Datasets:

myx160
/

rl-post-train-artifacts

Tags:

License:

Dataset card Files Files and versions

xet

Community

rl-post-train-artifacts / scripts /run_cfg_task0.sh

myx160

Add files using upload-large-folder tool

456c9d7 verified 20 days ago

Raw History Blame Contribute Delete

12.7 kB

#!/usr/bin/env bash
# Guarded launcher for staged Task 0 CFG preflights.
# It never auto-selects GPUs and never runs global Ray cleanup.

set -euo pipefail

PROJECT_ROOT="/data/myx/RL_post_train"
SCRIPT_DIR="${PROJECT_ROOT}/scripts"
CONFIG_PATH="${PROJECT_ROOT}/configs/recap/task0_cfg_recap_v1.yaml"
MODE="c2"
GPUS=""
RUN_ID=""
RESUME_DIR=""
DRY_RUN=0
C4_KIND="matrix"
C4_MICRO_BATCH_SIZE=1
C4_GLOBAL_BATCH_SIZE=2
C4_NUM_WORKERS=8
C4_MAX_STEPS=12
IDLE_CHECK_SECONDS="${CFG_GPU_IDLE_CHECK_SECONDS:-30}"
EXTRA_OVERRIDES=()

usage() {
  cat <<'EOF'
Usage: run_cfg_task0.sh --gpus PHYSICAL0,PHYSICAL1 [options] [-- hydra overrides]

Options:
  --mode c2|c3|c3_resume|c4
                     Select the bounded preflight (default: c2).
  --gpus A,B         Exactly two physical GPU indices; no auto-selection.
  --run-id ID        Project-local log/experiment identifier.
  --resume-dir PATH  Resume from an exact RLinf checkpoint directory.
  --config PATH      Project-local CFG YAML override.
  --c4-kind matrix|accumulation
                     Select the bounded C4 benchmark shape (default: matrix).
  --micro-batch-size N / --global-batch-size N
                     C4 batch sizes; accumulation uses global batch 512.
  --num-workers N    C4 dataloader workers (0..16, default: 8).
  --max-steps N      C4 total optimizer steps (matrix default: 12; accumulation: 1).
  --dry-run          Validate and print the command without launching Ray.
  -h, --help         Show this help.
EOF
}

while (($# > 0)); do
  case "$1" in
    --mode) MODE="$2"; shift 2 ;;
    --gpus) GPUS="$2"; shift 2 ;;
    --run-id) RUN_ID="$2"; shift 2 ;;
    --resume-dir) RESUME_DIR="$2"; shift 2 ;;
    --config) CONFIG_PATH="$2"; shift 2 ;;
    --c4-kind) C4_KIND="$2"; shift 2 ;;
    --micro-batch-size) C4_MICRO_BATCH_SIZE="$2"; shift 2 ;;
    --global-batch-size) C4_GLOBAL_BATCH_SIZE="$2"; shift 2 ;;
    --num-workers) C4_NUM_WORKERS="$2"; shift 2 ;;
    --max-steps) C4_MAX_STEPS="$2"; shift 2 ;;
    --dry-run) DRY_RUN=1; shift ;;
    -h|--help) usage; exit 0 ;;
    --) shift; EXTRA_OVERRIDES+=("$@"); break ;;
    *) echo "Unknown argument: $1" >&2; usage >&2; exit 2 ;;
  esac
done

if [[ -z "${GPUS}" ]]; then
  echo "CFG launcher requires --gpus PHYSICAL0,PHYSICAL1" >&2
  exit 2
fi
if [[ "${MODE}" != "c2" && "${MODE}" != "c3" && "${MODE}" != "c3_resume" && "${MODE}" != "c4" ]]; then
  echo "Only bounded c2/c3/c3_resume/c4 modes are enabled; got ${MODE}" >&2
  exit 2
fi
if [[ -n "${RAY_ADDRESS:-}" || -n "${CLUSTER_NAMESPACE:-}" ]]; then
  echo "Refusing inherited RAY_ADDRESS/CLUSTER_NAMESPACE." >&2
  exit 3
fi

IFS=',' read -r GPU0 GPU1 GPU_EXTRA <<< "${GPUS}"
if [[ -n "${GPU_EXTRA:-}" || -z "${GPU0:-}" || -z "${GPU1:-}" ]]; then
  echo "--gpus must contain exactly two comma-separated indices" >&2
  exit 2
fi
if [[ "${GPU0}" == "${GPU1}" || ! "${GPU0}" =~ ^[0-9]+$ || ! "${GPU1}" =~ ^[0-9]+$ ]]; then
  echo "--gpus must contain two distinct non-negative integers" >&2
  exit 2
fi
if [[ -z "${RUN_ID}" ]]; then
  RUN_ID="recap_cfg_${MODE}_$(date +%Y%m%d_%H%M%S)"
fi
if [[ ! "${RUN_ID}" =~ ^[A-Za-z0-9_.-]+$ ]]; then
  echo "run-id contains unsafe characters: ${RUN_ID}" >&2
  exit 2
fi

source "${PROJECT_ROOT}/scripts/activate_project_env.sh"
if [[ ! -f "${CONFIG_PATH}" ]]; then
  echo "missing CFG config: ${CONFIG_PATH}" >&2
  exit 2
fi
CONFIG_PATH="$(realpath -e "${CONFIG_PATH}")"
case "${CONFIG_PATH}" in
  "${PROJECT_ROOT}"/*) ;;
  *) echo "CFG config must remain under ${PROJECT_ROOT}: ${CONFIG_PATH}" >&2; exit 2 ;;
esac
if [[ -n "${RESUME_DIR}" ]]; then
  if [[ ! -d "${RESUME_DIR}" ]]; then
    echo "resume-dir does not exist: ${RESUME_DIR}" >&2
    exit 2
  fi
  RESUME_DIR="$(realpath -e "${RESUME_DIR}")"
  case "${RESUME_DIR}" in
    /data/myx/*) ;;
    *) echo "resume-dir must remain under /data/myx: ${RESUME_DIR}" >&2; exit 2 ;;
  esac
fi
if [[ "${MODE}" == "c3_resume" && -z "${RESUME_DIR}" ]]; then
  echo "c3_resume requires --resume-dir=global_step_20 directory" >&2
  exit 2
fi
if [[ "${MODE}" != "c3_resume" && -n "${RESUME_DIR}" ]]; then
  echo "--resume-dir is only valid with --mode c3_resume" >&2
  exit 2
fi
if [[ "${MODE}" != "c4" && ( "${C4_KIND}" != "matrix" || "${C4_MICRO_BATCH_SIZE}" != "1" || "${C4_GLOBAL_BATCH_SIZE}" != "2" || "${C4_NUM_WORKERS}" != "8" || "${C4_MAX_STEPS}" != "12" ) ]]; then
  echo "C4-only options require --mode c4" >&2
  exit 2
fi
if [[ "${MODE}" == "c4" ]]; then
  if [[ "${C4_KIND}" != "matrix" && "${C4_KIND}" != "accumulation" ]]; then
    echo "--c4-kind must be matrix or accumulation" >&2
    exit 2
  fi
  if [[ ! "${C4_MICRO_BATCH_SIZE}" =~ ^[124]$ || ! "${C4_GLOBAL_BATCH_SIZE}" =~ ^[0-9]+$ || ! "${C4_NUM_WORKERS}" =~ ^[0-9]+$ || ! "${C4_MAX_STEPS}" =~ ^[0-9]+$ ]]; then
    echo "invalid C4 numeric option" >&2
    exit 2
  fi
  if (( C4_NUM_WORKERS > 16 || C4_MAX_STEPS < 1 || C4_MAX_STEPS > 12 )); then
    echo "C4 workers must be 0..16 and max-steps must be 1..12" >&2
    exit 2
  fi
  if [[ "${C4_KIND}" == "matrix" ]]; then
    if (( C4_MAX_STEPS < 10 || C4_GLOBAL_BATCH_SIZE != C4_MICRO_BATCH_SIZE * 2 )); then
      echo "C4 matrix requires max-steps>=10 and global-batch=micro-batch*2" >&2
      exit 2
    fi
  elif (( C4_GLOBAL_BATCH_SIZE != 512 || C4_MAX_STEPS != 1 )); then
    echo "C4 accumulation requires global-batch=512 and max-steps=1" >&2
    exit 2
  fi
fi

for override in "${EXTRA_OVERRIDES[@]}"; do
  case "${override}" in
    runner.max_epochs=*|runner.max_steps=*|runner.save_interval=*|\
    data.num_workers=*|actor.micro_batch_size=*|actor.global_batch_size=*|\
    actor.fsdp_config.sharding_strategy=*|actor.fsdp_config.gradient_checkpointing=*|\
    actor.model.model_type=*|actor.model.model_path=*|actor.model.openpi.train_expert_only=*|\
    cluster.component_placement.actor=*)
      echo "Refusing override of guarded CFG field: ${override}" >&2
      exit 2
      ;;
  esac
done

LOG_ROOT="${PROJECT_ROOT}/logs/recap/cfg"
LOG_DIR="${LOG_ROOT}/${RUN_ID}"
RAY_TOKEN="$(printf '%s' "${RUN_ID}" | sha256sum | cut -c1-6)"
# Ray appends a timestamped session/socket suffix.  Keep the project-local
# prefix short enough for Linux's 107-byte AF_UNIX socket limit.
RAY_DIR="${PROJECT_ROOT}/tmp/r/${RAY_TOKEN}"
mkdir -p "${LOG_DIR}" "${RAY_DIR}"

check_gpu_idle() {
  local gpu="$1" line memory utilization compute_pids
  line="$(nvidia-smi -i "${gpu}" --query-gpu=memory.used,utilization.gpu --format=csv,noheader,nounits | tr -d ' ')"
  IFS=',' read -r memory utilization <<< "${line}"
  compute_pids="$(nvidia-smi -i "${gpu}" --query-compute-apps=pid --format=csv,noheader,nounits 2>/dev/null | tr -d '[:space:]')"
  echo "gpu=${gpu} memory_used_mb=${memory} utilization_pct=${utilization} compute_pids=${compute_pids:-none}" >> "${LOG_DIR}/gpu_idle_checks.txt"
  [[ "${memory}" =~ ^[0-9]+$ && "${utilization}" =~ ^[0-9]+$ ]]
  (( memory <= 1024 && utilization <= 5 )) && [[ -z "${compute_pids}" ]]
}

if ((DRY_RUN)); then
  echo "DRY_RUN: skipped NVIDIA driver, idle-window and Ray-process checks" | tee "${LOG_DIR}/status.txt"
else
  if ! command -v nvidia-smi >/dev/null 2>&1; then
    echo "BLOCKED_A800_DRIVER: nvidia-smi unavailable" | tee "${LOG_DIR}/status.txt" >&2
    exit 3
  fi
  if ! nvidia-smi >"${LOG_DIR}/nvidia-smi.before.txt" 2>&1; then
    echo "BLOCKED_A800_DRIVER: nvidia-smi cannot communicate with driver" | tee -a "${LOG_DIR}/status.txt" >&2
    exit 3
  fi
  if ! check_gpu_idle "${GPU0}" || ! check_gpu_idle "${GPU1}"; then
    echo "BLOCKED_GPU_AVAILABILITY: first two-card idle check failed" | tee "${LOG_DIR}/status.txt" >&2
    exit 3
  fi
  sleep "${IDLE_CHECK_SECONDS}"
  if ! nvidia-smi >"${LOG_DIR}/nvidia-smi.after.txt" 2>&1 || ! check_gpu_idle "${GPU0}" || ! check_gpu_idle "${GPU1}"; then
    echo "BLOCKED_GPU_AVAILABILITY: second two-card idle check failed" | tee "${LOG_DIR}/status.txt" >&2
    exit 3
  fi

  if pgrep -af '[r]aylet|[g]cs_server|[p]lasma_store' >"${LOG_DIR}/ray_processes.txt"; then
    echo "BLOCKED_RAY_ISOLATION: existing Ray process visible" | tee "${LOG_DIR}/status.txt" >&2
    exit 3
  fi
fi

export CUDA_VISIBLE_DEVICES="${GPUS}"
export REPO_PATH="${PROJECT_ROOT}/third_party/RLinf"
export MUJOCO_GL=egl
export PYOPENGL_PLATFORM=egl
export AV_LOG_FORCE_NOCOLOR=1
export LIBAV_LOG_LEVEL=quiet
export OPENCV_LOG_LEVEL=off
export RAY_TMPDIR="${RAY_DIR}"
export TMPDIR="${PROJECT_ROOT}/tmp"
if [[ -n "${PYTHONPATH:-}" ]]; then
  export PYTHONPATH="${PROJECT_ROOT}/third_party/RLinf:${PROJECT_ROOT}/third_party/openpi/src:${PROJECT_ROOT}/third_party/lerobot:${PYTHONPATH}"
else
  export PYTHONPATH="${PROJECT_ROOT}/third_party/RLinf:${PROJECT_ROOT}/third_party/openpi/src:${PROJECT_ROOT}/third_party/lerobot"
fi

# RLinf's placement ranks are physical node-local GPU IDs.  The parent
# CUDA_VISIBLE_DEVICES mask alone is insufficient because WorkerGroup writes
# its own per-worker mask; pin the selected physical pair explicitly.

OVERRIDES=(
  "runner.max_epochs=-1"
  "runner.val_check_interval=-1"
  "data.num_workers=0"
  "actor.micro_batch_size=1"
  "actor.global_batch_size=2"
)
OVERRIDES+=("${EXTRA_OVERRIDES[@]}")
# Escape Hydra's comma sweep separator while retaining an arbitrary physical
# pair (the pair need not be consecutive, so a range would be incorrect).
OVERRIDES+=("cluster.component_placement.actor=${GPU0}\\,${GPU1}")
if [[ "${MODE}" == "c2" ]]; then
  OVERRIDES+=("runner.max_steps=1" "runner.save_interval=-1")
elif [[ "${MODE}" == "c3" ]]; then
  OVERRIDES+=("runner.max_steps=20" "runner.save_interval=20")
elif [[ "${MODE}" == "c4" ]]; then
  OVERRIDES+=(
    "runner.max_steps=${C4_MAX_STEPS}"
    "runner.save_interval=-1"
    "actor.micro_batch_size=${C4_MICRO_BATCH_SIZE}"
    "actor.global_batch_size=${C4_GLOBAL_BATCH_SIZE}"
    "data.num_workers=${C4_NUM_WORKERS}"
  )
else
  OVERRIDES+=("runner.max_steps=21" "runner.save_interval=-1")
fi
if [[ -n "${RESUME_DIR}" ]]; then
  # The frozen YAML intentionally omits resume_dir; '+' is required under
  # Hydra's struct mode to add this run-only field.
  OVERRIDES+=("+runner.resume_dir=${RESUME_DIR}")
fi

VALIDATOR_ARGS=(--config "${CONFIG_PATH}" --mode "${MODE}" --require-two-gpus)
if [[ "${MODE}" == "c3" ]]; then
  VALIDATOR_ARGS+=(--effective-max-steps 20 --effective-save-interval 20)
elif [[ "${MODE}" == "c3_resume" ]]; then
  VALIDATOR_ARGS+=(--effective-max-steps 21 --effective-save-interval -1)
elif [[ "${MODE}" == "c4" ]]; then
  VALIDATOR_ARGS+=(
    --effective-max-steps "${C4_MAX_STEPS}"
    --effective-save-interval -1
    --effective-micro-batch-size "${C4_MICRO_BATCH_SIZE}"
    --effective-global-batch-size "${C4_GLOBAL_BATCH_SIZE}"
    --effective-num-workers "${C4_NUM_WORKERS}"
    --c4-kind "${C4_KIND}"
  )
fi
"${PROJECT_ROOT}/envs/rlinf-pi05/bin/python" "${SCRIPT_DIR}/validate_cfg_preflight.py" \
  "${VALIDATOR_ARGS[@]}" >"${LOG_DIR}/preflight.json"

CMD=(
  "${PROJECT_ROOT}/envs/rlinf-pi05/bin/python"
  "${PROJECT_ROOT}/third_party/RLinf/examples/recap/cfg/train_cfg.py"
  --config-path "$(dirname "${CONFIG_PATH}")"
  --config-name "$(basename "${CONFIG_PATH}" .yaml)"
  "runner.logger.log_path=${LOG_ROOT}"
  "runner.logger.experiment_name=${RUN_ID}"
  "hydra.run.dir=${LOG_DIR}"
  "${OVERRIDES[@]}"
)
printf '%q ' "${CMD[@]}" >"${LOG_DIR}/command.txt"
printf '\n' >>"${LOG_DIR}/command.txt"
echo "CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES}" >>"${LOG_DIR}/command.txt"
echo "RAY_TMPDIR=${RAY_TMPDIR}" >>"${LOG_DIR}/command.txt"
echo "launcher_pid=$$ pgid=$(ps -o pgid= -p $$ | tr -d ' ')" >>"${LOG_DIR}/command.txt"

if ((DRY_RUN)); then
  echo "CFG_LAUNCH_DRY_RUN_PASS log_dir=${LOG_DIR}"
  exit 0
fi

echo "GO_CFG_LAUNCH run_id=${RUN_ID} gpus=${GPUS}" | tee "${LOG_DIR}/status.txt"
echo "start_time=$(date --iso-8601=seconds)" >>"${LOG_DIR}/command.txt"
set +e
"${CMD[@]}" 2>&1 | tee "${LOG_DIR}/train.log"
CMD_RC=${PIPESTATUS[0]}
set -e

if ((CMD_RC == 0)) && [[ "${MODE}" == "c3" ]]; then
  C3_ACTOR_DIR="${LOG_DIR}/checkpoints/global_step_20/actor"
  if [[ ! -d "${C3_ACTOR_DIR}" ]]; then
    echo "C3 checkpoint directory missing: ${C3_ACTOR_DIR}" | tee -a "${LOG_DIR}/status.txt" >&2
    CMD_RC=4
  else
    set +e
    "${PROJECT_ROOT}/envs/rlinf-pi05/bin/python" \
      "${SCRIPT_DIR}/attach_cfg_norm_stats.py" \
      --checkpoint "${C3_ACTOR_DIR}" >>"${LOG_DIR}/train.log" 2>&1
    ATTACH_RC=$?
    set -e
    if ((ATTACH_RC != 0)); then
      echo "C3 norm-stats attachment failed rc=${ATTACH_RC}" | tee -a "${LOG_DIR}/status.txt" >&2
      CMD_RC=${ATTACH_RC}
    fi
  fi
fi

echo "end_time=$(date --iso-8601=seconds) exit_code=${CMD_RC}" >>"${LOG_DIR}/command.txt"
echo "CFG_LAUNCH_DONE run_id=${RUN_ID} exit_code=${CMD_RC}" | tee -a "${LOG_DIR}/status.txt"
exit "${CMD_RC}"

	#!/usr/bin/env bash
	# Guarded launcher for staged Task 0 CFG preflights.
	# It never auto-selects GPUs and never runs global Ray cleanup.

	set -euo pipefail

	PROJECT_ROOT="/data/myx/RL_post_train"
	SCRIPT_DIR="${PROJECT_ROOT}/scripts"
	CONFIG_PATH="${PROJECT_ROOT}/configs/recap/task0_cfg_recap_v1.yaml"
	MODE="c2"
	GPUS=""
	RUN_ID=""
	RESUME_DIR=""
	DRY_RUN=0
	C4_KIND="matrix"
	C4_MICRO_BATCH_SIZE=1
	C4_GLOBAL_BATCH_SIZE=2
	C4_NUM_WORKERS=8
	C4_MAX_STEPS=12
	IDLE_CHECK_SECONDS="${CFG_GPU_IDLE_CHECK_SECONDS:-30}"
	EXTRA_OVERRIDES=()

	usage() {
	cat <<'EOF'
	Usage: run_cfg_task0.sh --gpus PHYSICAL0,PHYSICAL1 [options] [-- hydra overrides]

	Options:
	--mode c2\|c3\|c3_resume\|c4
	Select the bounded preflight (default: c2).
	--gpus A,B Exactly two physical GPU indices; no auto-selection.
	--run-id ID Project-local log/experiment identifier.
	--resume-dir PATH Resume from an exact RLinf checkpoint directory.
	--config PATH Project-local CFG YAML override.
	--c4-kind matrix\|accumulation
	Select the bounded C4 benchmark shape (default: matrix).
	--micro-batch-size N / --global-batch-size N
	C4 batch sizes; accumulation uses global batch 512.
	--num-workers N C4 dataloader workers (0..16, default: 8).
	--max-steps N C4 total optimizer steps (matrix default: 12; accumulation: 1).
	--dry-run Validate and print the command without launching Ray.
	-h, --help Show this help.
	EOF
	}

	while (($# > 0)); do
	case "$1" in
	--mode) MODE="$2"; shift 2 ;;
	--gpus) GPUS="$2"; shift 2 ;;
	--run-id) RUN_ID="$2"; shift 2 ;;
	--resume-dir) RESUME_DIR="$2"; shift 2 ;;
	--config) CONFIG_PATH="$2"; shift 2 ;;
	--c4-kind) C4_KIND="$2"; shift 2 ;;
	--micro-batch-size) C4_MICRO_BATCH_SIZE="$2"; shift 2 ;;
	--global-batch-size) C4_GLOBAL_BATCH_SIZE="$2"; shift 2 ;;
	--num-workers) C4_NUM_WORKERS="$2"; shift 2 ;;
	--max-steps) C4_MAX_STEPS="$2"; shift 2 ;;
	--dry-run) DRY_RUN=1; shift ;;
	-h\|--help) usage; exit 0 ;;
	--) shift; EXTRA_OVERRIDES+=("$@"); break ;;
	*) echo "Unknown argument: $1" >&2; usage >&2; exit 2 ;;
	esac
	done

	if [[ -z "${GPUS}" ]]; then
	echo "CFG launcher requires --gpus PHYSICAL0,PHYSICAL1" >&2
	exit 2
	fi
	if [[ "${MODE}" != "c2" && "${MODE}" != "c3" && "${MODE}" != "c3_resume" && "${MODE}" != "c4" ]]; then
	echo "Only bounded c2/c3/c3_resume/c4 modes are enabled; got ${MODE}" >&2
	exit 2
	fi
	if [[ -n "${RAY_ADDRESS:-}" \|\| -n "${CLUSTER_NAMESPACE:-}" ]]; then
	echo "Refusing inherited RAY_ADDRESS/CLUSTER_NAMESPACE." >&2
	exit 3
	fi

	IFS=',' read -r GPU0 GPU1 GPU_EXTRA <<< "${GPUS}"
	if [[ -n "${GPU_EXTRA:-}" \|\| -z "${GPU0:-}" \|\| -z "${GPU1:-}" ]]; then
	echo "--gpus must contain exactly two comma-separated indices" >&2
	exit 2
	fi
	if [[ "${GPU0}" == "${GPU1}" \|\| ! "${GPU0}" =~ ^[0-9]+$ \|\| ! "${GPU1}" =~ ^[0-9]+$ ]]; then
	echo "--gpus must contain two distinct non-negative integers" >&2
	exit 2
	fi
	if [[ -z "${RUN_ID}" ]]; then
	RUN_ID="recap_cfg_${MODE}_$(date +%Y%m%d_%H%M%S)"
	fi
	if [[ ! "${RUN_ID}" =~ ^[A-Za-z0-9_.-]+$ ]]; then
	echo "run-id contains unsafe characters: ${RUN_ID}" >&2
	exit 2
	fi

	source "${PROJECT_ROOT}/scripts/activate_project_env.sh"
	if [[ ! -f "${CONFIG_PATH}" ]]; then
	echo "missing CFG config: ${CONFIG_PATH}" >&2
	exit 2
	fi
	CONFIG_PATH="$(realpath -e "${CONFIG_PATH}")"
	case "${CONFIG_PATH}" in
	"${PROJECT_ROOT}"/*) ;;
	*) echo "CFG config must remain under ${PROJECT_ROOT}: ${CONFIG_PATH}" >&2; exit 2 ;;
	esac
	if [[ -n "${RESUME_DIR}" ]]; then
	if [[ ! -d "${RESUME_DIR}" ]]; then
	echo "resume-dir does not exist: ${RESUME_DIR}" >&2
	exit 2
	fi
	RESUME_DIR="$(realpath -e "${RESUME_DIR}")"
	case "${RESUME_DIR}" in
	/data/myx/*) ;;
	*) echo "resume-dir must remain under /data/myx: ${RESUME_DIR}" >&2; exit 2 ;;
	esac
	fi
	if [[ "${MODE}" == "c3_resume" && -z "${RESUME_DIR}" ]]; then
	echo "c3_resume requires --resume-dir=global_step_20 directory" >&2
	exit 2
	fi
	if [[ "${MODE}" != "c3_resume" && -n "${RESUME_DIR}" ]]; then
	echo "--resume-dir is only valid with --mode c3_resume" >&2
	exit 2
	fi
	if [[ "${MODE}" != "c4" && ( "${C4_KIND}" != "matrix" \|\| "${C4_MICRO_BATCH_SIZE}" != "1" \|\| "${C4_GLOBAL_BATCH_SIZE}" != "2" \|\| "${C4_NUM_WORKERS}" != "8" \|\| "${C4_MAX_STEPS}" != "12" ) ]]; then
	echo "C4-only options require --mode c4" >&2
	exit 2
	fi
	if [[ "${MODE}" == "c4" ]]; then
	if [[ "${C4_KIND}" != "matrix" && "${C4_KIND}" != "accumulation" ]]; then
	echo "--c4-kind must be matrix or accumulation" >&2
	exit 2
	fi
	if [[ ! "${C4_MICRO_BATCH_SIZE}" =~ ^[124]$ \|\| ! "${C4_GLOBAL_BATCH_SIZE}" =~ ^[0-9]+$ \|\| ! "${C4_NUM_WORKERS}" =~ ^[0-9]+$ \|\| ! "${C4_MAX_STEPS}" =~ ^[0-9]+$ ]]; then
	echo "invalid C4 numeric option" >&2
	exit 2
	fi
	if (( C4_NUM_WORKERS > 16 \|\| C4_MAX_STEPS < 1 \|\| C4_MAX_STEPS > 12 )); then
	echo "C4 workers must be 0..16 and max-steps must be 1..12" >&2
	exit 2
	fi
	if [[ "${C4_KIND}" == "matrix" ]]; then
	if (( C4_MAX_STEPS < 10 \|\| C4_GLOBAL_BATCH_SIZE != C4_MICRO_BATCH_SIZE * 2 )); then
	echo "C4 matrix requires max-steps>=10 and global-batch=micro-batch*2" >&2
	exit 2
	fi
	elif (( C4_GLOBAL_BATCH_SIZE != 512 \|\| C4_MAX_STEPS != 1 )); then
	echo "C4 accumulation requires global-batch=512 and max-steps=1" >&2
	exit 2
	fi
	fi

	for override in "${EXTRA_OVERRIDES[@]}"; do
	case "${override}" in
	runner.max_epochs=\|runner.max_steps=\|runner.save_interval=*\|\
	data.num_workers=\|actor.micro_batch_size=\|actor.global_batch_size=*\|\
	actor.fsdp_config.sharding_strategy=\|actor.fsdp_config.gradient_checkpointing=\|\
	actor.model.model_type=\|actor.model.model_path=\|actor.model.openpi.train_expert_only=*\|\
	cluster.component_placement.actor=*)
	echo "Refusing override of guarded CFG field: ${override}" >&2
	exit 2
	;;
	esac
	done

	LOG_ROOT="${PROJECT_ROOT}/logs/recap/cfg"
	LOG_DIR="${LOG_ROOT}/${RUN_ID}"
	RAY_TOKEN="$(printf '%s' "${RUN_ID}" \| sha256sum \| cut -c1-6)"
	# Ray appends a timestamped session/socket suffix. Keep the project-local
	# prefix short enough for Linux's 107-byte AF_UNIX socket limit.
	RAY_DIR="${PROJECT_ROOT}/tmp/r/${RAY_TOKEN}"
	mkdir -p "${LOG_DIR}" "${RAY_DIR}"

	check_gpu_idle() {
	local gpu="$1" line memory utilization compute_pids
	line="$(nvidia-smi -i "${gpu}" --query-gpu=memory.used,utilization.gpu --format=csv,noheader,nounits \| tr -d ' ')"
	IFS=',' read -r memory utilization <<< "${line}"
	compute_pids="$(nvidia-smi -i "${gpu}" --query-compute-apps=pid --format=csv,noheader,nounits 2>/dev/null \| tr -d '[:space:]')"
	echo "gpu=${gpu} memory_used_mb=${memory} utilization_pct=${utilization} compute_pids=${compute_pids:-none}" >> "${LOG_DIR}/gpu_idle_checks.txt"
	[[ "${memory}" =~ ^[0-9]+$ && "${utilization}" =~ ^[0-9]+$ ]]
	(( memory <= 1024 && utilization <= 5 )) && [[ -z "${compute_pids}" ]]
	}

	if ((DRY_RUN)); then
	echo "DRY_RUN: skipped NVIDIA driver, idle-window and Ray-process checks" \| tee "${LOG_DIR}/status.txt"
	else
	if ! command -v nvidia-smi >/dev/null 2>&1; then
	echo "BLOCKED_A800_DRIVER: nvidia-smi unavailable" \| tee "${LOG_DIR}/status.txt" >&2
	exit 3
	fi
	if ! nvidia-smi >"${LOG_DIR}/nvidia-smi.before.txt" 2>&1; then
	echo "BLOCKED_A800_DRIVER: nvidia-smi cannot communicate with driver" \| tee -a "${LOG_DIR}/status.txt" >&2
	exit 3
	fi
	if ! check_gpu_idle "${GPU0}" \|\| ! check_gpu_idle "${GPU1}"; then
	echo "BLOCKED_GPU_AVAILABILITY: first two-card idle check failed" \| tee "${LOG_DIR}/status.txt" >&2
	exit 3
	fi
	sleep "${IDLE_CHECK_SECONDS}"
	if ! nvidia-smi >"${LOG_DIR}/nvidia-smi.after.txt" 2>&1 \|\| ! check_gpu_idle "${GPU0}" \|\| ! check_gpu_idle "${GPU1}"; then
	echo "BLOCKED_GPU_AVAILABILITY: second two-card idle check failed" \| tee "${LOG_DIR}/status.txt" >&2
	exit 3
	fi

	if pgrep -af '[r]aylet\|[g]cs_server\|[p]lasma_store' >"${LOG_DIR}/ray_processes.txt"; then
	echo "BLOCKED_RAY_ISOLATION: existing Ray process visible" \| tee "${LOG_DIR}/status.txt" >&2
	exit 3
	fi
	fi

	export CUDA_VISIBLE_DEVICES="${GPUS}"
	export REPO_PATH="${PROJECT_ROOT}/third_party/RLinf"
	export MUJOCO_GL=egl
	export PYOPENGL_PLATFORM=egl
	export AV_LOG_FORCE_NOCOLOR=1
	export LIBAV_LOG_LEVEL=quiet
	export OPENCV_LOG_LEVEL=off
	export RAY_TMPDIR="${RAY_DIR}"
	export TMPDIR="${PROJECT_ROOT}/tmp"
	if [[ -n "${PYTHONPATH:-}" ]]; then
	export PYTHONPATH="${PROJECT_ROOT}/third_party/RLinf:${PROJECT_ROOT}/third_party/openpi/src:${PROJECT_ROOT}/third_party/lerobot:${PYTHONPATH}"
	else
	export PYTHONPATH="${PROJECT_ROOT}/third_party/RLinf:${PROJECT_ROOT}/third_party/openpi/src:${PROJECT_ROOT}/third_party/lerobot"
	fi

	# RLinf's placement ranks are physical node-local GPU IDs. The parent
	# CUDA_VISIBLE_DEVICES mask alone is insufficient because WorkerGroup writes
	# its own per-worker mask; pin the selected physical pair explicitly.

	OVERRIDES=(
	"runner.max_epochs=-1"
	"runner.val_check_interval=-1"
	"data.num_workers=0"
	"actor.micro_batch_size=1"
	"actor.global_batch_size=2"
	)
	OVERRIDES+=("${EXTRA_OVERRIDES[@]}")
	# Escape Hydra's comma sweep separator while retaining an arbitrary physical
	# pair (the pair need not be consecutive, so a range would be incorrect).
	OVERRIDES+=("cluster.component_placement.actor=${GPU0}\\,${GPU1}")
	if [[ "${MODE}" == "c2" ]]; then
	OVERRIDES+=("runner.max_steps=1" "runner.save_interval=-1")
	elif [[ "${MODE}" == "c3" ]]; then
	OVERRIDES+=("runner.max_steps=20" "runner.save_interval=20")
	elif [[ "${MODE}" == "c4" ]]; then
	OVERRIDES+=(
	"runner.max_steps=${C4_MAX_STEPS}"
	"runner.save_interval=-1"
	"actor.micro_batch_size=${C4_MICRO_BATCH_SIZE}"
	"actor.global_batch_size=${C4_GLOBAL_BATCH_SIZE}"
	"data.num_workers=${C4_NUM_WORKERS}"
	)
	else
	OVERRIDES+=("runner.max_steps=21" "runner.save_interval=-1")
	fi
	if [[ -n "${RESUME_DIR}" ]]; then
	# The frozen YAML intentionally omits resume_dir; '+' is required under
	# Hydra's struct mode to add this run-only field.
	OVERRIDES+=("+runner.resume_dir=${RESUME_DIR}")
	fi

	VALIDATOR_ARGS=(--config "${CONFIG_PATH}" --mode "${MODE}" --require-two-gpus)
	if [[ "${MODE}" == "c3" ]]; then
	VALIDATOR_ARGS+=(--effective-max-steps 20 --effective-save-interval 20)
	elif [[ "${MODE}" == "c3_resume" ]]; then
	VALIDATOR_ARGS+=(--effective-max-steps 21 --effective-save-interval -1)
	elif [[ "${MODE}" == "c4" ]]; then
	VALIDATOR_ARGS+=(
	--effective-max-steps "${C4_MAX_STEPS}"
	--effective-save-interval -1
	--effective-micro-batch-size "${C4_MICRO_BATCH_SIZE}"
	--effective-global-batch-size "${C4_GLOBAL_BATCH_SIZE}"
	--effective-num-workers "${C4_NUM_WORKERS}"
	--c4-kind "${C4_KIND}"
	)
	fi
	"${PROJECT_ROOT}/envs/rlinf-pi05/bin/python" "${SCRIPT_DIR}/validate_cfg_preflight.py" \
	"${VALIDATOR_ARGS[@]}" >"${LOG_DIR}/preflight.json"

	CMD=(
	"${PROJECT_ROOT}/envs/rlinf-pi05/bin/python"
	"${PROJECT_ROOT}/third_party/RLinf/examples/recap/cfg/train_cfg.py"
	--config-path "$(dirname "${CONFIG_PATH}")"
	--config-name "$(basename "${CONFIG_PATH}" .yaml)"
	"runner.logger.log_path=${LOG_ROOT}"
	"runner.logger.experiment_name=${RUN_ID}"
	"hydra.run.dir=${LOG_DIR}"
	"${OVERRIDES[@]}"
	)
	printf '%q ' "${CMD[@]}" >"${LOG_DIR}/command.txt"
	printf '\n' >>"${LOG_DIR}/command.txt"
	echo "CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES}" >>"${LOG_DIR}/command.txt"
	echo "RAY_TMPDIR=${RAY_TMPDIR}" >>"${LOG_DIR}/command.txt"
	echo "launcher_pid=$$ pgid=$(ps -o pgid= -p $$ \| tr -d ' ')" >>"${LOG_DIR}/command.txt"

	if ((DRY_RUN)); then
	echo "CFG_LAUNCH_DRY_RUN_PASS log_dir=${LOG_DIR}"
	exit 0
	fi

	echo "GO_CFG_LAUNCH run_id=${RUN_ID} gpus=${GPUS}" \| tee "${LOG_DIR}/status.txt"
	echo "start_time=$(date --iso-8601=seconds)" >>"${LOG_DIR}/command.txt"
	set +e
	"${CMD[@]}" 2>&1 \| tee "${LOG_DIR}/train.log"
	CMD_RC=${PIPESTATUS[0]}
	set -e

	if ((CMD_RC == 0)) && [[ "${MODE}" == "c3" ]]; then
	C3_ACTOR_DIR="${LOG_DIR}/checkpoints/global_step_20/actor"
	if [[ ! -d "${C3_ACTOR_DIR}" ]]; then
	echo "C3 checkpoint directory missing: ${C3_ACTOR_DIR}" \| tee -a "${LOG_DIR}/status.txt" >&2
	CMD_RC=4
	else
	set +e
	"${PROJECT_ROOT}/envs/rlinf-pi05/bin/python" \
	"${SCRIPT_DIR}/attach_cfg_norm_stats.py" \
	--checkpoint "${C3_ACTOR_DIR}" >>"${LOG_DIR}/train.log" 2>&1
	ATTACH_RC=$?
	set -e
	if ((ATTACH_RC != 0)); then
	echo "C3 norm-stats attachment failed rc=${ATTACH_RC}" \| tee -a "${LOG_DIR}/status.txt" >&2
	CMD_RC=${ATTACH_RC}
	fi
	fi
	fi

	echo "end_time=$(date --iso-8601=seconds) exit_code=${CMD_RC}" >>"${LOG_DIR}/command.txt"
	echo "CFG_LAUNCH_DONE run_id=${RUN_ID} exit_code=${CMD_RC}" \| tee -a "${LOG_DIR}/status.txt"
	exit "${CMD_RC}"