Class ParallelismConfig

llm_analysis/config.py:115–126

@dataclass
class ParallelismConfig:
    tp_size: int = 1  # tensor parallelism size, Megatron-LM tensor parallelism implementation
    pp_size: int = 1  # pipeline parallelism size, Megatron-LM pipeline parallelism implementation
    dp_size: int = (
        1  # data parallelism size, DeepSpeed Zero parallelism implementation
    )
    ep_size: int = 1  # expert parallelism size
    sp_size: int = None  # sequence parallelism size, Megatron-LM sequence parallelism implementation

    def __post_init__(self):
        if self.sp_size is None:
            self.sp_size = self.tp_size


# model name and configurations mapping populated from MODEL_CONFIG_DIR_NAME