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

Function _add_memoryopt_args

codegeex/megatron/arguments.py:1485–1517  ·  view source on GitHub ↗

Memory optimization arguments.

(parser)

Source from the content-addressed store, hash-verified

1483
1484
1485def _add_memoryopt_args(parser):
1486 """Memory optimization arguments."""
1487
1488 group = parser.add_argument_group("Memory optimizations", "configurations")
1489 group.add_argument(
1490 "--scattered-embeddings",
1491 action="store_true",
1492 help="Save memory by scattering embedding activations. "
1493 "Introduces dropout differences across MP configurations.",
1494 )
1495 group.add_argument(
1496 "--split-transformers",
1497 action="store_true",
1498 help="Save memory by splitting transformer layers into two parts, "
1499 "allowing for more frequent activation checkpoint savings.",
1500 )
1501 group.add_argument(
1502 "--memory-centric-tiled-linear",
1503 action="store_true",
1504 help="Save memory by tiling with deepspeed.zero.TiledLinear.",
1505 )
1506 group.add_argument(
1507 "--tile-factor",
1508 type=int,
1509 default=1,
1510 help="Make all linear layers the same size of [hidden/tile_factor, hidden/tile_factor]. "
1511 "Must be enabled with --memory-centric-tiled-linear. "
1512 "Example A: if tile_factor=1, the qkv layer [hidden, 3* hidden] would be converted into [1,3] tiles of size [hidden,hidden]. "
1513 "Example B: if tile_factor=2, the intermediate layer [4*hidden, hidden] will be converted into [8, 2] tiles of size [hidden/2, hidden/2]. "
1514 "Default is 1.",
1515 )
1516
1517 return parser
1518
1519
1520def _add_activation_checkpoint_args(parser):

Callers 1

parse_argsFunction · 0.85

Calls

no outgoing calls

Tested by

no test coverage detected