playground/inference/watcher.sh

#!/bin/bash
# Orchestrator: runs eval sequentially on checkpoints, keeps GPUs busy,
# merges results, then shuts down MLX.
#
# Pipeline:
#   1. Wait for current ckpt-1700 eval (PIDs 13902-13909) to finish
#   2. Merge ckpt-1700 shards
#   3. Check if ckpt-5950 eval already done; if not, launch 8-GPU eval
#   4. Wait for ckpt-5950 eval to finish, then merge
#   5. Kill MLX session

PYTHON=/mlx/users/jiashuo.fan/miniconda3/envs/abbie/bin/python3
DATA_DIR=/mnt/hdfs/byte_tt_data_cu_vagcp/haogeng.liu/new_policy7w_v2_reformat
EXP_BASE=/mnt/bn/bohanzhainas1/jiashuo/exp
LOG_DIR=/mlx/users/jiashuo.fan/playground/inference/logs
EVAL_SCRIPT=/mlx/users/jiashuo.fan/playground/inference/eval.py
MERGE_SCRIPT=/mlx/users/jiashuo.fan/playground/inference/merge_results.py
MLX_PID=273

exec >> "$LOG_DIR/watcher.log" 2>&1
echo "=============================="
echo "[watcher] Started: $(date)"

# ── helpers ────────────────────────────────────────────────────────────────────

wait_pids() {
    local pids=("$@")
    echo "[watcher] Waiting for PIDs: ${pids[*]}"
    while true; do
        local alive=0
        for pid in "${pids[@]}"; do
            kill -0 "$pid" 2>/dev/null && alive=1 && break
        done
        [ $alive -eq 0 ] && break
        sleep 30
    done
    echo "[watcher] All done: $(date)"
}

merge_shards() {
    local prefix="$1"
    echo "[watcher] Merging → $prefix"
    $PYTHON -u "$MERGE_SCRIPT" --prefix "$prefix" --num-shards 8
    echo "[watcher] Merged: $(date)"
}

launch_eval() {
    local ckpt="$1"
    local out_prefix="$2"
    local ckpt_name
    ckpt_name=$(basename "$ckpt")
    echo "[watcher] Launching 8-GPU eval for $ckpt_name"
    local pids=()
    for GPU_ID in 0 1 2 3 4 5 6 7; do
        local log="$LOG_DIR/eval_${ckpt_name}_gpu${GPU_ID}.log"
        CUDA_VISIBLE_DEVICES=$GPU_ID $PYTHON -u "$EVAL_SCRIPT" \
            --model-path "$ckpt" \
            --data-dir   "$DATA_DIR" \
            --gpu-id     $GPU_ID \
            --shard-id   $GPU_ID \
            --num-shards 8 \
            --output     "$out_prefix" \
            > "$log" 2>&1 &
        pids+=($!)
        echo "[watcher]   GPU $GPU_ID → PID $! → $log"
    done
    # Return pids via global
    LAUNCHED_PIDS=("${pids[@]}")
}

eval_done() {
    # Returns 0 (true) if merged result file already exists with results
    local merged="$1_merged.json"
    [ -f "$merged" ] && python3 -c "
import json, sys
d = json.load(open('$merged'))
sys.exit(0 if d.get('evaluated', 0) > 0 else 1)
" 2>/dev/null
}

# ── Step 1: wait for ckpt-1700 (already running) ──────────────────────────────

CKPT1700_PIDS=(13902 13903 13904 13905 13906 13907 13908 13909)
wait_pids "${CKPT1700_PIDS[@]}"

merge_shards "$EXP_BASE/eval_ckpt1700_full"

# ── Step 2: ckpt-5950 ─────────────────────────────────────────────────────────

CKPT5950="$EXP_BASE/new_policy7w_v2_reformat/checkpoint-5950/hf_model"
OUT5950="$EXP_BASE/eval_ckpt5950_full"

if eval_done "$OUT5950"; then
    echo "[watcher] ckpt-5950 already evaluated, skipping."
else
    launch_eval "$CKPT5950" "$OUT5950"
    wait_pids "${LAUNCHED_PIDS[@]}"
    merge_shards "$OUT5950"
fi

# ── Step 3: print summary ─────────────────────────────────────────────────────

echo "[watcher] ===== FINAL RESULTS ====="
for f in "$EXP_BASE"/eval_ckpt*_merged.json; do
    [ -f "$f" ] || continue
    $PYTHON -c "
import json, sys
d = json.load(open('$f'))
print(f\"  {sys.argv[1]}: acc={d['accuracy']:.4f} ({d['correct']}/{d['evaluated']}) parse_fail={d['parse_failures']}\")
for lbl, s in d.get('per_class', {}).items():
    print(f\"    label={lbl}: precision={s['precision']:.3f} recall={s['recall']:.3f} f1={s['f1']:.3f}\")
" "$f"
done

echo "[watcher] All evals complete. Watcher exiting at $(date)"