#!/bin/bash
#SBATCH --job-name=sa_test_loss_sweep
#SBATCH --mail-type=FAIL,END
#SBATCH --mail-user=dingqy@umich.edu
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=4
#SBATCH --mem-per-cpu=11GB
#SBATCH --time=4:00:00
#SBATCH --account=ahowens1
#SBATCH --partition=spgpu
#SBATCH --qos=foe
#SBATCH --reservation=spgpu_foe_nodes
#SBATCH --gres=gpu:a40:1
#SBATCH --output=/nfs/turbo/coe-ahowens-nobackup/dingqy/sa_test_loss_sweep-%x-%j.log
#
# Test-loss sweep across the full lbkb1h5z + dbdxldk4 chain. Sharded into 4
# parallel jobs (one GPU each) to fit walltime; merge JSONs when all done.
#
# Submit one shard at a time:
#   SHARD=1 sbatch eval_sa_test_loss.sbat   # lbkb step 1k-9k
#   SHARD=2 sbatch eval_sa_test_loss.sbat   # lbkb step 10k-18k
#   SHARD=3 sbatch eval_sa_test_loss.sbat   # lbkb step 19k-26k
#   SHARD=4 sbatch eval_sa_test_loss.sbat   # dbdxldk4 step 27k-37k
set -eu

source /nfs/turbo/coe-ahowens-nobackup/dingqy/miniforge3/etc/profile.d/conda.sh
source /home/dingqy/.hf_token
export HF_CACHE=/nfs/turbo/coe-ahowens-nobackup/dingqy/.cache/huggingface

conda activate stable-audio

: "${SHARD:=1}"

# 11k ckpts split into 4 shards; one GPU per shard, ~3 hr each.
build_lbkb() { for s in "$@"; do echo "sa_open_bg2fg_rebalance/lbkb1h5z/epoch=0-step=${s}.ckpt"; done; }
build_dbdx() { for s in "$@"; do echo "sa_open_bg2fg_rebalance/dbdxldk4/epoch=0-step=${s}.ckpt"; done; }

case "$SHARD" in
    1) CKPTS=$(build_lbkb 1000 2000 3000 4000 5000 6000 7000 8000 9000) ;;
    2) CKPTS=$(build_lbkb 10000 11000 12000 13000 14000 15000 16000 17000 18000) ;;
    3) CKPTS=$(build_lbkb 19000 20000 21000 22000 23000 24000 25000 26000) ;;
    4) CKPTS=$(build_dbdx 27000 28000 29000 30000 31000 32000 33000 34000 35000 36000 37000) ;;
    *) echo "unknown SHARD=$SHARD (use 1/2/3/4)"; exit 1 ;;
esac

OUT=/nfs/turbo/coe-ahowens-nobackup/dingqy/sa_test_loss_shard${SHARD}.json
echo "shard=$SHARD  out=$OUT"
echo "$CKPTS"

python /nfs/turbo/coe-ahowens-nobackup/dingqy/eval_sa_test_loss.py \
    --ckpts $CKPTS \
    --out "$OUT" \
    --batch-size 8 --num-workers 4

echo ""
echo "=== shard $SHARD final ==="
python -c "
import json
r = json.load(open('$OUT'))
for k, v in sorted(r.items(), key=lambda x: x[1]['test_loss']):
    print(f'  {v[\"test_loss\"]:.5f}  {k}')
"