def __post_init__()

in src/nanotron/data/samplers.py [0:0]

• 33 lines of code
• 9 McCabe index (conditional complexity)

    def __post_init__(self):
        self.micro_batch_times_data_parallel_size = self.micro_batch_size * self.data_parallel_size

        # Sanity checks.
        if self.total_samples <= 0:
            raise RuntimeError("no sample to consume: {}".format(self.total_samples))
        if self.consumed_samples >= self.total_samples:
            raise RuntimeError("no samples left to consume: {}, {}".format(self.consumed_samples, self.total_samples))
        if self.micro_batch_size <= 0:
            raise RuntimeError(f"micro_batch_size size must be greater than 0, but {self.micro_batch_size}")
        if self.data_parallel_size <= 0:
            raise RuntimeError(f"data parallel size must be greater than 0, but {self.data_parallel_size}")
        if self.data_parallel_rank >= self.data_parallel_size:
            raise RuntimeError(
                "data_parallel_rank should be smaller than data size, but {} >= {}".format(
                    self.data_parallel_rank, self.data_parallel_size
                )
            )
        if self.global_batch_size % (self.micro_batch_size * self.data_parallel_size) != 0:
            raise RuntimeError(
                f"`global_batch_size` ({self.global_batch_size}) is not divisible by "
                f"`micro_batch_size ({self.micro_batch_size}) x data_parallel_size "
                f"({self.data_parallel_size})`"
            )
        if self.pad_samples_to_global_batch_size and self.global_batch_size is None:
            raise RuntimeError(
                "`pad_samples_to_global_batch_size` can be `True` only when "
                "`global_batch_size` is set to an integer value"
            )
        log_rank(
            f"Instantiating MegatronPretrainingSampler with total_samples: {self.total_samples} and consumed_samples: {self.consumed_samples}",
            logger=logger,
            level=logging.INFO,
            rank=0,
        )