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

export NCCL_DEBUG=INFO
# 如走 IB/RoCE，请按实际网卡开启：
# export NCCL_IB_HCA="mlx5_0,mlx5_1"
# export NCCL_SOCKET_IFNAME="ib0"
# 纯以太：
# export NCCL_SOCKET_IFNAME="eth0"

# ==== 超参数（本地路径；可用 VAR=xxx ./launch_ds.sh 覆写）====
MODEL_NAME_OR_PATH="${MODEL_NAME_OR_PATH:-/home/test/Qwen3-8B}"
DATA_GLOB="${DATA_GLOB:-/data/datasets/my_corpus/*.jsonl}"   # 每台机器都放相同路径
OUTDIR="${OUTDIR:-/data/checkpoints/run-qwen3-8b}"           # 每台机器各自本地输出
SEQ_LEN="${SEQ_LEN:-4096}"
LR="${LR:-2e-5}"
GAS="${GAS:-64}"
LOG_STEPS="${LOG_STEPS:-10}"
SAVE_STEPS="${SAVE_STEPS:-500}"
MAX_STEPS="${MAX_STEPS:-10000}"

mkdir -p "${OUTDIR}"

# ==== 多机 DeepSpeed ====
deepspeed --hostfile hostfile train_sft_ds.py \
  --model_name_or_path "${MODEL_NAME_OR_PATH}" \
  --data_glob "${DATA_GLOB}" \
  --output_dir "${OUTDIR}" \
  --seq_len "${SEQ_LEN}" \
  --learning_rate "${LR}" \
  --gradient_accumulation_steps "${GAS}" \
  --per_device_train_batch_size 1 \
  --warmup_ratio 0.02 \
  --weight_decay 0.1 \
  --max_steps "${MAX_STEPS}" \
  --log_interval "${LOG_STEPS}" \
  --save_steps "${SAVE_STEPS}" \
  --deepspeed ds_config_zero3.json \
  --gradient_checkpointing \
  --bf16