provenance/deepspeed.json

{
  "bf16": {"enabled": "auto"},
  "train_micro_batch_size_per_gpu": "auto",
  "train_batch_size": "auto",
  "gradient_accumulation_steps": "auto",
  "zero_force_ds_cpu_optimizer": false,
  "zero_optimization": {
    "stage": 2,
    "overlap_comm": true,
    "contiguous_gradients": true,
    "reduce_bucket_size": 50000000,
    "allgather_bucket_size": 50000000,
    "offload_optimizer": {"device": "cpu", "pin_memory": true}
  }
}