|
| 1 | +# Copyright (c) Meta Platforms, Inc. and affiliates. |
| 2 | +# All rights reserved. |
| 3 | +import argparse |
| 4 | +import sys |
| 5 | +from collections import defaultdict |
| 6 | +from typing import Union |
| 7 | + |
| 8 | +try: |
| 9 | + import tomllib |
| 10 | +except ModuleNotFoundError: |
| 11 | + import tomli as tomllib |
| 12 | + |
| 13 | + |
| 14 | +class JobConfig: |
| 15 | + """ |
| 16 | + A helper class to manage the train configuration. |
| 17 | + Semantics: |
| 18 | + - Default config is loaded from a toml file. If no toml file is provided, |
| 19 | + then the default config is loaded from argparse defaults. |
| 20 | + """ |
| 21 | + |
| 22 | + def parse_args(self, args_list: list = sys.argv[1:]): |
| 23 | + args = JobConfig.init_args_from_command_line(args_list) |
| 24 | + config_file = getattr(args, "job.config_file", None) |
| 25 | + if config_file is None: |
| 26 | + args_dict = self._args_to_two_level_dict(args) |
| 27 | + else: |
| 28 | + with open(config_file, "rb") as f: |
| 29 | + args_dict = tomllib.load(f) |
| 30 | + for k, v in args_dict.items(): |
| 31 | + class_type = type(k.title(), (), v) |
| 32 | + setattr(self, k, class_type()) |
| 33 | + self._validate_config() |
| 34 | + |
| 35 | + def _args_to_two_level_dict(self, args: argparse.Namespace) -> defaultdict: |
| 36 | + args_dict = defaultdict(defaultdict) |
| 37 | + for k, v in vars(args).items(): |
| 38 | + first_level_key, second_level_key = k.split(".", 1) |
| 39 | + args_dict[first_level_key][second_level_key] = v |
| 40 | + return args_dict |
| 41 | + |
| 42 | + def _validate_config(self): |
| 43 | + # TODO: Add more mandatory validations |
| 44 | + assert self.model.name and self.model.flavor and self.model.tokenizer_path |
| 45 | + return True |
| 46 | + |
| 47 | + @staticmethod |
| 48 | + def init_args_from_command_line( |
| 49 | + args_list: list = sys.argv[1:], |
| 50 | + ) -> argparse.Namespace: |
| 51 | + """ |
| 52 | + Each argument starts with <prefix>_ which is the section name in the toml file |
| 53 | + followed by name of the option in the toml file. For ex, |
| 54 | + model.name translates to: |
| 55 | + [model] |
| 56 | + name |
| 57 | + in the toml file |
| 58 | + """ |
| 59 | + parser = argparse.ArgumentParser(description="TorchTrain arg parser.") |
| 60 | + parser.add_argument( |
| 61 | + "--job.config_file", |
| 62 | + type=str, |
| 63 | + default=None, |
| 64 | + help="job config file", |
| 65 | + ) |
| 66 | + |
| 67 | + # misc configs |
| 68 | + parser.add_argument( |
| 69 | + "--job.dump_folder", |
| 70 | + type=str, |
| 71 | + default="./torchtrain/outputs", |
| 72 | + help="folder to dump job outputs", |
| 73 | + ) |
| 74 | + |
| 75 | + # profiling configs |
| 76 | + parser.add_argument( |
| 77 | + "--profiling.run_profiler", |
| 78 | + action="store_true", |
| 79 | + help="enable pytorch profiler", |
| 80 | + ) |
| 81 | + parser.add_argument( |
| 82 | + "--profiling.save_traces_folder", |
| 83 | + type=str, |
| 84 | + default="profiling/traces", |
| 85 | + help="trace file location", |
| 86 | + ) |
| 87 | + parser.add_argument( |
| 88 | + "--profiling.profile_every_x_iter", |
| 89 | + type=int, |
| 90 | + default=10, |
| 91 | + help="collect profiler traces every x iterations", |
| 92 | + ) |
| 93 | + # metrics configs |
| 94 | + parser.add_argument( |
| 95 | + "--metrics.log_freq", |
| 96 | + type=int, |
| 97 | + default=10, |
| 98 | + help="how often to log metrics to TensorBoard", |
| 99 | + ) |
| 100 | + parser.add_argument( |
| 101 | + "--metrics.enable_tensorboard", |
| 102 | + action="store_true", |
| 103 | + help="how often to log metrics to TensorBoard", |
| 104 | + ) |
| 105 | + parser.add_argument( |
| 106 | + "--metrics.save_tb_folder", |
| 107 | + type=str, |
| 108 | + default="tb", |
| 109 | + help="folder to dump tensorboard state", |
| 110 | + ) |
| 111 | + |
| 112 | + # model configs |
| 113 | + parser.add_argument( |
| 114 | + "--model.name", |
| 115 | + type=str, |
| 116 | + default="llama", |
| 117 | + help="which model to train", |
| 118 | + ) |
| 119 | + parser.add_argument( |
| 120 | + "--model.flavor", |
| 121 | + type=str, |
| 122 | + default="debugmodel", |
| 123 | + help="which model config to train", |
| 124 | + ) |
| 125 | + parser.add_argument( |
| 126 | + "--model.tokenizer_path", |
| 127 | + type=str, |
| 128 | + default="./torchtrain/datasets/tokenizer/tokenizer.model", |
| 129 | + help="tokenizer path", |
| 130 | + ) |
| 131 | + |
| 132 | + # optimizer configs |
| 133 | + parser.add_argument( |
| 134 | + "--optimizer.name", type=str, default="AdamW", help="optimizer to use" |
| 135 | + ) |
| 136 | + parser.add_argument( |
| 137 | + "--optimizer.lr", type=float, default=8e-4, help="learning rate to use" |
| 138 | + ) |
| 139 | + |
| 140 | + # training configs |
| 141 | + parser.add_argument( |
| 142 | + "--training.dataset", type=str, default="alpaca", help="dataset to use" |
| 143 | + ) |
| 144 | + parser.add_argument( |
| 145 | + "--training.batch_size", type=int, default=8, help="batch size" |
| 146 | + ) |
| 147 | + parser.add_argument( |
| 148 | + "--training.seq_len", type=int, default=2048, help="sequence length" |
| 149 | + ) |
| 150 | + parser.add_argument( |
| 151 | + "--training.warmup_pct", |
| 152 | + type=float, |
| 153 | + default=0.20, |
| 154 | + help="percentage of total training steps to use for warmup", |
| 155 | + ) |
| 156 | + parser.add_argument( |
| 157 | + "--training.max_norm", |
| 158 | + type=Union[float, int], |
| 159 | + default=1.0, |
| 160 | + help="max norm for gradient clipping", |
| 161 | + ) |
| 162 | + parser.add_argument( |
| 163 | + "--training.steps", type=int, default=-1, help="how many train steps to run" |
| 164 | + ) |
| 165 | + parser.add_argument( |
| 166 | + "--training.data_parallel_degree", |
| 167 | + type=int, |
| 168 | + default=-1, |
| 169 | + help="Data Parallelism degree. -1 means leftover ranks will be used (After SP/PP). 1 means disabled.", |
| 170 | + ) |
| 171 | + parser.add_argument( |
| 172 | + "--training.sequence_parallel_degree", |
| 173 | + type=int, |
| 174 | + default=1, |
| 175 | + help="Sequence Parallelism degree. 1 means disabled.", |
| 176 | + ) |
| 177 | + parser.add_argument( |
| 178 | + "--training.pipeline_parallel_degree", |
| 179 | + type=int, |
| 180 | + default=1, |
| 181 | + help="Pipeline Parallelism degree (default of 1 means disabled)", |
| 182 | + ) |
| 183 | + parser.add_argument( |
| 184 | + "--training.compile", |
| 185 | + action="store_true", |
| 186 | + help="Whether to compile the model.", |
| 187 | + ) |
| 188 | + parser.add_argument( |
| 189 | + "--training.checkpoint_interval", |
| 190 | + type=int, |
| 191 | + default=3600, |
| 192 | + help=( |
| 193 | + "Checkpointing interval. The unit of measurement is in seconds or " |
| 194 | + "steps depending on --training.checkpoint-internval-type." |
| 195 | + ), |
| 196 | + ) |
| 197 | + parser.add_argument( |
| 198 | + "--training.checkpoint_interval_type", |
| 199 | + type=str, |
| 200 | + default="steps", |
| 201 | + help=( |
| 202 | + "The checkpointing interval unit of measurement." |
| 203 | + "The default value is step." |
| 204 | + ), |
| 205 | + ) |
| 206 | + parser.add_argument( |
| 207 | + "--training.checkpoint_folder", |
| 208 | + type=str, |
| 209 | + default="", |
| 210 | + help=( |
| 211 | + "The folder to store the checkpoints. If this is not specified or " |
| 212 | + "is an empty string, checkpointing is disabled." |
| 213 | + ), |
| 214 | + ) |
| 215 | + return parser.parse_args(args_list) |
0 commit comments