MCPcopy Create free account
hub / github.com/zai-org/CodeGeeX / _add_training_args

Function _add_training_args

codegeex/megatron/arguments.py:589–745  ·  view source on GitHub ↗
(parser)

Source from the content-addressed store, hash-verified

587
588
589def _add_training_args(parser):
590 group = parser.add_argument_group(title="training")
591
592 group.add_argument(
593 "--micro-batch-size",
594 type=int,
595 default=None,
596 help="Batch size per model instance (local batch size). "
597 "Global batch size is local batch size times data "
598 "parallel size times number of micro batches.",
599 )
600 group.add_argument(
601 "--batch-size",
602 type=int,
603 default=None,
604 help="Old batch size parameter, do not use. " "Use --micro-batch-size instead",
605 )
606 group.add_argument(
607 "--global-batch-size",
608 type=int,
609 default=None,
610 help="Training batch size. If set, it should be a "
611 "multiple of micro-batch-size times data-parallel-size. "
612 "If this value is None, then "
613 "use micro-batch-size * data-parallel-size as the "
614 "global batch size. This choice will result in 1 for "
615 "number of micro-batches.",
616 )
617 group.add_argument(
618 "--rampup-batch-size",
619 nargs="*",
620 default=None,
621 help="Batch size ramp up with the following values:"
622 " --rampup-batch-size <start batch size> "
623 " <batch size incerement> "
624 " <ramp-up samples> "
625 "For example:"
626 " --rampup-batch-size 16 8 300000 \ "
627 " --global-batch-size 1024"
628 "will start with global batch size 16 and over "
629 " (1024 - 16) / 8 = 126 intervals will increase"
630 "the batch size linearly to 1024. In each interval"
631 "we will use approximately 300000 / 126 = 2380 samples.",
632 )
633 group.add_argument(
634 "--checkpoint-activations",
635 action="store_true",
636 help="Checkpoint activation to allow for training "
637 "with larger models, sequences, and batch sizes.",
638 )
639 group.add_argument(
640 "--distribute-checkpointed-activations",
641 action="store_true",
642 help="If set, distribute checkpointed activations "
643 "across model parallel group.",
644 )
645 group.add_argument(
646 "--checkpoint-num-layers",

Callers 1

parse_argsFunction · 0.85

Calls

no outgoing calls

Tested by

no test coverage detected