playground/inference/run_parallel.sh

#!/bin/bash
# Launch 8 parallel eval workers, one per A100.
# Each loads a full model copy (~16GB BF16) and processes its shard of parquet files.
# Supports resume: safe to re-run if interrupted.
set -e

PYTHON=/mlx/users/jiashuo.fan/miniconda3/envs/abbie/bin/python3
CKPT=/mnt/bn/bohanzhainas1/jiashuo/exp/new_policy7w_v2_reformat/checkpoint-1700/hf_model
DATA_DIR=/mnt/hdfs/byte_tt_data_cu_vagcp/haogeng.liu/new_policy7w_v2_reformat
OUT_PREFIX=/mnt/bn/bohanzhainas1/jiashuo/exp/eval_ckpt1700_full
LOG_DIR=/mlx/users/jiashuo.fan/playground/inference/logs

mkdir -p "$LOG_DIR"

echo "Launching 8 eval workers (full dataset, optimized)..."
PIDS=()
for GPU_ID in 0 1 2 3 4 5 6 7; do
    LOG="${LOG_DIR}/eval_gpu${GPU_ID}.log"
    CUDA_VISIBLE_DEVICES=$GPU_ID $PYTHON -u \
        /mlx/users/jiashuo.fan/playground/inference/eval.py \
        --model-path "$CKPT" \
        --data-dir   "$DATA_DIR" \
        --gpu-id     $GPU_ID \
        --shard-id   $GPU_ID \
        --num-shards 8 \
        --output     "$OUT_PREFIX" \
        >> "$LOG" 2>&1 &
    PIDS+=($!)
    echo "  GPU $GPU_ID → PID $! → $LOG"
done

echo ""
echo "Watch progress:  tail -f ${LOG_DIR}/eval_gpu*.log"
echo "Quick status:    tail -1 ${LOG_DIR}/eval_gpu*.log"
echo ""

ALL_OK=true
for i in "${!PIDS[@]}"; do
    wait "${PIDS[$i]}"
    CODE=$?
    GPU=$i
    if [ $CODE -ne 0 ]; then
        echo "GPU $GPU FAILED (exit $CODE) — check ${LOG_DIR}/eval_gpu${GPU}.log"
        ALL_OK=false
    else
        echo "GPU $GPU done ✓"
    fi
done

if $ALL_OK; then
    echo ""
    echo "Merging shards..."
    $PYTHON -u /mlx/users/jiashuo.fan/playground/inference/merge_results.py \
        --prefix     "$OUT_PREFIX" \
        --num-shards 8
fi