ds_zero3.json

{
    "train_micro_batch_size_per_gpu": 1,
    "gradient_accumulation_steps": "auto",
    "optimizer": {
        "type": "AdamW",
        "params": {
          "lr": "auto",
          "betas": "auto",
          "eps": "auto",
          "weight_decay": "auto"
        }
    },
    "fp16": {
        "enabled": true
    },
    "zero_optimization": {
        "stage": 3,
        "offload_optimizer": {
            "device": "cpu",
            "pin_memory": true
        },
        "offload_param": {
            "device": "cpu",
            "pin_memory": true
        },
        "overlap_comm": true,
        "contiguous_gradients": true,
        "round_robin_gradients": true,
        "sub_group_size": 1e9,
        "reduce_bucket_size": "auto",
        "stage3_prefetch_bucket_size": "auto",
        "stage3_param_persistence_threshold": "auto",
        "stage3_max_live_parameters": 1e9,
        "stage3_max_reuse_distance": 1e9,
        "stage3_gather_16bit_weights_on_model_save": true
    },
    "steps_per_print": 500
}