#!/usr/bin/env bash
# The MergeBench property sweep: one worker per GPU, each pinned, families smallest-first.
#
# Properties only -- no merged model is built and nothing is evaluated (RESULTS_MERGEBENCH.md).
# Each worker takes every n-th family from the size-ordered list, streams it (download -> measure ->
# delete), appends to its own CSV and redraws the figures after each family, so an interrupted run
# still leaves a usable figure on disk. Ledger-resumable: relaunching skips finished families.
#
#   bash scripts/run_mergebench.sh 4 5
#   EXTRA="--families gemma-2-9b gemma-2-9b-it" bash scripts/run_mergebench.sh 5
set -u
cd "$(dirname "$0")/.."
PY=${PYTHON:-/root/venvs/mergeability/bin/python}
GPUS=("$@"); [ ${#GPUS[@]} -eq 0 ] && GPUS=(4 5)
N=${#GPUS[@]}
mkdir -p results/mergebench/logs figures/mergebench
# Bounded thread pools. The box runs several concurrent agents on 128 cores; left to itself every
# numpy/torch process opens a pool per core and they preempt each other -- measured at load ~320
# with 311k threads and the GPUs near 0%. Must be set before python imports numpy.
export MB_THREADS=${MB_THREADS:-8}
export OMP_NUM_THREADS=$MB_THREADS
export MKL_NUM_THREADS=$MB_THREADS
export OPENBLAS_NUM_THREADS=$MB_THREADS
export NUMEXPR_NUM_THREADS=$MB_THREADS
export VECLIB_MAXIMUM_THREADS=$MB_THREADS
[ -f /root/.ms_hf_env ] && set -a && . /root/.ms_hf_env && set +a
for i in "${!GPUS[@]}"; do
  g=${GPUS[$i]}
  HF_HOME=/root/hf_cache_mergebench/w${i} \
  CUDA_VISIBLE_DEVICES=$g PYTHONPATH=src \
  setsid nohup "$PY" -u -m mergeschool.mergebench.sweep --worker "$i" --n-workers "$N" ${EXTRA:-} \
      > "results/mergebench/logs/w${i}_gpu${g}.log" 2>&1 &
  echo "worker $i -> GPU $g (pid $!)"
done
sleep 2
echo "logs: results/mergebench/logs/"