| import argparse |
|
|
| def parse_args(): |
| parser = argparse.ArgumentParser(description='Abbie trainer') |
| parser.add_argument("--pp_size", type=int, default=1, help="Pipeline parallelism size") |
| parser.add_argument("--ep_size", type=int, default=1, help="Data parallelism size") |
|
|
| |
| parser.add_argument("--model", type=str, required=True, help="Launching template") |
| |
| parser.add_argument("--trial_name", type=str, required=True, help="Namespace string for booking trial") |
| parser.add_argument("--train_home_path", type=str, required=True, help="Model path") |
|
|
| |
| parser.add_argument("--train_dataset", type=str, nargs="+", required=True, help="training dataset HDFS path") |
| parser.add_argument("--train_size", type=int, required=True, help="train_size in token amount") |
|
|
| |
| parser.add_argument("--tokenizer", type=str, required=True, help="HDFS path for tokenizer") |
| parser.add_argument("--pad_idx", type=int, default=1, help="tokenizer pad idx, default to 1") |
| parser.add_argument("--vocab_size", type=int, required=True, help="Vocab size") |
|
|
| |
| parser.add_argument("--global_batch_size", type=int, required=True, help="GBS setting") |
| parser.add_argument("--micro_batch_size", type=int, default=1, help="micro batch size setting") |
|
|
| parser.add_argument("--warmup_batch_ratio", type=float, required=True, help="Batch size") |
|
|
| |
| parser.add_argument("--max_seq_len", type=int, required=True, help="max_seq_len") |
| parser.add_argument("--max_position_embeddings", type=int, required=True, help="max_position_embeddings") |
| parser.add_argument("--stride", type=int, required=True, help="Data stride for text") |
|
|
| |
| parser.add_argument("--layer_number", type=int, default=None, help="Batch size") |
| parser.add_argument("--inner", type=int, default=None, help="Inner size") |
| parser.add_argument("--is_moe", type=bool, default=False, help="is moe enabled") |
| parser.add_argument("--expert_number", type=int, default=None, help="Moe expert number") |
| parser.add_argument("--num_shared_qheads", type=int, default=None, help="GQA shared head number") |
| parser.add_argument("--attention_dropout", type=float, default=None, help="Attention dropout in float, default to None") |
| parser.add_argument("--residual_dropout", type=float, default=None, help="Residual dropout in float, default to None") |
|
|
| |
| parser.add_argument("--lr_max", type=float, required=True, help="Learning rate maximum value") |
| parser.add_argument("--lr_min", type=float, required=True, help="Learning rate minimum value") |
| parser.add_argument("--lr_weight_decay", type=float, required=True, help="Learning rate weight decay ") |
| parser.add_argument("--lr_warmup_step_rate", type=float, required=True, help="Learning rate warmup step rate") |
|
|
| parser.add_argument("--vit_lr_max", type=float, help="Vision Transformer learning rate maximum value") |
| parser.add_argument("--vit_lr_min", type=float, help="Vision Transformer learning rate minimum value") |
|
|
| |
| parser.add_argument("--enable_profiler", type=int, default=False, help="Random seed") |
| parser.add_argument("--profiler_step", type=int, default=-1, help="Random seed") |
|
|
| |
| parser.add_argument("--ckpt_every_n_step", type=int, default=4000, help="Random seed") |
|
|
| |
| parser.add_argument("--resume_ckpt_step", type=int, default=None, help="Resuming step") |
|
|
| |
| parser.add_argument("--resume_from_sahara", type=str, default=None, help="Resume from sahara dumped weight") |
|
|
| |
| parser.add_argument("--pretrained_hf_path", type=str, default=None, help="Initialize with pretrained hf model") |
|
|
| return parser.parse_args() |
|
|
