ds_zero3_offload.json

{
  "train_batch_size": 32,
  "train_micro_batch_size_per_gpu": 4,
  "steps_per_print": 1,
  "zero_optimization": {
    "stage": 3,
    "offload_param": {
      "device": "cpu"
    },
    "offload_optimizer": {
      "device": "cpu"
    },
    "stage3_gather_16bit_weights_on_model_save": true
  },
  "bf16": {
    "enabled": false
  },
  "fp16": {
    "enabled": true,
    "loss_scale": 0,
    "loss_scale_window": 100
  },
  "gradient_clipping": 1.0,
  "prescale_gradients": false,
  "wall_clock_breakdown": false,
  "scheduler": {
        "type": "WarmupDecayLR",
        "params": {
            "last_batch_iteration": -1,
            "total_num_steps": "auto",
            "warmup_min_lr": "auto",
            "warmup_max_lr": "auto",
            "warmup_num_steps": "auto",
            "warmup_type": "cosine"
        }
    },
  "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": "auto",
            "betas": "auto",
            "eps": "auto",
            "weight_decay": "auto"
        }
    }
}