| #SBATCH --cpus-per-task=2 | |
| #SBATCH --mem=8G | |
| #SBATCH --time=00:10:00 | |
| #SBATCH --output=/home/tl356/LLaDA/logs/%j_tm_diffgate.out | |
| #SBATCH --error=/home/tl356/LLaDA/logs/%j_tm_diffgate.err | |
| # | |
| # HARD GATE between compression and eval. Runs summary_diff.py on the two | |
| # all-linear r=0.7 summaries; exits non-zero on any mismatch beyond noise fields. | |
| # Downstream BASE/OURS eval jobs are submitted with --dependency=afterok:<this>, | |
| # so a FAIL here auto-cancels them (Slurm never runs eval on a failed gate). | |
| # sbatch --dependency=afterok:<Jbase_all>:<Jours_all> jobs/diff_gate.sbatch weights/base/all_r07 weights/ours/all_r07 | |
| set -euo pipefail | |
| source /home/tl356/miniconda3/etc/profile.d/conda.sh | |
| conda activate llada | |
| cd /home/tl356/LLaDA/trajmc_main | |
| BASE=${1:-weights/base/all_r07} | |
| OURS=${2:-weights/ours/all_r07} | |
| echo "node=$(hostname) diff_gate base=$BASE ours=$OURS start=$(date)" | |
| python analysis/summary_diff.py --base "$BASE" --ours "$OURS" | |
| rc=$? | |
| echo "diff_gate exit=$rc (0=PASS -> eval proceeds; nonzero=FAIL -> eval cancelled) $(date)" | |
| exit $rc | |