#!/usr/bin/env bash
# Sequential per-checkpoint, 3-shard parallel per checkpoint.
set -e  # no -u to avoid conda conflict
REPO=/home/yurenh2/rrm/research/flossing
CKPT_ROOT="/home/yurenh2/rrm/hrm/checkpoints/Sudoku-extreme-1k-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV1 righteous-python"
source "$(conda info --base)/etc/profile.d/conda.sh"
conda activate rrm
cd "$REPO"

N=1024
K=8
mkdir -p ckpt_evolution

CKPTS=(step_2604 step_7812 step_13020 step_18228 step_20832)

for ckpt in "${CKPTS[@]}"; do
  if [[ -f "ckpt_evolution/${ckpt}.npz" ]]; then echo "skip $ckpt (merged exists)"; continue; fi
  echo "==> $ckpt"
  pids=()
  for shard in 0 1 2; do
    LOG=ckpt_evolution/${ckpt}_shard${shard}.log
    OUT=ckpt_evolution/${ckpt}_shard${shard}.npz
    if [[ -f "$OUT" ]]; then echo "  skip $OUT"; continue; fi
    nohup env CUDA_VISIBLE_DEVICES=$shard python diagnose_hrm.py \
      --ckpt-root "$CKPT_ROOT" --ckpt-name $ckpt \
      --n-samples $N --num-shards 3 --shard-id $shard \
      --batch-size 64 --k-lyap $K \
      --out "$OUT" > "$LOG" 2>&1 &
    pids+=($!)
  done
  for pid in "${pids[@]}"; do wait $pid; done
  echo "<== $ckpt done"
done

# Final merge per checkpoint
python - <<'PY'
import numpy as np, glob, os
out_dir = "/home/yurenh2/rrm/research/flossing/ckpt_evolution"
ckpts = ["step_2604","step_7812","step_13020","step_18228","step_20832","step_26040"]
for c in ckpts:
    files = sorted(glob.glob(f"{out_dir}/{c}_shard*.npz"))
    if not files:
        print(f"missing {c}"); continue
    m = {}
    for f in files:
        d = np.load(f)
        for k in d.files: m.setdefault(k, []).append(d[k])
    for k in list(m.keys()): m[k] = np.concatenate(m[k], 0)
    out = f"{out_dir}/{c}.npz"
    np.savez_compressed(out, **m)
    s = m['exact_correct']>0.5
    print(f"{c}: N={len(m['exact_correct'])} acc={m['exact_correct'].mean():.4f} "
          f"λ_1(s)={m['lyap_spec'][s,0].mean():+.3f} "
          f"λ_1(f)={m['lyap_spec'][~s,0].mean():+.3f}")
PY