pretrain_kaiyuan2b/scripts/kaiyuan2b-training/training_smoke_qwen3_1p7b.sh

#!/usr/bin/env bash
set -euo pipefail

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export CUDA_DEVICE_MAX_CONNECTIONS=1

MODE=${1:-qwen3_1p7b_smoke_yi}
TRAIN_NAME=${2:-qwen3_1p7b_smoke_yi}

MEGATRON_PATH=/apps/yi/model_training/Megatron-LM
ARTIFACT_ROOT=/apps/yi/model_training/artifacts
SCRIPT_DIR=/apps/yi/model_training/scripts/kaiyuan2b-training
PARAMS_DIR="${SCRIPT_DIR}/params"
TB_DIR="${ARTIFACT_ROOT}/tb_logs/${TRAIN_NAME}"
CKPT_DIR="${ARTIFACT_ROOT}/checkpoints/${TRAIN_NAME}"
CHECKPOINT_KEEP_RECENT=${CHECKPOINT_KEEP_RECENT:-3}
CHECKPOINT_CLEANUP_INTERVAL_SECONDS=${CHECKPOINT_CLEANUP_INTERVAL_SECONDS:-300}
EXTRA_ARGS=${EXTRA_ARGS:-}
NPROC_PER_NODE=${NPROC_PER_NODE:-8}
NNODES=${NNODES:-1}
NODE_RANK=${NODE_RANK:-0}
MASTER_ADDR=${MASTER_ADDR:-localhost}
MASTER_PORT=${MASTER_PORT:-6000}
ZERO_STAGE=${ZERO_STAGE:-0}

source "${PARAMS_DIR}/optim_common.sh"
source "${PARAMS_DIR}/qwen3_1p7b/model.sh"
source "${PARAMS_DIR}/qwen3_1p7b/data_phase1_smoke.sh"
source "${PARAMS_DIR}/qwen3_1p7b/hparams.sh"

LOGGING_ARGS="
  --tensorboard-dir ${TB_DIR}
  --tensorboard-log-interval 1
  --log-interval 1
  --log-timers-to-tensorboard
  --log-validation-ppl-to-tensorboard
  --log-memory-to-tensorboard
  --log-world-size-to-tensorboard
  --log-num-zeros-in-grad
  --log-device-memory-used
  --log-throughput
  --log-params-norm
"

PRECISION_ARGS="--bf16"
PARALLEL_ARGS="
    --tensor-model-parallel-size 1
    --pipeline-model-parallel-size 1
"
# PARALLEL_ARGS="
#     --tensor-model-parallel-size 2
#     --sequence-parallel
# "

if [ "$MODE" = "qwen3_1p7b_smoke_yi" ]; then
  source "${PARAMS_DIR}/qwen3_1p7b/hparams.sh"
  RUN_ARGS="$HPARAMS"
elif [ "$MODE" = "phase1" ]; then
  source "${PARAMS_DIR}/phase1_full.sh"
  RUN_ARGS="$PHASE_ARGS"
else
  echo "Unknown mode: $MODE"
  exit 1
fi

case "$ZERO_STAGE" in
  0)
    ZERO_ARGS=""
    ;;
  1)
    ZERO_ARGS="
      --use-distributed-optimizer
      --data-parallel-sharding-strategy optim
    "
    ;;
  2)
    ZERO_ARGS="
      --use-distributed-optimizer
      --data-parallel-sharding-strategy optim_grads
    "
    ;;
  *)
    echo "Unsupported ZERO_STAGE=${ZERO_STAGE}; expected 0, 1, or 2" >&2
    exit 1
    ;;
esac

mkdir -p "$CKPT_DIR" "$TB_DIR"

cleanup_old_checkpoints_once() {
  local ckpt_dir=$1
  local keep=$2

  if ! [[ "$keep" =~ ^[0-9]+$ ]] || [ "$keep" -le 0 ] || [ ! -d "$ckpt_dir" ]; then
    return 0
  fi

  local latest=""
  if [ -f "${ckpt_dir}/latest_checkpointed_iteration.txt" ]; then
    read -r latest < "${ckpt_dir}/latest_checkpointed_iteration.txt" || latest=""
    if [[ "$latest" =~ ^[0-9]+$ ]]; then
      latest=$(printf "iter_%07d" "$latest")
    else
      latest=""
    fi
  fi

  local checkpoints=()
  while IFS= read -r path; do
    checkpoints+=("$path")
  done < <(find "$ckpt_dir" -maxdepth 1 -type d -name 'iter_[0-9][0-9][0-9][0-9][0-9][0-9][0-9]' -print | sort)

  local delete_count=$((${#checkpoints[@]} - keep))
  if [ "$delete_count" -le 0 ]; then
    return 0
  fi

  local i base
  for ((i = 0; i < delete_count; i++)); do
    base=$(basename "${checkpoints[$i]}")
    if [ "$base" = "$latest" ]; then
      continue
    fi
    echo "[checkpoint-cleanup] deleting old checkpoint: ${checkpoints[$i]}"
    rm -rf -- "${checkpoints[$i]}"
  done
}

checkpoint_cleanup_loop() {
  local ckpt_dir=$1
  local keep=$2
  local interval=$3

  if ! [[ "$interval" =~ ^[0-9]+$ ]] || [ "$interval" -le 0 ]; then
    return 0
  fi

  while true; do
    sleep "$interval"
    cleanup_old_checkpoints_once "$ckpt_dir" "$keep"
  done
}

checkpoint_cleanup_loop "$CKPT_DIR" "$CHECKPOINT_KEEP_RECENT" "$CHECKPOINT_CLEANUP_INTERVAL_SECONDS" &
CHECKPOINT_CLEANUP_PID=$!
trap 'kill "$CHECKPOINT_CLEANUP_PID" 2>/dev/null || true; cleanup_old_checkpoints_once "$CKPT_DIR" "$CHECKPOINT_KEEP_RECENT"' EXIT

DISTRIBUTED_ARGS="
  --nproc_per_node ${NPROC_PER_NODE}
  --nnodes ${NNODES}
  --node_rank ${NODE_RANK}
  --master_addr ${MASTER_ADDR}
  --master_port ${MASTER_PORT}
"

torchrun $DISTRIBUTED_ARGS \
  $MEGATRON_PATH/pretrain_gpt.py \
  $MODEL_ARGS \
  $OPTIM_ARGS \
  $ZERO_ARGS \
  $PRECISION_ARGS \
  $PARALLEL_ARGS \
  $DATA_ARGS \
  $RUN_ARGS \
  $LOGGING_ARGS\
  --save "$CKPT_DIR" \
  --load "$CKPT_DIR" \
  --enable-cuda-graph \
  --cuda-graph-warmup-steps 3 \
  --transformer-impl transformer_engine \
  --cross-entropy-loss-fusion \
  --cross-entropy-fusion-impl te \
  $EXTRA_ARGS