#!/usr/bin/bash

set -x
umask 007

# ─── Config ───
GPUS_PER_NODE=${GPUS_PER_NODE:-8}
LOG_RANK=${LOG_RANK:-"0"}
CONFIG_NAME=${CONFIG_NAME:-"robotwin_pretrain"}
SAVE_ROOT=${SAVE_ROOT:-"/c20250502/szh/train_out_ego2robotwin_postrain"}
LOG_DIR=${LOG_DIR:-"${SAVE_ROOT}/logs/torchrun"}
DEBUG=${DEBUG:-""}
cache_root=${CACHE_ROOT:-"/vepfs-mlp2/c20250502/haoce/szh/lingbot-va/cache_v2"}

debug_flag=""
if [ -n "${DEBUG}" ]; then
    debug_flag="--debug"
fi

overrides=""
if [ $# -ne 0 ]; then
    overrides="$*"
fi

# ─── Environment ───
export DATASET_VERSION="v2"
export PYTHONUNBUFFERED=1
export PYTHONWARNINGS="ignore::UserWarning"
export TOKENIZERS_PARALLELISM=false
export PYTORCH_ALLOC_CONF="${PYTORCH_ALLOC_CONF:-expandable_segments:True}"
# export TORCH_DISTRIBUTED_DEBUG=DETAIL

export NCCL_IB_DISABLE=0
export NCCL_P2P_DISABLE=0
export NCCL_SHM_DISABLE=0
export NCCL_TIMEOUT=36000

export HF_HOME="${HF_HOME:-${cache_root}/hf_home}"
export HF_DATASETS_CACHE="${HF_DATASETS_CACHE:-${cache_root}/hf_datasets}"
export TMPDIR="${TMPDIR:-${cache_root}/tmp/node_${MLP_ROLE_INDEX}}"
export TORCHINDUCTOR_CACHE_DIR="${TORCHINDUCTOR_CACHE_DIR:-${cache_root}/torchinductor}"
export TRITON_CACHE_DIR="${TRITON_CACHE_DIR:-${cache_root}/triton}"
export TORCH_EXTENSIONS_DIR="${TORCH_EXTENSIONS_DIR:-${cache_root}/torch_extensions}"

export WANDB_API_KEY="${WANDB_API_KEY:-"705019b40922ce96942fbdca07c2b22ebc3891fe"}"
export WANDB_BASE_URL="${WANDB_BASE_URL:-"https://api.wandb.ai"}"
export WANDB_TEAM_NAME="${WANDB_TEAM_NAME:-"screatorx"}"
export WANDB_PROJECT="${WANDB_PROJECT:-"lingbot-va-ego2robotwin-postrain"}"

# ─── Setup dirs ───
mkdir -p "${HF_HOME}" "${HF_DATASETS_CACHE}" "${TMPDIR}" \
         "${TORCHINDUCTOR_CACHE_DIR}" "${TRITON_CACHE_DIR}" \
         "${TORCH_EXTENSIONS_DIR}" "${LOG_DIR}"

# ─── Launch ───
echo "GPUS_PER_NODE: ${GPUS_PER_NODE}"

LOG_FILE="${LOG_DIR}/train_node${MLP_ROLE_INDEX}_$(date +%Y%m%d_%H%M%S).log"
mkdir -p ${LOG_DIR}

torchrun \
    --nnodes=$MLP_WORKER_NUM \
    --node_rank=$MLP_ROLE_INDEX \
    --master_addr=$MLP_WORKER_0_HOST \
    --master_port=$MLP_WORKER_0_PORT \
    --nproc_per_node=${GPUS_PER_NODE} \
    -m wan_va.pretrain_v2 \
    --config-name ${CONFIG_NAME} \
    --save-root ${SAVE_ROOT} \
    ${debug_flag} \
    $overrides \
    2>&1 | tee ${LOG_FILE}