ds_config.json

{
  "train_batch_size": 4,
  "train_micro_batch_size_per_gpu": 1,
  "gradient_accumulation_steps": 4,
  
  "optimizer": {
    "type": "AdamW",
    "params": {
      "lr": 1.5e-4,
      "betas": [0.9, 0.999],
      "eps": 1e-8,
      "weight_decay": 0.01,
      "torch_adam": true
    }
  },
  
  "scheduler": {
    "type": "WarmupLR",
    "params": {
      "warmup_min_lr": 0,
      "warmup_max_lr": 1.5e-4,
      "warmup_num_steps": 500
    }
  },
  
  "zero_optimization": {
    "stage": 2,
    "offload_optimizer": {
      "device": "cpu",
      "pin_memory": false
    },
    "overlap_comm": true,
    "contiguous_gradients": true,
    "reduce_bucket_size": 1e6,
    "gather_16bit_weights_on_model_save": false
  },
  
  "bf16": {
    "enabled": true
  },
  
  "gradient_clipping": 1.0,
  
  "activation_checkpointing": {
    "partition_activations": true,
    "contiguous_memory_optimization": true,
    "number_checkpoints": 12,
    "synchronize_checkpoint_boundary": false,
    "cpu_checkpointing": true
  },
  
  "wall_clock_breakdown": false,
  
  "steps_per_print": 100,
  
  "fp16": {
    "enabled": false
  },
  
  "amp": {
    "enabled": false,
    "amp_master_weights": false,
    "loss_scale_window": 1000
  }
}