deepspeed_config.json

{
  "train_micro_batch_size_per_gpu": 2,
  "gradient_accumulation_steps": 1,
  "steps_per_print": 100,
  "gradient_clipping": 1.0,
  "fp16": {
    "enabled": true,
    "loss_scale": 0,
    "loss_scale_window": 2000,
    "hysteresis": 2,
    "min_loss_scale": 0.0
  },
  "zero_optimization": {
    "stage": 2,
    "reduce_bucket_size": 50000000,
    "overlap_comm": true
  },
  "sparse_attention": {
    "mode": "fixed",
    "block": 16,
    "different_layout_per_head": true,
    "num_local_blocks": 8,
    "num_global_blocks": 1,
    "attention": "unidirectional",
    "horizontal_global_attention": false,
    "num_different_global_patterns": 8
  }
}