#!/usr/bin/env bash
set -uo pipefail
# ─────────────────────────────────────────────────────────────────────────────
# run_pi0_seed.sh — evaluate ONE pi0 ckpt for ONE seed on ONE GPU, using the
# EXACT GR00T conflict definition (200 episodes, always-distractor, dual-axis
# color/size or color/spatial success).  Reuses GR00T's own job-gen+batch:
#   run_ood_groot_inference.sh  +  pi0_pairwise_main.py (= groot_main.py with
#   the policy boundary swapped to the pi0 openpi websocket server).
#
# Usage: run_pi0_seed.sh <ckpt_name> <gpu> <seed> [port]
# One (ckpt, seed) per GPU.  sim_backend=cpu (pi0 server holds the GPU; pi0 +
# ManiSkill-GPU-sim on one card → CUDA illegal access, so physics on CPU).
# ─────────────────────────────────────────────────────────────────────────────
ROOT=/workspace/groot_eval
HARNESS="${ROOT}/harness"
SIM_ROOT=/workspace/eval_simulation/simulation
CKPT="${1:?ckpt_name}"
GPU="${2:?gpu}"
SEED="${3:?seed}"
PORT="${4:-$((8200 + GPU))}"
STEP="${STEP:-17999}"
TOTAL_EPISODES="${TOTAL_EPISODES:-200}"
SIM_BACKEND="${SIM_BACKEND:-cpu}"
MAX_EPISODE_STEPS="${MAX_EPISODE_STEPS:-150}"
REPLAN_STEPS="${REPLAN_STEPS:-10}"

MODEL_DIR="/workspace/pi0_ckpt/${CKPT}/${STEP}"
OUT_DIR="${ROOT}/results_pi0_pairwise/${CKPT}"
LOG_DIR="${ROOT}/logs/pi0_pairwise"
mkdir -p "${OUT_DIR}" "${LOG_DIR}"
slog="${LOG_DIR}/server_${CKPT}_gpu${GPU}_s${SEED}.log"
[[ -d "${MODEL_DIR}" ]] || { echo "[${CKPT}] MODEL_DIR not found: ${MODEL_DIR}"; exit 1; }

EXPERIMENT="$(python3 - "${CKPT}" <<'PY'
import sys
name=sys.argv[1]
exps=["spatial_object","spatial_size","color_spatial","color_object","color_size",
      "verb_spatial","verb_object","verb_color","verb_size","size_object"]
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"; exit 1; }
echo "[$(date +%H:%M:%S)] ${CKPT}: experiment=${EXPERIMENT} gpu=${GPU} seed=${SEED} port=${PORT} (GR00T-def, ${TOTAL_EPISODES}ep, cpu-sim)"

# ── start pi0 openpi websocket server (uses the GPU) ──
( cd /workspace/eval_pi0 && CUDA_VISIBLE_DEVICES="${GPU}" \
    UV_PROJECT_ENVIRONMENT=/venv/pi0_eval uv run --no-sync scripts/serve_policy.py \
      --port "${PORT}" policy:checkpoint \
      --policy.config pi0_maniskill --policy.dir "${MODEL_DIR}" ) > "${slog}" 2>&1 &
spid=$!
ok=0
for _ in $(seq 1 200); do
  kill -0 "${spid}" 2>/dev/null || { echo "[${CKPT}] SERVER DIED"; break; }
  grep -qiE "Creating server|websocket_policy_server|server listening" "${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}: pi0 server ready (pid ${spid})"

rt="${OUT_DIR}/pi0_${EXPERIMENT}_${CKPT}_seed${SEED}.txt"
HOST=127.0.0.1 PORT="${PORT}" SIM_BACKEND="${SIM_BACKEND}" \
MAX_EPISODE_STEPS="${MAX_EPISODE_STEPS}" REPLAN_STEPS="${REPLAN_STEPS}" \
SEED_BASE=0 \
EXPERIMENT_ROOT="${OUT_DIR}/video_seed${SEED}" \
GROOT_MAIN="${HARNESS}/pi0_pairwise_main.py" \
MS_PY="/venv/pi05_ms/bin/python" \
MANISKILL_CONFLICT_ROOT="${SIM_ROOT}" \
bash "${HARNESS}/run_ood_groot_inference.sh" "${EXPERIMENT}" "${SEED}" "${TOTAL_EPISODES}" "${rt}" \
  > "${LOG_DIR}/client_${CKPT}_seed${SEED}.log" 2>&1
rc=$?

kill -9 "${spid}" 2>/dev/null; pkill -9 -P "${spid}" 2>/dev/null
echo "[$(date +%H:%M:%S)] ${CKPT} seed${SEED}: DONE rc=${rc}"
grep -E "^overall_" "${rt}" 2>/dev/null || echo "[${CKPT} s${SEED}] no overall_ (see client log)"
exit ${rc}