zero-2.json

{
  "bf16": { "enabled": "auto" },
  "optimizer": {
    "type": "AdamW",
    "params": {
      "lr": "auto",
      "weight_decay": "auto",
      "torch_adam": "false",
      "adam_w_mode": "true"
    }
  },
  "scheduler": {
    "type": "WarmupLR",
    "params": {
        "warmup_min_lr": 0,
        "warmup_max_lr": "auto",
        "warmup_num_steps": "auto"
    }
  },
  "zero_optimization": {
    "stage": 2,
    "allgather_partitions": true,
    "allgather_bucket_size": 5e6,
    "reduce_scatter": true,
    "reduce_bucket_size": 5e6,
    "overlap_comm": true,
    "contiguous_gradients": true
  },
  "activation_checkpointing": {
    "partition_activations": true,
    "contiguous_memory_optimization": true,
    "synchronize_checkpoint_boundary": true
  },
  "gradient_checkpointing": true,
  "wall_clock_breakdown": true,
  "steps_per_print": 1,
  "dump_state": true,
  "gradient_accumulation_steps": "auto",
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto"
}