# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import abc import warnings from typing import Tuple import torch from nemo.utils.decorators import experimental __all__ = [ "MegatronPretrainingBatchSampler", "MegatronPretrainingRandomBatchSampler", ] class BaseMegatronBatchSampler: """Megatron style BatchSampler. Let mbs, gbs, tp, pp, and dp stand for "micro batch size", "global batch size", "tensor model parallel world size", "pipeline model parallel world size", and "data parallel world size", the number of micro batches (hereafter, nmb) is defined as :math:`nmb = gbs \\div (mbs \\times dp)`. See `apex/transformer/microbatches.py#L91-L98 `_ for the initial settings of the number of micro batches and `apex/transformer/microbatches.py#L160-L177 _`. for warming up of global batch size. e.g.) `(mbs, gbs, tp, pp, dp) = (1, 16, 1, 1, 2)`, then the number of micro batches is :math:`gbs \\div (mbs \\times dp) = 16 \\div (1 \\times 2) = 8`. In this case, an instance of Megatron Batch Sampler on each data parallel rank is expected returns :math:`nmb \\times mbs = 8` indices. """ _global_batch_size: int _num_micro_batches: int _global_batch_size_on_this_data_parallel_rank: int def __init__( self, total_samples: int, consumed_samples: int, micro_batch_size: int, global_batch_size: int, data_parallel_rank: int, data_parallel_size: int, drop_last: bool, pad_samples_to_global_batch_size=False, ) -> None: """Constructor of Megatron-LM style Batch Sampler. Args: total_samples: The size of dataset. consumed_samples: The number of samples that have been used. micro_batch_size: The size of each micro batch. global_batch_size: The size of global batch. data_parallel_rank: The value you can obtain via `parallel_state.get_data_parallel_rank()` of megatron.core. data_parallel_size: The value you can obtain via `parallel_state.get_data_parallel_world_size()` of megatron.core. """ # Sanity checks. if total_samples <= 0: raise RuntimeError("no sample to consume: {}".format(total_samples)) if micro_batch_size <= 0: raise RuntimeError(f"micro_batch_size size must be greater than 0, but {micro_batch_size}") if data_parallel_size <= 0: raise RuntimeError(f"data parallel size must be greater than 0, but {data_parallel_size}") if data_parallel_rank >= data_parallel_size: raise RuntimeError( "data_parallel_rank should be smaller than data size, but {} >= {}".format( data_parallel_rank, data_parallel_size ) ) # Keep a copy of input params for later use. self.total_samples: int = total_samples self.consumed_samples: int = consumed_samples self.micro_batch_size: int = micro_batch_size self.data_parallel_rank: int = data_parallel_rank self.data_parallel_size: int = data_parallel_size self.drop_last: bool = drop_last self.pad_samples_to_global_batch_size = pad_samples_to_global_batch_size self.micro_batch_times_data_parallel_size = self.micro_batch_size * self.data_parallel_size self.update_global_batch_size(global_batch_size) def update_global_batch_size(self, new_global_batch_size: int) -> None: """Update the global batch size.""" self._global_batch_size = new_global_batch_size if self._global_batch_size % self.micro_batch_times_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})`" ) self._num_micro_batches = self._global_batch_size // self.micro_batch_times_data_parallel_size self._global_batch_size_on_this_data_parallel_rank = self._num_micro_batches * self.micro_batch_size @property def global_batch_size(self) -> int: return self._global_batch_size @global_batch_size.setter def global_batch_size(self, new_global_batch_size: int) -> None: warnings.warn("`self.update_global_batch_size(new_global_batch_size)` is recommended.") self.update_global_batch_size(new_global_batch_size=new_global_batch_size) def __len__(self) -> int: """Length of Batch Sampler. ..note:: When `rampup_batch_size` is enabled, the return value can be not exactly precise. """ num_available_samples: int = self.total_samples - self.consumed_samples % self.total_samples if self.drop_last: return num_available_samples // self.global_batch_size else: return (num_available_samples + self.global_batch_size - 1) // self.global_batch_size @abc.abstractmethod def __iter__(self): ... class MegatronPretrainingBatchSampler(BaseMegatronBatchSampler): def get_start_end_idx(self) -> Tuple[int, int]: start_idx = self.data_parallel_rank * self._global_batch_size_on_this_data_parallel_rank end_idx = start_idx + self._global_batch_size_on_this_data_parallel_rank return start_idx, end_idx def __iter__(self): batch = [] # Last batch will be dropped if drop_last is not set False for idx in range(self.consumed_samples % self.total_samples, self.total_samples): batch.append(idx) if len(batch) == self._global_batch_size: # start_idx, end_idx = self.get_start_end_idx() indices = [ batch[i] for i in range( self.data_parallel_rank, self._global_batch_size, self.data_parallel_size, ) ] assert len(indices) == self._global_batch_size_on_this_data_parallel_rank yield indices # yield batch[start_idx:end_idx] batch = [] # Check the last partial batch and see drop_last is set if len(batch) > 0 and not self.drop_last: # start_idx, end_idx = self.get_start_end_idx() indices = [batch[i] for i in range(self.data_parallel_rank, len(batch), self.data_parallel_size)] if self.pad_samples_to_global_batch_size: num_pad = self._global_batch_size // self.data_parallel_size - len(indices) indices = indices + [-1] * num_pad yield indices @experimental class MegatronPretrainingRandomBatchSampler(BaseMegatronBatchSampler): # NOTE (mkozuki): [[Argument of `dataset` and `data_sharding`]] # From the commit below, it seems like `dataset` argument and `data_sharding` argument # are necessary for ViT training. However, to keep this simple, # I omit those two arguments. # commit: https://github.com/NVIDIA/Megatron-LM/commit/7a77abd9b6267dc0020a60b424b4748fc22790bb # # NOTE (degert): I have re-written this class somewhat to give the length correctly when consumed_samples # are larger than total_samples, which happens with epochs > 1 training when using this Sampler # I have also added an explicit seed which allows us to remove Dataset-side shuffling in Nemo-Aligner # # This class does not currently work with pad_samples_to_global_batch_size=True def __init__( self, total_samples: int, consumed_samples: int, micro_batch_size: int, global_batch_size: int, data_parallel_rank: int, data_parallel_size: int, drop_last: bool, pad_samples_to_global_batch_size: bool = False, seed: int = 0, ) -> None: super().__init__( total_samples=total_samples, consumed_samples=consumed_samples, micro_batch_size=micro_batch_size, data_parallel_rank=data_parallel_rank, data_parallel_size=data_parallel_size, drop_last=drop_last, global_batch_size=global_batch_size, pad_samples_to_global_batch_size=pad_samples_to_global_batch_size, ) assert ( not pad_samples_to_global_batch_size ), "`MegatronPretrainingRandomBatchSampler` does not support sample padding" if (not drop_last) and self.micro_batch_times_data_parallel_size > 1: raise RuntimeError( "`MegatronPretrainingRandomBatchSampler` does not support drop_last=False when micro_batch_size * data_parallel_size > 1. \ please reduce your MBS and data parallelism to 1 if you want to use drop_last=False, or switch to drop_last=True to avoid this error" ) self.last_batch_size = self.total_samples % self._global_batch_size self.seed = seed def __len__(self) -> int: """Length of Random Batch Sampler. ..note:: When `rampup_batch_size` is enabled, the return value can be not exactly precise. """ active_total_samples = self.total_samples - (self.last_batch_size if self.drop_last else 0) num_available_samples = active_total_samples - self.consumed_samples % active_total_samples if self.drop_last: return num_available_samples // self.global_batch_size else: return (num_available_samples + self.global_batch_size - 1) // self.global_batch_size def __iter__(self): active_total_samples = self.total_samples - self.last_batch_size self.epoch = self.consumed_samples // active_total_samples current_epoch_samples = self.consumed_samples % active_total_samples assert current_epoch_samples % self.micro_batch_times_data_parallel_size == 0 # data sharding and random sampling bucket_size = (self.total_samples // self.micro_batch_times_data_parallel_size) * self.micro_batch_size bucket_offset = current_epoch_samples // self.data_parallel_size start_idx = self.data_parallel_rank * bucket_size g = torch.Generator() g.manual_seed(self.seed + self.epoch) random_idx = torch.randperm(bucket_size, generator=g).tolist() idx_range = [start_idx + x for x in random_idx[bucket_offset:]] batch = [] # Last batch if not complete will be dropped. for idx in idx_range: batch.append(idx) if len(batch) == self._global_batch_size_on_this_data_parallel_rank: self.consumed_samples += self._global_batch_size yield batch batch = [] # Check the last partial batch and see drop_last is set if len(batch) > 0 and not self.drop_last: yield batch