Fix minor issues with config option defaults, and indentation error i…

…n calculate_derived

megatron/neox_arguments/arguments.py

@@ -784,7 +784,7 @@ def calculate_batch_parameters(
  else:
  assert (
  False
- ), "Either train_batch_size or micro_batch_per_gpu needs to be provided"
+ ), "Either train_batch_size or train_micro_batch_size_per_gpu needs to be provided"
  return int(train_batch), int(micro_batch), int(grad_acc)
 
  @staticmethod
@@ -907,11 +907,11 @@ def calculate_derived(self):
  save_iters = list(save_iters)
  save_iters.sort()
 
- self.update_values(
- {
- "save_iters": save_iters,
- }
- )
+  self.update_values(
+  {
+  "save_iters": save_iters,
+  }
+  )
 
  # derive precision
  if (self.fp16 or {}).get("type", self.precision) == "bfloat16":

megatron/neox_arguments/neox_args.py

@@ -490,7 +490,7 @@ class NeoXArgsLogging(NeoXArgsTemplate):
  Write TensorBoard logs to this directory.
  """
 
- log_interval: int = None
+ log_interval: int = 100
  """
  Interval between logging.
  """

megatron/training.py

@@ -447,6 +447,10 @@ def get_optimizer(model, neox_args):
  """Set up the optimizer."""
  if neox_args.no_load_optim:
  return None, None
+
+ if neox_args.optimizer is None:
+ print_rank_0(f'ERROR: Optimizer is None. Either set the optimizer dict in your config (if training) or set no_load_optim in your config (if inference)')
+ exit()
  # Build parameter groups (weight decay and non-decay).
  param_groups = get_params_for_weight_decay_optimization(model, neox_args)
  print_rank_0(