ds_config.json

{
    "train_micro_batch_size_per_gpu": 256,
    "gradient_accumulation_steps": 1,
    "optimizer": {
        "type": "Adam",
        "params": {
            "lr": 3e-4,
            "weight_decay": 0
        }
    },
    "scheduler": {
        "type": "WarmupLR",
        "params": {
            "warmup_min_lr": 0,
            "warmup_max_lr": 3e-4,
            "warmup_num_steps": 1000
        }
    },
    "zero_optimization": {
        "stage": 1,
        "allgather_partitions": true,
        "allgather_bucket_size": 50000000,
        "reduce_scatter": true,
        "reduce_bucket_size": 50000000,
        "overlap_comm": false,
        "contiguous_gradients": true,
        "elastic_checkpoint": true
    },
    "bf16": {
        "enabled": false
    },
    "amp": {
        "enabled": false,
        "opt_level": "O1",
        "min_loss_scale": 32.0
    },
    "flops_profiler": {
        "enabled": false,
        "profile_step": 10,
        "module_depth": -1,
        "top_modules": 1,
        "detailed": true,
        "output_file": null,
        "end_profile_step": 5
    },
    "wall_clock_breakdown": false,
    "gradient_clipping": 0.5,
    "prescale_gradients": false,
    "steps_per_print": 10000
}