scrapegoat-lora/ds_config_zero3_nvme.json

{
    "bf16": {
        "enabled": true
    },
    "zero_optimization": {
        "stage": 3,
        "offload_optimizer": {
            "device": "nvme",
            "nvme_path": "/ephemeral/ds_offload",
            "pin_memory": true,
            "buffer_count": 4,
            "fast_init": false
        },
        "offload_param": {
            "device": "nvme",
            "nvme_path": "/ephemeral/ds_offload",
            "pin_memory": true,
            "buffer_count": 4,
            "buffer_size": 6e9,
            "max_in_cpu": 1e10
        },
        "overlap_comm": true,
        "contiguous_gradients": true,
        "sub_group_size": 1e10,
        "reduce_bucket_size": "auto",
        "stage3_prefetch_bucket_size": "auto",
        "stage3_param_persistence_threshold": "auto",
        "stage3_max_live_parameters": 1e10,
        "stage3_max_reuse_distance": 1e10,
        "stage3_gather_16bit_weights_on_model_save": true
    },
    "gradient_accumulation_steps": 8,
    "gradient_clipping": 1.0,
    "steps_per_print": 10,
    "train_micro_batch_size_per_gpu": 1,
    "wall_clock_breakdown": false
}