unilm/dit/classification/deepspeed_configs/config.json

{
    "train_batch_size": 2048,
    "train_micro_batch_size_per_gpu": 32,
    "steps_per_print": 1000,
    "optimizer": {
        "type": "Adam",
        "adam_w_mode": true,
        "params": {
            "lr": 0.001,
            "weight_decay": 0.05,
            "bias_correction": true,
            "betas": [
                0.9,
                0.999
            ]
        }
    },
    "gradient_clipping": 1,
    "fp16": {
        "enabled": true,
        "loss_scale": 0,
        "initial_scale_power": 7,
        "loss_scale_window": 128
    },
    "zero_optimization": {
        "stage": 1
    }
}