llama13b_poc/ds_config.json

{
  "train_batch_size" : 256,
  "train_micro_batch_size_per_gpu": 1,
  "steps_per_print": 10,
  "gradient_clipping": 1.0,
  "zero_optimization": {
    "stage": 0
  },
  "bf16": {
    "enabled": true,
    "accumulate_grads_via_hooks": true
  },
  "fp16": {"enabled": false},
  "wall_clock_breakdown": false,
  "pipeline": {
    "pipe_partitioned": false,
    "grad_partitioned": false
  }
}