{
    "bf16": { "enabled": true },
  
    "zero_optimization": {
      "stage": 3,
      "overlap_comm": false,
      "contiguous_gradients": true,
  
      /* 下面三项是按 40GB 显存收紧的桶大小与持久化阈值 */
      "reduce_bucket_size": 200000000,
      "stage3_prefetch_bucket_size": 50000000,
      "stage3_param_persistence_threshold": 1000000,
  
      "stage3_gather_16bit_weights_on_model_save": true
    },
  
    "aio": {
      "block_size": 1048576,
      "queue_depth": 8,
      "single_submit": false,
      "overlap_events": true,
      "thread_count": 1
    },
  
    /* 交由命令行的 per_device_train_batch_size / gradient_accumulation_steps 控制 */
    "train_micro_batch_size_per_gpu": "auto",
    "gradient_accumulation_steps": "auto",
  
    "zero_force_ds_cpu_optimizer": false,
    "zero_allow_untested_optimizer": true,
    "wall_clock_breakdown": false
  }