#!/usr/bin/env bash
set -uo pipefail
# ─────────────────────────────────────────────────────────────────────────────
# run_groot_one_ckpt.sh — evaluate ONE GR00T ckpt on the HARD pairwise grid
# (color_size or color_spatial), ONE GPU, ONE or more seeds.  Mirrors
# run_pi0_one_ckpt.sh structure; policy=GR00T zmq server (gr00t venv).
#
# Speed-matched to pi0:  MAX_EPISODE_STEPS=150, REPLAN_STEPS=10, gpu-sim.
# Output ONLY under results_gr00t_pair_color_size/<ckpt>/ (never touches
# canonical results/ or genie's results_af/).
#
# Usage:  run_groot_one_ckpt.sh <ckpt_name> <gpu> <port>
# ─────────────────────────────────────────────────────────────────────────────
ROOT=/workspace/groot_eval
HARNESS="${ROOT}/harness"
SIM_ROOT=/workspace/eval_simulation/simulation
CKPT="${1:?ckpt_name (dir under /workspace/gr00t_pair_ckpt)}"
GPU="${2:?gpu}"
PORT="${3:?port}"
SEEDS=(${SEEDS_OVERRIDE:-42 40 41})
TASK_DIFFICULTY="${TASK_DIFFICULTY:-1.5}"
SIM_BACKEND="${SIM_BACKEND:-gpu}"
MAX_EPISODE_STEPS="${MAX_EPISODE_STEPS:-150}"
REPLAN_STEPS="${REPLAN_STEPS:-10}"
TARGET_EPISODES="${TARGET_EPISODES:-200}"
COLOR_SIZE_MODE="${COLOR_SIZE_MODE:-hard}"

MODEL_DIR="/workspace/gr00t_pair_ckpt/${CKPT}"
# OUT_DIR depends on parsed EXPERIMENT (set just below);  we re-set it after parsing.
LOG_DIR="${ROOT}/logs/gr00t_pair"
[[ -d "${MODEL_DIR}" ]] || { echo "[${CKPT}] MODEL_DIR not found: ${MODEL_DIR}"; exit 1; }

# experiment from ckpt name (longest matching pair prefix)
EXPERIMENT="$(python3 - "${CKPT}" <<'PY'
import sys
name = sys.argv[1]
exps = ["color_spatial","color_size","color_object","spatial_size","spatial_object",
        "size_object","verb_spatial","verb_object","verb_color","verb_size"]
hit = [e for e in exps if name.startswith(e)]
print(max(hit, key=len) if hit else "")
PY
)"
[[ -n "${EXPERIMENT}" ]] || { echo "[${CKPT}] cannot parse experiment from name"; exit 1; }
# OUT_DIR: experiment-aware; color_size_easy goes to a separate dir so HARD/EASY don't collide
if [[ "${EXPERIMENT}" == "color_size" && "${COLOR_SIZE_MODE}" == "easy" ]]; then
  OUT_DIR="${ROOT}/results_gr00t_pair_color_size_easy/${CKPT}"
else
  OUT_DIR="${ROOT}/results_gr00t_pair_${EXPERIMENT}/${CKPT}"
fi
mkdir -p "${OUT_DIR}" "${LOG_DIR}"
slog="${LOG_DIR}/server_${CKPT}_gpu${GPU}.log"
echo "[$(date +%H:%M:%S)] ${CKPT}: experiment=${EXPERIMENT} gpu=${GPU} port=${PORT} difficulty=${TASK_DIFFICULTY} sim=${SIM_BACKEND}"

# start GR00T zmq inference server (uses GPU)
GROOT_VENV_PY="${ROOT}/.venv_groot/bin/python"
GROOT_CODE="${ROOT}/gr00t_repo/codebase"
HF_TOKEN_FILE="${ROOT}/.hf_token"
HF_TOKEN_ARG=""
[[ -f "${HF_TOKEN_FILE}" ]] && HF_TOKEN_ARG="$(cat ${HF_TOKEN_FILE})"
( cd "${GROOT_CODE}" && CUDA_VISIBLE_DEVICES="${GPU}" \
    HF_HOME="${ROOT}/.hf_cache" HF_TOKEN="${HF_TOKEN_ARG}" \
    NO_ALBUMENTATIONS_UPDATE=1 TOKENIZERS_PARALLELISM=false \
    "${GROOT_VENV_PY}" -m gr00t.eval.run_gr00t_server \
      --model-path "${MODEL_DIR}" \
      --embodiment-tag new_embodiment \
      --device cuda:0 \
      --host 127.0.0.1 \
      --port "${PORT}" ) > "${slog}" 2>&1 &
spid=$!
ok=0
for _ in $(seq 1 200); do
  kill -0 "${spid}" 2>/dev/null || { echo "[${CKPT}] SERVER DIED during load"; break; }
  grep -qE "Server ready|Server is ready and listening|listening on" "${slog}" 2>/dev/null && { ok=1; break; }
  sleep 3
done
[[ "${ok}" == "1" ]] || { echo "[${CKPT}] server not ready"; tail -n 25 "${slog}"; kill -9 "${spid}" 2>/dev/null; exit 1; }
echo "[$(date +%H:%M:%S)] ${CKPT}: GR00T server ready (pid ${spid})"

export MANISKILL_CONFLICT_ROOT="${SIM_ROOT}"
MS_PY="${ROOT}/.venv_ms/bin/python"
[[ -x "${MS_PY}" ]] || MS_PY="/venv/pi05_ms/bin/python"
_MS_TORCH_LIB="$(${MS_PY} -c 'import torch,os;print(os.path.join(os.path.dirname(torch.__file__),"lib"))' 2>/dev/null || true)"
export LD_LIBRARY_PATH="${_MS_TORCH_LIB}:${LD_LIBRARY_PATH:-}"

rc_all=0
for sb in "${SEEDS[@]}"; do
  rt="${OUT_DIR}/gr00t_${EXPERIMENT}_${CKPT}_seed${sb}.txt"
  echo "[$(date +%H:%M:%S)] ${CKPT}: seed${sb} experiment=${EXPERIMENT} → ${rt}"
  CUDA_VISIBLE_DEVICES="${GPU}" SIM_ROOT="${SIM_ROOT}" \
  "${MS_PY}" "${HARNESS}/groot_grid_eval.py" \
    --experiment "${EXPERIMENT}" --host 127.0.0.1 --port "${PORT}" \
    --seed "${sb}" --sim-backend "${SIM_BACKEND}" \
    --task-difficulty "${TASK_DIFFICULTY}" \
    --max-episode-steps "${MAX_EPISODE_STEPS}" --replan-steps "${REPLAN_STEPS}" \
    --target-episodes "${TARGET_EPISODES}" \
    ${MAX_CELLS:+--max-cells "${MAX_CELLS}"} \
    --results-txt "${rt}" \
    > "${LOG_DIR}/client_${CKPT}_seed${sb}.log" 2>&1
  src=$?
  [[ "${src}" -ne 0 ]] && rc_all=1
  grep -E "^overall_" "${rt}" 2>/dev/null || echo "[${CKPT} seed${sb}] no overall_ line (see client log)"
done

kill -9 "${spid}" 2>/dev/null; pkill -9 -P "${spid}" 2>/dev/null

python3 - "${OUT_DIR}" "${EXPERIMENT}" "${CKPT}" "${SEEDS[@]}" > "${OUT_DIR}/SUMMARY.txt" <<'PY'
import re, sys
from pathlib import Path
out_dir, exp, ckpt, *seeds = sys.argv[1:]
acc = {}
for sb in seeds:
    p = Path(out_dir) / f"gr00t_{exp}_{ckpt}_seed{sb}.txt"
    if not p.exists():
        print(f"seed{sb}: MISSING"); continue
    txt = p.read_text()
    line = []
    for m in re.finditer(r"^(overall_\w+)=(\d+)/(\d+) \(([\d.]+)%\)", txt, re.M):
        k, _s, _n, r = m.group(1), m.group(2), m.group(3), float(m.group(4))
        acc.setdefault(k, []).append(r); line.append(f"{k}={r:.1f}%")
    print(f"seed{sb}: " + "  ".join(line) if line else f"seed{sb}: PARSE_FAILED")
print(f"\n# {ckpt}  experiment={exp}  (gr00t HARD-def, {len(seeds)} seeds)")
for k, v in acc.items():
    print(f"AVG {k} = {sum(v)/len(v):.1f}%  (n={len(v)} seeds)")
PY
echo "[$(date +%H:%M:%S)] ${CKPT}: DONE rc=${rc_all}"
cat "${OUT_DIR}/SUMMARY.txt"
exit ${rc_all}