| from typing import List, Any, Optional | |
| from zencfg import ConfigBase | |
| class DistributedConfig(ConfigBase): | |
| """ | |
| DistributedConfig provides config options for multi-GPU | |
| and multi-node settings. Our current setup depends on ``torchrun``-based | |
| Elastic Launch and the ``nccl`` distributed backend for communication. | |
| Parameters | |
| ---------- | |
| use_distributed: bool, default False | |
| Whether to use distributed data/model parallelism | |
| model_parallel_size: Optional[int], default 1 | |
| number of GPUs across which to spread model layers, | |
| by default 1. If 1, does not perform **any** model parallelism. | |
| seed: Optional[int] = None | |
| special distributed random torch seed for reproducibility. | |
| """ | |
| use_distributed: bool = False | |
| model_parallel_size: Optional[int] = 1 | |
| seed: Optional[int] = None | |