MCPcopy Create free account
hub / github.com/deepspeedai/DeepSpeedExamples / add_training_args

Function add_training_args

Megatron-LM/arguments.py:96–173  ·  view source on GitHub ↗

Training arguments.

(parser)

Source from the content-addressed store, hash-verified

94
95
96def add_training_args(parser):
97 """Training arguments."""
98
99 group = parser.add_argument_group('train', 'training configurations')
100
101 group.add_argument('--batch-size', type=int, default=4,
102 help='Data Loader batch size')
103 group.add_argument('--weight-decay', type=float, default=0.01,
104 help='weight decay coefficient for L2 regularization')
105 group.add_argument('--checkpoint-activations', action='store_true',
106 help='checkpoint activation to allow for training '
107 'with larger models and sequences')
108 group.add_argument('--checkpoint-num-layers', type=int, default=1,
109 help='chunk size (number of layers) for checkpointing')
110 group.add_argument('--partition-activations', action='store_true',
111 help='partition the checkpoint activation across model parallel process')
112 group.add_argument('--clip-grad', type=float, default=1.0,
113 help='gradient clipping')
114 group.add_argument('--train-iters', type=int, default=1000000,
115 help='total number of iterations to train over all training runs')
116 group.add_argument('--log-interval', type=int, default=100,
117 help='report interval')
118 group.add_argument('--exit-interval', type=int, default=None,
119 help='Exit the program after this many new iterations.')
120
121 group.add_argument('--seed', type=int, default=1234,
122 help='random seed')
123 # Batch prodecuer arguments
124 group.add_argument('--reset-position-ids', action='store_true',
125 help='Reset posistion ids after end-of-document token.')
126 group.add_argument('--reset-attention-mask', action='store_true',
127 help='Reset self attention maske after '
128 'end-of-document token.')
129
130 # Learning rate.
131 group.add_argument('--lr-decay-iters', type=int, default=None,
132 help='number of iterations to decay LR over,'
133 ' If None defaults to `--train-iters`*`--epochs`')
134 group.add_argument('--lr-decay-style', type=str, default='linear',
135 choices=['constant', 'linear', 'cosine', 'exponential'],
136 help='learning rate decay function')
137 group.add_argument('--lr', type=float, default=1.0e-4,
138 help='initial learning rate')
139 group.add_argument('--warmup', type=float, default=0.01,
140 help='percentage of data to warmup on (.01 = 1% of all '
141 'training iters). Default 0.01')
142 # model checkpointing
143 group.add_argument('--save', type=str, default=None,
144 help='Output directory to save checkpoints to.')
145 group.add_argument('--save-interval', type=int, default=5000,
146 help='number of iterations between saves')
147 group.add_argument('--no-save-optim', action='store_true',
148 help='Do not save current optimizer.')
149 group.add_argument('--no-save-rng', action='store_true',
150 help='Do not save current rng state.')
151 group.add_argument('--load', type=str, default=None,
152 help='Path to a directory containing a model checkpoint.')
153 group.add_argument('--no-load-optim', action='store_true',

Callers 1

get_argsFunction · 0.85

Calls

no outgoing calls

Tested by

no test coverage detected