Megatron-DeepSpeed/examples_deepspeed/compression/ds_config_gpt_TEMPLATE.json

{
  "train_batch_size" : CONFIG_BATCH_SIZE,
  "train_micro_batch_size_per_gpu": CONFIG_MBSIZE,
  "steps_per_print": LOG_INTERVAL,

  "zero_optimization": {
    "stage": ZERO_STAGE
  },

  "gradient_clipping": 1.0,
  "prescale_gradients": PRESCALE_GRAD,

  "fp16": {
    "enabled": CONFIG_FP16_ENABLED,
    "loss_scale": 0,
    "loss_scale_window": 500,
    "hysteresis": 2,
    "min_loss_scale": 1,
    "initial_scale_power": 11
  },

  "bf16": {
    "enabled": CONFIG_BF16_ENABLED
  },
  "curriculum_learning": {
    "enabled": CONFIG_CL_ENABLED,
    "curriculum_type": "seqlen",
    "min_difficulty": CONFIG_CL_MIN,
    "max_difficulty": CONFIG_CL_MAX,
    "schedule_type": "fixed_linear",
    "schedule_config": {
      "total_curriculum_step": CONFIG_CL_DURATION,
      "difficulty_step": 8
    }
  },

  "wall_clock_breakdown" : false
}