code/run_af_one_ckpt_fast.sh

#!/usr/bin/env bash
set -uo pipefail
# ─────────────────────────────────────────────────────────────────────────────
# run_af_one_ckpt_fast.sh — method-A (single-process batch) version of
# run_af_one_ckpt.sh.  Starts ONE GR00T server for the ckpt, then runs
# groot_full_factor_batch.py once per seed (200 cells in one process → no
# per-cell cold start).  Output identical layout, only faster.
#
# Usage: run_af_one_ckpt_fast.sh <ckpt_name> <gpu> <port>
# Env: SEEDS_OVERRIDE ("40" / "40 41 42"), SAMPLE_N (200), TOTAL_EPISODES (200),
#      SIM_BACKEND (gpu), RENDER_BACKEND (gpu)
# ─────────────────────────────────────────────────────────────────────────────
ROOT=/workspace/groot_eval
HARNESS="${ROOT}/harness"
CKPT="${1:?ckpt_name}"
GPU="${2:?gpu}"
PORT="${3:?port}"
SEEDS=(${SEEDS_OVERRIDE:-40 41 42})
SAMPLE_N="${SAMPLE_N:-200}"
TOTAL_EPISODES="${TOTAL_EPISODES:-200}"
SIM_BACKEND="${SIM_BACKEND:-gpu}"
RENDER_BACKEND="${RENDER_BACKEND:-gpu}"
NO_DISTRACTOR_PROB="${NO_DISTRACTOR_PROB:-0.70}"

# Auto-detect ckpt layout: either <ckpt>/checkpoint-10000/ or <ckpt>/ directly (HF layout)
MODEL_PATH="${ROOT}/gr00t_af_ckpts/${CKPT}/checkpoint-10000"
[[ -d "${MODEL_PATH}" ]] || MODEL_PATH="${ROOT}/gr00t_af_ckpts/${CKPT}"
OUT_DIR="${RESULTS_ROOT:-${ROOT}/results_af}/${CKPT}"
LOG_DIR="${ROOT}/logs/gr00t_af"
mkdir -p "${OUT_DIR}" "${LOG_DIR}"
slog="${LOG_DIR}/server_fast_${CKPT}_gpu${GPU}.log"
[[ -d "${MODEL_PATH}" ]] || { echo "[${CKPT}] MODEL_PATH not found: ${MODEL_PATH}"; exit 1; }
ls "${MODEL_PATH}"/*.safetensors >/dev/null 2>&1 || { echo "[${CKPT}] no safetensors in ${MODEL_PATH}"; exit 1; }

echo "[$(date +%H:%M:%S)] ${CKPT}: start GR00T server gpu=${GPU} port=${PORT}"
( cd "${ROOT}/gr00t_repo/codebase" && CUDA_VISIBLE_DEVICES="${GPU}" \
    HF_HOME="${ROOT}/.hf_cache" HF_TOKEN="$(cat ${ROOT}/.hf_token)" \
    NO_ALBUMENTATIONS_UPDATE=1 TOKENIZERS_PARALLELISM=false \
    "${ROOT}/.venv_groot/bin/python" -m gr00t.eval.run_gr00t_server \
      --model-path "${MODEL_PATH}" \
      --embodiment-tag new_embodiment --device cuda:0 \
      --host 127.0.0.1 --port "${PORT}" ) > "${slog}" 2>&1 &
spid=$!

ok=0
for _ in $(seq 1 300); do
  kill -0 "${spid}" 2>/dev/null || { echo "[${CKPT}] SERVER DIED during load"; break; }
  grep -q "Server ready\|Server is ready" "${slog}" 2>/dev/null && { ok=1; break; }
  sleep 3
done
[[ "${ok}" == "1" ]] || { echo "[${CKPT}] server not ready"; tail -n 25 "${slog}"; kill "${spid}" 2>/dev/null; exit 1; }
echo "[$(date +%H:%M:%S)] ${CKPT}: server ready (pid ${spid})"

_MS_TORCH_LIB="$("${ROOT}/.venv_ms/bin/python" -c 'import torch,os;print(os.path.join(os.path.dirname(torch.__file__),"lib"))' 2>/dev/null || true)"
export LD_LIBRARY_PATH="${_MS_TORCH_LIB}:${LD_LIBRARY_PATH:-}"
export MANISKILL_CONFLICT_ROOT="${MANISKILL_CONFLICT_ROOT:-${ROOT}/genie_repo/maniskill_conflict}"

rc_all=0
for sb in "${SEEDS[@]}"; do
  rt="${OUT_DIR}/full_factor_${CKPT}_seed${sb}.txt"
  echo "[$(date +%H:%M:%S)] ${CKPT}: seed_base=${sb} → ${rt}"
  CUDA_VISIBLE_DEVICES="${GPU}" \
  "${ROOT}/.venv_ms/bin/python" "${HARNESS}/groot_full_factor_batch.py" \
    --host 127.0.0.1 --port "${PORT}" \
    --results-txt "${rt}" \
    --video-root "${OUT_DIR}/videos_seed${sb}" \
    --sample-n "${SAMPLE_N}" --sample-seed 42 --seed-base "${sb}" \
    --total-episodes "${TOTAL_EPISODES}" --max-episode-steps 500 \
    --no-distractor-prob "${NO_DISTRACTOR_PROB}" --replan-steps 5 \
    --sim-backend "${SIM_BACKEND}" --render-backend "${RENDER_BACKEND}" \
    > "${LOG_DIR}/client_fast_${CKPT}_seed${sb}.log" 2>&1
  src=$?
  [[ "${src}" -ne 0 ]] && rc_all=1
  grep -E "^overall_success" "${rt}" 2>/dev/null || echo "[${CKPT} seed${sb}] no overall_ line"
done

kill "${spid}" 2>/dev/null; wait "${spid}" 2>/dev/null

python3 - "${OUT_DIR}" "${CKPT}" "${SEEDS[@]}" > "${OUT_DIR}/SUMMARY.txt" <<'PY'
import re, sys
from pathlib import Path
out_dir, ckpt, *seeds = sys.argv[1:]
rates, line = [], []
for sb in seeds:
    p = Path(out_dir) / f"full_factor_{ckpt}_seed{sb}.txt"
    if not p.exists():
        line.append(f"seed{sb}: MISSING"); continue
    m = re.search(r"overall_success=(\d+)/(\d+) \(([\d.]+)%\)", p.read_text())
    if m:
        s, n, r = int(m.group(1)), int(m.group(2)), float(m.group(3))
        rates.append(r); line.append(f"seed{sb}: {s}/{n} ({r:.1f}%)")
    else:
        line.append(f"seed{sb}: PARSE_FAILED")
avg = sum(rates)/len(rates) if rates else 0.0
print(f"# {ckpt} — full-factor TASK-aligned, single-process batch "
      f"(sample_n=200 seed=42, 200 eps, max_steps=500, no_distractor=0.70, "
      f"default difficulty, sim=gpu)")
for l in line: print(l)
print(f"AVG over {len(rates)} seed(s): {avg:.1f}%")
PY

echo "[$(date +%H:%M:%S)] ${CKPT}: DONE rc=${rc_all}"
cat "${OUT_DIR}/SUMMARY.txt"
exit ${rc_all}