#!/bin/bash
#SBATCH --job-name=EHD_v2_loyo
#SBATCH --nodes=5
#SBATCH --exclusive
#SBATCH --time=06:00:00
#SBATCH --output=/home/tanpv/Progs/claude/QG_1/logs/loyo_v2_%j.out
#SBATCH --error=/home/tanpv/Progs/claude/QG_1/logs/loyo_v2_%j.err
set -euo pipefail
ROOT=/home/tanpv/Progs/claude/QG_1
PY=/home/tanpv/.conda/envs/ehd/bin/python
# torch/sklearn need the conda libstdc++ ahead of the system one (no CXXABI_1.3.9 in /lib64)
export LD_LIBRARY_PATH=/home/tanpv/.conda/envs/ehd/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}
cd "$ROOT"
mkdir -p logs outputs/loyo_pred_v2

CUTOFFS=${CUTOFFS:-"1.0 0.90 0.99"}
CAP=${CAP:-200000}
YEARLIST=$(seq 2003 2022)                       # 20 LOYO folds
NFOLD=$(echo $YEARLIST | wc -w)

# Each fold is one srun task on 5 cores; 4 tasks per node x 5 nodes = 20 folds at once.
# Bound each task's math libs to its 5 cores so tasks don't oversubscribe the node.
export OMP_NUM_THREADS=5 MKL_NUM_THREADS=5 OPENBLAS_NUM_THREADS=5

echo "=== v2 LOYO: $NFOLD folds | cutoffs=[$CUTOFFS] cap=$CAP | 5 nodes x 4 tasks x 5 cpus ==="
srun --nodes=5 --ntasks=$NFOLD --ntasks-per-node=4 --cpus-per-task=5 \
     --kill-on-bad-exit=1 bash -c \
     "YEARS=($YEARLIST); YR=\${YEARS[\$SLURM_PROCID]}; \
      echo \"[proc \$SLURM_PROCID @ \$(hostname)] fold \$YR start \$(date +%T)\"; \
      $PY scripts/12_run_loyo_v2.py --folds \$YR --cutoffs $CUTOFFS --train-cap $CAP; \
      echo \"[proc \$SLURM_PROCID] fold \$YR done \$(date +%T)\""

echo "=== all folds returned; pred files: $(ls outputs/loyo_pred_v2/pred_v2_*.npz 2>/dev/null | wc -l)/$NFOLD ==="
echo "=== running stage 13 verification on the pooled dataset ==="
$PY scripts/13_verify_v2.py
echo "=== done $(date +%T) ==="