onf-v3/scripts/remeasure_long_spread.sh

#!/usr/bin/env bash
# Settle whether the long x {Objects_Layout, Language_Instructions} sentinel numbers moved, or whether
# the run-to-run spread of a stochastic policy was being read as a trend.
#
#   bash scripts/remeasure_long_spread.sh <WAIT_PID> [GPU] [PORT]
#
# Waits for WAIT_PID (the in-flight long bench) to exit, then re-runs the SAME sentinel config twice
# more on the freed GPU. Two independent samples taken back to back measure the run-to-run spread
# directly, which is what decides whether an apparent drop is a real change or sampling noise.
# Each repeat gets its own --variant so nothing overwrites an existing result directory.
#
# GR00T-N1.7 is flow-matching and unseeded: only same-session arms are comparable.
set -uo pipefail

REPO="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
PY="${ONF_PY:-/home/quang/miniconda3/envs/stablevla/bin/python}"
WAIT_PID="${1:?usage: remeasure_long_spread.sh <WAIT_PID> [GPU] [PORT]}"
GPU="${2:-2}"
PORT="${3:-10095}"
SESSION="${SESSION_TAG:-s0806}"

export OMP_NUM_THREADS=4 PYTHONPATH="$REPO/src"
cd "$REPO"; mkdir -p logs/_drivers

echo "[$(date -u +%FT%TZ)] waiting for pid $WAIT_PID to release cuda:$GPU"
while kill -0 "$WAIT_PID" 2>/dev/null; do sleep 60; done
echo "[$(date -u +%FT%TZ)] free; starting re-measurement on cuda:$GPU port $PORT"

run () {  # run <mode> <variant>
    local mode="$1" variant="$2"
    echo "[$(date -u +%FT%TZ)] === mode=$mode variant=$variant"
    $PY evals/gr00t/run_gr00t.py --suites long --gpu "$GPU" --port "$PORT" --num-clients 6 \
        --axes "Objects Layout" "Language Instructions" \
        --mode "$mode" --variant "$variant" \
        >>"logs/_drivers/remeasure_${mode}_${variant}.log" 2>&1
    echo "[$(date -u +%FT%TZ)] === mode=$mode variant=$variant exit $?"
}

run sentinel "rep2_$SESSION"
run sentinel "rep3_$SESSION"
echo "[$(date -u +%FT%TZ)] re-measurement finished"