#!/bin/bash
#SBATCH --job-name=EHD_v3zoo
#SBATCH --nodes=5
#SBATCH --exclusive
#SBATCH --time=08:00:00
#SBATCH --output=/home/tanpv/Progs/claude/QG_1/logs/v3zoo_%j.out
#SBATCH --error=/home/tanpv/Progs/claude/QG_1/logs/v3zoo_%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_v3zoo

CAP=${CAP:-200000}
INNER=${INNER:-4}
YEARLIST=$(seq 2003 2022)                       # 20 LOYO folds
NFOLD=$(echo $YEARLIST | wc -w)

# Each fold = one srun task on 5 cores; 4 tasks/node x 5 nodes = 20 folds at once.
# Pin every math/tree library to the 5 cpus of its task so tasks don't oversubscribe the node.
export OMP_NUM_THREADS=5 MKL_NUM_THREADS=5 OPENBLAS_NUM_THREADS=5 EHD_NJOBS=5

echo "=== v3zoo LOYO: $NFOLD folds | cap=$CAP inner=$INNER | 5 nodes x 4 tasks x 5 cpus ==="
srun --nodes=5 --ntasks=$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/21_run_loyo_v3zoo.py --folds \$YR --cap $CAP --inner $INNER; \
      echo \"[proc \$SLURM_PROCID] fold \$YR done \$(date +%T)\""

echo "=== all folds returned; pred files: $(ls outputs/loyo_pred_v3zoo/pred_v3zoo_*.npz 2>/dev/null | wc -l)/$NFOLD ==="
echo "=== stage 25 whole-dataset verification (per-family binary inventory) ==="
$PY scripts/25_verify_v3zoo.py
echo "=== stage 23 diversity-driven lead-dependent stack ==="
$PY scripts/23_v3_diversity_stack.py
echo "=== done $(date +%T) ==="