magicciv/scripts/rl-train.sh

#!/usr/bin/env bash
# rl-train.sh — manage the RL self-play training run on apricot.
#
# Subcommands:
#   status      Show PID, elapsed, fps, last train block, eval dir state
#   logs [N]    Tail N (default 60) lines of the active training log
#   gpu         Show nvidia-smi snapshot
#   procs       Count godot/python procs related to training
#   launch      Start training (env-driven; see ENV below)
#   kill        Stop training cleanly (SIGTERM, then SIGKILL after 5s)
#   sync        scp local rl_self_play/ to the apricot worktree
#
# Env vars (with defaults):
#   RL_HOST=apricot               SSH alias for the training box
#   RL_WORKTREE=/var/home/lilith/.cache/mc-rl-train-1779015795
#   RL_RUN_NAME=duel-v1b-cuda1
#   RL_DEVICE=cuda:1
#   RL_ENVS=4
#   RL_TOTAL_STEPS=200000
#   RL_EVAL_FREQ=20000
#   RL_EVAL_EPS=10
#   RL_MAX_TURNS=100
#   RL_PIDFILE=/tmp/rl-train.pid  (on the remote)

set -euo pipefail

: "${RL_HOST:=apricot}"
: "${RL_WORKTREE:=/var/home/lilith/.cache/mc-rl-train-1779015795}"
: "${RL_RUN_NAME:=duel-v1b-cuda1}"
: "${RL_DEVICE:=cuda:1}"
: "${RL_ENVS:=4}"
: "${RL_TOTAL_STEPS:=200000}"
: "${RL_EVAL_FREQ:=20000}"
: "${RL_EVAL_EPS:=10}"
: "${RL_MAX_TURNS:=100}"
: "${RL_PIDFILE:=/tmp/rl-train.pid}"

LOG_REMOTE="${RL_WORKTREE}/training-${RL_RUN_NAME}.log"
EVAL_DIR_REMOTE="${RL_WORKTREE}/tooling/rl_self_play/runs/${RL_RUN_NAME}/eval"

cmd="${1:-status}"
shift || true

remote() { ssh "${RL_HOST}" "$1"; }

case "$cmd" in
  status)
    remote "
      set +e
      echo '---PYTHON PID---'
      py=\$(pgrep -f 'python3 -m tooling.rl_self_play.train' | head -1)
      if [ -z \"\$py\" ]; then
        echo 'no training process'
        echo '---EVAL DIR (${EVAL_DIR_REMOTE})---'
        ls -la ${EVAL_DIR_REMOTE} 2>/dev/null || echo 'missing'
        exit 0
      fi
      ps -p \$py -o pid,etime,pcpu,pmem,cmd
      log=\$(readlink /proc/\$py/fd/1 2>/dev/null)
      echo \"---LOG (\$log)---\"
      tail -40 \"\$log\" 2>/dev/null || echo 'log not readable'
      echo '---EVAL DIR (${EVAL_DIR_REMOTE})---'
      ls -la ${EVAL_DIR_REMOTE} 2>/dev/null || echo 'missing'
    "
    ;;

  logs)
    n="${1:-60}"
    remote "
      py=\$(pgrep -f 'python3 -m tooling.rl_self_play.train' | head -1)
      if [ -z \"\$py\" ]; then echo 'no training process'; exit 1; fi
      tail -${n} \"\$(readlink /proc/\$py/fd/1)\"
    "
    ;;

  gpu)
    remote 'nvidia-smi --query-gpu=index,name,memory.used,memory.total,utilization.gpu --format=csv'
    ;;

  procs)
    remote "
      printf 'python train procs: '; pgrep -af 'python3 -m tooling.rl_self_play.train' | wc -l
      printf 'godot-bin procs: ';   pgrep -af 'godot-bin' | wc -l
      uptime
    "
    ;;

  launch)
    remote "
      cd ${RL_WORKTREE} || exit 1
      if pgrep -f 'python3 -m tooling.rl_self_play.train' >/dev/null; then
        echo 'training already running; run kill first'
        pgrep -af 'python3 -m tooling.rl_self_play.train'
        exit 1
      fi
      nohup python3 -m tooling.rl_self_play.train \
        --device ${RL_DEVICE} --num-envs ${RL_ENVS} \
        --total-steps ${RL_TOTAL_STEPS} --eval-freq ${RL_EVAL_FREQ} \
        --eval-episodes ${RL_EVAL_EPS} --max-turns ${RL_MAX_TURNS} \
        --run-name ${RL_RUN_NAME} > ${LOG_REMOTE} 2>&1 &
      echo \$! > ${RL_PIDFILE}
      sleep 3
      pgrep -af 'python3 -m tooling.rl_self_play.train' || (echo 'launch failed; check log'; tail -20 ${LOG_REMOTE})
    "
    ;;

  kill)
    remote "
      pkill -f 'rl_self_play.train' 2>/dev/null || true
      sleep 5
      pkill -9 -f 'rl_self_play.train' 2>/dev/null || true
      pkill -9 -f 'godot-bin --path' 2>/dev/null || true
      printf 'remaining: '
      pgrep -f 'rl_self_play.train|godot-bin' | wc -l
    "
    ;;

  sync)
    here="$(cd "$(dirname "$0")/.." && pwd)"
    scp -q "${here}/tooling/rl_self_play/"*.py \
      "${RL_HOST}:${RL_WORKTREE}/tooling/rl_self_play/"
    echo "synced tooling/rl_self_play/*.py to ${RL_HOST}:${RL_WORKTREE}"
    ;;

  *)
    sed -n '2,28p' "$0"
    exit 2
    ;;
esac