config.json

{
    "train_micro_batch_size_per_gpu": 2,
    "gradient_accumulation_steps": 2,
    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": "auto",
            "betas": "auto",
            "eps": "auto",
            "weight_decay": "auto"
        }
    },
    "scheduler": {
        "type": "WarmupLR",
        "params": {
            "warmup_min_lr": "auto",
            "warmup_max_lr": "auto",
            "warmup_num_steps": "auto"
        }
    },
    "zero_optimization": {
        "stage": 3,
        "allgather_partitions": true,
        "allgather_bucket_size": 5e8,
        "overlap_comm": false,
        "sub_group_size": 1e9,
        "reduce_bucket_size": "auto",
        "contiguous_gradients" : true,
        "stage3_prefetch_bucket_size": "auto",
        "stage3_params_persistence_threshold": "auto",
        "stage3_max_live_parameters": 1e9,
        "stage3_max_reuse_distance": 1e9,
        "stage3_gather_16bit_weights_on_model_save": false,
        "offload_param": {
            "device":"cpu",
            "pin_memory":true
        },
        "offload_optimizer": {
            "device":"cpu",
            "pin_memory":true
        }
    },
    "bf16":{
        "enabled": "auto"
    },
    "profiling": {
        "enabled": true,
        "output_path": "deepspeed_profile.json"
    },
    "checkpoint": {
        "parallel_checkpoint": true
    },
    "communication_data_compression": true
}