GPT2 base on megatron-deepspeed

megatron-deepspeed_dtk22.10/megatron/__init__.py

+# coding=utf-8
+# Copyright (c) 2020, 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 torch
+
+from .package_info import (
+    __description__,
+    __contact_names__,
+    __url__,
+    __download_url__,
+    __keywords__,
+    __license__,
+    __package_name__,
+    __version__,
+)
+
+from .global_vars import get_args
+from .global_vars import get_current_global_batch_size
+from .global_vars import get_num_microbatches
+from .global_vars import update_num_microbatches
+from .global_vars import get_tokenizer
+from .global_vars import get_tensorboard_writer
+from .global_vars import get_adlr_autoresume
+from .global_vars import get_timers
+from .initialize  import initialize_megatron
+
+def print_rank_0(message):
+    """If distributed is initialized, print only on rank 0."""
+    if torch.distributed.is_initialized():
+        if torch.distributed.get_rank() == 0:
+            print(message, flush=True)
+    else:
+        print(message, flush=True)
+
+def is_last_rank():
+    return torch.distributed.get_rank() == (
+        torch.distributed.get_world_size() - 1)
+
+def print_rank_last(message):
+    """If distributed is initialized, print only on last rank."""
+    if torch.distributed.is_initialized():
+        if is_last_rank():
+            print(message, flush=True)
+    else:
+        print(message, flush=True)

megatron-deepspeed_dtk22.10/megatron/arguments.py

+# coding=utf-8
+# Copyright (c) 2020, 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.
+
+"""Megatron arguments."""
+
+import argparse
+import collections
+import os
+import re
+import time
+
+import torch
+import deepspeed
+
+from megatron.enums import PositionEmbeddingType
+import megatron
+from megatron.logging import log_levels
+
+
+def parse_args(extra_args_provider=None, defaults={},
+               ignore_unknown_args=False):
+    """Parse all arguments."""
+    parser = argparse.ArgumentParser(description='Megatron-LM Arguments',
+                                     allow_abbrev=False)
+
+    # Standard arguments.
+    parser = _add_network_size_args(parser)
+    parser = _add_regularization_args(parser)
+    parser = _add_training_args(parser)
+    parser = _add_initialization_args(parser)
+    parser = _add_learning_rate_args(parser)
+    parser = _add_checkpointing_args(parser)
+    parser = _add_mixed_precision_args(parser)
+    parser = _add_distributed_args(parser)
+    parser = _add_validation_args(parser)
+    parser = _add_data_args(parser)
+    parser = _add_autoresume_args(parser)
+    parser = _add_biencoder_args(parser)
+    parser = _add_vit_args(parser)
+    parser = _add_logging_args(parser)
+    parser = _add_zero_args(parser)
+    parser = _add_memoryopt_args(parser)
+    parser = _add_activation_checkpoint_args(parser)
+
+    # Custom arguments.
+    if extra_args_provider is not None:
+        parser = extra_args_provider(parser)
+
+    parser = deepspeed.add_config_arguments(parser)
+
+    # Parse.
+    if ignore_unknown_args:
+        args, _ = parser.parse_known_args()
+    else:
+        args = parser.parse_args()
+
+    # Distributed args.
+    args.rank = int(os.getenv('RANK', '0'))
+    args.world_size = int(os.getenv("WORLD_SIZE", '1'))
+    # Tensor model parallel size.
+    args.tensor_model_parallel_size = min(
+        args.tensor_model_parallel_size, args.world_size)
+    assert args.world_size % args.tensor_model_parallel_size == 0, 'world size'\
+        ' ({}) is not divisible by tensor model parallel size ({})'.format(
+            args.world_size, args.tensor_model_parallel_size)
+    # Pipeline model parallel size.
+    args.pipeline_model_parallel_size = min(
+        args.pipeline_model_parallel_size,
+        (args.world_size // args.tensor_model_parallel_size))
+    # Checks.
+    model_parallel_size = args.pipeline_model_parallel_size * \
+                          args.tensor_model_parallel_size
+    assert args.world_size % model_parallel_size == 0, 'world size is not'\
+        ' divisible by tensor parallel size ({}) times pipeline parallel ' \
+        'size ({})'.format(args.world_size, args.tensor_model_parallel_size,
+                           args.pipeline_model_parallel_size)
+    args.data_parallel_size = args.world_size // model_parallel_size
+    if args.rank == 0:
+        print('using world size: {}, data-parallel-size: {}, '
+              'tensor-model-parallel size: {}, '
+              'pipeline-model-parallel size: {} '.format(
+                  args.world_size, args.data_parallel_size,
+                  args.tensor_model_parallel_size,
+                  args.pipeline_model_parallel_size), flush=True)
+
+    # --data-path and --train-weighted-splits-paths
+    message = "Data loading Mode 1: --data-path and --split "\
+            "and Mode 2: --(train|valid|test)-weighted-split-paths"\
+            "are mutually exclusive i.e. cannot be set together."
+
+    if args.data_path:
+        assert args.train_weighted_split_paths is None, message
+        setattr(args, "valid_weighted_split_names", None)
+        setattr(args, "valid_weighted_split_weights", None)
+        setattr(args, "valid_weighted_split_splits", None)
+
+        setattr(args, "test_weighted_split_names", None)
+        setattr(args, "test_weighted_split_weights", None)
+        setattr(args, "test_weighted_split_splits", None)
+
+        # args.split default value in the args is None it is set here in order
+        # to check that it does not to overlap with the 2nd mode of data loading
+        if args.split is None:
+            args.split = "969, 30, 1"
+
+    if args.train_weighted_split_paths or args.valid_weighted_split_paths or \
+                args.test_weighted_split_paths:
+        assert args.data_path is None and args.split is None, message
+
+
+
+    # Deprecated arguments
+    assert args.batch_size is None, '--batch-size argument is no longer ' \
+        'valid, use --micro-batch-size instead'
+    del args.batch_size
+    assert args.warmup is None, '--warmup argument is no longer valid, use ' \
+        '--lr-warmup-fraction instead'
+    del args.warmup
+    assert args.model_parallel_size is None, '--model-parallel-size is no ' \
+        'longer valid, use --tensor-model-parallel-size instead'
+    del args.model_parallel_size
+
+    # Set input defaults.
+    for key in defaults:
+        # For default to be valid, it should not be provided in the
+        # arguments that are passed to the program. We check this by
+        # ensuring the arg is set to None.
+        if getattr(args, key) is not None:
+            if args.rank == 0:
+                print('WARNING: overriding default arguments for {key}:{v} \
+                       with {key}:{v2}'.format(key=key, v=defaults[key],
+                                               v2=getattr(args, key)),
+                                               flush=True)
+        else:
+            setattr(args, key, defaults[key])
+
+    # Batch size.
+    assert args.micro_batch_size is not None
+    assert args.micro_batch_size > 0
+    if args.global_batch_size is None:
+        args.global_batch_size = args.micro_batch_size * args.data_parallel_size
+        if args.rank == 0:
+            print('setting global batch size to {}'.format(
+                args.global_batch_size), flush=True)
+    assert args.global_batch_size > 0
+    if args.num_layers_per_virtual_pipeline_stage is not None:
+        assert args.pipeline_model_parallel_size > 2, \
+            'pipeline-model-parallel size should be greater than 2 with ' \
+            'interleaved schedule'
+        assert args.num_layers % args.num_layers_per_virtual_pipeline_stage == 0, \
+            'number of layers is not divisible by number of layers per virtual ' \
+            'pipeline stage'
+        args.virtual_pipeline_model_parallel_size = \
+            (args.num_layers // args.pipeline_model_parallel_size) // \
+            args.num_layers_per_virtual_pipeline_stage
+    else:
+        args.virtual_pipeline_model_parallel_size = None
+
+    # Parameters dtype.
+    args.params_dtype = torch.float
+    if args.fp16:
+        assert not args.bf16
+        args.params_dtype = torch.half
+    if args.bf16:
+        assert not args.fp16
+        args.params_dtype = torch.bfloat16
+        # bfloat16 requires gradient accumulation and all-reduce to
+        # be done in fp32.
+        if not args.accumulate_allreduce_grads_in_fp32:
+            args.accumulate_allreduce_grads_in_fp32 = True
+            if args.rank == 0:
+                print('accumulate and all-reduce gradients in fp32 for '
+                      'bfloat16 data type.', flush=True)
+
+    if args.rank == 0:
+        print('using {} for parameters ...'.format(args.params_dtype),
+              flush=True)
+
+    # If we do accumulation and all-reduces in fp32, we need to have
+    # local DDP and we should set the use-contiguous-buffers-in-ddp.
+    if args.accumulate_allreduce_grads_in_fp32:
+        assert args.DDP_impl == 'local'
+        args.use_contiguous_buffers_in_ddp = True
+
+    if args.dataloader_type is None:
+        args.dataloader_type = 'single'
+
+    # Consumed tokens.
+    args.consumed_train_samples = 0
+    args.consumed_valid_samples = 0
+    args.consumed_train_tokens = 0
+    args.gigaflos_no_embeds = 0
+
+    # Iteration-based training.
+    if args.train_iters:
+        # If we use iteration-based training, make sure the
+        # sample-based options are off.
+        assert args.train_samples is None, \
+            'expected iteration-based training'
+        assert args.lr_decay_samples is None, \
+            'expected iteration-based learning rate decay'
+        assert args.lr_warmup_samples == 0, \
+            'expected iteration-based learning rate warmup'
+        assert args.rampup_batch_size is None, \
+            'expected no batch-size rampup for iteration-based training'
+        if args.lr_warmup_fraction is not None:
+            assert args.lr_warmup_iters == 0, \
+                'can only specify one of lr-warmup-fraction and lr-warmup-iters'
+
+    # Sample-based training.
+    if args.train_samples:
+        # If we use sample-based training, make sure the
+        # iteration-based options are off.
+        assert args.train_iters is None, \
+            'expected sample-based training'
+        assert args.lr_decay_iters is None, \
+            'expected sample-based learning rate decay'
+        assert args.lr_warmup_iters == 0, \
+            'expected sample-based learnig rate warmup'
+        if args.lr_warmup_fraction is not None:
+            assert args.lr_warmup_samples == 0, \
+                'can only specify one of lr-warmup-fraction ' \
+                'and lr-warmup-samples'
+
+    # Check required arguments.
+    required_args = ['num_layers', 'hidden_size', 'num_attention_heads']
+    for req_arg in required_args:
+        _check_arg_is_not_none(args, req_arg)
+
+    # Checks.
+    if args.ffn_hidden_size is None:
+        args.ffn_hidden_size = 4 * args.hidden_size
+
+    if args.kv_channels is None:
+        assert args.hidden_size % args.num_attention_heads == 0
+        args.kv_channels = args.hidden_size // args.num_attention_heads
+
+    if args.seq_length is not None:
+        assert args.encoder_seq_length is None
+        args.encoder_seq_length = args.seq_length
+    else:
+        assert args.encoder_seq_length is not None
+        args.seq_length = args.encoder_seq_length
+
+    if args.position_embedding_type == PositionEmbeddingType.absolute or args.position_embedding_type == PositionEmbeddingType.alibi:
+        assert args.max_position_embeddings is not None
+        if args.seq_length is not None:
+            assert args.max_position_embeddings >= args.seq_length
+        if args.decoder_seq_length is not None:
+            assert args.max_position_embeddings >= args.decoder_seq_length
+    else:
+        assert args.max_position_embeddings is None
+
+    if args.lr is not None:
+        assert args.min_lr <= args.lr
+    if args.save is not None:
+        assert args.save_interval is not None
+    # Mixed precision checks.
+    if args.fp16_lm_cross_entropy:
+        assert args.fp16, 'lm cross entropy in fp16 only support in fp16 mode.'
+    if args.fp32_residual_connection:
+        assert args.fp16 or args.bf16, \
+            'residual connection in fp32 only supported when using fp16 or bf16.'
+    # Activation checkpointing.
+    if args.distribute_checkpointed_activations:
+        assert args.checkpoint_activations, \
+            'for distribute-checkpointed-activations to work you '\
+            'need to enable checkpoint-activations'
+
+    args.curriculum_learning = False
+
+    # Activation function
+    if args.glu_activation is not None and args.bias_gelu_fusion:
+        raise ValueError("if glu-activation is used, please set --no-bias-gelu-fusion")
+
+    # Skip train iterations
+    if args.skip_train_iteration_range is not None:
+        args.skip_train_iteration_range = [
+            list(map(int, range_.split("-"))) for range_ in args.skip_train_iteration_range
+        ]
+        args.skip_train_iteration_range.sort()
+        skip_train_iteration_range = collections.deque()
+        for range_ in args.skip_train_iteration_range:
+            if len(range_) == 2:
+                start, end = range_
+                assert end >= start, \
+                "end of skip range cannot be smaller than start of skip range"
+                # merge overlapping intervals (e.g. 1-5 2-6 -> 1-6)
+                if not skip_train_iteration_range:
+                    skip_train_iteration_range.append([start, end])
+                elif skip_train_iteration_range[-1][1] >= start:
+                    skip_train_iteration_range[-1][1] = max(end, skip_train_iteration_range[-1][1])
+                else:
+                    skip_train_iteration_range.append([start, end])
+            else:
+                raise ValueError(
+                    "skip train iterations should be specified as two numbers, i.e. start-end"
+                )
+        args.skip_train_iteration_range = skip_train_iteration_range
+
+    if args.use_bnb_optimizer:
+        try:
+            import bitsandbytes as bnb
+        except ModuleNotFoundError:
+            raise ModuleNotFoundError("Please install bitsandbytes from https://github.com/facebookresearch/bitsandbytes.")
+
+    _print_args(args)
+    return args
+
+
+def _print_args(args):
+    """Print arguments."""
+    if args.rank == 0:
+        print('------------------------ arguments ------------------------',
+              flush=True)
+        str_list = []
+        for arg in vars(args):
+            dots = '.' * (48 - len(arg))
+            str_list.append('  {} {} {}'.format(arg, dots, getattr(args, arg)))
+
+        if args.log_path is not None:
+            with open(os.path.join(args.log_path,f'args_{time.strftime("%Y-%m-%dT%H:%M:%S")}.txt'), 'w') as f:
+                for arg in sorted(str_list, key=lambda x: x.lower()):
+                    f.write(arg+"\n")
+                    print(arg, flush=True)
+        else:
+            for arg in sorted(str_list, key=lambda x: x.lower()):
+                print(arg, flush=True)
+        print('-------------------- end of arguments ---------------------',
+              flush=True)
+
+
+def _check_arg_is_not_none(args, arg):
+    assert getattr(args, arg) is not None, '{} argument is None'.format(arg)
+
+
+def _add_network_size_args(parser):
+    group = parser.add_argument_group(title='network size')
+
+    group.add_argument('--num-layers', type=int, default=None,
+                       help='Number of transformer layers.')
+    group.add_argument('--hidden-size', type=int, default=None,
+                       help='Tansformer hidden size.')
+    group.add_argument('--ffn-hidden-size', type=int, default=None,
+                       help='Transformer Feed-Forward Network hidden size. '
+                       'This is set to 4*hidden-size if not provided')
+    group.add_argument('--num-attention-heads', type=int, default=None,
+                       help='Number of transformer attention heads.')
+    group.add_argument('--kv-channels', type=int, default=None,
+                       help='Projection weights dimension in multi-head '
+                       'attention. This is set to '
+                       '   args.hidden_size // args.num_attention_heads '
+                       'if not provided.')
+    group.add_argument('--max-position-embeddings', type=int, default=None,
+                       help='Maximum number of position embeddings to use. '
+                       'This is the size of position embedding.')
+    group.add_argument('--make-vocab-size-divisible-by', type=int, default=128,
+                       help='Pad the vocab size to be divisible by this value.'
+                       'This is added for computational efficieny reasons.')
+    group.add_argument('--pad-vocab-size-to', type=int, default=None,
+                       help='Pad the vocab size to this value.'
+                       'This value must be greater than the initial size of the tokenizer'
+                       ', needs to be divisible by TP size and `make-vocab-size-divisible-by`.')
+    group.add_argument('--layernorm-epsilon', type=float, default=1e-5,
+                       help='Layer norm epsilon.')
+    group.add_argument('--sync-tp-duplicated-parameters', action='store_true',
+                       help='Force syncing duplicated params across TP ranks in forward. '
+                       'This is a workaround for an unresolved bug leading to TP ranks '
+                       'getting out of sync with each other.')
+    group.add_argument('--apply-residual-connection-post-layernorm',
+                       action='store_true',
+                       help='If set, use original BERT residula connection '
+                       'ordering.')
+    group.add_argument('--embed-layernorm', action='store_true',
+                       help='use layernorm for embedding')
+    group.add_argument('--openai-gelu', action='store_true',
+                       help='Use OpenAIs GeLU implementation. This option'
+                       'should not be used unless for backward compatibility'
+                       'reasons.')
+    group.add_argument('--onnx-safe', type=bool, required=False,
+                       help='Use workarounds for known problems with '
+                       'Torch ONNX exporter')
+    group.add_argument('--bert-no-binary-head', action='store_false',
+                       help='Disable BERT binary head.',
+                       dest='bert_binary_head')
+    group.add_argument('--position-embedding-type', type=lambda x: PositionEmbeddingType[x],
+                       choices=list(PositionEmbeddingType),
+                       default=PositionEmbeddingType.absolute,
+                       help='Define position embedding type ("absolute" | "rotary" | "alibi"). "absolute" by default.'
+                       )
+    group.add_argument('--glu-activation', type=str,
+                       choices=megatron.model.glu_activations.GLU_ACTIVATIONS.keys(),
+                       help='GLU activations to use.'
+                       )
+
+    group.add_argument('--kill-switch-path', type=str,
+                       help='path to look for a kill switch, which if found will automatically exit the program'
+                       )
+
+
+    group.add_argument('--log-level', type=str, choices=list(log_levels.keys()),
+                       help="Logger log level to use on the main process. Possible choices are the log levels as strings: 'debug', "
+                       "'info', 'warning', 'error' and 'critical', plus a 'passive' level which doesn't set anything and lets the "
+                       "application set the level."
+                       )
+    group.add_argument('--log-level-replica', type=str, choices=list(log_levels.keys()),
+                       help="Logger log level to use on replicas. Same choices as ``log_level``"
+                       )
+    return parser
+
+
+def _add_logging_args(parser):
+    group = parser.add_argument_group(title='logging')
+
+    group.add_argument('--log-params-norm', action='store_true',
+                       help='If set, calculate and log parameters norm.')
+    group.add_argument('--log-num-zeros-in-grad', action='store_true',
+                       help='If set, calculate and log the number of zeros in gradient.')
+    group.add_argument('--tensorboard-log-interval', type=int, default=1,
+                       help='Report to tensorboard interval.')
+    group.add_argument('--tensorboard-queue-size', type=int, default=1000,
+                       help='Size of the tensorboard queue for pending events '
+                       'and summaries before one of the ‘add’ calls forces a '
+                       'flush to disk.')
+    group.add_argument('--log-timers-to-tensorboard', action='store_true',
+                       help='If set, write timers to tensorboard.')
+    group.add_argument('--log-batch-size-to-tensorboard', action='store_true',
+                       help='If set, write batch-size to tensorboard.')
+    group.add_argument('--no-log-learnig-rate-to-tensorboard',
+                       action='store_false',
+                       help='Disable learning rate logging to tensorboard.',
+                       dest='log_learning_rate_to_tensorboard')
+    group.add_argument('--no-log-loss-scale-to-tensorboard',
+                       action='store_false',
+                       help='Disable loss-scale logging to tensorboard.',
+                       dest='log_loss_scale_to_tensorboard')
+    group.add_argument('--log-validation-ppl-to-tensorboard',
+                       action='store_true',
+                       help='If set, write validation perplexity to '
+                       'tensorboard.')
+
+    return parser
+
+
+def _add_regularization_args(parser):
+    group = parser.add_argument_group(title='regularization')
+
+    group.add_argument('--attention-dropout', type=float, default=0.1,
+                       help='Post attention dropout probability.')
+    group.add_argument('--hidden-dropout', type=float, default=0.1,
+                       help='Dropout probability for hidden state transformer.')
+    group.add_argument('--weight-decay', type=float, default=0.01,
+                       help='Weight decay coefficient for L2 regularization.')
+    group.add_argument('--clip-grad', type=float, default=1.0,
+                       help='Gradient clipping based on global L2 norm.')
+    group.add_argument('--adam-beta1', type=float, default=0.9,
+                       help='First coefficient for computing running averages '
+                       'of gradient and its square')
+    group.add_argument('--adam-beta2', type=float, default=0.999,
+                       help='Second coefficient for computing running averages '
+                       'of gradient and its square')
+    group.add_argument('--adam-eps', type=float, default=1e-08,
+                       help='Term added to the denominator to improve'
+                       'numerical stability')
+    group.add_argument('--sgd-momentum', type=float, default=0.9,
+                       help='Momentum factor for sgd')
+
+    return parser
+
+
+def _add_training_args(parser):
+    group = parser.add_argument_group(title='training')
+
+    group.add_argument('--micro-batch-size', type=int, default=None,
+                       help='Batch size per model instance (local batch size). '
+                       'Global batch size is local batch size times data '
+                       'parallel size times number of micro batches.')
+    group.add_argument('--batch-size', type=int, default=None,
+                       help='Old batch size parameter, do not use. '
+                       'Use --micro-batch-size instead')
+    group.add_argument('--global-batch-size', type=int, default=None,
+                       help='Training batch size. If set, it should be a '
+                       'multiple of micro-batch-size times data-parallel-size. '
+                       'If this value is None, then '
+                       'use micro-batch-size * data-parallel-size as the '
+                       'global batch size. This choice will result in 1 for '
+                       'number of micro-batches.')
+    group.add_argument('--rampup-batch-size', nargs='*', default=None,
+                       help='Batch size ramp up with the following values:'
+                       '  --rampup-batch-size <start batch size> '
+                       '                      <batch size increment> '
+                       '                      <ramp-up samples> '
+                       'For example: '
+                       '   --rampup-batch-size 16 8 300000 '
+                       '   --global-batch-size 1024 '
+                       'will start with global batch size 16 and over '
+                       ' (1024 - 16) / 8 = 126 intervals will increase '
+                       'the batch size linearly to 1024. In each interval '
+                       'we will use approximately 300000 / 126 = 2380 samples.')
+    group.add_argument('--checkpoint-activations', action='store_true',
+                       help='Checkpoint activation to allow for training '
+                       'with larger models, sequences, and batch sizes.')
+    group.add_argument('--distribute-checkpointed-activations',
+                       action='store_true',
+                       help='If set, distribute checkpointed activations '
+                       'across model parallel group.')
+    group.add_argument('--checkpoint-num-layers', type=int, default=1,
+                       help='chunk size (number of layers) for checkpointing.')
+    group.add_argument('--train-iters', type=int, default=None,
+                       help='Total number of iterations to train over all '
+                       'training runs. Note that either train-iters or '
+                       'train-samples should be provided.')
+    group.add_argument('--train-samples', type=int, default=None,
+                       help='Total number of samples to train over all '
+                       'training runs. Note that either train-iters or '
+                       'train-samples should be provided.')
+    group.add_argument('--train-tokens', type=int, default=None,
+                       help='Total number of tokens to train over all '
+                       'training runs.')
+    group.add_argument('--log-interval', type=int, default=100,
+                       help='Report loss and timing interval.')
+    group.add_argument('--exit-interval', type=int, default=None,
+                       help='Exit the program after the iteration is divisible '
+                       'by this value.')
+    group.add_argument('--exit-duration-in-mins', type=int, default=None,
+                       help='Exit the program after this many minutes.')
+    group.add_argument('--tensorboard-dir', type=str, default=None,
+                       help='Write TensorBoard logs to this directory.')
+    group.add_argument('--no-masked-softmax-fusion',
+                       action='store_false',
+                       help='Disable fusion of query_key_value scaling, '
+                       'masking, and softmax.',
+                       dest='masked_softmax_fusion')
+    group.add_argument('--no-bias-gelu-fusion', action='store_false',
+                       help='Disable bias and gelu fusion.',
+                       dest='bias_gelu_fusion')
+    group.add_argument('--no-bias-dropout-fusion', action='store_false',
+                       help='Disable bias and dropout fusion.',
+                       dest='bias_dropout_fusion')
+    group.add_argument('--optimizer', type=str, default='adam',
+                       choices=['adam', 'sgd'],
+                       help='Optimizer function')
+    group.add_argument('--use-bnb-optimizer', action='store_true',
+                       help='Use bitsandbytes optimizer for efficient training,'
+                       'please refer https://github.com/facebookresearch/bitsandbytes.',
+                       dest='use_bnb_optimizer')
+    group.add_argument('--dataloader-type', type=str, default=None,
+                       choices=['single', 'cyclic'],
+                       help='Single pass vs multiple pass data loader')
+    group.add_argument('--cpu-optimizer', action='store_true',
+                       help='Run optimizer on CPU')
+    group.add_argument('--cpu_torch_adam', action='store_true',
+                       help='Use Torch Adam as optimizer on CPU.')
+    group.add_argument('--codecarbon-dir', type=str, default=None,
+                       help='Write CodeCarbon logs to this directory.')
+    group.add_argument('--eval-only', type=bool, required=False,
+                       help='If set to True, no train step will be performed.'
+                       'and only the evaluation on the `valid` and `test` sets '
+                       'will be performed' )
+    group.add_argument('--skip-train-iteration-range', type=str, nargs='+', default=None,
+                       help='Iteration ranges to skip. The values are one or more dash-separated ranges. e.g., 101-200 251-300.')
+    group.add_argument('--inference', action='store_true',
+                       help='Very basic inference mode: not allocating optim/lr - requires ZERO_STAGE=0')
+    group.add_argument('--abort-on-unmet-fused-kernel-constraints', action='store_true',
+                       help="If set to True, the program will abort if the constraints for loading a fused kernel aren't met")
+    group.add_argument('--pp-partition-method', type=str, default=None,
+                       help="Use to override the pipeline stages partitioning method. e.g., 'type:transformer|embedding'")
+
+    return parser
+
+
+def _add_initialization_args(parser):
+    group = parser.add_argument_group(title='initialization')
+
+    group.add_argument('--seed', type=int, default=1234,
+                       help='Random seed used for python, numpy, '
+                       'pytorch, and cuda.')
+    group.add_argument('--init-method-std', type=float, default=0.02,
+                       help='Standard deviation of the zero mean normal '
+                       'distribution used for weight initialization.')
+    group.add_argument('--init-method-xavier-uniform', action='store_true',
+                       help='Enable Xavier uniform parameter initialization')
+
+    return parser
+
+
+def _add_learning_rate_args(parser):
+    group = parser.add_argument_group(title='learning rate')
+
+    group.add_argument('--lr', type=float, default=None,
+                       help='Initial learning rate. Depending on decay style '
+                       'and initial warmup, the learing rate at each '
+                       'iteration would be different.')
+    group.add_argument('--lr-decay-style', type=str, default='linear',
+                       choices=['constant', 'linear', 'cosine'],
+                       help='Learning rate decay function.')
+    group.add_argument('--lr-decay-iters', type=int, default=None,
+                       help='number of iterations to decay learning rate over,'
+                       ' If None defaults to `--train-iters`')
+    group.add_argument('--lr-decay-samples', type=int, default=None,
+                       help='number of samples to decay learning rate over,'
+                       ' If None defaults to `--train-samples`')
+    group.add_argument('--lr-decay-tokens', type=int, default=None,
+                       help='number of tokens to decay learning rate over,'
+                       ' If not None will override iter/sample-based decay')
+    group.add_argument('--lr-warmup-fraction', type=float, default=None,
+                       help='fraction of lr-warmup-(iters/samples) to use '
+                       'for warmup (as a float)')
+    group.add_argument('--lr-warmup-iters', type=int, default=0,
+                       help='number of iterations to linearly warmup '
+                       'learning rate over.')
+    group.add_argument('--lr-warmup-samples', type=int, default=0,
+                       help='number of samples to linearly warmup '
+                       'learning rate over.')
+    group.add_argument('--warmup', type=int, default=None,
+                       help='Old lr warmup argument, do not use. Use one of the'
+                       '--lr-warmup-* arguments above')
+    group.add_argument('--min-lr', type=float, default=0.0,
+                       help='Minumum value for learning rate. The scheduler'
+                       'clip values below this threshold.')
+    group.add_argument('--override-lr-scheduler', action='store_true',
+                       help='Reset the values of the scheduler (learning rate,'
+                       'warmup iterations, minimum learning rate, maximum '
+                       'number of iterations, and decay style from input '
+                       'arguments and ignore values from checkpoints. Note'
+                       'that all the above values will be reset.')
+    group.add_argument('--use-checkpoint-lr-scheduler', action='store_true',
+                       help='Use checkpoint to set the values of the scheduler '
+                       '(learning rate, warmup iterations, minimum learning '
+                       'rate, maximum number of iterations, and decay style '
+                       'from checkpoint and ignore input arguments.')
+    group.add_argument('--universal-checkpoint', action='store_true',
+                        help='Loading a universal format checkpoint.')
+
+    return parser
+
+
+def _add_checkpointing_args(parser):
+    group = parser.add_argument_group(title='checkpointing')
+
+    group.add_argument('--save', type=str, default=None,
+                       help='Output directory to save checkpoints to.')
+    group.add_argument('--save-interval', type=int, default=None,
+                       help='Number of iterations between checkpoint saves.')
+    group.add_argument('--no-save-optim', action='store_true', default=None,
+                       help='Do not save current optimizer.')
+    group.add_argument('--no-save-rng', action='store_true', default=None,
+                       help='Do not save current rng state.')
+    group.add_argument('--load', type=str, default=None,
+                       help='Directory containing a model checkpoint.')
+    group.add_argument('--no-load-optim', action='store_true', default=None,
+                       help='Do not load optimizer when loading checkpoint.')
+    group.add_argument('--no-load-rng', action='store_true', default=None,
+                       help='Do not load rng state when loading checkpoint.')
+    group.add_argument('--finetune', action='store_true',
+                       help='Load model for finetuning. Do not load optimizer '
+                       'or rng state from checkpoint and set iteration to 0. '
+                       'Assumed when loading a release checkpoint.')
+
+    return parser
+
+
+def _add_mixed_precision_args(parser):
+    group = parser.add_argument_group(title='mixed precision')
+
+    group.add_argument('--fp16', action='store_true',
+                       help='Run model in fp16 mode.')
+    group.add_argument('--bf16', action='store_true',
+                       help='Run model in bfloat16 mode.')
+    group.add_argument('--loss-scale', type=float, default=None,
+                       help='Static loss scaling, positive power of 2 '
+                       'values can improve fp16 convergence. If None, dynamic'
+                       'loss scaling is used.')
+    group.add_argument('--initial-loss-scale', type=float, default=2**32,
+                       help='Initial loss-scale for dynamic loss scaling.')
+    group.add_argument('--min-loss-scale', type=float, default=1.0,
+                       help='Minimum loss scale for dynamic loss scale.')
+    group.add_argument('--loss-scale-window', type=float, default=1000,
+                       help='Window over which to raise/lower dynamic scale.')
+    group.add_argument('--hysteresis', type=int, default=2,
+                       help='hysteresis for dynamic loss scaling')
+    group.add_argument('--fp32-residual-connection', action='store_true',
+                       help='Move residual connections to fp32.')
+    group.add_argument('--no-query-key-layer-scaling', action='store_false',
+                       help='Do not scale Q * K^T by 1 / layer-number.',
+                       dest='apply_query_key_layer_scaling')
+    group.add_argument('--attention-softmax-in-fp32', action='store_true',
+                       help='Run attention masking and softmax in fp32. '
+                       'This flag is ignored unless '
+                       '--no-query-key-layer-scaling is specified.')
+    group.add_argument('--accumulate-allreduce-grads-in-fp32',
+                       action='store_true',
+                       help='Gradient accumulation and all-reduce in fp32.')
+    group.add_argument('--fp16-lm-cross-entropy', action='store_true',
+                       help='Move the cross entropy unreduced loss calculation'
+                       'for lm head to fp16.')
+
+    return parser
+
+
+def _add_distributed_args(parser):
+    group = parser.add_argument_group(title='distributed')
+
+    group.add_argument('--tensor-model-parallel-size', type=int, default=1,
+                       help='Degree of tensor model parallelism.')
+    group.add_argument('--pipeline-model-parallel-size', type=int, default=1,
+                       help='Degree of pipeline model parallelism.')
+    group.add_argument('--model-parallel-size', type=int, default=None,
+                       help='Old model parallel argument, do not use. Use '
+                       '--tensor-model-parallel-size instead.')
+    group.add_argument('--num-layers-per-virtual-pipeline-stage', type=int, default=None,
+                       help='Number of layers per virtual pipeline stage')
+    group.add_argument('--distributed-backend', default='nccl',
+                       choices=['nccl', 'gloo'],
+                       help='Which backend to use for distributed training.')
+    group.add_argument('--DDP-impl', default='local',
+                       choices=['local', 'torch'],
+                       help='which DistributedDataParallel implementation '
+                       'to use.')
+    group.add_argument('--use-contiguous-buffers-in-ddp', action='store_true',
+                       help='If set, use contiguous buffer in DDP. Note that '
+                       'this option only works woth local DDP.' )
+    group.add_argument('--no-scatter-gather-tensors-in-pipeline', action='store_false',
+                       help='Use scatter/gather to optimize communication of tensors in pipeline',
+                       dest='scatter_gather_tensors_in_pipeline')
+    group.add_argument('--local_rank', type=int, default=None,
+                       help='local rank passed from distributed launcher.')
+    group.add_argument('--lazy-mpu-init', type=bool, required=False,
+                       help='If set to True, initialize_megatron() '
+                       'skips DDP initialization and returns function to '
+                       'complete it instead.Also turns on '
+                       '--use-cpu-initialization flag. This is for '
+                       'external DDP manager.' )
+    group.add_argument('--use-cpu-initialization', action='store_true',
+                       default=None, help='If set, affine parallel weights '
+                       'initialization uses CPU' )
+    return parser
+
+
+def _add_validation_args(parser):
+    group = parser.add_argument_group(title='validation')
+
+    group.add_argument('--eval-iters', type=int, default=100,
+                       help='Number of iterations to run for evaluation'
+                       'validation/test for.')
+    group.add_argument('--eval-interval', type=int, default=1000,
+                       help='Interval between running evaluation on '
+                       'validation set.')
+
+    return parser
+
+
+def _add_data_args(parser):
+    group = parser.add_argument_group(title='data and dataloader')
+
+
+    # option 1 for data loading  (mutually exclusive with option2)
+    group.add_argument('--data-path', nargs='*', default=None,
+                       help='Path to the training dataset. Accepted format:'
+                       '1) a single data path, 2) multiple datasets in the'
+                       'form: dataset1-weight dataset1-path dataset2-weight '
+                       'dataset2-path ...')
+
+    group.add_argument('--split', type=str, default=None,
+                       help='Comma-separated list of proportions for training,'
+                       ' validation, and test split. For example the split '
+                       '`90,5,5` will use 90%% of data for training, 5%% for '
+                       'validation and 5%% for test.')
+
+    # option 2 for data loading (mutually exclusive with option1)
+
+    # helper class to parse the --xxx-weighted-split-paths
+    # note here two args are set: extra valid dataset paths and names
+    class parse_data_paths(argparse.Action):
+        def __call__(self, parser, args, values, option_string=None):
+
+            if option_string == "--train-weighted-split-paths":
+                assert len(values) == 1, 'Only 1 dataset group is allowed to'
+                'be passed for the argument --train-weighted-split-paths'
+
+            # make sure string given in the correct format
+            err_message = 'Each data group should be input on the following format'
+            '"GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2"'
+            'where START < END'
+            for v in values:
+                # each prefix consists several datasets separated by commas
+                prefix = ":".join(v.split(":")[1:]) # remove GIVEN_NAME
+                datasets = prefix.split(",")
+                # check if each dataset is formatted like `WEIGHT START:END PATH`
+                for d in datasets:
+                    assert len(d.split()) == 3, err_message
+                    start, end = d.split()[1].split(":")
+                    assert float(start) < float(end), err_message
+
+            names = [v.split(":")[0] for v in values]
+
+            prefixes = [":".join(v.split(":")[1:]).strip() for v in values]
+            weights = [[d.split()[0] for d in p.split(",")] for p in prefixes]
+            splits = [[d.split()[1] for d in p.split(",")] for p in prefixes]
+            paths = [[d.split()[2] for d in p.split(",")] for p in prefixes]
+
+            # # to keep consistency with Option 1 of data loading (through --data-path)
+            # #  paths will contain strings on the following form
+            # # "WEIGHTS1 PATH1 WEIGHTS2 PATH2 WEIGHTS3 PATH3" for each dataset group
+            # # while data will be parsed in additional arguments below
+            # paths_option1_style = []
+            # for p, w in zip(paths, weights):
+            #   paths_option1_style.append(" ".join([f"{w_i} {p_i}" for p_i, w_i in zip(p,w)]))
+            # setattr(args, self.dest, paths_option1_style)
+            setattr(args, self.dest, paths)
+            setattr(args, self.dest.replace("paths", "weights"), weights)
+            setattr(args, self.dest.replace("paths", "splits"), splits)
+            setattr(args, self.dest.replace("paths","names"), names)
+
+
+    group.add_argument('--train-weighted-split-paths', nargs='*', default=None,
+                    help='Weights, splits and paths to groups of datasets'
+                    'Accepted format: ONE dataset groups could be'
+                    'submitted in the following form between double quotes'
+                    '"GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2"'
+                    'e.g.: "NAME_ABC: 0.6 0:0.6 A, 0.3 0:1 B, 0.1 0:1 C" '
+                    'WEIGHT is used to up and down sample each dataset A,B,C in the group'
+                    'START:END indicates the split portion of the dataset',
+                    action=parse_data_paths)
+
+    group.add_argument('--valid-weighted-split-paths', nargs='*', default=None,
+                    help='Weights, splits and paths to groups of datasets'
+                    'Accepted format: one or many dataset groups could be'
+                    'submitted in the following form each between double quotes'
+                    '"GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2"'
+                    'e.g.: "NAME_ABC: 0.6 0.6:0.8 A, 0.3 0:1 B, 0.1 0:1 C" '
+                    '"NAME_CDE: 0.6 0.6:0.8 C, 0.3 0:1 D, 0.1 0:1 E" '
+                    'validation will be run on each of those groups independently',
+                    action=parse_data_paths)
+
+    group.add_argument('--test-weighted-split-paths', nargs='*', default=None,
+                    help='Weights, splits and paths to groups of datasets'
+                    'Accepted format: one or many dataset groups could be'
+                    'submitted in the following form each between double quotes'
+                    '"GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2"'
+                    'e.g.: "NAME_ABC: 0.6 0.6:0.8 A, 0.3 0:1 B, 0.1 0:1 C" '
+                    '"NAME_CDE: 0.6 0.6:0.8 C, 0.3 0:1 D, 0.1 0:1 E" '
+                    'test will be run on each of those groups independently',
+                    action=parse_data_paths)
+
+    class parse_data_paths_path(argparse.Action):
+        def __call__(self, parser, args, values, option_string=None):
+            expected_option_strings = ["--train-weighted-split-paths-path", "--valid-weighted-split-paths-path", "--test-weighted-split-paths-path"]
+            assert option_string in expected_option_strings, f"Expected {option_string} to be in {expected_option_strings}"
+
+            with open(values, "r") as fi:
+                lines = fi.readlines()
+                assert len(lines) == 1, f"Got multiple lines {len(lines)} instead of 1 expected"
+                assert lines[0][-2:] == "\"\n" and lines[0][0] == "\"", f"Invalid input format, got {lines}"
+                values = lines[0][1:-2].split("\" \"")
+                weighted_split_paths_dest = re.sub(r"_path$", "", self.dest)
+                weighted_split_paths_option = re.sub(r"-path$", "", self.option_strings[0])
+                setattr(args, weighted_split_paths_dest, values)
+                parse_data_paths(option_strings=[weighted_split_paths_option], dest=weighted_split_paths_dest)(parser, args, values, option_string=weighted_split_paths_option)
+
+
+    group.add_argument('--train-weighted-split-paths-path', type=str, action=parse_data_paths_path ,default=None)
+    group.add_argument('--valid-weighted-split-paths-path', type=str, action=parse_data_paths_path, default=None)
+    group.add_argument('--test-weighted-split-paths-path', type=str, action=parse_data_paths_path, default=None)
+
+    group.add_argument('--log-path', type=str, default=None,
+                       help='Path to the save arguments file.')
+    group.add_argument('--vocab-file', type=str, default=None,
+                       help='Path to the vocab file.')
+    group.add_argument('--merge-file', type=str, default=None,
+                       help='Path to the BPE merge file.')
+    group.add_argument('--vocab-extra-ids', type=int, default=0,
+                       help='Number of additional vocabulary tokens. '
+                            'They are used for span masking in the T5 model')
+    group.add_argument('--seq-length', type=int, default=None,
+                       help='Maximum sequence length to process.')
+    group.add_argument('--encoder-seq-length', type=int, default=None,
+                       help='Maximum encoder sequence length to process.'
+                       'This should be exclusive of --seq-length')
+    group.add_argument('--decoder-seq-length', type=int, default=None,
+                       help="Maximum decoder sequence length to process.")
+    group.add_argument('--retriever-seq-length', type=int, default=256,
+                       help='Maximum sequence length for the biencoder model '
+                        ' for retriever')
+    group.add_argument('--sample-rate', type=float, default=1.0,
+                       help='sample rate for training data. Supposed to be 0 '
+                            ' < sample_rate < 1')
+    group.add_argument('--mask-prob', type=float, default=0.15,
+                       help='Probability of replacing a token with mask.')
+    group.add_argument('--short-seq-prob', type=float, default=0.1,
+                       help='Probability of producing a short sequence.')
+    group.add_argument('--mmap-warmup', action='store_true',
+                       help='Warm up mmap files.')
+    group.add_argument('--num-workers', type=int, default=2,
+                       help="Dataloader number of workers.")
+    group.add_argument('--valid-num-workers', type=int, default=2,
+                       help="Dataloader number of workers for validation.")
+    group.add_argument('--tokenizer-type', type=str,
+                       default=None,
+                       choices=['BertWordPieceLowerCase',
+                                'BertWordPieceCase',
+                                'GPT2BPETokenizer',
+                                'PretrainedFromHF'],
+                       help='What type of tokenizer to use.')
+    group.add_argument("--tokenizer-name-or-path", type=str, default=None,
+                       help="Name or path of the huggingface tokenizer.")
+    group.add_argument('--data-impl', type=str, default='infer',
+                       choices=['lazy', 'cached', 'mmap', 'infer'],
+                       help='Implementation of indexed datasets.')
+    group.add_argument('--reset-position-ids', action='store_true',
+                       help='Reset posistion ids after end-of-document token.')
+    group.add_argument('--reset-attention-mask', action='store_true',
+                       help='Reset self attention maske after '
+                       'end-of-document token. Attention between tokens from different documents is null.')
+    group.add_argument('--eod-mask-loss', action='store_true',
+                       help='Mask loss for the end of document tokens.')
+    group.add_argument('--loss-on-targets-only', action='store_true',
+                       help='Mask loss on input sequence.')
+    group.add_argument('--reweight-loss-based-on-position-frequency', action="store_true",
+                       help='Some objectives require us to sample loss_mask. This might introduce bias towards '
+                       'specific positions. This option tries to un-bias the loss by reweighting loss on specific '
+                       'positions based on how frequently we train on that position.'
+                       'This is mostly used for prefix_lm training')
+    group.add_argument("--noise-density", type=float, default=None, help="Span corruption noise density")
+    group.add_argument("--mean-noise-span-length", type=int, default=None, help="Span corruption mean noise span length")
+
+
+    return parser
+
+
+def _add_autoresume_args(parser):
+    group = parser.add_argument_group(title='autoresume')
+
+    group.add_argument('--adlr-autoresume', action='store_true',
+                       help='Enable autoresume on adlr cluster.')
+    group.add_argument('--adlr-autoresume-interval', type=int, default=1000,
+                       help='Intervals over which check for autoresume'
+                       'termination signal')
+
+    return parser
+
+
+def _add_biencoder_args(parser):
+    group = parser.add_argument_group(title='biencoder')
+
+    # network size
+    group.add_argument('--ict-head-size', type=int, default=None,
+                       help='Size of block embeddings to be used in ICT and '
+                        'REALM (paper default: 128)')
+    group.add_argument('--biencoder-projection-dim', type=int, default=0,
+                       help='Size of projection head used in biencoder (paper'
+                        ' default: 128)')
+    group.add_argument('--biencoder-shared-query-context-model', action='store_true',
+                        help='Whether to share the parameters of the query '
+                        'and context models or not')
+
+    # checkpointing
+    group.add_argument('--ict-load', type=str, default=None,
+                       help='Directory containing an ICTBertModel checkpoint')
+    group.add_argument('--bert-load', type=str, default=None,
+                       help='Directory containing an BertModel checkpoint '
+                       '(needed to start ICT and REALM)')
+
+    # data
+    group.add_argument('--titles-data-path', type=str, default=None,
+                       help='Path to titles dataset used for ICT')
+    group.add_argument('--query-in-block-prob', type=float, default=0.1,
+                       help='Probability of keeping query in block for '
+                       'ICT dataset')
+    group.add_argument('--use-one-sent-docs', action='store_true',
+                       help='Whether to use one sentence documents in ICT')
+    group.add_argument('--evidence-data-path', type=str, default=None,
+                       help='Path to Wikipedia Evidence frm DPR paper')
+
+    # training
+    group.add_argument('--retriever-report-topk-accuracies', nargs='+', type=int,
+                        default=[], help="Which top-k accuracies to report "
+                        "(e.g. '1 5 20')")
+    group.add_argument('--retriever-score-scaling', action='store_true',
+                       help='Whether to scale retriever scores by inverse '
+                        'square root of hidden size')
+
+    # faiss index
+    group.add_argument('--block-data-path', type=str, default=None,
+                       help='Where to save/load BlockData to/from')
+    group.add_argument('--embedding-path', type=str, default=None,
+                       help='Where to save/load Open-Retrieval Embedding'
+                        ' data to/from')
+
+    # indexer
+    group.add_argument('--indexer-batch-size', type=int, default=128,
+                       help='How large of batches to use when doing indexing '
+                       'jobs')
+    group.add_argument('--indexer-log-interval', type=int, default=1000,
+                       help='After how many batches should the indexer '
+                       'report progress')
+    return parser
+
+
+def _add_vit_args(parser):
+    group = parser.add_argument_group(title="vit")
+
+    group.add_argument('--num-classes', type=int, default=1000,
+                       help='num of classes in vision classificaiton task')
+    group.add_argument('--img-dim', type=int, default=224,
+                       help='Image size for vision classification task')
+    group.add_argument('--num-channels', type=int, default=3,
+                       help='Number of channels in input image data')
+    group.add_argument('--patch-dim', type=int, default=16,
+                       help='patch dimension used in vit')
+
+    return parser
+
+
+def _add_zero_args(parser):
+    """Text generate arguments."""
+
+    group = parser.add_argument_group('ZeRO configurations', 'configurations')
+    group.add_argument("--zero-stage", type=int, default=1.0)
+    group.add_argument('--zero-reduce-scatter', action='store_true',
+                       help='Use reduce scatter if specified')
+    group.add_argument('--zero-contigious-gradients', action='store_true',
+                       help='Use contigious memory optimizaiton if specified')
+    group.add_argument("--zero-reduce-bucket-size", type=int, default=0.0)
+    group.add_argument("--zero-allgather-bucket-size", type=int, default=0.0)
+    group.add_argument('--remote-device', type=str, default='none', choices=['none', 'cpu', 'nvme'],
+                      help='Remote device for ZeRO-3 initialized parameters.')
+    group.add_argument('--use-pin-memory', action='store_true',
+                     help='Use pinned CPU memory for ZeRO-3 initialized model parameters.')
+    return parser
+
+def _add_memoryopt_args(parser):
+    """Memory optimization arguments."""
+
+    group = parser.add_argument_group('Memory optimizations', 'configurations')
+    group.add_argument("--scattered-embeddings", action='store_true',
+                       help='Save memory by scattering embedding activations. '
+                            'Introduces dropout differences across MP configurations.')
+    group.add_argument("--split-transformers", action='store_true',
+                       help='Save memory by splitting transformer layers into two parts, '
+                       'allowing for more frequent activation checkpoint savings.')
+    group.add_argument("--memory-centric-tiled-linear", action="store_true",
+                       help='Save memory by tiling with deepspeed.zero.TiledLinear.')
+    group.add_argument("--tile-factor", type=int, default=1,
+                       help='Make all linear layers the same size of [hidden/tile_factor, hidden/tile_factor]. '
+                            'Must be enabled with --memory-centric-tiled-linear. '
+                            'Example A: if tile_factor=1, the qkv layer [hidden, 3* hidden] would be converted into [1,3] tiles of size [hidden,hidden]. '
+                            'Example B: if tile_factor=2, the intermediate layer [4*hidden, hidden] will be converted into [8, 2] tiles of size [hidden/2, hidden/2]. '
+                            'Default is 1.')
+
+    return parser
+
+def _add_activation_checkpoint_args(parser):
+    group = parser.add_argument_group('Activation Checkpointing',
+                                      'Checkpointing Configurations')
+    group.add_argument('--deepspeed-activation-checkpointing', action='store_true',
+                       help='uses activation checkpointing from deepspeed')
+    group.add_argument('--partition-activations', action='store_true',
+                       help='partition Activations across GPUs before checkpointing.')
+    group.add_argument('--contigious-checkpointing', action='store_true',
+                       help='Contigious memory checkpointing for activatoins.')
+    group.add_argument('--checkpoint-in-cpu', action='store_true',
+                       help='Move the activation checkpoints to CPU.')
+    group.add_argument('--synchronize-each-layer', action='store_true',
+                       help='does a synchronize at the beginning and end of each checkpointed layer.')
+    group.add_argument('--profile-backward', action='store_true',
+                       help='Enables backward pass profiling for checkpointed layers.')
+    return parser

megatron-deepspeed_dtk22.10/megatron/arguments.py-nodes

+# coding=utf-8
+# Copyright (c) 2020, 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.
+
+"""Megatron arguments."""
+
+import argparse
+import collections
+import os
+import re
+import time
+
+import torch
+import deepspeed
+
+from megatron.enums import PositionEmbeddingType
+import megatron
+from megatron.logging import log_levels
+
+
+def parse_args(extra_args_provider=None, defaults={},
+               ignore_unknown_args=False):
+    """Parse all arguments."""
+    parser = argparse.ArgumentParser(description='Megatron-LM Arguments',
+                                     allow_abbrev=False)
+
+    # Standard arguments.
+    parser = _add_network_size_args(parser)
+    parser = _add_regularization_args(parser)
+    parser = _add_training_args(parser)
+    parser = _add_initialization_args(parser)
+    parser = _add_learning_rate_args(parser)
+    parser = _add_checkpointing_args(parser)
+    parser = _add_mixed_precision_args(parser)
+    parser = _add_distributed_args(parser)
+    parser = _add_validation_args(parser)
+    parser = _add_data_args(parser)
+    parser = _add_autoresume_args(parser)
+    parser = _add_biencoder_args(parser)
+    parser = _add_vit_args(parser)
+    parser = _add_logging_args(parser)
+    parser = _add_zero_args(parser)
+    parser = _add_memoryopt_args(parser)
+    parser = _add_activation_checkpoint_args(parser)
+
+    # Custom arguments.
+    if extra_args_provider is not None:
+        parser = extra_args_provider(parser)
+
+    parser = deepspeed.add_config_arguments(parser)
+
+    # Parse.
+    if ignore_unknown_args:
+        args, _ = parser.parse_known_args()
+    else:
+        args = parser.parse_args()
+
+    # Distributed args.
+    #args.rank = int(os.getenv('RANK', '0'))
+    #args.world_size = int(os.getenv("WORLD_SIZE", '1'))
+    # Tensor model parallel size.
+    args.tensor_model_parallel_size = min(
+        args.tensor_model_parallel_size, args.world_size)
+    assert args.world_size % args.tensor_model_parallel_size == 0, 'world size'\
+        ' ({}) is not divisible by tensor model parallel size ({})'.format(
+            args.world_size, args.tensor_model_parallel_size)
+    # Pipeline model parallel size.
+    args.pipeline_model_parallel_size = min(
+        args.pipeline_model_parallel_size,
+        (args.world_size // args.tensor_model_parallel_size))
+    # Checks.
+    model_parallel_size = args.pipeline_model_parallel_size * \
+                          args.tensor_model_parallel_size
+    assert args.world_size % model_parallel_size == 0, 'world size is not'\
+        ' divisible by tensor parallel size ({}) times pipeline parallel ' \
+        'size ({})'.format(args.world_size, args.tensor_model_parallel_size,
+                           args.pipeline_model_parallel_size)
+    args.data_parallel_size = args.world_size // model_parallel_size
+    if args.rank == 0:
+        print('using world size: {}, data-parallel-size: {}, '
+              'tensor-model-parallel size: {}, '
+              'pipeline-model-parallel size: {} '.format(
+                  args.world_size, args.data_parallel_size,
+                  args.tensor_model_parallel_size,
+                  args.pipeline_model_parallel_size), flush=True)
+
+    # --data-path and --train-weighted-splits-paths
+    message = "Data loading Mode 1: --data-path and --split "\
+            "and Mode 2: --(train|valid|test)-weighted-split-paths"\
+            "are mutually exclusive i.e. cannot be set together."
+
+    if args.data_path:
+        assert args.train_weighted_split_paths is None, message
+        setattr(args, "valid_weighted_split_names", None)
+        setattr(args, "valid_weighted_split_weights", None)
+        setattr(args, "valid_weighted_split_splits", None)
+
+        setattr(args, "test_weighted_split_names", None)
+        setattr(args, "test_weighted_split_weights", None)
+        setattr(args, "test_weighted_split_splits", None)
+
+        # args.split default value in the args is None it is set here in order
+        # to check that it does not to overlap with the 2nd mode of data loading
+        if args.split is None:
+            args.split = "969, 30, 1"
+
+    if args.train_weighted_split_paths or args.valid_weighted_split_paths or \
+                args.test_weighted_split_paths:
+        assert args.data_path is None and args.split is None, message
+
+
+
+    # Deprecated arguments
+    assert args.batch_size is None, '--batch-size argument is no longer ' \
+        'valid, use --micro-batch-size instead'
+    del args.batch_size
+    assert args.warmup is None, '--warmup argument is no longer valid, use ' \
+        '--lr-warmup-fraction instead'
+    del args.warmup
+    assert args.model_parallel_size is None, '--model-parallel-size is no ' \
+        'longer valid, use --tensor-model-parallel-size instead'
+    del args.model_parallel_size
+
+    # Set input defaults.
+    for key in defaults:
+        # For default to be valid, it should not be provided in the
+        # arguments that are passed to the program. We check this by
+        # ensuring the arg is set to None.
+        if getattr(args, key) is not None:
+            if args.rank == 0:
+                print('WARNING: overriding default arguments for {key}:{v} \
+                       with {key}:{v2}'.format(key=key, v=defaults[key],
+                                               v2=getattr(args, key)),
+                                               flush=True)
+        else:
+            setattr(args, key, defaults[key])
+
+    # Batch size.
+    assert args.micro_batch_size is not None
+    assert args.micro_batch_size > 0
+    if args.global_batch_size is None:
+        args.global_batch_size = args.micro_batch_size * args.data_parallel_size
+        if args.rank == 0:
+            print('setting global batch size to {}'.format(
+                args.global_batch_size), flush=True)
+    assert args.global_batch_size > 0
+    if args.num_layers_per_virtual_pipeline_stage is not None:
+        assert args.pipeline_model_parallel_size > 2, \
+            'pipeline-model-parallel size should be greater than 2 with ' \
+            'interleaved schedule'
+        assert args.num_layers % args.num_layers_per_virtual_pipeline_stage == 0, \
+            'number of layers is not divisible by number of layers per virtual ' \
+            'pipeline stage'
+        args.virtual_pipeline_model_parallel_size = \
+            (args.num_layers // args.pipeline_model_parallel_size) // \
+            args.num_layers_per_virtual_pipeline_stage
+    else:
+        args.virtual_pipeline_model_parallel_size = None
+
+    # Parameters dtype.
+    args.params_dtype = torch.float
+    if args.fp16:
+        assert not args.bf16
+        args.params_dtype = torch.half
+    if args.bf16:
+        assert not args.fp16
+        args.params_dtype = torch.bfloat16
+        # bfloat16 requires gradient accumulation and all-reduce to
+        # be done in fp32.
+        if not args.accumulate_allreduce_grads_in_fp32:
+            args.accumulate_allreduce_grads_in_fp32 = True
+            if args.rank == 0:
+                print('accumulate and all-reduce gradients in fp32 for '
+                      'bfloat16 data type.', flush=True)
+
+    if args.rank == 0:
+        print('using {} for parameters ...'.format(args.params_dtype),
+              flush=True)
+
+    # If we do accumulation and all-reduces in fp32, we need to have
+    # local DDP and we should set the use-contiguous-buffers-in-ddp.
+    if args.accumulate_allreduce_grads_in_fp32:
+        assert args.DDP_impl == 'local'
+        args.use_contiguous_buffers_in_ddp = True
+
+    if args.dataloader_type is None:
+        args.dataloader_type = 'single'
+
+    # Consumed tokens.
+    args.consumed_train_samples = 0
+    args.consumed_valid_samples = 0
+    args.consumed_train_tokens = 0
+    args.gigaflos_no_embeds = 0
+
+    # Iteration-based training.
+    if args.train_iters:
+        # If we use iteration-based training, make sure the
+        # sample-based options are off.
+        assert args.train_samples is None, \
+            'expected iteration-based training'
+        assert args.lr_decay_samples is None, \
+            'expected iteration-based learning rate decay'
+        assert args.lr_warmup_samples == 0, \
+            'expected iteration-based learning rate warmup'
+        assert args.rampup_batch_size is None, \
+            'expected no batch-size rampup for iteration-based training'
+        if args.lr_warmup_fraction is not None:
+            assert args.lr_warmup_iters == 0, \
+                'can only specify one of lr-warmup-fraction and lr-warmup-iters'
+
+    # Sample-based training.
+    if args.train_samples:
+        # If we use sample-based training, make sure the
+        # iteration-based options are off.
+        assert args.train_iters is None, \
+            'expected sample-based training'
+        assert args.lr_decay_iters is None, \
+            'expected sample-based learning rate decay'
+        assert args.lr_warmup_iters == 0, \
+            'expected sample-based learnig rate warmup'
+        if args.lr_warmup_fraction is not None:
+            assert args.lr_warmup_samples == 0, \
+                'can only specify one of lr-warmup-fraction ' \
+                'and lr-warmup-samples'
+
+    # Check required arguments.
+    required_args = ['num_layers', 'hidden_size', 'num_attention_heads']
+    for req_arg in required_args:
+        _check_arg_is_not_none(args, req_arg)
+
+    # Checks.
+    if args.ffn_hidden_size is None:
+        args.ffn_hidden_size = 4 * args.hidden_size
+
+    if args.kv_channels is None:
+        assert args.hidden_size % args.num_attention_heads == 0
+        args.kv_channels = args.hidden_size // args.num_attention_heads
+
+    if args.seq_length is not None:
+        assert args.encoder_seq_length is None
+        args.encoder_seq_length = args.seq_length
+    else:
+        assert args.encoder_seq_length is not None
+        args.seq_length = args.encoder_seq_length
+
+    if args.position_embedding_type == PositionEmbeddingType.absolute or args.position_embedding_type == PositionEmbeddingType.alibi:
+        assert args.max_position_embeddings is not None
+        if args.seq_length is not None:
+            assert args.max_position_embeddings >= args.seq_length
+        if args.decoder_seq_length is not None:
+            assert args.max_position_embeddings >= args.decoder_seq_length
+    else:
+        assert args.max_position_embeddings is None
+
+    if args.lr is not None:
+        assert args.min_lr <= args.lr
+    if args.save is not None:
+        assert args.save_interval is not None
+    # Mixed precision checks.
+    if args.fp16_lm_cross_entropy:
+        assert args.fp16, 'lm cross entropy in fp16 only support in fp16 mode.'
+    if args.fp32_residual_connection:
+        assert args.fp16 or args.bf16, \
+            'residual connection in fp32 only supported when using fp16 or bf16.'
+    # Activation checkpointing.
+    if args.distribute_checkpointed_activations:
+        assert args.checkpoint_activations, \
+            'for distribute-checkpointed-activations to work you '\
+            'need to enable checkpoint-activations'
+
+    args.curriculum_learning = False
+
+    # Activation function
+    if args.glu_activation is not None and args.bias_gelu_fusion:
+        raise ValueError("if glu-activation is used, please set --no-bias-gelu-fusion")
+
+    # Skip train iterations
+    if args.skip_train_iteration_range is not None:
+        args.skip_train_iteration_range = [
+            list(map(int, range_.split("-"))) for range_ in args.skip_train_iteration_range
+        ]
+        args.skip_train_iteration_range.sort()
+        skip_train_iteration_range = collections.deque()
+        for range_ in args.skip_train_iteration_range:
+            if len(range_) == 2:
+                start, end = range_
+                assert end >= start, \
+                "end of skip range cannot be smaller than start of skip range"
+                # merge overlapping intervals (e.g. 1-5 2-6 -> 1-6)
+                if not skip_train_iteration_range:
+                    skip_train_iteration_range.append([start, end])
+                elif skip_train_iteration_range[-1][1] >= start:
+                    skip_train_iteration_range[-1][1] = max(end, skip_train_iteration_range[-1][1])
+                else:
+                    skip_train_iteration_range.append([start, end])
+            else:
+                raise ValueError(
+                    "skip train iterations should be specified as two numbers, i.e. start-end"
+                )
+        args.skip_train_iteration_range = skip_train_iteration_range
+
+    if args.use_bnb_optimizer:
+        try:
+            import bitsandbytes as bnb
+        except ModuleNotFoundError:
+            raise ModuleNotFoundError("Please install bitsandbytes from https://github.com/facebookresearch/bitsandbytes.")
+
+    _print_args(args)
+    return args
+
+
+def _print_args(args):
+    """Print arguments."""
+    if args.rank == 0:
+        print('------------------------ arguments ------------------------',
+              flush=True)
+        str_list = []
+        for arg in vars(args):
+            dots = '.' * (48 - len(arg))
+            str_list.append('  {} {} {}'.format(arg, dots, getattr(args, arg)))
+
+        if args.log_path is not None:
+            with open(os.path.join(args.log_path,f'args_{time.strftime("%Y-%m-%dT%H:%M:%S")}.txt'), 'w') as f:
+                for arg in sorted(str_list, key=lambda x: x.lower()):
+                    f.write(arg+"\n")
+                    print(arg, flush=True)
+        else:
+            for arg in sorted(str_list, key=lambda x: x.lower()):
+                print(arg, flush=True)
+        print('-------------------- end of arguments ---------------------',
+              flush=True)
+
+
+def _check_arg_is_not_none(args, arg):
+    assert getattr(args, arg) is not None, '{} argument is None'.format(arg)
+
+
+def _add_network_size_args(parser):
+    group = parser.add_argument_group(title='network size')
+
+    group.add_argument('--num-layers', type=int, default=None,
+                       help='Number of transformer layers.')
+    group.add_argument('--hidden-size', type=int, default=None,
+                       help='Tansformer hidden size.')
+    group.add_argument('--ffn-hidden-size', type=int, default=None,
+                       help='Transformer Feed-Forward Network hidden size. '
+                       'This is set to 4*hidden-size if not provided')
+    group.add_argument('--num-attention-heads', type=int, default=None,
+                       help='Number of transformer attention heads.')
+    group.add_argument('--kv-channels', type=int, default=None,
+                       help='Projection weights dimension in multi-head '
+                       'attention. This is set to '
+                       '   args.hidden_size // args.num_attention_heads '
+                       'if not provided.')
+    group.add_argument('--max-position-embeddings', type=int, default=None,
+                       help='Maximum number of position embeddings to use. '
+                       'This is the size of position embedding.')
+    group.add_argument('--make-vocab-size-divisible-by', type=int, default=128,
+                       help='Pad the vocab size to be divisible by this value.'
+                       'This is added for computational efficieny reasons.')
+    group.add_argument('--pad-vocab-size-to', type=int, default=None,
+                       help='Pad the vocab size to this value.'
+                       'This value must be greater than the initial size of the tokenizer'
+                       ', needs to be divisible by TP size and `make-vocab-size-divisible-by`.')
+    group.add_argument('--layernorm-epsilon', type=float, default=1e-5,
+                       help='Layer norm epsilon.')
+    group.add_argument('--sync-tp-duplicated-parameters', action='store_true',
+                       help='Force syncing duplicated params across TP ranks in forward. '
+                       'This is a workaround for an unresolved bug leading to TP ranks '
+                       'getting out of sync with each other.')
+    group.add_argument('--apply-residual-connection-post-layernorm',
+                       action='store_true',
+                       help='If set, use original BERT residula connection '
+                       'ordering.')
+    group.add_argument('--embed-layernorm', action='store_true',
+                       help='use layernorm for embedding')
+    group.add_argument('--openai-gelu', action='store_true',
+                       help='Use OpenAIs GeLU implementation. This option'
+                       'should not be used unless for backward compatibility'
+                       'reasons.')
+    group.add_argument('--onnx-safe', type=bool, required=False,
+                       help='Use workarounds for known problems with '
+                       'Torch ONNX exporter')
+    group.add_argument('--bert-no-binary-head', action='store_false',
+                       help='Disable BERT binary head.',
+                       dest='bert_binary_head')
+    group.add_argument('--position-embedding-type', type=lambda x: PositionEmbeddingType[x],
+                       choices=list(PositionEmbeddingType),
+                       default=PositionEmbeddingType.absolute,
+                       help='Define position embedding type ("absolute" | "rotary" | "alibi"). "absolute" by default.'
+                       )
+    group.add_argument('--glu-activation', type=str,
+                       choices=megatron.model.glu_activations.GLU_ACTIVATIONS.keys(),
+                       help='GLU activations to use.'
+                       )
+
+    group.add_argument('--kill-switch-path', type=str,
+                       help='path to look for a kill switch, which if found will automatically exit the program'
+                       )
+
+
+    group.add_argument('--log-level', type=str, choices=list(log_levels.keys()),
+                       help="Logger log level to use on the main process. Possible choices are the log levels as strings: 'debug', "
+                       "'info', 'warning', 'error' and 'critical', plus a 'passive' level which doesn't set anything and lets the "
+                       "application set the level."
+                       )
+    group.add_argument('--log-level-replica', type=str, choices=list(log_levels.keys()),
+                       help="Logger log level to use on replicas. Same choices as ``log_level``"
+                       )
+    return parser
+
+
+def _add_logging_args(parser):
+    group = parser.add_argument_group(title='logging')
+
+    group.add_argument('--log-params-norm', action='store_true',
+                       help='If set, calculate and log parameters norm.')
+    group.add_argument('--log-num-zeros-in-grad', action='store_true',
+                       help='If set, calculate and log the number of zeros in gradient.')
+    group.add_argument('--tensorboard-log-interval', type=int, default=1,
+                       help='Report to tensorboard interval.')
+    group.add_argument('--tensorboard-queue-size', type=int, default=1000,
+                       help='Size of the tensorboard queue for pending events '
+                       'and summaries before one of the ‘add’ calls forces a '
+                       'flush to disk.')
+    group.add_argument('--log-timers-to-tensorboard', action='store_true',
+                       help='If set, write timers to tensorboard.')
+    group.add_argument('--log-batch-size-to-tensorboard', action='store_true',
+                       help='If set, write batch-size to tensorboard.')
+    group.add_argument('--no-log-learnig-rate-to-tensorboard',
+                       action='store_false',
+                       help='Disable learning rate logging to tensorboard.',
+                       dest='log_learning_rate_to_tensorboard')
+    group.add_argument('--no-log-loss-scale-to-tensorboard',
+                       action='store_false',
+                       help='Disable loss-scale logging to tensorboard.',
+                       dest='log_loss_scale_to_tensorboard')
+    group.add_argument('--log-validation-ppl-to-tensorboard',
+                       action='store_true',
+                       help='If set, write validation perplexity to '
+                       'tensorboard.')
+
+    return parser
+
+
+def _add_regularization_args(parser):
+    group = parser.add_argument_group(title='regularization')
+
+    group.add_argument('--attention-dropout', type=float, default=0.1,
+                       help='Post attention dropout probability.')
+    group.add_argument('--hidden-dropout', type=float, default=0.1,
+                       help='Dropout probability for hidden state transformer.')
+    group.add_argument('--weight-decay', type=float, default=0.01,
+                       help='Weight decay coefficient for L2 regularization.')
+    group.add_argument('--clip-grad', type=float, default=1.0,
+                       help='Gradient clipping based on global L2 norm.')
+    group.add_argument('--adam-beta1', type=float, default=0.9,
+                       help='First coefficient for computing running averages '
+                       'of gradient and its square')
+    group.add_argument('--adam-beta2', type=float, default=0.999,
+                       help='Second coefficient for computing running averages '
+                       'of gradient and its square')
+    group.add_argument('--adam-eps', type=float, default=1e-08,
+                       help='Term added to the denominator to improve'
+                       'numerical stability')
+    group.add_argument('--sgd-momentum', type=float, default=0.9,
+                       help='Momentum factor for sgd')
+
+    return parser
+
+
+def _add_training_args(parser):
+    group = parser.add_argument_group(title='training')
+
+    group.add_argument('--micro-batch-size', type=int, default=None,
+                       help='Batch size per model instance (local batch size). '
+                       'Global batch size is local batch size times data '
+                       'parallel size times number of micro batches.')
+    group.add_argument('--batch-size', type=int, default=None,
+                       help='Old batch size parameter, do not use. '
+                       'Use --micro-batch-size instead')
+    group.add_argument('--global-batch-size', type=int, default=None,
+                       help='Training batch size. If set, it should be a '
+                       'multiple of micro-batch-size times data-parallel-size. '
+                       'If this value is None, then '
+                       'use micro-batch-size * data-parallel-size as the '
+                       'global batch size. This choice will result in 1 for '
+                       'number of micro-batches.')
+    group.add_argument('--rampup-batch-size', nargs='*', default=None,
+                       help='Batch size ramp up with the following values:'
+                       '  --rampup-batch-size <start batch size> '
+                       '                      <batch size increment> '
+                       '                      <ramp-up samples> '
+                       'For example: '
+                       '   --rampup-batch-size 16 8 300000 '
+                       '   --global-batch-size 1024 '
+                       'will start with global batch size 16 and over '
+                       ' (1024 - 16) / 8 = 126 intervals will increase '
+                       'the batch size linearly to 1024. In each interval '
+                       'we will use approximately 300000 / 126 = 2380 samples.')
+    group.add_argument('--checkpoint-activations', action='store_true',
+                       help='Checkpoint activation to allow for training '
+                       'with larger models, sequences, and batch sizes.')
+    group.add_argument('--distribute-checkpointed-activations',
+                       action='store_true',
+                       help='If set, distribute checkpointed activations '
+                       'across model parallel group.')
+    group.add_argument('--checkpoint-num-layers', type=int, default=1,
+                       help='chunk size (number of layers) for checkpointing.')
+    group.add_argument('--train-iters', type=int, default=None,
+                       help='Total number of iterations to train over all '
+                       'training runs. Note that either train-iters or '
+                       'train-samples should be provided.')
+    group.add_argument('--train-samples', type=int, default=None,
+                       help='Total number of samples to train over all '
+                       'training runs. Note that either train-iters or '
+                       'train-samples should be provided.')
+    group.add_argument('--train-tokens', type=int, default=None,
+                       help='Total number of tokens to train over all '
+                       'training runs.')
+    group.add_argument('--log-interval', type=int, default=100,
+                       help='Report loss and timing interval.')
+    group.add_argument('--exit-interval', type=int, default=None,
+                       help='Exit the program after the iteration is divisible '
+                       'by this value.')
+    group.add_argument('--exit-duration-in-mins', type=int, default=None,
+                       help='Exit the program after this many minutes.')
+    group.add_argument('--tensorboard-dir', type=str, default=None,
+                       help='Write TensorBoard logs to this directory.')
+    group.add_argument('--no-masked-softmax-fusion',
+                       action='store_false',
+                       help='Disable fusion of query_key_value scaling, '
+                       'masking, and softmax.',
+                       dest='masked_softmax_fusion')
+    group.add_argument('--no-bias-gelu-fusion', action='store_false',
+                       help='Disable bias and gelu fusion.',
+                       dest='bias_gelu_fusion')
+    group.add_argument('--no-bias-dropout-fusion', action='store_false',
+                       help='Disable bias and dropout fusion.',
+                       dest='bias_dropout_fusion')
+    group.add_argument('--optimizer', type=str, default='adam',
+                       choices=['adam', 'sgd'],
+                       help='Optimizer function')
+    group.add_argument('--use-bnb-optimizer', action='store_true',
+                       help='Use bitsandbytes optimizer for efficient training,'
+                       'please refer https://github.com/facebookresearch/bitsandbytes.',
+                       dest='use_bnb_optimizer')
+    group.add_argument('--dataloader-type', type=str, default=None,
+                       choices=['single', 'cyclic'],
+                       help='Single pass vs multiple pass data loader')
+    group.add_argument('--cpu-optimizer', action='store_true',
+                       help='Run optimizer on CPU')
+    group.add_argument('--cpu_torch_adam', action='store_true',
+                       help='Use Torch Adam as optimizer on CPU.')
+    group.add_argument('--codecarbon-dir', type=str, default=None,
+                       help='Write CodeCarbon logs to this directory.')
+    group.add_argument('--eval-only', type=bool, required=False,
+                       help='If set to True, no train step will be performed.'
+                       'and only the evaluation on the `valid` and `test` sets '
+                       'will be performed' )
+    group.add_argument('--skip-train-iteration-range', type=str, nargs='+', default=None,
+                       help='Iteration ranges to skip. The values are one or more dash-separated ranges. e.g., 101-200 251-300.')
+    group.add_argument('--inference', action='store_true',
+                       help='Very basic inference mode: not allocating optim/lr - requires ZERO_STAGE=0')
+    group.add_argument('--abort-on-unmet-fused-kernel-constraints', action='store_true',
+                       help="If set to True, the program will abort if the constraints for loading a fused kernel aren't met")
+    group.add_argument('--pp-partition-method', type=str, default=None,
+                       help="Use to override the pipeline stages partitioning method. e.g., 'type:transformer|embedding'")
+
+    return parser
+
+
+def _add_initialization_args(parser):
+    group = parser.add_argument_group(title='initialization')
+
+    group.add_argument('--seed', type=int, default=1234,
+                       help='Random seed used for python, numpy, '
+                       'pytorch, and cuda.')
+    group.add_argument('--init-method-std', type=float, default=0.02,
+                       help='Standard deviation of the zero mean normal '
+                       'distribution used for weight initialization.')
+    group.add_argument('--init-method-xavier-uniform', action='store_true',
+                       help='Enable Xavier uniform parameter initialization')
+
+    return parser
+
+
+def _add_learning_rate_args(parser):
+    group = parser.add_argument_group(title='learning rate')
+
+    group.add_argument('--lr', type=float, default=None,
+                       help='Initial learning rate. Depending on decay style '
+                       'and initial warmup, the learing rate at each '
+                       'iteration would be different.')
+    group.add_argument('--lr-decay-style', type=str, default='linear',
+                       choices=['constant', 'linear', 'cosine'],
+                       help='Learning rate decay function.')
+    group.add_argument('--lr-decay-iters', type=int, default=None,
+                       help='number of iterations to decay learning rate over,'
+                       ' If None defaults to `--train-iters`')
+    group.add_argument('--lr-decay-samples', type=int, default=None,
+                       help='number of samples to decay learning rate over,'
+                       ' If None defaults to `--train-samples`')
+    group.add_argument('--lr-decay-tokens', type=int, default=None,
+                       help='number of tokens to decay learning rate over,'
+                       ' If not None will override iter/sample-based decay')
+    group.add_argument('--lr-warmup-fraction', type=float, default=None,
+                       help='fraction of lr-warmup-(iters/samples) to use '
+                       'for warmup (as a float)')
+    group.add_argument('--lr-warmup-iters', type=int, default=0,
+                       help='number of iterations to linearly warmup '
+                       'learning rate over.')
+    group.add_argument('--lr-warmup-samples', type=int, default=0,
+                       help='number of samples to linearly warmup '
+                       'learning rate over.')
+    group.add_argument('--warmup', type=int, default=None,
+                       help='Old lr warmup argument, do not use. Use one of the'
+                       '--lr-warmup-* arguments above')
+    group.add_argument('--min-lr', type=float, default=0.0,
+                       help='Minumum value for learning rate. The scheduler'
+                       'clip values below this threshold.')
+    group.add_argument('--override-lr-scheduler', action='store_true',
+                       help='Reset the values of the scheduler (learning rate,'
+                       'warmup iterations, minimum learning rate, maximum '
+                       'number of iterations, and decay style from input '
+                       'arguments and ignore values from checkpoints. Note'
+                       'that all the above values will be reset.')
+    group.add_argument('--use-checkpoint-lr-scheduler', action='store_true',
+                       help='Use checkpoint to set the values of the scheduler '
+                       '(learning rate, warmup iterations, minimum learning '
+                       'rate, maximum number of iterations, and decay style '
+                       'from checkpoint and ignore input arguments.')
+    group.add_argument('--universal-checkpoint', action='store_true',
+                        help='Loading a universal format checkpoint.')
+
+    return parser
+
+
+def _add_checkpointing_args(parser):
+    group = parser.add_argument_group(title='checkpointing')
+
+    group.add_argument('--save', type=str, default=None,
+                       help='Output directory to save checkpoints to.')
+    group.add_argument('--save-interval', type=int, default=None,
+                       help='Number of iterations between checkpoint saves.')
+    group.add_argument('--no-save-optim', action='store_true', default=None,
+                       help='Do not save current optimizer.')
+    group.add_argument('--no-save-rng', action='store_true', default=None,
+                       help='Do not save current rng state.')
+    group.add_argument('--load', type=str, default=None,
+                       help='Directory containing a model checkpoint.')
+    group.add_argument('--no-load-optim', action='store_true', default=None,
+                       help='Do not load optimizer when loading checkpoint.')
+    group.add_argument('--no-load-rng', action='store_true', default=None,
+                       help='Do not load rng state when loading checkpoint.')
+    group.add_argument('--finetune', action='store_true',
+                       help='Load model for finetuning. Do not load optimizer '
+                       'or rng state from checkpoint and set iteration to 0. '
+                       'Assumed when loading a release checkpoint.')
+
+    return parser
+
+
+def _add_mixed_precision_args(parser):
+    group = parser.add_argument_group(title='mixed precision')
+
+    group.add_argument('--fp16', action='store_true',
+                       help='Run model in fp16 mode.')
+    group.add_argument('--bf16', action='store_true',
+                       help='Run model in bfloat16 mode.')
+    group.add_argument('--loss-scale', type=float, default=None,
+                       help='Static loss scaling, positive power of 2 '
+                       'values can improve fp16 convergence. If None, dynamic'
+                       'loss scaling is used.')
+    group.add_argument('--initial-loss-scale', type=float, default=2**32,
+                       help='Initial loss-scale for dynamic loss scaling.')
+    group.add_argument('--min-loss-scale', type=float, default=1.0,
+                       help='Minimum loss scale for dynamic loss scale.')
+    group.add_argument('--loss-scale-window', type=float, default=1000,
+                       help='Window over which to raise/lower dynamic scale.')
+    group.add_argument('--hysteresis', type=int, default=2,
+                       help='hysteresis for dynamic loss scaling')
+    group.add_argument('--fp32-residual-connection', action='store_true',
+                       help='Move residual connections to fp32.')
+    group.add_argument('--no-query-key-layer-scaling', action='store_false',
+                       help='Do not scale Q * K^T by 1 / layer-number.',
+                       dest='apply_query_key_layer_scaling')
+    group.add_argument('--attention-softmax-in-fp32', action='store_true',
+                       help='Run attention masking and softmax in fp32. '
+                       'This flag is ignored unless '
+                       '--no-query-key-layer-scaling is specified.')
+    group.add_argument('--accumulate-allreduce-grads-in-fp32',
+                       action='store_true',
+                       help='Gradient accumulation and all-reduce in fp32.')
+    group.add_argument('--fp16-lm-cross-entropy', action='store_true',
+                       help='Move the cross entropy unreduced loss calculation'
+                       'for lm head to fp16.')
+
+    return parser
+
+
+def _add_distributed_args(parser):
+    group = parser.add_argument_group(title='distributed')
+
+    group.add_argument('--tensor-model-parallel-size', type=int, default=1,
+                       help='Degree of tensor model parallelism.')
+    group.add_argument('--pipeline-model-parallel-size', type=int, default=1,
+                       help='Degree of pipeline model parallelism.')
+    group.add_argument('--model-parallel-size', type=int, default=None,
+                       help='Old model parallel argument, do not use. Use '
+                       '--tensor-model-parallel-size instead.')
+    group.add_argument('--num-layers-per-virtual-pipeline-stage', type=int, default=None,
+                       help='Number of layers per virtual pipeline stage')
+    group.add_argument('--distributed-backend', default='nccl',
+                       choices=['nccl', 'gloo'],
+                       help='Which backend to use for distributed training.')
+    group.add_argument('--DDP-impl', default='local',
+                       choices=['local', 'torch'],
+                       help='which DistributedDataParallel implementation '
+                       'to use.')
+    group.add_argument('--use-contiguous-buffers-in-ddp', action='store_true',
+                       help='If set, use contiguous buffer in DDP. Note that '
+                       'this option only works woth local DDP.' )
+    group.add_argument('--no-scatter-gather-tensors-in-pipeline', action='store_false',
+                       help='Use scatter/gather to optimize communication of tensors in pipeline',
+                       dest='scatter_gather_tensors_in_pipeline')
+    group.add_argument('--local_rank', type=int, default=None,
+                       help='local rank passed from distributed launcher.')
+    group.add_argument('--lazy-mpu-init', type=bool, required=False,
+                       help='If set to True, initialize_megatron() '
+                       'skips DDP initialization and returns function to '
+                       'complete it instead.Also turns on '
+                       '--use-cpu-initialization flag. This is for '
+                       'external DDP manager.' )
+    group.add_argument('--use-cpu-initialization', action='store_true',
+                       default=None, help='If set, affine parallel weights '
+                       'initialization uses CPU' )
+    group.add_argument('--rank', default=-1, type=int, help='node rank for distributed training')
+    group.add_argument('--dist_url', type=str, default="env://127.0.0.1:23456")
+    group.add_argument('--world_size', type=int, default=-1, help='number of nodes for distributed training')
+    group.add_argument('--dist_backend', default='nccl', type=str, help='distributed backend')
+    return parser
+
+
+def _add_validation_args(parser):
+    group = parser.add_argument_group(title='validation')
+
+    group.add_argument('--eval-iters', type=int, default=100,
+                       help='Number of iterations to run for evaluation'
+                       'validation/test for.')
+    group.add_argument('--eval-interval', type=int, default=1000,
+                       help='Interval between running evaluation on '
+                       'validation set.')
+
+    return parser
+
+
+def _add_data_args(parser):
+    group = parser.add_argument_group(title='data and dataloader')
+
+
+    # option 1 for data loading  (mutually exclusive with option2)
+    group.add_argument('--data-path', nargs='*', default=None,
+                       help='Path to the training dataset. Accepted format:'
+                       '1) a single data path, 2) multiple datasets in the'
+                       'form: dataset1-weight dataset1-path dataset2-weight '
+                       'dataset2-path ...')
+
+    group.add_argument('--split', type=str, default=None,
+                       help='Comma-separated list of proportions for training,'
+                       ' validation, and test split. For example the split '
+                       '`90,5,5` will use 90%% of data for training, 5%% for '
+                       'validation and 5%% for test.')
+
+    # option 2 for data loading (mutually exclusive with option1)
+
+    # helper class to parse the --xxx-weighted-split-paths
+    # note here two args are set: extra valid dataset paths and names
+    class parse_data_paths(argparse.Action):
+        def __call__(self, parser, args, values, option_string=None):
+
+            if option_string == "--train-weighted-split-paths":
+                assert len(values) == 1, 'Only 1 dataset group is allowed to'
+                'be passed for the argument --train-weighted-split-paths'
+
+            # make sure string given in the correct format
+            err_message = 'Each data group should be input on the following format'
+            '"GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2"'
+            'where START < END'
+            for v in values:
+                # each prefix consists several datasets separated by commas
+                prefix = ":".join(v.split(":")[1:]) # remove GIVEN_NAME
+                datasets = prefix.split(",")
+                # check if each dataset is formatted like `WEIGHT START:END PATH`
+                for d in datasets:
+                    assert len(d.split()) == 3, err_message
+                    start, end = d.split()[1].split(":")
+                    assert float(start) < float(end), err_message
+
+            names = [v.split(":")[0] for v in values]
+
+            prefixes = [":".join(v.split(":")[1:]).strip() for v in values]
+            weights = [[d.split()[0] for d in p.split(",")] for p in prefixes]
+            splits = [[d.split()[1] for d in p.split(",")] for p in prefixes]
+            paths = [[d.split()[2] for d in p.split(",")] for p in prefixes]
+
+            # # to keep consistency with Option 1 of data loading (through --data-path)
+            # #  paths will contain strings on the following form
+            # # "WEIGHTS1 PATH1 WEIGHTS2 PATH2 WEIGHTS3 PATH3" for each dataset group
+            # # while data will be parsed in additional arguments below
+            # paths_option1_style = []
+            # for p, w in zip(paths, weights):
+            #   paths_option1_style.append(" ".join([f"{w_i} {p_i}" for p_i, w_i in zip(p,w)]))
+            # setattr(args, self.dest, paths_option1_style)
+            setattr(args, self.dest, paths)
+            setattr(args, self.dest.replace("paths", "weights"), weights)
+            setattr(args, self.dest.replace("paths", "splits"), splits)
+            setattr(args, self.dest.replace("paths","names"), names)
+
+
+    group.add_argument('--train-weighted-split-paths', nargs='*', default=None,
+                    help='Weights, splits and paths to groups of datasets'
+                    'Accepted format: ONE dataset groups could be'
+                    'submitted in the following form between double quotes'
+                    '"GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2"'
+                    'e.g.: "NAME_ABC: 0.6 0:0.6 A, 0.3 0:1 B, 0.1 0:1 C" '
+                    'WEIGHT is used to up and down sample each dataset A,B,C in the group'
+                    'START:END indicates the split portion of the dataset',
+                    action=parse_data_paths)
+
+    group.add_argument('--valid-weighted-split-paths', nargs='*', default=None,
+                    help='Weights, splits and paths to groups of datasets'
+                    'Accepted format: one or many dataset groups could be'
+                    'submitted in the following form each between double quotes'
+                    '"GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2"'
+                    'e.g.: "NAME_ABC: 0.6 0.6:0.8 A, 0.3 0:1 B, 0.1 0:1 C" '
+                    '"NAME_CDE: 0.6 0.6:0.8 C, 0.3 0:1 D, 0.1 0:1 E" '
+                    'validation will be run on each of those groups independently',
+                    action=parse_data_paths)
+
+    group.add_argument('--test-weighted-split-paths', nargs='*', default=None,
+                    help='Weights, splits and paths to groups of datasets'
+                    'Accepted format: one or many dataset groups could be'
+                    'submitted in the following form each between double quotes'
+                    '"GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2"'
+                    'e.g.: "NAME_ABC: 0.6 0.6:0.8 A, 0.3 0:1 B, 0.1 0:1 C" '
+                    '"NAME_CDE: 0.6 0.6:0.8 C, 0.3 0:1 D, 0.1 0:1 E" '
+                    'test will be run on each of those groups independently',
+                    action=parse_data_paths)
+
+    class parse_data_paths_path(argparse.Action):
+        def __call__(self, parser, args, values, option_string=None):
+            expected_option_strings = ["--train-weighted-split-paths-path", "--valid-weighted-split-paths-path", "--test-weighted-split-paths-path"]
+            assert option_string in expected_option_strings, f"Expected {option_string} to be in {expected_option_strings}"
+
+            with open(values, "r") as fi:
+                lines = fi.readlines()
+                assert len(lines) == 1, f"Got multiple lines {len(lines)} instead of 1 expected"
+                assert lines[0][-2:] == "\"\n" and lines[0][0] == "\"", f"Invalid input format, got {lines}"
+                values = lines[0][1:-2].split("\" \"")
+                weighted_split_paths_dest = re.sub(r"_path$", "", self.dest)
+                weighted_split_paths_option = re.sub(r"-path$", "", self.option_strings[0])
+                setattr(args, weighted_split_paths_dest, values)
+                parse_data_paths(option_strings=[weighted_split_paths_option], dest=weighted_split_paths_dest)(parser, args, values, option_string=weighted_split_paths_option)
+
+
+    group.add_argument('--train-weighted-split-paths-path', type=str, action=parse_data_paths_path ,default=None)
+    group.add_argument('--valid-weighted-split-paths-path', type=str, action=parse_data_paths_path, default=None)
+    group.add_argument('--test-weighted-split-paths-path', type=str, action=parse_data_paths_path, default=None)
+
+    group.add_argument('--log-path', type=str, default=None,
+                       help='Path to the save arguments file.')
+    group.add_argument('--vocab-file', type=str, default=None,
+                       help='Path to the vocab file.')
+    group.add_argument('--merge-file', type=str, default=None,
+                       help='Path to the BPE merge file.')
+    group.add_argument('--vocab-extra-ids', type=int, default=0,
+                       help='Number of additional vocabulary tokens. '
+                            'They are used for span masking in the T5 model')
+    group.add_argument('--seq-length', type=int, default=None,
+                       help='Maximum sequence length to process.')
+    group.add_argument('--encoder-seq-length', type=int, default=None,
+                       help='Maximum encoder sequence length to process.'
+                       'This should be exclusive of --seq-length')
+    group.add_argument('--decoder-seq-length', type=int, default=None,
+                       help="Maximum decoder sequence length to process.")
+    group.add_argument('--retriever-seq-length', type=int, default=256,
+                       help='Maximum sequence length for the biencoder model '
+                        ' for retriever')
+    group.add_argument('--sample-rate', type=float, default=1.0,
+                       help='sample rate for training data. Supposed to be 0 '
+                            ' < sample_rate < 1')
+    group.add_argument('--mask-prob', type=float, default=0.15,
+                       help='Probability of replacing a token with mask.')
+    group.add_argument('--short-seq-prob', type=float, default=0.1,
+                       help='Probability of producing a short sequence.')
+    group.add_argument('--mmap-warmup', action='store_true',
+                       help='Warm up mmap files.')
+    group.add_argument('--num-workers', type=int, default=2,
+                       help="Dataloader number of workers.")
+    group.add_argument('--valid-num-workers', type=int, default=2,
+                       help="Dataloader number of workers for validation.")
+    group.add_argument('--tokenizer-type', type=str,
+                       default=None,
+                       choices=['BertWordPieceLowerCase',
+                                'BertWordPieceCase',
+                                'GPT2BPETokenizer',
+                                'PretrainedFromHF'],
+                       help='What type of tokenizer to use.')
+    group.add_argument("--tokenizer-name-or-path", type=str, default=None,
+                       help="Name or path of the huggingface tokenizer.")
+    group.add_argument('--data-impl', type=str, default='infer',
+                       choices=['lazy', 'cached', 'mmap', 'infer'],
+                       help='Implementation of indexed datasets.')
+    group.add_argument('--reset-position-ids', action='store_true',
+                       help='Reset posistion ids after end-of-document token.')
+    group.add_argument('--reset-attention-mask', action='store_true',
+                       help='Reset self attention maske after '
+                       'end-of-document token. Attention between tokens from different documents is null.')
+    group.add_argument('--eod-mask-loss', action='store_true',
+                       help='Mask loss for the end of document tokens.')
+    group.add_argument('--loss-on-targets-only', action='store_true',
+                       help='Mask loss on input sequence.')
+    group.add_argument('--reweight-loss-based-on-position-frequency', action="store_true",
+                       help='Some objectives require us to sample loss_mask. This might introduce bias towards '
+                       'specific positions. This option tries to un-bias the loss by reweighting loss on specific '
+                       'positions based on how frequently we train on that position.'
+                       'This is mostly used for prefix_lm training')
+    group.add_argument("--noise-density", type=float, default=None, help="Span corruption noise density")
+    group.add_argument("--mean-noise-span-length", type=int, default=None, help="Span corruption mean noise span length")
+
+
+    return parser
+
+
+def _add_autoresume_args(parser):
+    group = parser.add_argument_group(title='autoresume')
+
+    group.add_argument('--adlr-autoresume', action='store_true',
+                       help='Enable autoresume on adlr cluster.')
+    group.add_argument('--adlr-autoresume-interval', type=int, default=1000,
+                       help='Intervals over which check for autoresume'
+                       'termination signal')
+
+    return parser
+
+
+def _add_biencoder_args(parser):
+    group = parser.add_argument_group(title='biencoder')
+
+    # network size
+    group.add_argument('--ict-head-size', type=int, default=None,
+                       help='Size of block embeddings to be used in ICT and '
+                        'REALM (paper default: 128)')
+    group.add_argument('--biencoder-projection-dim', type=int, default=0,
+                       help='Size of projection head used in biencoder (paper'
+                        ' default: 128)')
+    group.add_argument('--biencoder-shared-query-context-model', action='store_true',
+                        help='Whether to share the parameters of the query '
+                        'and context models or not')
+
+    # checkpointing
+    group.add_argument('--ict-load', type=str, default=None,
+                       help='Directory containing an ICTBertModel checkpoint')
+    group.add_argument('--bert-load', type=str, default=None,
+                       help='Directory containing an BertModel checkpoint '
+                       '(needed to start ICT and REALM)')
+
+    # data
+    group.add_argument('--titles-data-path', type=str, default=None,
+                       help='Path to titles dataset used for ICT')
+    group.add_argument('--query-in-block-prob', type=float, default=0.1,
+                       help='Probability of keeping query in block for '
+                       'ICT dataset')
+    group.add_argument('--use-one-sent-docs', action='store_true',
+                       help='Whether to use one sentence documents in ICT')
+    group.add_argument('--evidence-data-path', type=str, default=None,
+                       help='Path to Wikipedia Evidence frm DPR paper')
+
+    # training
+    group.add_argument('--retriever-report-topk-accuracies', nargs='+', type=int,
+                        default=[], help="Which top-k accuracies to report "
+                        "(e.g. '1 5 20')")
+    group.add_argument('--retriever-score-scaling', action='store_true',
+                       help='Whether to scale retriever scores by inverse '
+                        'square root of hidden size')
+
+    # faiss index
+    group.add_argument('--block-data-path', type=str, default=None,
+                       help='Where to save/load BlockData to/from')
+    group.add_argument('--embedding-path', type=str, default=None,
+                       help='Where to save/load Open-Retrieval Embedding'
+                        ' data to/from')
+
+    # indexer
+    group.add_argument('--indexer-batch-size', type=int, default=128,
+                       help='How large of batches to use when doing indexing '
+                       'jobs')
+    group.add_argument('--indexer-log-interval', type=int, default=1000,
+                       help='After how many batches should the indexer '
+                       'report progress')
+    return parser
+
+
+def _add_vit_args(parser):
+    group = parser.add_argument_group(title="vit")
+
+    group.add_argument('--num-classes', type=int, default=1000,
+                       help='num of classes in vision classificaiton task')
+    group.add_argument('--img-dim', type=int, default=224,
+                       help='Image size for vision classification task')
+    group.add_argument('--num-channels', type=int, default=3,
+                       help='Number of channels in input image data')
+    group.add_argument('--patch-dim', type=int, default=16,
+                       help='patch dimension used in vit')
+
+    return parser
+
+
+def _add_zero_args(parser):
+    """Text generate arguments."""
+
+    group = parser.add_argument_group('ZeRO configurations', 'configurations')
+    group.add_argument("--zero-stage", type=int, default=1.0)
+    group.add_argument('--zero-reduce-scatter', action='store_true',
+                       help='Use reduce scatter if specified')
+    group.add_argument('--zero-contigious-gradients', action='store_true',
+                       help='Use contigious memory optimizaiton if specified')
+    group.add_argument("--zero-reduce-bucket-size", type=int, default=0.0)
+    group.add_argument("--zero-allgather-bucket-size", type=int, default=0.0)
+    group.add_argument('--remote-device', type=str, default='none', choices=['none', 'cpu', 'nvme'],
+                      help='Remote device for ZeRO-3 initialized parameters.')
+    group.add_argument('--use-pin-memory', action='store_true',
+                     help='Use pinned CPU memory for ZeRO-3 initialized model parameters.')
+    return parser
+
+def _add_memoryopt_args(parser):
+    """Memory optimization arguments."""
+
+    group = parser.add_argument_group('Memory optimizations', 'configurations')
+    group.add_argument("--scattered-embeddings", action='store_true',
+                       help='Save memory by scattering embedding activations. '
+                            'Introduces dropout differences across MP configurations.')
+    group.add_argument("--split-transformers", action='store_true',
+                       help='Save memory by splitting transformer layers into two parts, '
+                       'allowing for more frequent activation checkpoint savings.')
+    group.add_argument("--memory-centric-tiled-linear", action="store_true",
+                       help='Save memory by tiling with deepspeed.zero.TiledLinear.')
+    group.add_argument("--tile-factor", type=int, default=1,
+                       help='Make all linear layers the same size of [hidden/tile_factor, hidden/tile_factor]. '
+                            'Must be enabled with --memory-centric-tiled-linear. '
+                            'Example A: if tile_factor=1, the qkv layer [hidden, 3* hidden] would be converted into [1,3] tiles of size [hidden,hidden]. '
+                            'Example B: if tile_factor=2, the intermediate layer [4*hidden, hidden] will be converted into [8, 2] tiles of size [hidden/2, hidden/2]. '
+                            'Default is 1.')
+
+    return parser
+
+def _add_activation_checkpoint_args(parser):
+    group = parser.add_argument_group('Activation Checkpointing',
+                                      'Checkpointing Configurations')
+    group.add_argument('--deepspeed-activation-checkpointing', action='store_true',
+                       help='uses activation checkpointing from deepspeed')
+    group.add_argument('--partition-activations', action='store_true',
+                       help='partition Activations across GPUs before checkpointing.')
+    group.add_argument('--contigious-checkpointing', action='store_true',
+                       help='Contigious memory checkpointing for activatoins.')
+    group.add_argument('--checkpoint-in-cpu', action='store_true',
+                       help='Move the activation checkpoints to CPU.')
+    group.add_argument('--synchronize-each-layer', action='store_true',
+                       help='does a synchronize at the beginning and end of each checkpointed layer.')
+    group.add_argument('--profile-backward', action='store_true',
+                       help='Enables backward pass profiling for checkpointed layers.')
+    return parser

megatron-deepspeed_dtk22.10/megatron/arguments.py-one_node

+# coding=utf-8
+# Copyright (c) 2020, 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.
+
+"""Megatron arguments."""
+
+import argparse
+import collections
+import os
+import re
+import time
+
+import torch
+import deepspeed
+
+from megatron.enums import PositionEmbeddingType
+import megatron
+from megatron.logging import log_levels
+
+
+def parse_args(extra_args_provider=None, defaults={},
+               ignore_unknown_args=False):
+    """Parse all arguments."""
+    parser = argparse.ArgumentParser(description='Megatron-LM Arguments',
+                                     allow_abbrev=False)
+
+    # Standard arguments.
+    parser = _add_network_size_args(parser)
+    parser = _add_regularization_args(parser)
+    parser = _add_training_args(parser)
+    parser = _add_initialization_args(parser)
+    parser = _add_learning_rate_args(parser)
+    parser = _add_checkpointing_args(parser)
+    parser = _add_mixed_precision_args(parser)
+    parser = _add_distributed_args(parser)
+    parser = _add_validation_args(parser)
+    parser = _add_data_args(parser)
+    parser = _add_autoresume_args(parser)
+    parser = _add_biencoder_args(parser)
+    parser = _add_vit_args(parser)
+    parser = _add_logging_args(parser)
+    parser = _add_zero_args(parser)
+    parser = _add_memoryopt_args(parser)
+    parser = _add_activation_checkpoint_args(parser)
+
+    # Custom arguments.
+    if extra_args_provider is not None:
+        parser = extra_args_provider(parser)
+
+    parser = deepspeed.add_config_arguments(parser)
+
+    # Parse.
+    if ignore_unknown_args:
+        args, _ = parser.parse_known_args()
+    else:
+        args = parser.parse_args()
+
+    # Distributed args.
+    args.rank = int(os.getenv('RANK', '0'))
+    args.world_size = int(os.getenv("WORLD_SIZE", '1'))
+    # Tensor model parallel size.
+    args.tensor_model_parallel_size = min(
+        args.tensor_model_parallel_size, args.world_size)
+    assert args.world_size % args.tensor_model_parallel_size == 0, 'world size'\
+        ' ({}) is not divisible by tensor model parallel size ({})'.format(
+            args.world_size, args.tensor_model_parallel_size)
+    # Pipeline model parallel size.
+    args.pipeline_model_parallel_size = min(
+        args.pipeline_model_parallel_size,
+        (args.world_size // args.tensor_model_parallel_size))
+    # Checks.
+    model_parallel_size = args.pipeline_model_parallel_size * \
+                          args.tensor_model_parallel_size
+    assert args.world_size % model_parallel_size == 0, 'world size is not'\
+        ' divisible by tensor parallel size ({}) times pipeline parallel ' \
+        'size ({})'.format(args.world_size, args.tensor_model_parallel_size,
+                           args.pipeline_model_parallel_size)
+    args.data_parallel_size = args.world_size // model_parallel_size
+    if args.rank == 0:
+        print('using world size: {}, data-parallel-size: {}, '
+              'tensor-model-parallel size: {}, '
+              'pipeline-model-parallel size: {} '.format(
+                  args.world_size, args.data_parallel_size,
+                  args.tensor_model_parallel_size,
+                  args.pipeline_model_parallel_size), flush=True)
+
+    # --data-path and --train-weighted-splits-paths
+    message = "Data loading Mode 1: --data-path and --split "\
+            "and Mode 2: --(train|valid|test)-weighted-split-paths"\
+            "are mutually exclusive i.e. cannot be set together."
+
+    if args.data_path:
+        assert args.train_weighted_split_paths is None, message
+        setattr(args, "valid_weighted_split_names", None)
+        setattr(args, "valid_weighted_split_weights", None)
+        setattr(args, "valid_weighted_split_splits", None)
+
+        setattr(args, "test_weighted_split_names", None)
+        setattr(args, "test_weighted_split_weights", None)
+        setattr(args, "test_weighted_split_splits", None)
+
+        # args.split default value in the args is None it is set here in order
+        # to check that it does not to overlap with the 2nd mode of data loading
+        if args.split is None:
+            args.split = "969, 30, 1"
+
+    if args.train_weighted_split_paths or args.valid_weighted_split_paths or \
+                args.test_weighted_split_paths:
+        assert args.data_path is None and args.split is None, message
+
+
+
+    # Deprecated arguments
+    assert args.batch_size is None, '--batch-size argument is no longer ' \
+        'valid, use --micro-batch-size instead'
+    del args.batch_size
+    assert args.warmup is None, '--warmup argument is no longer valid, use ' \
+        '--lr-warmup-fraction instead'
+    del args.warmup
+    assert args.model_parallel_size is None, '--model-parallel-size is no ' \
+        'longer valid, use --tensor-model-parallel-size instead'
+    del args.model_parallel_size
+
+    # Set input defaults.
+    for key in defaults:
+        # For default to be valid, it should not be provided in the
+        # arguments that are passed to the program. We check this by
+        # ensuring the arg is set to None.
+        if getattr(args, key) is not None:
+            if args.rank == 0:
+                print('WARNING: overriding default arguments for {key}:{v} \
+                       with {key}:{v2}'.format(key=key, v=defaults[key],
+                                               v2=getattr(args, key)),
+                                               flush=True)
+        else:
+            setattr(args, key, defaults[key])
+
+    # Batch size.
+    assert args.micro_batch_size is not None
+    assert args.micro_batch_size > 0
+    if args.global_batch_size is None:
+        args.global_batch_size = args.micro_batch_size * args.data_parallel_size
+        if args.rank == 0:
+            print('setting global batch size to {}'.format(
+                args.global_batch_size), flush=True)
+    assert args.global_batch_size > 0
+    if args.num_layers_per_virtual_pipeline_stage is not None:
+        assert args.pipeline_model_parallel_size > 2, \
+            'pipeline-model-parallel size should be greater than 2 with ' \
+            'interleaved schedule'
+        assert args.num_layers % args.num_layers_per_virtual_pipeline_stage == 0, \
+            'number of layers is not divisible by number of layers per virtual ' \
+            'pipeline stage'
+        args.virtual_pipeline_model_parallel_size = \
+            (args.num_layers // args.pipeline_model_parallel_size) // \
+            args.num_layers_per_virtual_pipeline_stage
+    else:
+        args.virtual_pipeline_model_parallel_size = None
+
+    # Parameters dtype.
+    args.params_dtype = torch.float
+    if args.fp16:
+        assert not args.bf16
+        args.params_dtype = torch.half
+    if args.bf16:
+        assert not args.fp16
+        args.params_dtype = torch.bfloat16
+        # bfloat16 requires gradient accumulation and all-reduce to
+        # be done in fp32.
+        if not args.accumulate_allreduce_grads_in_fp32:
+            args.accumulate_allreduce_grads_in_fp32 = True
+            if args.rank == 0:
+                print('accumulate and all-reduce gradients in fp32 for '
+                      'bfloat16 data type.', flush=True)
+
+    if args.rank == 0:
+        print('using {} for parameters ...'.format(args.params_dtype),
+              flush=True)
+
+    # If we do accumulation and all-reduces in fp32, we need to have
+    # local DDP and we should set the use-contiguous-buffers-in-ddp.
+    if args.accumulate_allreduce_grads_in_fp32:
+        assert args.DDP_impl == 'local'
+        args.use_contiguous_buffers_in_ddp = True
+
+    if args.dataloader_type is None:
+        args.dataloader_type = 'single'
+
+    # Consumed tokens.
+    args.consumed_train_samples = 0
+    args.consumed_valid_samples = 0
+    args.consumed_train_tokens = 0
+    args.gigaflos_no_embeds = 0
+
+    # Iteration-based training.
+    if args.train_iters:
+        # If we use iteration-based training, make sure the
+        # sample-based options are off.
+        assert args.train_samples is None, \
+            'expected iteration-based training'
+        assert args.lr_decay_samples is None, \
+            'expected iteration-based learning rate decay'
+        assert args.lr_warmup_samples == 0, \
+            'expected iteration-based learning rate warmup'
+        assert args.rampup_batch_size is None, \
+            'expected no batch-size rampup for iteration-based training'
+        if args.lr_warmup_fraction is not None:
+            assert args.lr_warmup_iters == 0, \
+                'can only specify one of lr-warmup-fraction and lr-warmup-iters'
+
+    # Sample-based training.
+    if args.train_samples:
+        # If we use sample-based training, make sure the
+        # iteration-based options are off.
+        assert args.train_iters is None, \
+            'expected sample-based training'
+        assert args.lr_decay_iters is None, \
+            'expected sample-based learning rate decay'
+        assert args.lr_warmup_iters == 0, \
+            'expected sample-based learnig rate warmup'
+        if args.lr_warmup_fraction is not None:
+            assert args.lr_warmup_samples == 0, \
+                'can only specify one of lr-warmup-fraction ' \
+                'and lr-warmup-samples'
+
+    # Check required arguments.
+    required_args = ['num_layers', 'hidden_size', 'num_attention_heads']
+    for req_arg in required_args:
+        _check_arg_is_not_none(args, req_arg)
+
+    # Checks.
+    if args.ffn_hidden_size is None:
+        args.ffn_hidden_size = 4 * args.hidden_size
+
+    if args.kv_channels is None:
+        assert args.hidden_size % args.num_attention_heads == 0
+        args.kv_channels = args.hidden_size // args.num_attention_heads
+
+    if args.seq_length is not None:
+        assert args.encoder_seq_length is None
+        args.encoder_seq_length = args.seq_length
+    else:
+        assert args.encoder_seq_length is not None
+        args.seq_length = args.encoder_seq_length
+
+    if args.position_embedding_type == PositionEmbeddingType.absolute or args.position_embedding_type == PositionEmbeddingType.alibi:
+        assert args.max_position_embeddings is not None
+        if args.seq_length is not None:
+            assert args.max_position_embeddings >= args.seq_length
+        if args.decoder_seq_length is not None:
+            assert args.max_position_embeddings >= args.decoder_seq_length
+    else:
+        assert args.max_position_embeddings is None
+
+    if args.lr is not None:
+        assert args.min_lr <= args.lr
+    if args.save is not None:
+        assert args.save_interval is not None
+    # Mixed precision checks.
+    if args.fp16_lm_cross_entropy:
+        assert args.fp16, 'lm cross entropy in fp16 only support in fp16 mode.'
+    if args.fp32_residual_connection:
+        assert args.fp16 or args.bf16, \
+            'residual connection in fp32 only supported when using fp16 or bf16.'
+    # Activation checkpointing.
+    if args.distribute_checkpointed_activations:
+        assert args.checkpoint_activations, \
+            'for distribute-checkpointed-activations to work you '\
+            'need to enable checkpoint-activations'
+
+    args.curriculum_learning = False
+
+    # Activation function
+    if args.glu_activation is not None and args.bias_gelu_fusion:
+        raise ValueError("if glu-activation is used, please set --no-bias-gelu-fusion")
+
+    # Skip train iterations
+    if args.skip_train_iteration_range is not None:
+        args.skip_train_iteration_range = [
+            list(map(int, range_.split("-"))) for range_ in args.skip_train_iteration_range
+        ]
+        args.skip_train_iteration_range.sort()
+        skip_train_iteration_range = collections.deque()
+        for range_ in args.skip_train_iteration_range:
+            if len(range_) == 2:
+                start, end = range_
+                assert end >= start, \
+                "end of skip range cannot be smaller than start of skip range"
+                # merge overlapping intervals (e.g. 1-5 2-6 -> 1-6)
+                if not skip_train_iteration_range:
+                    skip_train_iteration_range.append([start, end])
+                elif skip_train_iteration_range[-1][1] >= start:
+                    skip_train_iteration_range[-1][1] = max(end, skip_train_iteration_range[-1][1])
+                else:
+                    skip_train_iteration_range.append([start, end])
+            else:
+                raise ValueError(
+                    "skip train iterations should be specified as two numbers, i.e. start-end"
+                )
+        args.skip_train_iteration_range = skip_train_iteration_range
+
+    if args.use_bnb_optimizer:
+        try:
+            import bitsandbytes as bnb
+        except ModuleNotFoundError:
+            raise ModuleNotFoundError("Please install bitsandbytes from https://github.com/facebookresearch/bitsandbytes.")
+
+    _print_args(args)
+    return args
+
+
+def _print_args(args):
+    """Print arguments."""
+    if args.rank == 0:
+        print('------------------------ arguments ------------------------',
+              flush=True)
+        str_list = []
+        for arg in vars(args):
+            dots = '.' * (48 - len(arg))
+            str_list.append('  {} {} {}'.format(arg, dots, getattr(args, arg)))
+
+        if args.log_path is not None:
+            with open(os.path.join(args.log_path,f'args_{time.strftime("%Y-%m-%dT%H:%M:%S")}.txt'), 'w') as f:
+                for arg in sorted(str_list, key=lambda x: x.lower()):
+                    f.write(arg+"\n")
+                    print(arg, flush=True)
+        else:
+            for arg in sorted(str_list, key=lambda x: x.lower()):
+                print(arg, flush=True)
+        print('-------------------- end of arguments ---------------------',
+              flush=True)
+
+
+def _check_arg_is_not_none(args, arg):
+    assert getattr(args, arg) is not None, '{} argument is None'.format(arg)
+
+
+def _add_network_size_args(parser):
+    group = parser.add_argument_group(title='network size')
+
+    group.add_argument('--num-layers', type=int, default=None,
+                       help='Number of transformer layers.')
+    group.add_argument('--hidden-size', type=int, default=None,
+                       help='Tansformer hidden size.')
+    group.add_argument('--ffn-hidden-size', type=int, default=None,
+                       help='Transformer Feed-Forward Network hidden size. '
+                       'This is set to 4*hidden-size if not provided')
+    group.add_argument('--num-attention-heads', type=int, default=None,
+                       help='Number of transformer attention heads.')
+    group.add_argument('--kv-channels', type=int, default=None,
+                       help='Projection weights dimension in multi-head '
+                       'attention. This is set to '
+                       '   args.hidden_size // args.num_attention_heads '
+                       'if not provided.')
+    group.add_argument('--max-position-embeddings', type=int, default=None,
+                       help='Maximum number of position embeddings to use. '
+                       'This is the size of position embedding.')
+    group.add_argument('--make-vocab-size-divisible-by', type=int, default=128,
+                       help='Pad the vocab size to be divisible by this value.'
+                       'This is added for computational efficieny reasons.')
+    group.add_argument('--pad-vocab-size-to', type=int, default=None,
+                       help='Pad the vocab size to this value.'
+                       'This value must be greater than the initial size of the tokenizer'
+                       ', needs to be divisible by TP size and `make-vocab-size-divisible-by`.')
+    group.add_argument('--layernorm-epsilon', type=float, default=1e-5,
+                       help='Layer norm epsilon.')
+    group.add_argument('--sync-tp-duplicated-parameters', action='store_true',
+                       help='Force syncing duplicated params across TP ranks in forward. '
+                       'This is a workaround for an unresolved bug leading to TP ranks '
+                       'getting out of sync with each other.')
+    group.add_argument('--apply-residual-connection-post-layernorm',
+                       action='store_true',
+                       help='If set, use original BERT residula connection '
+                       'ordering.')
+    group.add_argument('--embed-layernorm', action='store_true',
+                       help='use layernorm for embedding')
+    group.add_argument('--openai-gelu', action='store_true',
+                       help='Use OpenAIs GeLU implementation. This option'
+                       'should not be used unless for backward compatibility'
+                       'reasons.')
+    group.add_argument('--onnx-safe', type=bool, required=False,
+                       help='Use workarounds for known problems with '
+                       'Torch ONNX exporter')
+    group.add_argument('--bert-no-binary-head', action='store_false',
+                       help='Disable BERT binary head.',
+                       dest='bert_binary_head')
+    group.add_argument('--position-embedding-type', type=lambda x: PositionEmbeddingType[x],
+                       choices=list(PositionEmbeddingType),
+                       default=PositionEmbeddingType.absolute,
+                       help='Define position embedding type ("absolute" | "rotary" | "alibi"). "absolute" by default.'
+                       )
+    group.add_argument('--glu-activation', type=str,
+                       choices=megatron.model.glu_activations.GLU_ACTIVATIONS.keys(),
+                       help='GLU activations to use.'
+                       )
+
+    group.add_argument('--kill-switch-path', type=str,
+                       help='path to look for a kill switch, which if found will automatically exit the program'
+                       )
+
+
+    group.add_argument('--log-level', type=str, choices=list(log_levels.keys()),
+                       help="Logger log level to use on the main process. Possible choices are the log levels as strings: 'debug', "
+                       "'info', 'warning', 'error' and 'critical', plus a 'passive' level which doesn't set anything and lets the "
+                       "application set the level."
+                       )
+    group.add_argument('--log-level-replica', type=str, choices=list(log_levels.keys()),
+                       help="Logger log level to use on replicas. Same choices as ``log_level``"
+                       )
+    return parser
+
+
+def _add_logging_args(parser):
+    group = parser.add_argument_group(title='logging')
+
+    group.add_argument('--log-params-norm', action='store_true',
+                       help='If set, calculate and log parameters norm.')
+    group.add_argument('--log-num-zeros-in-grad', action='store_true',
+                       help='If set, calculate and log the number of zeros in gradient.')
+    group.add_argument('--tensorboard-log-interval', type=int, default=1,
+                       help='Report to tensorboard interval.')
+    group.add_argument('--tensorboard-queue-size', type=int, default=1000,
+                       help='Size of the tensorboard queue for pending events '
+                       'and summaries before one of the ‘add’ calls forces a '
+                       'flush to disk.')
+    group.add_argument('--log-timers-to-tensorboard', action='store_true',
+                       help='If set, write timers to tensorboard.')
+    group.add_argument('--log-batch-size-to-tensorboard', action='store_true',
+                       help='If set, write batch-size to tensorboard.')
+    group.add_argument('--no-log-learnig-rate-to-tensorboard',
+                       action='store_false',
+                       help='Disable learning rate logging to tensorboard.',
+                       dest='log_learning_rate_to_tensorboard')
+    group.add_argument('--no-log-loss-scale-to-tensorboard',
+                       action='store_false',
+                       help='Disable loss-scale logging to tensorboard.',
+                       dest='log_loss_scale_to_tensorboard')
+    group.add_argument('--log-validation-ppl-to-tensorboard',
+                       action='store_true',
+                       help='If set, write validation perplexity to '
+                       'tensorboard.')
+
+    return parser
+
+
+def _add_regularization_args(parser):
+    group = parser.add_argument_group(title='regularization')
+
+    group.add_argument('--attention-dropout', type=float, default=0.1,
+                       help='Post attention dropout probability.')
+    group.add_argument('--hidden-dropout', type=float, default=0.1,
+                       help='Dropout probability for hidden state transformer.')
+    group.add_argument('--weight-decay', type=float, default=0.01,
+                       help='Weight decay coefficient for L2 regularization.')
+    group.add_argument('--clip-grad', type=float, default=1.0,
+                       help='Gradient clipping based on global L2 norm.')
+    group.add_argument('--adam-beta1', type=float, default=0.9,
+                       help='First coefficient for computing running averages '
+                       'of gradient and its square')
+    group.add_argument('--adam-beta2', type=float, default=0.999,
+                       help='Second coefficient for computing running averages '
+                       'of gradient and its square')
+    group.add_argument('--adam-eps', type=float, default=1e-08,
+                       help='Term added to the denominator to improve'
+                       'numerical stability')
+    group.add_argument('--sgd-momentum', type=float, default=0.9,
+                       help='Momentum factor for sgd')
+
+    return parser
+
+
+def _add_training_args(parser):
+    group = parser.add_argument_group(title='training')
+
+    group.add_argument('--micro-batch-size', type=int, default=None,
+                       help='Batch size per model instance (local batch size). '
+                       'Global batch size is local batch size times data '
+                       'parallel size times number of micro batches.')
+    group.add_argument('--batch-size', type=int, default=None,
+                       help='Old batch size parameter, do not use. '
+                       'Use --micro-batch-size instead')
+    group.add_argument('--global-batch-size', type=int, default=None,
+                       help='Training batch size. If set, it should be a '
+                       'multiple of micro-batch-size times data-parallel-size. '
+                       'If this value is None, then '
+                       'use micro-batch-size * data-parallel-size as the '
+                       'global batch size. This choice will result in 1 for '
+                       'number of micro-batches.')
+    group.add_argument('--rampup-batch-size', nargs='*', default=None,
+                       help='Batch size ramp up with the following values:'
+                       '  --rampup-batch-size <start batch size> '
+                       '                      <batch size increment> '
+                       '                      <ramp-up samples> '
+                       'For example: '
+                       '   --rampup-batch-size 16 8 300000 '
+                       '   --global-batch-size 1024 '
+                       'will start with global batch size 16 and over '
+                       ' (1024 - 16) / 8 = 126 intervals will increase '
+                       'the batch size linearly to 1024. In each interval '
+                       'we will use approximately 300000 / 126 = 2380 samples.')
+    group.add_argument('--checkpoint-activations', action='store_true',
+                       help='Checkpoint activation to allow for training '
+                       'with larger models, sequences, and batch sizes.')
+    group.add_argument('--distribute-checkpointed-activations',
+                       action='store_true',
+                       help='If set, distribute checkpointed activations '
+                       'across model parallel group.')
+    group.add_argument('--checkpoint-num-layers', type=int, default=1,
+                       help='chunk size (number of layers) for checkpointing.')
+    group.add_argument('--train-iters', type=int, default=None,
+                       help='Total number of iterations to train over all '
+                       'training runs. Note that either train-iters or '
+                       'train-samples should be provided.')
+    group.add_argument('--train-samples', type=int, default=None,
+                       help='Total number of samples to train over all '
+                       'training runs. Note that either train-iters or '
+                       'train-samples should be provided.')
+    group.add_argument('--train-tokens', type=int, default=None,
+                       help='Total number of tokens to train over all '
+                       'training runs.')
+    group.add_argument('--log-interval', type=int, default=100,
+                       help='Report loss and timing interval.')
+    group.add_argument('--exit-interval', type=int, default=None,
+                       help='Exit the program after the iteration is divisible '
+                       'by this value.')
+    group.add_argument('--exit-duration-in-mins', type=int, default=None,
+                       help='Exit the program after this many minutes.')
+    group.add_argument('--tensorboard-dir', type=str, default=None,
+                       help='Write TensorBoard logs to this directory.')
+    group.add_argument('--no-masked-softmax-fusion',
+                       action='store_false',
+                       help='Disable fusion of query_key_value scaling, '
+                       'masking, and softmax.',
+                       dest='masked_softmax_fusion')
+    group.add_argument('--no-bias-gelu-fusion', action='store_false',
+                       help='Disable bias and gelu fusion.',
+                       dest='bias_gelu_fusion')
+    group.add_argument('--no-bias-dropout-fusion', action='store_false',
+                       help='Disable bias and dropout fusion.',
+                       dest='bias_dropout_fusion')
+    group.add_argument('--optimizer', type=str, default='adam',
+                       choices=['adam', 'sgd'],
+                       help='Optimizer function')
+    group.add_argument('--use-bnb-optimizer', action='store_true',
+                       help='Use bitsandbytes optimizer for efficient training,'
+                       'please refer https://github.com/facebookresearch/bitsandbytes.',
+                       dest='use_bnb_optimizer')
+    group.add_argument('--dataloader-type', type=str, default=None,
+                       choices=['single', 'cyclic'],
+                       help='Single pass vs multiple pass data loader')
+    group.add_argument('--cpu-optimizer', action='store_true',
+                       help='Run optimizer on CPU')
+    group.add_argument('--cpu_torch_adam', action='store_true',
+                       help='Use Torch Adam as optimizer on CPU.')
+    group.add_argument('--codecarbon-dir', type=str, default=None,
+                       help='Write CodeCarbon logs to this directory.')
+    group.add_argument('--eval-only', type=bool, required=False,
+                       help='If set to True, no train step will be performed.'
+                       'and only the evaluation on the `valid` and `test` sets '
+                       'will be performed' )
+    group.add_argument('--skip-train-iteration-range', type=str, nargs='+', default=None,
+                       help='Iteration ranges to skip. The values are one or more dash-separated ranges. e.g., 101-200 251-300.')
+    group.add_argument('--inference', action='store_true',
+                       help='Very basic inference mode: not allocating optim/lr - requires ZERO_STAGE=0')
+    group.add_argument('--abort-on-unmet-fused-kernel-constraints', action='store_true',
+                       help="If set to True, the program will abort if the constraints for loading a fused kernel aren't met")
+    group.add_argument('--pp-partition-method', type=str, default=None,
+                       help="Use to override the pipeline stages partitioning method. e.g., 'type:transformer|embedding'")
+
+    return parser
+
+
+def _add_initialization_args(parser):
+    group = parser.add_argument_group(title='initialization')
+
+    group.add_argument('--seed', type=int, default=1234,
+                       help='Random seed used for python, numpy, '
+                       'pytorch, and cuda.')
+    group.add_argument('--init-method-std', type=float, default=0.02,
+                       help='Standard deviation of the zero mean normal '
+                       'distribution used for weight initialization.')
+    group.add_argument('--init-method-xavier-uniform', action='store_true',
+                       help='Enable Xavier uniform parameter initialization')
+
+    return parser
+
+
+def _add_learning_rate_args(parser):
+    group = parser.add_argument_group(title='learning rate')
+
+    group.add_argument('--lr', type=float, default=None,
+                       help='Initial learning rate. Depending on decay style '
+                       'and initial warmup, the learing rate at each '
+                       'iteration would be different.')
+    group.add_argument('--lr-decay-style', type=str, default='linear',
+                       choices=['constant', 'linear', 'cosine'],
+                       help='Learning rate decay function.')
+    group.add_argument('--lr-decay-iters', type=int, default=None,
+                       help='number of iterations to decay learning rate over,'
+                       ' If None defaults to `--train-iters`')
+    group.add_argument('--lr-decay-samples', type=int, default=None,
+                       help='number of samples to decay learning rate over,'
+                       ' If None defaults to `--train-samples`')
+    group.add_argument('--lr-decay-tokens', type=int, default=None,
+                       help='number of tokens to decay learning rate over,'
+                       ' If not None will override iter/sample-based decay')
+    group.add_argument('--lr-warmup-fraction', type=float, default=None,
+                       help='fraction of lr-warmup-(iters/samples) to use '
+                       'for warmup (as a float)')
+    group.add_argument('--lr-warmup-iters', type=int, default=0,
+                       help='number of iterations to linearly warmup '
+                       'learning rate over.')
+    group.add_argument('--lr-warmup-samples', type=int, default=0,
+                       help='number of samples to linearly warmup '
+                       'learning rate over.')
+    group.add_argument('--warmup', type=int, default=None,
+                       help='Old lr warmup argument, do not use. Use one of the'
+                       '--lr-warmup-* arguments above')
+    group.add_argument('--min-lr', type=float, default=0.0,
+                       help='Minumum value for learning rate. The scheduler'
+                       'clip values below this threshold.')
+    group.add_argument('--override-lr-scheduler', action='store_true',
+                       help='Reset the values of the scheduler (learning rate,'
+                       'warmup iterations, minimum learning rate, maximum '
+                       'number of iterations, and decay style from input '
+                       'arguments and ignore values from checkpoints. Note'
+                       'that all the above values will be reset.')
+    group.add_argument('--use-checkpoint-lr-scheduler', action='store_true',
+                       help='Use checkpoint to set the values of the scheduler '
+                       '(learning rate, warmup iterations, minimum learning '
+                       'rate, maximum number of iterations, and decay style '
+                       'from checkpoint and ignore input arguments.')
+    group.add_argument('--universal-checkpoint', action='store_true',
+                        help='Loading a universal format checkpoint.')
+
+    return parser
+
+
+def _add_checkpointing_args(parser):
+    group = parser.add_argument_group(title='checkpointing')
+
+    group.add_argument('--save', type=str, default=None,
+                       help='Output directory to save checkpoints to.')
+    group.add_argument('--save-interval', type=int, default=None,
+                       help='Number of iterations between checkpoint saves.')
+    group.add_argument('--no-save-optim', action='store_true', default=None,
+                       help='Do not save current optimizer.')
+    group.add_argument('--no-save-rng', action='store_true', default=None,
+                       help='Do not save current rng state.')
+    group.add_argument('--load', type=str, default=None,
+                       help='Directory containing a model checkpoint.')
+    group.add_argument('--no-load-optim', action='store_true', default=None,
+                       help='Do not load optimizer when loading checkpoint.')
+    group.add_argument('--no-load-rng', action='store_true', default=None,
+                       help='Do not load rng state when loading checkpoint.')
+    group.add_argument('--finetune', action='store_true',
+                       help='Load model for finetuning. Do not load optimizer '
+                       'or rng state from checkpoint and set iteration to 0. '
+                       'Assumed when loading a release checkpoint.')
+
+    return parser
+
+
+def _add_mixed_precision_args(parser):
+    group = parser.add_argument_group(title='mixed precision')
+
+    group.add_argument('--fp16', action='store_true',
+                       help='Run model in fp16 mode.')
+    group.add_argument('--bf16', action='store_true',
+                       help='Run model in bfloat16 mode.')
+    group.add_argument('--loss-scale', type=float, default=None,
+                       help='Static loss scaling, positive power of 2 '
+                       'values can improve fp16 convergence. If None, dynamic'
+                       'loss scaling is used.')
+    group.add_argument('--initial-loss-scale', type=float, default=2**32,
+                       help='Initial loss-scale for dynamic loss scaling.')
+    group.add_argument('--min-loss-scale', type=float, default=1.0,
+                       help='Minimum loss scale for dynamic loss scale.')
+    group.add_argument('--loss-scale-window', type=float, default=1000,
+                       help='Window over which to raise/lower dynamic scale.')
+    group.add_argument('--hysteresis', type=int, default=2,
+                       help='hysteresis for dynamic loss scaling')
+    group.add_argument('--fp32-residual-connection', action='store_true',
+                       help='Move residual connections to fp32.')
+    group.add_argument('--no-query-key-layer-scaling', action='store_false',
+                       help='Do not scale Q * K^T by 1 / layer-number.',
+                       dest='apply_query_key_layer_scaling')
+    group.add_argument('--attention-softmax-in-fp32', action='store_true',
+                       help='Run attention masking and softmax in fp32. '
+                       'This flag is ignored unless '
+                       '--no-query-key-layer-scaling is specified.')
+    group.add_argument('--accumulate-allreduce-grads-in-fp32',
+                       action='store_true',
+                       help='Gradient accumulation and all-reduce in fp32.')
+    group.add_argument('--fp16-lm-cross-entropy', action='store_true',
+                       help='Move the cross entropy unreduced loss calculation'
+                       'for lm head to fp16.')
+
+    return parser
+
+
+def _add_distributed_args(parser):
+    group = parser.add_argument_group(title='distributed')
+
+    group.add_argument('--tensor-model-parallel-size', type=int, default=1,
+                       help='Degree of tensor model parallelism.')
+    group.add_argument('--pipeline-model-parallel-size', type=int, default=1,
+                       help='Degree of pipeline model parallelism.')
+    group.add_argument('--model-parallel-size', type=int, default=None,
+                       help='Old model parallel argument, do not use. Use '
+                       '--tensor-model-parallel-size instead.')
+    group.add_argument('--num-layers-per-virtual-pipeline-stage', type=int, default=None,
+                       help='Number of layers per virtual pipeline stage')
+    group.add_argument('--distributed-backend', default='nccl',
+                       choices=['nccl', 'gloo'],
+                       help='Which backend to use for distributed training.')
+    group.add_argument('--DDP-impl', default='local',
+                       choices=['local', 'torch'],
+                       help='which DistributedDataParallel implementation '
+                       'to use.')
+    group.add_argument('--use-contiguous-buffers-in-ddp', action='store_true',
+                       help='If set, use contiguous buffer in DDP. Note that '
+                       'this option only works woth local DDP.' )
+    group.add_argument('--no-scatter-gather-tensors-in-pipeline', action='store_false',
+                       help='Use scatter/gather to optimize communication of tensors in pipeline',
+                       dest='scatter_gather_tensors_in_pipeline')
+    group.add_argument('--local_rank', type=int, default=None,
+                       help='local rank passed from distributed launcher.')
+    group.add_argument('--lazy-mpu-init', type=bool, required=False,
+                       help='If set to True, initialize_megatron() '
+                       'skips DDP initialization and returns function to '
+                       'complete it instead.Also turns on '
+                       '--use-cpu-initialization flag. This is for '
+                       'external DDP manager.' )
+    group.add_argument('--use-cpu-initialization', action='store_true',
+                       default=None, help='If set, affine parallel weights '
+                       'initialization uses CPU' )
+    return parser
+
+
+def _add_validation_args(parser):
+    group = parser.add_argument_group(title='validation')
+
+    group.add_argument('--eval-iters', type=int, default=100,
+                       help='Number of iterations to run for evaluation'
+                       'validation/test for.')
+    group.add_argument('--eval-interval', type=int, default=1000,
+                       help='Interval between running evaluation on '
+                       'validation set.')
+
+    return parser
+
+
+def _add_data_args(parser):
+    group = parser.add_argument_group(title='data and dataloader')
+
+
+    # option 1 for data loading  (mutually exclusive with option2)
+    group.add_argument('--data-path', nargs='*', default=None,
+                       help='Path to the training dataset. Accepted format:'
+                       '1) a single data path, 2) multiple datasets in the'
+                       'form: dataset1-weight dataset1-path dataset2-weight '
+                       'dataset2-path ...')
+
+    group.add_argument('--split', type=str, default=None,
+                       help='Comma-separated list of proportions for training,'
+                       ' validation, and test split. For example the split '
+                       '`90,5,5` will use 90%% of data for training, 5%% for '
+                       'validation and 5%% for test.')
+
+    # option 2 for data loading (mutually exclusive with option1)
+
+    # helper class to parse the --xxx-weighted-split-paths
+    # note here two args are set: extra valid dataset paths and names
+    class parse_data_paths(argparse.Action):
+        def __call__(self, parser, args, values, option_string=None):
+
+            if option_string == "--train-weighted-split-paths":
+                assert len(values) == 1, 'Only 1 dataset group is allowed to'
+                'be passed for the argument --train-weighted-split-paths'
+
+            # make sure string given in the correct format
+            err_message = 'Each data group should be input on the following format'
+            '"GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2"'
+            'where START < END'
+            for v in values:
+                # each prefix consists several datasets separated by commas
+                prefix = ":".join(v.split(":")[1:]) # remove GIVEN_NAME
+                datasets = prefix.split(",")
+                # check if each dataset is formatted like `WEIGHT START:END PATH`
+                for d in datasets:
+                    assert len(d.split()) == 3, err_message
+                    start, end = d.split()[1].split(":")
+                    assert float(start) < float(end), err_message
+
+            names = [v.split(":")[0] for v in values]
+
+            prefixes = [":".join(v.split(":")[1:]).strip() for v in values]
+            weights = [[d.split()[0] for d in p.split(",")] for p in prefixes]
+            splits = [[d.split()[1] for d in p.split(",")] for p in prefixes]
+            paths = [[d.split()[2] for d in p.split(",")] for p in prefixes]
+
+            # # to keep consistency with Option 1 of data loading (through --data-path)
+            # #  paths will contain strings on the following form
+            # # "WEIGHTS1 PATH1 WEIGHTS2 PATH2 WEIGHTS3 PATH3" for each dataset group
+            # # while data will be parsed in additional arguments below
+            # paths_option1_style = []
+            # for p, w in zip(paths, weights):
+            #   paths_option1_style.append(" ".join([f"{w_i} {p_i}" for p_i, w_i in zip(p,w)]))
+            # setattr(args, self.dest, paths_option1_style)
+            setattr(args, self.dest, paths)
+            setattr(args, self.dest.replace("paths", "weights"), weights)
+            setattr(args, self.dest.replace("paths", "splits"), splits)
+            setattr(args, self.dest.replace("paths","names"), names)
+
+
+    group.add_argument('--train-weighted-split-paths', nargs='*', default=None,
+                    help='Weights, splits and paths to groups of datasets'
+                    'Accepted format: ONE dataset groups could be'
+                    'submitted in the following form between double quotes'
+                    '"GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2"'
+                    'e.g.: "NAME_ABC: 0.6 0:0.6 A, 0.3 0:1 B, 0.1 0:1 C" '
+                    'WEIGHT is used to up and down sample each dataset A,B,C in the group'
+                    'START:END indicates the split portion of the dataset',
+                    action=parse_data_paths)
+
+    group.add_argument('--valid-weighted-split-paths', nargs='*', default=None,
+                    help='Weights, splits and paths to groups of datasets'
+                    'Accepted format: one or many dataset groups could be'
+                    'submitted in the following form each between double quotes'
+                    '"GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2"'
+                    'e.g.: "NAME_ABC: 0.6 0.6:0.8 A, 0.3 0:1 B, 0.1 0:1 C" '
+                    '"NAME_CDE: 0.6 0.6:0.8 C, 0.3 0:1 D, 0.1 0:1 E" '
+                    'validation will be run on each of those groups independently',
+                    action=parse_data_paths)
+
+    group.add_argument('--test-weighted-split-paths', nargs='*', default=None,
+                    help='Weights, splits and paths to groups of datasets'
+                    'Accepted format: one or many dataset groups could be'
+                    'submitted in the following form each between double quotes'
+                    '"GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2"'
+                    'e.g.: "NAME_ABC: 0.6 0.6:0.8 A, 0.3 0:1 B, 0.1 0:1 C" '
+                    '"NAME_CDE: 0.6 0.6:0.8 C, 0.3 0:1 D, 0.1 0:1 E" '
+                    'test will be run on each of those groups independently',
+                    action=parse_data_paths)
+
+    class parse_data_paths_path(argparse.Action):
+        def __call__(self, parser, args, values, option_string=None):
+            expected_option_strings = ["--train-weighted-split-paths-path", "--valid-weighted-split-paths-path", "--test-weighted-split-paths-path"]
+            assert option_string in expected_option_strings, f"Expected {option_string} to be in {expected_option_strings}"
+
+            with open(values, "r") as fi:
+                lines = fi.readlines()
+                assert len(lines) == 1, f"Got multiple lines {len(lines)} instead of 1 expected"
+                assert lines[0][-2:] == "\"\n" and lines[0][0] == "\"", f"Invalid input format, got {lines}"
+                values = lines[0][1:-2].split("\" \"")
+                weighted_split_paths_dest = re.sub(r"_path$", "", self.dest)
+                weighted_split_paths_option = re.sub(r"-path$", "", self.option_strings[0])
+                setattr(args, weighted_split_paths_dest, values)
+                parse_data_paths(option_strings=[weighted_split_paths_option], dest=weighted_split_paths_dest)(parser, args, values, option_string=weighted_split_paths_option)
+
+
+    group.add_argument('--train-weighted-split-paths-path', type=str, action=parse_data_paths_path ,default=None)
+    group.add_argument('--valid-weighted-split-paths-path', type=str, action=parse_data_paths_path, default=None)
+    group.add_argument('--test-weighted-split-paths-path', type=str, action=parse_data_paths_path, default=None)
+
+    group.add_argument('--log-path', type=str, default=None,
+                       help='Path to the save arguments file.')
+    group.add_argument('--vocab-file', type=str, default=None,
+                       help='Path to the vocab file.')
+    group.add_argument('--merge-file', type=str, default=None,
+                       help='Path to the BPE merge file.')
+    group.add_argument('--vocab-extra-ids', type=int, default=0,
+                       help='Number of additional vocabulary tokens. '
+                            'They are used for span masking in the T5 model')
+    group.add_argument('--seq-length', type=int, default=None,
+                       help='Maximum sequence length to process.')
+    group.add_argument('--encoder-seq-length', type=int, default=None,
+                       help='Maximum encoder sequence length to process.'
+                       'This should be exclusive of --seq-length')
+    group.add_argument('--decoder-seq-length', type=int, default=None,
+                       help="Maximum decoder sequence length to process.")
+    group.add_argument('--retriever-seq-length', type=int, default=256,
+                       help='Maximum sequence length for the biencoder model '
+                        ' for retriever')
+    group.add_argument('--sample-rate', type=float, default=1.0,
+                       help='sample rate for training data. Supposed to be 0 '
+                            ' < sample_rate < 1')
+    group.add_argument('--mask-prob', type=float, default=0.15,
+                       help='Probability of replacing a token with mask.')
+    group.add_argument('--short-seq-prob', type=float, default=0.1,
+                       help='Probability of producing a short sequence.')
+    group.add_argument('--mmap-warmup', action='store_true',
+                       help='Warm up mmap files.')
+    group.add_argument('--num-workers', type=int, default=2,
+                       help="Dataloader number of workers.")
+    group.add_argument('--valid-num-workers', type=int, default=2,
+                       help="Dataloader number of workers for validation.")
+    group.add_argument('--tokenizer-type', type=str,
+                       default=None,
+                       choices=['BertWordPieceLowerCase',
+                                'BertWordPieceCase',
+                                'GPT2BPETokenizer',
+                                'PretrainedFromHF'],
+                       help='What type of tokenizer to use.')
+    group.add_argument("--tokenizer-name-or-path", type=str, default=None,
+                       help="Name or path of the huggingface tokenizer.")
+    group.add_argument('--data-impl', type=str, default='infer',
+                       choices=['lazy', 'cached', 'mmap', 'infer'],
+                       help='Implementation of indexed datasets.')
+    group.add_argument('--reset-position-ids', action='store_true',
+                       help='Reset posistion ids after end-of-document token.')
+    group.add_argument('--reset-attention-mask', action='store_true',
+                       help='Reset self attention maske after '
+                       'end-of-document token. Attention between tokens from different documents is null.')
+    group.add_argument('--eod-mask-loss', action='store_true',
+                       help='Mask loss for the end of document tokens.')
+    group.add_argument('--loss-on-targets-only', action='store_true',
+                       help='Mask loss on input sequence.')
+    group.add_argument('--reweight-loss-based-on-position-frequency', action="store_true",
+                       help='Some objectives require us to sample loss_mask. This might introduce bias towards '
+                       'specific positions. This option tries to un-bias the loss by reweighting loss on specific '
+                       'positions based on how frequently we train on that position.'
+                       'This is mostly used for prefix_lm training')
+    group.add_argument("--noise-density", type=float, default=None, help="Span corruption noise density")
+    group.add_argument("--mean-noise-span-length", type=int, default=None, help="Span corruption mean noise span length")
+
+
+    return parser
+
+
+def _add_autoresume_args(parser):
+    group = parser.add_argument_group(title='autoresume')
+
+    group.add_argument('--adlr-autoresume', action='store_true',
+                       help='Enable autoresume on adlr cluster.')
+    group.add_argument('--adlr-autoresume-interval', type=int, default=1000,
+                       help='Intervals over which check for autoresume'
+                       'termination signal')
+
+    return parser
+
+
+def _add_biencoder_args(parser):
+    group = parser.add_argument_group(title='biencoder')
+
+    # network size
+    group.add_argument('--ict-head-size', type=int, default=None,
+                       help='Size of block embeddings to be used in ICT and '
+                        'REALM (paper default: 128)')
+    group.add_argument('--biencoder-projection-dim', type=int, default=0,
+                       help='Size of projection head used in biencoder (paper'
+                        ' default: 128)')
+    group.add_argument('--biencoder-shared-query-context-model', action='store_true',
+                        help='Whether to share the parameters of the query '
+                        'and context models or not')
+
+    # checkpointing
+    group.add_argument('--ict-load', type=str, default=None,
+                       help='Directory containing an ICTBertModel checkpoint')
+    group.add_argument('--bert-load', type=str, default=None,
+                       help='Directory containing an BertModel checkpoint '
+                       '(needed to start ICT and REALM)')
+
+    # data
+    group.add_argument('--titles-data-path', type=str, default=None,
+                       help='Path to titles dataset used for ICT')
+    group.add_argument('--query-in-block-prob', type=float, default=0.1,
+                       help='Probability of keeping query in block for '
+                       'ICT dataset')
+    group.add_argument('--use-one-sent-docs', action='store_true',
+                       help='Whether to use one sentence documents in ICT')
+    group.add_argument('--evidence-data-path', type=str, default=None,
+                       help='Path to Wikipedia Evidence frm DPR paper')
+
+    # training
+    group.add_argument('--retriever-report-topk-accuracies', nargs='+', type=int,
+                        default=[], help="Which top-k accuracies to report "
+                        "(e.g. '1 5 20')")
+    group.add_argument('--retriever-score-scaling', action='store_true',
+                       help='Whether to scale retriever scores by inverse '
+                        'square root of hidden size')
+
+    # faiss index
+    group.add_argument('--block-data-path', type=str, default=None,
+                       help='Where to save/load BlockData to/from')
+    group.add_argument('--embedding-path', type=str, default=None,
+                       help='Where to save/load Open-Retrieval Embedding'
+                        ' data to/from')
+
+    # indexer
+    group.add_argument('--indexer-batch-size', type=int, default=128,
+                       help='How large of batches to use when doing indexing '
+                       'jobs')
+    group.add_argument('--indexer-log-interval', type=int, default=1000,
+                       help='After how many batches should the indexer '
+                       'report progress')
+    return parser
+
+
+def _add_vit_args(parser):
+    group = parser.add_argument_group(title="vit")
+
+    group.add_argument('--num-classes', type=int, default=1000,
+                       help='num of classes in vision classificaiton task')
+    group.add_argument('--img-dim', type=int, default=224,
+                       help='Image size for vision classification task')
+    group.add_argument('--num-channels', type=int, default=3,
+                       help='Number of channels in input image data')
+    group.add_argument('--patch-dim', type=int, default=16,
+                       help='patch dimension used in vit')
+
+    return parser
+
+
+def _add_zero_args(parser):
+    """Text generate arguments."""
+
+    group = parser.add_argument_group('ZeRO configurations', 'configurations')
+    group.add_argument("--zero-stage", type=int, default=1.0)
+    group.add_argument('--zero-reduce-scatter', action='store_true',
+                       help='Use reduce scatter if specified')
+    group.add_argument('--zero-contigious-gradients', action='store_true',
+                       help='Use contigious memory optimizaiton if specified')
+    group.add_argument("--zero-reduce-bucket-size", type=int, default=0.0)
+    group.add_argument("--zero-allgather-bucket-size", type=int, default=0.0)
+    group.add_argument('--remote-device', type=str, default='none', choices=['none', 'cpu', 'nvme'],
+                      help='Remote device for ZeRO-3 initialized parameters.')
+    group.add_argument('--use-pin-memory', action='store_true',
+                     help='Use pinned CPU memory for ZeRO-3 initialized model parameters.')
+    return parser
+
+def _add_memoryopt_args(parser):
+    """Memory optimization arguments."""
+
+    group = parser.add_argument_group('Memory optimizations', 'configurations')
+    group.add_argument("--scattered-embeddings", action='store_true',
+                       help='Save memory by scattering embedding activations. '
+                            'Introduces dropout differences across MP configurations.')
+    group.add_argument("--split-transformers", action='store_true',
+                       help='Save memory by splitting transformer layers into two parts, '
+                       'allowing for more frequent activation checkpoint savings.')
+    group.add_argument("--memory-centric-tiled-linear", action="store_true",
+                       help='Save memory by tiling with deepspeed.zero.TiledLinear.')
+    group.add_argument("--tile-factor", type=int, default=1,
+                       help='Make all linear layers the same size of [hidden/tile_factor, hidden/tile_factor]. '
+                            'Must be enabled with --memory-centric-tiled-linear. '
+                            'Example A: if tile_factor=1, the qkv layer [hidden, 3* hidden] would be converted into [1,3] tiles of size [hidden,hidden]. '
+                            'Example B: if tile_factor=2, the intermediate layer [4*hidden, hidden] will be converted into [8, 2] tiles of size [hidden/2, hidden/2]. '
+                            'Default is 1.')
+
+    return parser
+
+def _add_activation_checkpoint_args(parser):
+    group = parser.add_argument_group('Activation Checkpointing',
+                                      'Checkpointing Configurations')
+    group.add_argument('--deepspeed-activation-checkpointing', action='store_true',
+                       help='uses activation checkpointing from deepspeed')
+    group.add_argument('--partition-activations', action='store_true',
+                       help='partition Activations across GPUs before checkpointing.')
+    group.add_argument('--contigious-checkpointing', action='store_true',
+                       help='Contigious memory checkpointing for activatoins.')
+    group.add_argument('--checkpoint-in-cpu', action='store_true',
+                       help='Move the activation checkpoints to CPU.')
+    group.add_argument('--synchronize-each-layer', action='store_true',
+                       help='does a synchronize at the beginning and end of each checkpointed layer.')
+    group.add_argument('--profile-backward', action='store_true',
+                       help='Enables backward pass profiling for checkpointed layers.')
+    return parser

megatron-deepspeed_dtk22.10/megatron/checkpointing.py

+# coding=utf-8
+# Copyright (c) 2020, 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.
+
+"""Input/output checkpointing."""
+
+import os
+import random
+import sys
+import numpy as np
+
+import torch
+
+from megatron.global_vars import codecarbon_tracker_flush
+from megatron import (get_args,
+                      mpu,
+                      print_rank_0,
+                      update_num_microbatches,
+                      utils,
+                      get_tokenizer)
+from megatron.enums import PositionEmbeddingType
+
+_CHECKPOINT_VERSION = None
+
+def set_checkpoint_version(value):
+    global _CHECKPOINT_VERSION
+    if _CHECKPOINT_VERSION is not None:
+        assert _CHECKPOINT_VERSION == value, \
+            "checkpoint versions do not match"
+    _CHECKPOINT_VERSION = value
+
+def get_checkpoint_version():
+    global _CHECKPOINT_VERSION
+    return _CHECKPOINT_VERSION
+
+def check_checkpoint_args(checkpoint_args):
+    """Ensure fixed arguments for a model are the same for the input
+    arguments and the one retrieved from checkpoint."""
+    args = get_args()
+
+    def _compare(arg_name, old_arg_name=None):
+        if old_arg_name is not None:
+            checkpoint_value = getattr(checkpoint_args, old_arg_name)
+        else:
+            checkpoint_value = getattr(checkpoint_args, arg_name)
+        args_value = getattr(args, arg_name)
+        error_message = '{} value from checkpoint ({}) is not equal to the ' \
+                        'input argument value ({}).'.format(
+                            arg_name, checkpoint_value, args_value)
+        assert checkpoint_value == args_value, error_message
+
+    _compare('num_layers')
+    _compare('hidden_size')
+    _compare('num_attention_heads')
+    _compare('position_embedding_type')
+    # with alibi we can change `max_position_embeddings`
+    if args.position_embedding_type != PositionEmbeddingType.alibi:
+        _compare('max_position_embeddings')
+    if args.vocab_file:
+        _compare('make_vocab_size_divisible_by')
+        _compare('padded_vocab_size')
+        _compare('tokenizer_type')
+    if get_checkpoint_version() < 3.0:
+        _compare('tensor_model_parallel_size',
+                 old_arg_name='model_parallel_size')
+    if get_checkpoint_version() >= 3.0:
+        _compare('tensor_model_parallel_size')
+        _compare('pipeline_model_parallel_size')
+
+
+def ensure_directory_exists(filename):
+    """Build filename's path if it does not already exists."""
+    dirname = os.path.dirname(filename)
+    if not os.path.exists(dirname):
+        os.makedirs(dirname)
+
+
+def get_checkpoint_name(checkpoints_path, iteration,
+                        release=False):
+    """A unified checkpoint name."""
+    if release:
+        directory = 'release'
+    else:
+        directory = 'iter_{:07d}'.format(iteration)
+    # Use both the tensor and pipeline MP rank.
+    if mpu.get_pipeline_model_parallel_world_size() == 1:
+        return os.path.join(checkpoints_path, directory,
+                            'mp_rank_{:02d}'.format(
+                                mpu.get_tensor_model_parallel_rank()),
+                            'model_optim_rng.pt')
+    return os.path.join(checkpoints_path, directory,
+                        'mp_rank_{:02d}_{:03d}'.format(
+                            mpu.get_tensor_model_parallel_rank(),
+                            mpu.get_pipeline_model_parallel_rank()),
+                        'model_optim_rng.pt')
+
+
+def get_checkpoint_tracker_filename(checkpoints_path):
+    """Tracker file rescords the latest chckpoint during
+    training to restart from."""
+    return os.path.join(checkpoints_path, 'latest_checkpointed_iteration.txt')
+
+
+def save_checkpoint(iteration, model, optimizer, lr_scheduler):
+    """Save a model checkpoint."""
+    args = get_args()
+
+    # Only rank zero of the data parallel writes to the disk.
+    if not args.deepspeed:
+        model = utils.unwrap_model(model)
+
+    print_rank_0('saving checkpoint at iteration {:7d} to {}'.format(
+        iteration, args.save))
+
+    if not torch.distributed.is_initialized() or mpu.get_data_parallel_rank() == 0 \
+        or args.deepspeed:
+
+        # Arguments, iteration, and model.
+        state_dict = {}
+        state_dict['args'] = args
+        state_dict['checkpoint_version'] = 3.0
+        state_dict['iteration'] = iteration
+        state_dict['tokens'] = args.consumed_train_tokens
+        state_dict['checkpoint_info'] = _checkpoint_info()
+
+        # DeepSpeed saves the model/optimizer/scheduler
+        if not args.deepspeed:
+            if len(model) == 1:
+                state_dict['model'] = model[0].state_dict_for_save_checkpoint()
+            else:
+                for i in range(len(model)):
+                    mpu.set_virtual_pipeline_model_parallel_rank(i)
+                    state_dict['model%d' % i] = model[i].state_dict_for_save_checkpoint()
+
+            # Optimizer stuff.
+            if not args.no_save_optim:
+                if optimizer is not None:
+                    state_dict['optimizer'] = optimizer.state_dict()
+                if lr_scheduler is not None:
+                    state_dict['lr_scheduler'] = lr_scheduler.state_dict()
+
+        # RNG states.
+        if not args.no_save_rng:
+            state_dict['random_rng_state'] = random.getstate()
+            state_dict['np_rng_state'] = np.random.get_state()
+            state_dict['torch_rng_state'] = torch.get_rng_state()
+            state_dict['cuda_rng_state'] = torch.cuda.get_rng_state()
+            state_dict['rng_tracker_states'] \
+                = mpu.get_cuda_rng_tracker().get_states()
+
+        # Save.
+        checkpoint_name = get_checkpoint_name(args.save, iteration)
+        if not args.deepspeed:
+            ensure_directory_exists(checkpoint_name)
+            torch.save(state_dict, checkpoint_name)
+
+    if args.deepspeed:
+        # Saving is a collective communication
+        checkpoint_name = get_checkpoint_name(args.save, iteration)
+        # Trim off the filename and mp_rank_* directory.
+        for _ in range(3):
+            checkpoint_name = os.path.dirname(checkpoint_name)
+        model[0].save_checkpoint(checkpoint_name, client_state=state_dict)
+
+    # Wait so everyone is done (necessary)
+    if torch.distributed.is_initialized():
+        torch.distributed.barrier()
+
+    print_rank_0('  successfully saved checkpoint at iteration {:7d} to {}'.format(
+        iteration, args.save))
+
+    # And update the latest iteration
+    if (not args.deepspeed
+        and (not torch.distributed.is_initialized() or torch.distributed.get_rank() == 0)):
+        tracker_filename = get_checkpoint_tracker_filename(args.save)
+        with open(tracker_filename, 'w') as f:
+            f.write(str(iteration))
+
+    # Wait so everyone is done (not necessary)
+    if torch.distributed.is_initialized():
+        torch.distributed.barrier()
+
+    # since the code can be exited or aborted in various places we use the checkpoint saving as
+    # a save saving point for the codecarbon tracker. If the program doesn't run to its normal
+    # end, then only the data since the last saved checkpoint will be lost.
+    codecarbon_tracker_flush()
+
+def _transpose_first_dim(t, num_splits, num_splits_first, model):
+    input_shape = t.size()
+    # We use a self_attention module but the values extracted aren't
+    # specific to self attention so should work for cross attention as well
+    while hasattr(model, 'module'):
+        model = model.module
+    attention_module = model.language_model.encoder.layers[0].self_attention
+    hidden_size_per_attention_head = attention_module.hidden_size_per_attention_head
+    num_attention_heads_per_partition = attention_module.num_attention_heads_per_partition
+    if num_splits_first:
+        """[num_splits * np * hn, h]
+        -->(view) [num_splits, np, hn, h]
+        -->(tranpose) [np, num_splits, hn, h]
+        -->(view) [np * num_splits * hn, h] """
+
+        intermediate_shape = \
+            (num_splits, num_attention_heads_per_partition,
+             hidden_size_per_attention_head) + input_shape[1:]
+
+        t = t.view(*intermediate_shape)
+        t = t.transpose(0, 1).contiguous()
+    else:
+        """[np * hn * num_splits, h]
+        -->(view) [np, hn, num_splits, h]
+        -->(tranpose) [np, num_splits, hn, h]
+        -->(view) [np * num_splits * hn, h] """
+
+        intermediate_shape = \
+            (num_attention_heads_per_partition,
+             hidden_size_per_attention_head, num_splits) +\
+             input_shape[1:]
+
+        t = t.view(*intermediate_shape)
+        t = t.transpose(1, 2).contiguous()
+    t = t.view(*input_shape)
+
+    return t
+
+def fix_query_key_value_ordering(model, checkpoint_version):
+    """Fix up query/key/value matrix ordering if checkpoint
+    version is smaller than 2.0
+    """
+    if checkpoint_version < 2.0:
+        if isinstance(model, list):
+            assert len(model)==1
+            model = model[0]
+        for name, param in model.named_parameters():
+            if name.endswith(('.query_key_value.weight', '.query_key_value.bias')):
+                if checkpoint_version == 0:
+                    fixed_param = _transpose_first_dim(param.data, 3, True, model)
+                elif checkpoint_version == 1.0:
+                    fixed_param = _transpose_first_dim(param.data, 3, False, model)
+                else:
+                    print_rank_0(f"Invalid checkpoint version {checkpoint_version}.")
+                    sys.exit()
+                param.data.copy_(fixed_param)
+            if name.endswith(('.key_value.weight', '.key_value.bias')):
+                if checkpoint_version == 0:
+                    fixed_param = _transpose_first_dim(param.data, 2, True, model)
+                elif checkpoint_version == 1.0:
+                    fixed_param = _transpose_first_dim(param.data, 2, False, model)
+                else:
+                    print_rank_0(f"Invalid checkpoint version {checkpoint_version}.")
+                    sys.exit()
+                param.data.copy_(fixed_param)
+        print_rank_0(" succesfully fixed query-key-values ordering for"
+                    " checkpoint version {}".format(checkpoint_version))
+
+def load_checkpoint(model, optimizer, lr_scheduler, load_arg='load', strict=True):
+    """Load a model checkpoint and return the iteration.
+    strict (bool): whether to strictly enforce that the keys in
+        :attr:`state_dict` of the checkpoint match the names of
+        parameters and buffers in model.
+    """
+    args = get_args()
+    load_dir = getattr(args, load_arg)
+
+    if args.deepspeed:
+        load_optimizer_states = False if args.no_load_optim else True
+        loaded_dir, state_dict = model[0].load_checkpoint(load_dir, load_optimizer_states=load_optimizer_states)
+        if loaded_dir is None:
+            print_rank_0('WARNING: could not find the metadata file {} '.format(
+                load_dir))
+            print_rank_0('    will not load any checkpoints and will start from '
+                        'random')
+            return 0
+        release = False
+    else:
+        model = utils.unwrap_model(model)
+
+        # Read the tracker file and set the iteration.
+        tracker_filename = get_checkpoint_tracker_filename(load_dir)
+
+        # If no tracker file, return iretation zero.
+        if not os.path.isfile(tracker_filename):
+            print_rank_0('WARNING: could not find the metadata file {} '.format(
+                tracker_filename))
+            print_rank_0('    will not load any checkpoints and will start from '
+                        'random')
+            return 0
+
+        # Otherwise, read the tracker file and either set the iteration or
+        # mark it as a release checkpoint.
+        iteration = 0
+        release = False
+        with open(tracker_filename, 'r') as f:
+            metastring = f.read().strip()
+            try:
+                iteration = int(metastring)
+            except ValueError:
+                release = metastring == 'release'
+                if not release:
+                    print_rank_0('ERROR: Invalid metadata file {}. Exiting'.format(
+                        tracker_filename))
+                    sys.exit()
+
+        assert iteration > 0 or release, 'error parsing metadata file {}'.format(
+            tracker_filename)
+
+        # Checkpoint.
+        checkpoint_name = get_checkpoint_name(load_dir, iteration, release)
+        print_rank_0(f' loading checkpoint from {args.load} at iteration {iteration}')
+
+        # Load the checkpoint.
+        try:
+            state_dict = torch.load(checkpoint_name, map_location='cpu')
+        except ModuleNotFoundError:
+            from megatron.fp16_deprecated import loss_scaler
+            # For backward compatibility.
+            print_rank_0(' > deserializing using the old code structure ...')
+            sys.modules['fp16.loss_scaler'] = sys.modules[
+                'megatron.fp16_deprecated.loss_scaler']
+            sys.modules['megatron.fp16.loss_scaler'] = sys.modules[
+                'megatron.fp16_deprecated.loss_scaler']
+            state_dict = torch.load(checkpoint_name, map_location='cpu')
+            sys.modules.pop('fp16.loss_scaler', None)
+            sys.modules.pop('megatron.fp16.loss_scaler', None)
+        except BaseException as e:
+            print_rank_0('could not load the checkpoint')
+            print_rank_0(e)
+            sys.exit()
+
+    # set checkpoint version
+    set_checkpoint_version(state_dict.get('checkpoint_version', 0))
+
+    # Set iteration.
+    if args.finetune or release:
+        iteration = 0
+    else:
+        try:
+            iteration = state_dict['iteration']
+            if 'tokens' in state_dict:
+                args.consumed_train_tokens = state_dict['tokens']
+        except KeyError:
+            try:  # Backward compatible with older checkpoints
+                iteration = state_dict['total_iters']
+            except KeyError:
+                print_rank_0('A metadata file exists but unable to load '
+                             'iteration from checkpoint {}, exiting'.format(
+                                 checkpoint_name))
+                sys.exit()
+
+    # Check arguments.
+    assert args.consumed_train_samples == 0
+    assert args.consumed_valid_samples == 0
+    if 'args' in state_dict:
+        checkpoint_args = state_dict['args']
+        if not args.universal_checkpoint:
+            check_checkpoint_args(checkpoint_args)
+        args.consumed_train_samples = getattr(checkpoint_args,
+                                              'consumed_train_samples', 0)
+        update_num_microbatches(consumed_samples=args.consumed_train_samples)
+        args.consumed_valid_samples = getattr(checkpoint_args,
+                                              'consumed_valid_samples', 0)
+        args.gigaflos_no_embeds = getattr(checkpoint_args,
+                                          'gigaflos_no_embeds', 0)
+    else:
+        print_rank_0('could not find arguments in the checkpoint ...')
+
+    # Model.
+    if not args.deepspeed:
+        if len(model) == 1:
+            model[0].load_state_dict(state_dict['model'], strict=strict)
+        else:
+            for i in range(len(model)):
+                mpu.set_virtual_pipeline_model_parallel_rank(i)
+                model[i].load_state_dict(state_dict['model%d' % i], strict=strict)
+
+    # Fix up query/key/value matrix ordering if needed
+    checkpoint_version = get_checkpoint_version()
+    print_rank_0(f' checkpoint version {checkpoint_version}')
+    fix_query_key_value_ordering(model, checkpoint_version)
+
+    # Optimizer.
+    if not args.deepspeed:
+        if not release and not args.finetune and not args.no_load_optim:
+            try:
+                if optimizer is not None:
+                    optimizer.load_state_dict(state_dict['optimizer'])
+                if lr_scheduler is not None:
+                    lr_scheduler.load_state_dict(state_dict['lr_scheduler'])
+            except KeyError:
+                print_rank_0('Unable to load optimizer from checkpoint {}. '
+                            'Specify --no-load-optim or --finetune to prevent '
+                            'attempting to load the optimizer state, '
+                            'exiting ...'.format(checkpoint_name))
+                sys.exit()
+
+    # rng states.
+    if not release and not args.finetune and not args.no_load_rng:
+        try:
+            random.setstate(state_dict['random_rng_state'])
+            np.random.set_state(state_dict['np_rng_state'])
+            torch.set_rng_state(state_dict['torch_rng_state'])
+            torch.cuda.set_rng_state(state_dict['cuda_rng_state'])
+            # Check for empty states array
+            if not state_dict['rng_tracker_states']:
+                raise KeyError
+            mpu.get_cuda_rng_tracker().set_states(
+                state_dict['rng_tracker_states'])
+        except KeyError:
+            print_rank_0('Unable to load rng state from checkpoint {}. '
+                         'Specify --no-load-rng or --finetune to prevent '
+                         'attempting to load the rng state, '
+                         'exiting ...'.format(checkpoint_name))
+            sys.exit()
+
+    # Some utilities want to load a checkpoint without distributed being initialized
+    if torch.distributed.is_initialized():
+        torch.distributed.barrier()
+
+    print_rank_0(f'  successfully loaded checkpoint from {args.load} '
+                 f'at iteration {iteration}')
+
+    return iteration
+
+
+def load_biencoder_checkpoint(model, only_query_model=False,
+        only_context_model=False, custom_load_path=None):
+    """
+    selectively load retrieval models for indexing/retrieving
+    from saved checkpoints
+    """
+
+    args = get_args()
+
+    model = utils.unwrap_model(model)
+
+    load_path = custom_load_path if custom_load_path is not None else args.load
+
+    tracker_filename = get_checkpoint_tracker_filename(load_path)
+    with open(tracker_filename, 'r') as f:
+        iteration = int(f.read().strip())
+
+    checkpoint_name = get_checkpoint_name(load_path, iteration, False)
+    if mpu.get_data_parallel_rank() == 0:
+        print('global rank {} is loading checkpoint {}'.format(
+            torch.distributed.get_rank(), checkpoint_name))
+
+    state_dict = torch.load(checkpoint_name, map_location='cpu')
+    ret_state_dict = state_dict['model']
+
+    if only_query_model:
+        ret_state_dict.pop('context_model')
+    if only_context_model:
+        ret_state_dict.pop('query_model')
+
+    assert len(model) == 1
+    model[0].load_state_dict(ret_state_dict)
+    torch.distributed.barrier()
+
+    if mpu.get_data_parallel_rank() == 0:
+        print(' successfully loaded {}'.format(checkpoint_name))
+
+    return model
+
+
+def _checkpoint_info():
+    args = get_args()
+    tokenizer = get_tokenizer()
+
+    return {
+        "padded_vocab_size": args.padded_vocab_size,
+        "original_vocab_size": tokenizer.vocab_size,
+    }
\ No newline at end of file

megatron-deepspeed_dtk22.10/megatron/data/Makefile

+
+PYTHON3CONFIG := $(shell command -v python3-config 2> /dev/null)
+
+ifndef PYTHON3CONFIG
+    $(error "python3-config is not available. Please install it. It may be in a python-dev or another package")
+endif
+
+CXXFLAGS += -O3 -Wall -shared -std=c++11 -fPIC -fdiagnostics-color
+CPPFLAGS += $(shell python3 -m pybind11 --includes)
+LIBNAME = helpers
+LIBEXT = $(shell python3-config --extension-suffix)
+
+default: $(LIBNAME)$(LIBEXT)
+
+%$(LIBEXT): %.cpp
+	$(CXX) $(CXXFLAGS) $(CPPFLAGS) $< -o $@

megatron-deepspeed_dtk22.10/megatron/data/__init__.py

+from . import indexed_dataset

megatron-deepspeed_dtk22.10/megatron/data/autoaugment.py

+"""AutoAugment data augmentation policy for ImageNet.
+
+-- Begin license text.
+
+MIT License
+
+Copyright (c) 2018 Philip Popien
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+-- End license text.
+
+Code adapted from https://github.com/DeepVoltaire/AutoAugment.
+
+This module implements the fixed AutoAugment data augmentation policy for ImageNet provided in
+Appendix A, Table 9 of reference [1]. It does not include any of the search code for augmentation
+policies.
+
+Reference:
+[1] https://arxiv.org/abs/1805.09501
+"""
+
+import random
+
+import numpy as np
+from PIL import Image
+from PIL import ImageEnhance
+from PIL import ImageOps
+
+_MAX_LEVEL = 10  # Maximum integer strength of an augmentation, if applicable.
+
+
+class ImageNetPolicy:
+    """Definition of an ImageNetPolicy.
+
+    Implements a fixed AutoAugment data augmentation policy targeted at
+    ImageNet training by randomly applying at runtime one of the 25 pre-defined
+    data augmentation sub-policies provided in Reference [1].
+
+    Usage example as a Pytorch Transform:
+    >>> transform=transforms.Compose([transforms.Resize(256),
+    >>>                               ImageNetPolicy(),
+    >>>                               transforms.ToTensor()])
+    """
+
+    def __init__(self, fillcolor=(128, 128, 128)):
+        """Initialize an ImageNetPolicy.
+
+        Args:
+            fillcolor (tuple): RGB color components of the color to be used for
+            filling when needed (default: (128, 128, 128), which
+            corresponds to gray).
+        """
+        # Instantiate a list of sub-policies.
+        # Each entry of the list is a SubPolicy which consists of
+        # two augmentation operations,
+        # each of those parametrized as operation, probability, magnitude.
+        # Those two operations are applied sequentially on the image upon call.
+        self.policies = [
+            SubPolicy("posterize", 0.4, 8, "rotate", 0.6, 9, fillcolor),
+            SubPolicy("solarize", 0.6, 5, "autocontrast", 0.6, 5, fillcolor),
+            SubPolicy("equalize", 0.8, 8, "equalize", 0.6, 3, fillcolor),
+            SubPolicy("posterize", 0.6, 7, "posterize", 0.6, 6, fillcolor),
+            SubPolicy("equalize", 0.4, 7, "solarize", 0.2, 4, fillcolor),
+            SubPolicy("equalize", 0.4, 4, "rotate", 0.8, 8, fillcolor),
+            SubPolicy("solarize", 0.6, 3, "equalize", 0.6, 7, fillcolor),
+            SubPolicy("posterize", 0.8, 5, "equalize", 1.0, 2, fillcolor),
+            SubPolicy("rotate", 0.2, 3, "solarize", 0.6, 8, fillcolor),
+            SubPolicy("equalize", 0.6, 8, "posterize", 0.4, 6, fillcolor),
+            SubPolicy("rotate", 0.8, 8, "color", 0.4, 0, fillcolor),
+            SubPolicy("rotate", 0.4, 9, "equalize", 0.6, 2, fillcolor),
+            SubPolicy("equalize", 0.0, 7, "equalize", 0.8, 8, fillcolor),
+            SubPolicy("invert", 0.6, 4, "equalize", 1.0, 8, fillcolor),
+            SubPolicy("color", 0.6, 4, "contrast", 1.0, 8, fillcolor),
+            SubPolicy("rotate", 0.8, 8, "color", 1.0, 2, fillcolor),
+            SubPolicy("color", 0.8, 8, "solarize", 0.8, 7, fillcolor),
+            SubPolicy("sharpness", 0.4, 7, "invert", 0.6, 8, fillcolor),
+            SubPolicy("shearX", 0.6, 5, "equalize", 1.0, 9, fillcolor),
+            SubPolicy("color", 0.4, 0, "equalize", 0.6, 3, fillcolor),
+            SubPolicy("equalize", 0.4, 7, "solarize", 0.2, 4, fillcolor),
+            SubPolicy("solarize", 0.6, 5, "autocontrast", 0.6, 5, fillcolor),
+            SubPolicy("invert", 0.6, 4, "equalize", 1.0, 8, fillcolor),
+            SubPolicy("color", 0.6, 4, "contrast", 1.0, 8, fillcolor),
+            SubPolicy("equalize", 0.8, 8, "equalize", 0.6, 3, fillcolor),
+        ]
+
+    def __call__(self, img):
+        """Define call method for ImageNetPolicy class."""
+        policy_idx = random.randint(0, len(self.policies) - 1)
+        return self.policies[policy_idx](img)
+
+    def __repr__(self):
+        """Define repr method for ImageNetPolicy class."""
+        return "ImageNetPolicy"
+
+
+class SubPolicy:
+    """Definition of a SubPolicy.
+
+    A SubPolicy consists of two augmentation operations,
+    each of those parametrized as operation, probability, magnitude.
+    The two operations are applied sequentially on the image upon call.
+    """
+
+    def __init__(
+        self,
+        operation1,
+        probability1,
+        magnitude_idx1,
+        operation2,
+        probability2,
+        magnitude_idx2,
+        fillcolor,
+    ):
+        """Initialize a SubPolicy.
+
+        Args:
+            operation1 (str): Key specifying the first augmentation operation.
+            There are fourteen key values altogether (see supported_ops below
+            listing supported operations). probability1 (float): Probability
+            within [0., 1.] of applying the first augmentation operation.
+            magnitude_idx1 (int): Integer specifiying the strength of the first
+            operation as an index further used to derive the magnitude from a
+            range of possible values.
+            operation2 (str): Key specifying the second augmentation operation.
+            probability2 (float): Probability within [0., 1.] of applying the
+            second augmentation operation.
+            magnitude_idx2 (int): Integer specifiying the strength of the
+            second operation as an index further used to derive the magnitude
+            from a range of possible values.
+            fillcolor (tuple): RGB color components of the color to be used for
+            filling.
+        Returns:
+        """
+        # List of supported operations for operation1 and operation2.
+        supported_ops = [
+            "shearX",
+            "shearY",
+            "translateX",
+            "translateY",
+            "rotate",
+            "color",
+            "posterize",
+            "solarize",
+            "contrast",
+            "sharpness",
+            "brightness",
+            "autocontrast",
+            "equalize",
+            "invert",
+        ]
+        assert (operation1 in supported_ops) and (
+            operation2 in supported_ops
+        ), "SubPolicy:one of oper1 or oper2 refers to an unsupported operation."
+
+        assert (
+            0.0 <= probability1 <= 1.0 and 0.0 <= probability2 <= 1.0
+        ), "SubPolicy: prob1 and prob2 should be within [0., 1.]."
+
+        assert (
+            isinstance(magnitude_idx1, int) and 0 <= magnitude_idx1 <= 10
+        ), "SubPolicy: idx1 should be specified as an integer within [0, 10]."
+
+        assert (
+            isinstance(magnitude_idx2, int) and 0 <= magnitude_idx2 <= 10
+        ), "SubPolicy: idx2 should be specified as an integer within [0, 10]."
+
+        # Define a dictionary where each key refers to a specific type of
+        # augmentation and the corresponding value is a range of ten possible
+        # magnitude values for that augmentation.
+        num_levels = _MAX_LEVEL + 1
+        ranges = {
+            "shearX": np.linspace(0, 0.3, num_levels),
+            "shearY": np.linspace(0, 0.3, num_levels),
+            "translateX": np.linspace(0, 150 / 331, num_levels),
+            "translateY": np.linspace(0, 150 / 331, num_levels),
+            "rotate": np.linspace(0, 30, num_levels),
+            "color": np.linspace(0.0, 0.9, num_levels),
+            "posterize": np.round(np.linspace(8, 4, num_levels), 0).astype(
+                np.int
+            ),
+            "solarize": np.linspace(256, 0, num_levels),  # range [0, 256]
+            "contrast": np.linspace(0.0, 0.9, num_levels),
+            "sharpness": np.linspace(0.0, 0.9, num_levels),
+            "brightness": np.linspace(0.0, 0.9, num_levels),
+            "autocontrast": [0]
+            * num_levels,  # This augmentation doesn't use magnitude parameter.
+            "equalize": [0]
+            * num_levels,  # This augmentation doesn't use magnitude parameter.
+            "invert": [0]
+            * num_levels,  # This augmentation doesn't use magnitude parameter.
+        }
+
+        def rotate_with_fill(img, magnitude):
+            """Define rotation transformation with fill.
+
+            The input image is first rotated, then it is blended together with
+            a gray mask of the same size. Note that fillcolor as defined
+            elsewhere in this module doesn't apply here.
+
+            Args:
+                magnitude (float): rotation angle in degrees.
+            Returns:
+                rotated_filled (PIL Image): rotated image with gray filling for
+                disoccluded areas unveiled by the rotation.
+            """
+            rotated = img.convert("RGBA").rotate(magnitude)
+            rotated_filled = Image.composite(
+                rotated, Image.new("RGBA", rotated.size, (128,) * 4), rotated
+            )
+            return rotated_filled.convert(img.mode)
+
+        # Define a dictionary of augmentation functions where each key refers
+        # to a specific type of augmentation and the corresponding value defines
+        # the augmentation itself using a lambda function.
+        # pylint: disable=unnecessary-lambda
+        func_dict = {
+            "shearX": lambda img, magnitude: img.transform(
+                img.size,
+                Image.AFFINE,
+                (1, magnitude * random.choice([-1, 1]), 0, 0, 1, 0),
+                Image.BICUBIC,
+                fillcolor=fillcolor,
+            ),
+            "shearY": lambda img, magnitude: img.transform(
+                img.size,
+                Image.AFFINE,
+                (1, 0, 0, magnitude * random.choice([-1, 1]), 1, 0),
+                Image.BICUBIC,
+                fillcolor=fillcolor,
+            ),
+            "translateX": lambda img, magnitude: img.transform(
+                img.size,
+                Image.AFFINE,
+                (
+                    1,
+                    0,
+                    magnitude * img.size[0] * random.choice([-1, 1]),
+                    0,
+                    1,
+                    0,
+                ),
+                fillcolor=fillcolor,
+            ),
+            "translateY": lambda img, magnitude: img.transform(
+                img.size,
+                Image.AFFINE,
+                (
+                    1,
+                    0,
+                    0,
+                    0,
+                    1,
+                    magnitude * img.size[1] * random.choice([-1, 1]),
+                ),
+                fillcolor=fillcolor,
+            ),
+            "rotate": lambda img, magnitude: rotate_with_fill(img, magnitude),
+            "color": lambda img, magnitude: ImageEnhance.Color(img).enhance(
+                1 + magnitude * random.choice([-1, 1])
+            ),
+            "posterize": lambda img, magnitude: ImageOps.posterize(
+                img, magnitude
+            ),
+            "solarize": lambda img, magnitude: ImageOps.solarize(
+                img, magnitude
+            ),
+            "contrast": lambda img, magnitude: ImageEnhance.Contrast(
+                img
+            ).enhance(1 + magnitude * random.choice([-1, 1])),
+            "sharpness": lambda img, magnitude: ImageEnhance.Sharpness(
+                img
+            ).enhance(1 + magnitude * random.choice([-1, 1])),
+            "brightness": lambda img, magnitude: ImageEnhance.Brightness(
+                img
+            ).enhance(1 + magnitude * random.choice([-1, 1])),
+            "autocontrast": lambda img, magnitude: ImageOps.autocontrast(img),
+            "equalize": lambda img, magnitude: ImageOps.equalize(img),
+            "invert": lambda img, magnitude: ImageOps.invert(img),
+        }
+
+        # Store probability, function and magnitude of the first augmentation
+        # for the sub-policy.
+        self.probability1 = probability1
+        self.operation1 = func_dict[operation1]
+        self.magnitude1 = ranges[operation1][magnitude_idx1]
+
+        # Store probability, function and magnitude of the second augmentation
+        # for the sub-policy.
+        self.probability2 = probability2
+        self.operation2 = func_dict[operation2]
+        self.magnitude2 = ranges[operation2][magnitude_idx2]
+
+    def __call__(self, img):
+        """Define call method for SubPolicy class."""
+        # Randomly apply operation 1.
+        if random.random() < self.probability1:
+            img = self.operation1(img, self.magnitude1)
+
+        # Randomly apply operation 2.
+        if random.random() < self.probability2:
+            img = self.operation2(img, self.magnitude2)
+
+        return img

megatron-deepspeed_dtk22.10/megatron/data/bert_dataset.py

+# coding=utf-8
+# Copyright (c) 2020, 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.
+
+"""BERT Style dataset."""
+
+import numpy as np
+import torch
+
+from megatron import (
+    get_args,
+    get_tokenizer,
+    mpu,
+    print_rank_0
+)
+from megatron.data.dataset_utils import (
+    get_samples_mapping,
+    get_a_and_b_segments,
+    truncate_segments,
+    create_tokens_and_tokentypes,
+    create_masked_lm_predictions
+)
+
+
+class BertDataset(torch.utils.data.Dataset):
+
+    def __init__(self, name, indexed_dataset, data_prefix,
+                 num_epochs, max_num_samples, masked_lm_prob,
+                 max_seq_length, short_seq_prob, seed, binary_head):
+
+        # Params to store.
+        self.name = name
+        self.seed = seed
+        self.masked_lm_prob = masked_lm_prob
+        self.max_seq_length = max_seq_length
+        self.binary_head = binary_head
+
+        # Dataset.
+        self.indexed_dataset = indexed_dataset
+
+        # Build the samples mapping.
+        self.samples_mapping = get_samples_mapping(self.indexed_dataset,
+                                                   data_prefix,
+                                                   num_epochs,
+                                                   max_num_samples,
+                                                   self.max_seq_length - 3, # account for added tokens
+                                                   short_seq_prob,
+                                                   self.seed,
+                                                   self.name,
+                                                   self.binary_head)
+
+        # Vocab stuff.
+        tokenizer = get_tokenizer()
+        self.vocab_id_list = list(tokenizer.inv_vocab.keys())
+        self.vocab_id_to_token_dict = tokenizer.inv_vocab
+        self.cls_id = tokenizer.cls
+        self.sep_id = tokenizer.sep
+        self.mask_id = tokenizer.mask
+        self.pad_id = tokenizer.pad
+
+    def __len__(self):
+        return self.samples_mapping.shape[0]
+
+    def __getitem__(self, idx):
+        start_idx, end_idx, seq_length = self.samples_mapping[idx]
+        sample = [self.indexed_dataset[i] for i in range(start_idx, end_idx)]
+        # Note that this rng state should be numpy and not python since
+        # python randint is inclusive whereas the numpy one is exclusive.
+        # We % 2**32 since numpy requres the seed to be between 0 and 2**32 - 1
+        np_rng = np.random.RandomState(seed=((self.seed + idx) % 2**32))
+        return build_training_sample(sample, seq_length,
+                                     self.max_seq_length,  # needed for padding
+                                     self.vocab_id_list,
+                                     self.vocab_id_to_token_dict,
+                                     self.cls_id, self.sep_id,
+                                     self.mask_id, self.pad_id,
+                                     self.masked_lm_prob, np_rng,
+                                     self.binary_head)
+
+
+
+
+def build_training_sample(sample,
+                          target_seq_length, max_seq_length,
+                          vocab_id_list, vocab_id_to_token_dict,
+                          cls_id, sep_id, mask_id, pad_id,
+                          masked_lm_prob, np_rng, binary_head):
+    """Biuld training sample.
+
+    Arguments:
+        sample: A list of sentences in which each sentence is a list token ids.
+        target_seq_length: Desired sequence length.
+        max_seq_length: Maximum length of the sequence. All values are padded to
+            this length.
+        vocab_id_list: List of vocabulary ids. Used to pick a random id.
+        vocab_id_to_token_dict: A dictionary from vocab ids to text tokens.
+        cls_id: Start of example id.
+        sep_id: Separator id.
+        mask_id: Mask token id.
+        pad_id: Padding token id.
+        masked_lm_prob: Probability to mask tokens.
+        np_rng: Random number genenrator. Note that this rng state should be
+              numpy and not python since python randint is inclusive for
+              the opper bound whereas the numpy one is exclusive.
+    """
+
+    if binary_head:
+        # We assume that we have at least two sentences in the sample
+        assert len(sample) > 1
+    assert target_seq_length <= max_seq_length
+
+    # Divide sample into two segments (A and B).
+    if binary_head:
+        tokens_a, tokens_b, is_next_random = get_a_and_b_segments(sample,
+                                                                  np_rng)
+    else:
+        tokens_a = []
+        for j in range(len(sample)):
+            tokens_a.extend(sample[j])
+        tokens_b = []
+        is_next_random = False
+
+    # Truncate to `target_sequence_length`.
+    max_num_tokens = target_seq_length
+    truncated = truncate_segments(tokens_a, tokens_b, len(tokens_a),
+                                  len(tokens_b), max_num_tokens, np_rng)
+
+    # Build tokens and toketypes.
+    tokens, tokentypes = create_tokens_and_tokentypes(tokens_a, tokens_b,
+                                                      cls_id, sep_id)
+
+    # Masking.
+    max_predictions_per_seq = masked_lm_prob * max_num_tokens
+    (tokens, masked_positions, masked_labels, _, _) = create_masked_lm_predictions(
+        tokens, vocab_id_list, vocab_id_to_token_dict, masked_lm_prob,
+        cls_id, sep_id, mask_id, max_predictions_per_seq, np_rng)
+
+    # Padding.
+    tokens_np, tokentypes_np, labels_np, padding_mask_np, loss_mask_np \
+        = pad_and_convert_to_numpy(tokens, tokentypes, masked_positions,
+                                   masked_labels, pad_id, max_seq_length)
+
+    train_sample = {
+        'text': tokens_np,
+        'types': tokentypes_np,
+        'labels': labels_np,
+        'is_random': int(is_next_random),
+        'loss_mask': loss_mask_np,
+        'padding_mask': padding_mask_np,
+        'truncated': int(truncated)}
+    return train_sample
+
+
+def pad_and_convert_to_numpy(tokens, tokentypes, masked_positions,
+                             masked_labels, pad_id, max_seq_length):
+    """Pad sequences and convert them to numpy."""
+
+    # Some checks.
+    num_tokens = len(tokens)
+    padding_length = max_seq_length - num_tokens
+    assert padding_length >= 0
+    assert len(tokentypes) == num_tokens
+    assert len(masked_positions) == len(masked_labels)
+
+    # Tokens and token types.
+    filler = [pad_id] * padding_length
+    tokens_np = np.array(tokens + filler, dtype=np.int64)
+    tokentypes_np = np.array(tokentypes + filler, dtype=np.int64)
+
+    # Padding mask.
+    padding_mask_np = np.array([1] * num_tokens + [0] * padding_length,
+                               dtype=np.int64)
+
+    # Lables and loss mask.
+    labels = [-1] * max_seq_length
+    loss_mask = [0] * max_seq_length
+    for i in range(len(masked_positions)):
+        assert masked_positions[i] < num_tokens
+        labels[masked_positions[i]] = masked_labels[i]
+        loss_mask[masked_positions[i]] = 1
+    labels_np = np.array(labels, dtype=np.int64)
+    loss_mask_np = np.array(loss_mask, dtype=np.int64)
+
+    return tokens_np, tokentypes_np, labels_np, padding_mask_np, loss_mask_np

megatron-deepspeed_dtk22.10/megatron/data/biencoder_dataset_utils.py

+import os
+import time
+
+import numpy as np
+import torch
+
+from megatron import get_args, get_tokenizer, mpu, print_rank_0
+from megatron.data.dataset_utils import create_masked_lm_predictions, \
+                                            pad_and_convert_to_numpy
+from megatron.data.data_samplers import MegatronPretrainingSampler
+
+def make_attention_mask(source_block, target_block):
+    """
+    Returns a 2-dimensional (2-D) attention mask
+    :param source_block: 1-D array
+    :param target_block: 1-D array
+    """
+    mask = (target_block[None, :] >= 1) * (source_block[:, None] >= 1)
+    mask = mask.astype(np.int64)
+    # (source_length, target_length)
+    return mask
+
+def get_one_epoch_dataloader(dataset, micro_batch_size=None):
+    """Specifically one epoch to be used in an indexing job."""
+    args = get_args()
+
+    if micro_batch_size is None:
+        micro_batch_size = args.micro_batch_size
+    num_workers = args.num_workers
+
+    # Use megatron's sampler with consumed samples set to 0 as
+    # this is only for evaluation and don't intend to resume half way.
+    # Also, set the drop last to false as don't intend to remove
+    # the last batch
+    batch_sampler = MegatronPretrainingSampler(
+        total_samples=len(dataset),
+        consumed_samples=0,
+        micro_batch_size=args.micro_batch_size,
+        data_parallel_rank=mpu.get_data_parallel_rank(),
+        data_parallel_size=mpu.get_data_parallel_world_size(),
+        drop_last=False)
+
+    return torch.utils.data.DataLoader(dataset,
+                                       batch_sampler=batch_sampler,
+                                       num_workers=num_workers,
+                                       pin_memory=True)
+
+
+def get_ict_batch(data_iterator):
+    # Items and their type.
+    keys = ['query_tokens', 'query_mask',
+            'context_tokens', 'context_mask', 'block_data']
+    datatype = torch.int64
+
+    # Broadcast data.
+    if data_iterator is None:
+        data = None
+    else:
+        data = next(data_iterator)
+    data_b = mpu.broadcast_data(keys, data, datatype)
+
+    # Unpack.
+    query_tokens = data_b['query_tokens'].long()
+    query_mask = data_b['query_mask'] < 0.5
+    context_tokens = data_b['context_tokens'].long()
+    context_mask = data_b['context_mask'] < 0.5
+    block_indices = data_b['block_data'].long()
+
+    return query_tokens, query_mask,\
+           context_tokens, context_mask, block_indices
+
+
+def join_str_list(str_list):
+    """Join a list of strings, handling spaces appropriately"""
+    result = ""
+    for s in str_list:
+        if s.startswith("##"):
+            result += s[2:]
+        else:
+            result += " " + s
+    return result
+
+
+class BlockSampleData(object):
+    """A struct for fully describing a fixed-size block of data as used in REALM
+
+    :param start_idx: for first sentence of the block
+    :param end_idx: for last sentence of the block (may be partially truncated in sample construction)
+    :param doc_idx: the index of the document from which the block comes in the original indexed dataset
+    :param block_idx: a unique integer identifier given to every block.
+    """
+    def __init__(self, start_idx, end_idx, doc_idx, block_idx):
+        self.start_idx = start_idx
+        self.end_idx = end_idx
+        self.doc_idx = doc_idx
+        self.block_idx = block_idx
+
+    def as_array(self):
+        return np.array([self.start_idx, self.end_idx, self.doc_idx, self.block_idx]).astype(np.int64)
+
+    def as_tuple(self):
+        return self.start_idx, self.end_idx, self.doc_idx, self.block_idx
+
+
+class BlockSamplesMapping(object):
+    def __init__(self, mapping_array):
+        # make sure that the array is compatible with BlockSampleData
+        assert mapping_array.shape[1] == 4
+        self.mapping_array = mapping_array
+
+    def __len__(self):
+        return self.mapping_array.shape[0]
+
+    def __getitem__(self, idx):
+        """Get the data associated with an indexed sample."""
+        sample_data = BlockSampleData(*self.mapping_array[idx])
+        return sample_data
+
+
+def get_block_samples_mapping(block_dataset, title_dataset, data_prefix, num_epochs,
+                              max_num_samples, max_seq_length, seed, name, use_one_sent_docs=False):
+    """Get samples mapping for a dataset over fixed size blocks. This function also requires
+    a dataset of the titles for the source documents since their lengths must be taken into account.
+
+    :return: samples_mapping (BlockSamplesMapping)
+    """
+
+    if not num_epochs:
+        if not max_num_samples:
+            raise ValueError("Need to specify either max_num_samples "
+                             "or num_epochs")
+        num_epochs = np.iinfo(np.int32).max - 1
+    if not max_num_samples:
+        max_num_samples = np.iinfo(np.int64).max - 1
+
+    # Filename of the index mapping
+    indexmap_filename = data_prefix
+    indexmap_filename += '_{}_indexmap'.format(name)
+    if num_epochs != (np.iinfo(np.int32).max - 1):
+        indexmap_filename += '_{}ep'.format(num_epochs)
+    if max_num_samples != (np.iinfo(np.int64).max - 1):
+        indexmap_filename += '_{}mns'.format(max_num_samples)
+    indexmap_filename += '_{}msl'.format(max_seq_length)
+    indexmap_filename += '_{}s'.format(seed)
+    if use_one_sent_docs:
+        indexmap_filename += '_1sentok'
+    indexmap_filename += '.npy'
+
+    # Build the indexed mapping if not exist.
+    if mpu.get_data_parallel_rank() == 0 and \
+            not os.path.isfile(indexmap_filename):
+        print(' > WARNING: could not find index map file {}, building '
+              'the indices on rank 0 ...'.format(indexmap_filename))
+
+        # Make sure the types match the helpers input types.
+        assert block_dataset.doc_idx.dtype == np.int64
+        assert block_dataset.sizes.dtype == np.int32
+
+        # Build samples mapping
+        verbose = torch.distributed.get_rank() == 0
+        start_time = time.time()
+        print_rank_0(' > building samples index mapping for {} ...'.format(
+            name))
+
+        from megatron.data import helpers
+        mapping_array = helpers.build_blocks_mapping(
+            block_dataset.doc_idx,
+            block_dataset.sizes,
+            title_dataset.sizes,
+            num_epochs,
+            max_num_samples,
+            max_seq_length - 3,  # account for added tokens
+            seed,
+            verbose,
+            use_one_sent_docs)
+
+
+        print_rank_0(' > done building samples index mapping')
+        np.save(indexmap_filename, mapping_array, allow_pickle=True)
+        print_rank_0(' > saved the index mapping in {}'.format(
+            indexmap_filename))
+        # Make sure all the ranks have built the mapping
+        print_rank_0(' > elapsed time to build and save samples mapping '
+                     '(seconds): {:4f}'.format(
+            time.time() - start_time))
+
+    # This should be a barrier but nccl barrier assumes
+    # device_index=rank which is not the case for model
+    # parallel case
+    counts = torch.cuda.LongTensor([1])
+    torch.distributed.all_reduce(counts, group=mpu.get_data_parallel_group())
+    assert counts[0].item() == torch.distributed.get_world_size(
+        group=mpu.get_data_parallel_group())
+
+    # Load indexed dataset.
+    print_rank_0(' > loading indexed mapping from {}'.format(
+        indexmap_filename))
+    start_time = time.time()
+
+    mapping_array = np.load(indexmap_filename, allow_pickle=True, mmap_mode='r')
+    samples_mapping = BlockSamplesMapping(mapping_array)
+
+    print_rank_0('    loaded indexed file in {:3.3f} seconds'.format(
+        time.time() - start_time))
+    print_rank_0('    total number of samples: {}'.format(
+        mapping_array.shape[0]))
+
+    return samples_mapping

megatron-deepspeed_dtk22.10/megatron/data/blendable_dataset.py

+# coding=utf-8
+# Copyright (c) 2020, 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.
+
+"""Blendable dataset."""
+
+import time
+
+import numpy as np
+import torch
+
+from megatron import print_rank_0
+from megatron import mpu
+
+
+class BlendableDataset(torch.utils.data.Dataset):
+
+
+    def __init__(self, datasets, weights):
+
+        self.datasets = datasets
+        num_datasets = len(datasets)
+        assert num_datasets == len(weights)
+
+        self.size = 0
+        for dataset in self.datasets:
+            self.size += len(dataset)
+
+        # Normalize weights.
+        weights = np.array(weights, dtype=np.float64)
+        sum_weights = np.sum(weights)
+        assert sum_weights > 0.0
+        weights /= sum_weights
+
+        # Build indecies.
+        start_time = time.time()
+        assert num_datasets < 255
+        self.dataset_index = np.zeros(self.size, dtype=np.uint8)
+        self.dataset_sample_index = np.zeros(self.size, dtype=np.int64)
+
+        from megatron.data import helpers
+        helpers.build_blending_indices(self.dataset_index,
+                                       self.dataset_sample_index,
+                                       weights, num_datasets, self.size,
+                                       torch.distributed.get_rank() == 0)
+        print_rank_0('> elapsed time for building blendable dataset indices: '
+                     '{:.2f} (sec)'.format(time.time() - start_time))
+
+
+    def __len__(self):
+        return self.size
+
+
+    def __getitem__(self, idx):
+        dataset_idx = self.dataset_index[idx]
+        sample_idx = self.dataset_sample_index[idx]
+        return self.datasets[dataset_idx][sample_idx]

megatron-deepspeed_dtk22.10/megatron/data/data_samplers.py

+# coding=utf-8
+# Copyright (c) 2020, 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.
+
+"""Dataloaders."""
+
+import torch
+
+from megatron import get_args
+from megatron import mpu
+from megatron.data.decoder_packed_mtf_dataset import DecoderPackedMTFDataset
+
+
+def build_pretraining_data_loader(dataset, consumed_samples, num_workers=None):
+    """Buld dataloader given an input dataset."""
+
+    if dataset is None:
+        return None
+    args = get_args()
+
+    # Megatron sampler
+    if args.dataloader_type == 'single':
+        batch_sampler = MegatronPretrainingSampler(
+            total_samples=len(dataset),
+            consumed_samples=consumed_samples,
+            micro_batch_size=args.micro_batch_size,
+            data_parallel_rank=mpu.get_data_parallel_rank(),
+            data_parallel_size=mpu.get_data_parallel_world_size())
+    elif args.dataloader_type == 'cyclic':
+        batch_sampler = MegatronPretrainingRandomSampler(
+            total_samples=len(dataset),
+            consumed_samples=consumed_samples,
+            micro_batch_size=args.micro_batch_size,
+            data_parallel_rank=mpu.get_data_parallel_rank(),
+            data_parallel_size=mpu.get_data_parallel_world_size())
+    else:
+        raise Exception('{} dataloader type is not supported.'.format(
+            args.dataloader_type))
+
+    if num_workers is None:
+        num_workers = args.num_workers
+
+    # Torch dataloader.
+    return torch.utils.data.DataLoader(
+        dataset,
+        batch_sampler=batch_sampler,
+        num_workers=num_workers,
+        generator=torch.Generator().manual_seed(args.seed),
+        collate_fn=None,
+        pin_memory=True
+    )
+
+class MegatronPretrainingSampler:
+
+    def __init__(self, total_samples, consumed_samples, micro_batch_size,
+                 data_parallel_rank, data_parallel_size, drop_last=True):
+        # Keep a copy of input params for later use.
+        self.total_samples = total_samples
+        self.consumed_samples = consumed_samples
+        self.micro_batch_size = micro_batch_size
+        self.data_parallel_rank = data_parallel_rank
+        self.micro_batch_times_data_parallel_size = \
+            self.micro_batch_size * data_parallel_size
+        self.drop_last = drop_last
+
+        # Sanity checks.
+        assert self.total_samples > 0, \
+            'no sample to consume: {}'.format(self.total_samples)
+        assert self.consumed_samples < self.total_samples, \
+            'no samples left to consume: {}, {}'.format(self.consumed_samples,
+                                                        self.total_samples)
+        assert self.micro_batch_size > 0
+        assert data_parallel_size > 0
+        assert self.data_parallel_rank < data_parallel_size, \
+            'data_parallel_rank should be smaller than data size: {}, ' \
+            '{}'.format(self.data_parallel_rank, data_parallel_size)
+
+    def __len__(self):
+        return self.total_samples
+
+    def get_start_end_idx(self):
+        start_idx = self.data_parallel_rank * self.micro_batch_size
+        end_idx = start_idx + self.micro_batch_size
+        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):
+            batch.append(idx)
+            if len(batch) == self.micro_batch_times_data_parallel_size:
+                start_idx, end_idx = self.get_start_end_idx()
+                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()
+            yield batch[start_idx:end_idx]
+
+
+class MegatronPretrainingRandomSampler:
+
+    def __init__(self, total_samples, consumed_samples, micro_batch_size,
+                 data_parallel_rank, data_parallel_size):
+        # Keep a copy of input params for later use.
+        self.total_samples = total_samples
+        self.consumed_samples = consumed_samples
+        self.micro_batch_size = micro_batch_size
+        self.data_parallel_rank = data_parallel_rank
+        self.data_parallel_size = data_parallel_size
+        self.micro_batch_times_data_parallel_size = \
+            self.micro_batch_size * data_parallel_size
+        self.last_batch_size = \
+            self.total_samples % self.micro_batch_times_data_parallel_size
+
+        # Sanity checks.
+        assert self.total_samples > 0, \
+            'no sample to consume: {}'.format(self.total_samples)
+        assert self.micro_batch_size > 0
+        assert data_parallel_size > 0
+        assert self.data_parallel_rank < data_parallel_size, \
+            'data_parallel_rank should be smaller than data size: {}, ' \
+            '{}'.format(self.data_parallel_rank, data_parallel_size)
+
+    def __len__(self):
+        return self.total_samples
+
+    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.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.micro_batch_size:
+                self.consumed_samples += self.micro_batch_times_data_parallel_size
+                yield batch
+                batch = []

megatron-deepspeed_dtk22.10/megatron/data/dataset_utils.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors, and NVIDIA.
+#
+# 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.
+
+
+# Most of the code here has been copied from:
+#   https://github.com/google-research/albert/blob/master/create_pretraining_data.py
+# with some modifications.
+
+import math
+import os
+import time
+import collections
+
+import numpy as np
+import torch
+
+from megatron import (
+    get_args,
+    mpu,
+    print_rank_0
+)
+from megatron.data.blendable_dataset import BlendableDataset
+from megatron.data.indexed_dataset import make_dataset as make_indexed_dataset
+
+DSET_TYPE_BERT = 'standard_bert'
+DSET_TYPE_ICT = 'ict'
+DSET_TYPE_T5  = 't5'
+
+DSET_TYPES = [DSET_TYPE_BERT, DSET_TYPE_ICT, DSET_TYPE_T5]
+
+
+def analyze_data_prefix(data_prefix):
+
+    # The data prefix should be in the format of:
+    #   weight-1, data-prefix-1, weight-2, data-prefix-2, ..
+    assert len(data_prefix) % 2 == 0
+    num_datasets = len(data_prefix) // 2
+    weights = [0]*num_datasets
+    prefixes = [0]*num_datasets
+    for i in range(num_datasets):
+        weights[i] = float(data_prefix[2*i])
+        prefixes[i] = (data_prefix[2*i+1]).strip()
+    # Normalize weights
+    weight_sum = 0.0
+    for weight in weights:
+        weight_sum += weight
+    assert weight_sum > 0.0
+    weights = [weight / weight_sum for weight in weights]
+    return prefixes, weights
+
+
+def get_datasets_weights_and_num_samples(data_prefix,
+                                         train_valid_test_num_samples):
+
+    # Add 0.5% (the 1.005 factor) so in case the blending dataset does
+    # not uniformly distribute the number of samples, we still have
+    # samples left to feed to the network.
+    prefixes, weights = analyze_data_prefix(data_prefix)
+    datasets_train_valid_test_num_samples = []
+    for weight in weights:
+        datasets_train_valid_test_num_samples.append(
+            [int(math.ceil(val * weight * 1.005))
+             for val in train_valid_test_num_samples])
+
+
+    return prefixes, weights, datasets_train_valid_test_num_samples
+
+
+def compile_helper():
+    """Compile helper function ar runtime. Make sure this
+    is invoked on a single process."""
+    import os
+    import subprocess
+    path = os.path.abspath(os.path.dirname(__file__))
+    ret = subprocess.run(['make', '-C', path])
+    if ret.returncode != 0:
+        print("Making C++ dataset helpers module failed, exiting.")
+        import sys
+        sys.exit(1)
+
+
+def get_a_and_b_segments(sample, np_rng):
+    """Divide sample into a and b segments."""
+
+    # Number of sentences in the sample.
+    n_sentences = len(sample)
+    # Make sure we always have two sentences.
+    assert n_sentences > 1, 'make sure each sample has at least two sentences.'
+
+    # First part:
+    # `a_end` is how many sentences go into the `A`.
+    a_end = 1
+    if n_sentences >= 3:
+        # Note that randin in numpy is exclusive.
+        a_end = np_rng.randint(1, n_sentences)
+    tokens_a = []
+    for j in range(a_end):
+        tokens_a.extend(sample[j])
+
+    # Second part:
+    tokens_b = []
+    for j in range(a_end, n_sentences):
+        tokens_b.extend(sample[j])
+
+    # Random next:
+    is_next_random = False
+    if np_rng.random() < 0.5:
+        is_next_random = True
+        tokens_a, tokens_b = tokens_b, tokens_a
+
+    return tokens_a, tokens_b, is_next_random
+
+
+def truncate_segments(tokens_a, tokens_b, len_a, len_b, max_num_tokens, np_rng):
+    """Truncates a pair of sequences to a maximum sequence length."""
+    #print(len_a, len_b, max_num_tokens)
+    assert len_a > 0
+    if len_a + len_b <= max_num_tokens:
+        return False
+    while len_a + len_b > max_num_tokens:
+        if len_a > len_b:
+            len_a -= 1
+            tokens = tokens_a
+        else:
+            len_b -= 1
+            tokens = tokens_b
+        if np_rng.random() < 0.5:
+            del tokens[0]
+        else:
+            tokens.pop()
+    return True
+
+
+def create_tokens_and_tokentypes(tokens_a, tokens_b, cls_id, sep_id):
+    """Merge segments A and B, add [CLS] and [SEP] and build tokentypes."""
+
+    tokens = []
+    tokentypes = []
+    # [CLS].
+    tokens.append(cls_id)
+    tokentypes.append(0)
+    # Segment A.
+    for token in tokens_a:
+        tokens.append(token)
+        tokentypes.append(0)
+    # [SEP].
+    tokens.append(sep_id)
+    tokentypes.append(0)
+    # Segment B.
+    for token in tokens_b:
+        tokens.append(token)
+        tokentypes.append(1)
+    if tokens_b:
+        # [SEP].
+        tokens.append(sep_id)
+        tokentypes.append(1)
+
+    return tokens, tokentypes
+
+
+MaskedLmInstance = collections.namedtuple("MaskedLmInstance",
+                                          ["index", "label"])
+
+
+def is_start_piece(piece):
+    """Check if the current word piece is the starting piece (BERT)."""
+    # When a word has been split into
+    # WordPieces, the first token does not have any marker and any subsequence
+    # tokens are prefixed with ##. So whenever we see the ## token, we
+    # append it to the previous set of word indexes.
+    return not piece.startswith("##")
+
+
+def create_masked_lm_predictions(tokens,
+                                 vocab_id_list, vocab_id_to_token_dict,
+                                 masked_lm_prob,
+                                 cls_id, sep_id, mask_id,
+                                 max_predictions_per_seq,
+                                 np_rng,
+                                 max_ngrams=3,
+                                 do_whole_word_mask=True,
+                                 favor_longer_ngram=False,
+                                 do_permutation=False,
+                                 geometric_dist=False,
+                                 masking_style="bert"):
+    """Creates the predictions for the masked LM objective.
+    Note: Tokens here are vocab ids and not text tokens."""
+
+    cand_indexes = []
+    # Note(mingdachen): We create a list for recording if the piece is
+    # the starting piece of current token, where 1 means true, so that
+    # on-the-fly whole word masking is possible.
+    token_boundary = [0] * len(tokens)
+
+    for (i, token) in enumerate(tokens):
+        if token == cls_id or token == sep_id:
+            token_boundary[i] = 1
+            continue
+        # Whole Word Masking means that if we mask all of the wordpieces
+        # corresponding to an original word.
+        #
+        # Note that Whole Word Masking does *not* change the training code
+        # at all -- we still predict each WordPiece independently, softmaxed
+        # over the entire vocabulary.
+        if (do_whole_word_mask and len(cand_indexes) >= 1 and
+                not is_start_piece(vocab_id_to_token_dict[token])):
+            cand_indexes[-1].append(i)
+        else:
+            cand_indexes.append([i])
+            if is_start_piece(vocab_id_to_token_dict[token]):
+                token_boundary[i] = 1
+
+    output_tokens = list(tokens)
+
+    masked_lm_positions = []
+    masked_lm_labels = []
+
+    if masked_lm_prob == 0:
+        return (output_tokens, masked_lm_positions,
+                masked_lm_labels, token_boundary)
+
+    num_to_predict = min(max_predictions_per_seq,
+                         max(1, int(round(len(tokens) * masked_lm_prob))))
+
+    ngrams = np.arange(1, max_ngrams + 1, dtype=np.int64)
+    if not geometric_dist:
+        # Note(mingdachen):
+        # By default, we set the probilities to favor shorter ngram sequences.
+        pvals = 1. / np.arange(1, max_ngrams + 1)
+        pvals /= pvals.sum(keepdims=True)
+        if favor_longer_ngram:
+            pvals = pvals[::-1]
+
+    ngram_indexes = []
+    for idx in range(len(cand_indexes)):
+        ngram_index = []
+        for n in ngrams:
+            ngram_index.append(cand_indexes[idx:idx + n])
+        ngram_indexes.append(ngram_index)
+
+    np_rng.shuffle(ngram_indexes)
+
+    (masked_lms, masked_spans) = ([], [])
+    covered_indexes = set()
+    for cand_index_set in ngram_indexes:
+        if len(masked_lms) >= num_to_predict:
+            break
+        if not cand_index_set:
+            continue
+        # Note(mingdachen):
+        # Skip current piece if they are covered in lm masking or previous ngrams.
+        for index_set in cand_index_set[0]:
+            for index in index_set:
+                if index in covered_indexes:
+                    continue
+
+        if not geometric_dist:
+            n = np_rng.choice(ngrams[:len(cand_index_set)],
+                              p=pvals[:len(cand_index_set)] /
+                              pvals[:len(cand_index_set)].sum(keepdims=True))
+        else:
+            # Sampling "n" from the geometric distribution and clipping it to
+            # the max_ngrams. Using p=0.2 default from the SpanBERT paper
+            # https://arxiv.org/pdf/1907.10529.pdf (Sec 3.1)
+            n = min(np_rng.geometric(0.2), max_ngrams)
+
+        index_set = sum(cand_index_set[n - 1], [])
+        n -= 1
+        # Note(mingdachen):
+        # Repeatedly looking for a candidate that does not exceed the
+        # maximum number of predictions by trying shorter ngrams.
+        while len(masked_lms) + len(index_set) > num_to_predict:
+            if n == 0:
+                break
+            index_set = sum(cand_index_set[n - 1], [])
+            n -= 1
+        # If adding a whole-word mask would exceed the maximum number of
+        # predictions, then just skip this candidate.
+        if len(masked_lms) + len(index_set) > num_to_predict:
+            continue
+        is_any_index_covered = False
+        for index in index_set:
+            if index in covered_indexes:
+                is_any_index_covered = True
+                break
+        if is_any_index_covered:
+            continue
+        for index in index_set:
+            covered_indexes.add(index)
+            masked_token = None
+            if masking_style == "bert":
+                # 80% of the time, replace with [MASK]
+                if np_rng.random() < 0.8:
+                    masked_token = mask_id
+                else:
+                    # 10% of the time, keep original
+                    if np_rng.random() < 0.5:
+                        masked_token = tokens[index]
+                    # 10% of the time, replace with random word
+                    else:
+                        masked_token = vocab_id_list[np_rng.randint(0, len(vocab_id_list))]
+            elif masking_style == "t5":
+                masked_token = mask_id
+            else:
+                raise ValueError("invalid value of masking style")
+
+            output_tokens[index] = masked_token
+            masked_lms.append(MaskedLmInstance(index=index, label=tokens[index]))
+
+        masked_spans.append(MaskedLmInstance(
+            index=index_set,
+            label=[tokens[index] for index in index_set]))
+
+    assert len(masked_lms) <= num_to_predict
+    np_rng.shuffle(ngram_indexes)
+
+    select_indexes = set()
+    if do_permutation:
+        for cand_index_set in ngram_indexes:
+            if len(select_indexes) >= num_to_predict:
+                break
+            if not cand_index_set:
+                continue
+            # Note(mingdachen):
+            # Skip current piece if they are covered in lm masking or previous ngrams.
+            for index_set in cand_index_set[0]:
+                for index in index_set:
+                    if index in covered_indexes or index in select_indexes:
+                        continue
+
+            n = np.random.choice(ngrams[:len(cand_index_set)],
+                                 p=pvals[:len(cand_index_set)] /
+                                 pvals[:len(cand_index_set)].sum(keepdims=True))
+            index_set = sum(cand_index_set[n - 1], [])
+            n -= 1
+
+            while len(select_indexes) + len(index_set) > num_to_predict:
+                if n == 0:
+                    break
+                index_set = sum(cand_index_set[n - 1], [])
+                n -= 1
+            # If adding a whole-word mask would exceed the maximum number of
+            # predictions, then just skip this candidate.
+            if len(select_indexes) + len(index_set) > num_to_predict:
+                continue
+            is_any_index_covered = False
+            for index in index_set:
+                if index in covered_indexes or index in select_indexes:
+                    is_any_index_covered = True
+                    break
+            if is_any_index_covered:
+                continue
+            for index in index_set:
+                select_indexes.add(index)
+        assert len(select_indexes) <= num_to_predict
+
+        select_indexes = sorted(select_indexes)
+        permute_indexes = list(select_indexes)
+        np_rng.shuffle(permute_indexes)
+        orig_token = list(output_tokens)
+
+        for src_i, tgt_i in zip(select_indexes, permute_indexes):
+            output_tokens[src_i] = orig_token[tgt_i]
+            masked_lms.append(MaskedLmInstance(index=src_i, label=orig_token[src_i]))
+
+    masked_lms = sorted(masked_lms, key=lambda x: x.index)
+    # Sort the spans by the index of the first span
+    masked_spans = sorted(masked_spans, key=lambda x: x.index[0])
+
+    for p in masked_lms:
+        masked_lm_positions.append(p.index)
+        masked_lm_labels.append(p.label)
+    return (output_tokens, masked_lm_positions, masked_lm_labels, token_boundary, masked_spans)
+
+
+def pad_and_convert_to_numpy(tokens, tokentypes, masked_positions,
+                             masked_labels, pad_id, max_seq_length):
+    """Pad sequences and convert them to numpy."""
+
+    # Some checks.
+    num_tokens = len(tokens)
+    padding_length = max_seq_length - num_tokens
+    assert padding_length >= 0
+    assert len(tokentypes) == num_tokens
+    assert len(masked_positions) == len(masked_labels)
+
+    # Tokens and token types.
+    filler = [pad_id] * padding_length
+    tokens_np = np.array(tokens + filler, dtype=np.int64)
+    tokentypes_np = np.array(tokentypes + filler, dtype=np.int64)
+
+    # Padding mask.
+    padding_mask_np = np.array([1] * num_tokens + [0] * padding_length,
+                               dtype=np.int64)
+
+    # Lables and loss mask.
+    labels = [-1] * max_seq_length
+    loss_mask = [0] * max_seq_length
+    for i in range(len(masked_positions)):
+        assert masked_positions[i] < num_tokens
+        labels[masked_positions[i]] = masked_labels[i]
+        loss_mask[masked_positions[i]] = 1
+    labels_np = np.array(labels, dtype=np.int64)
+    loss_mask_np = np.array(loss_mask, dtype=np.int64)
+
+    return tokens_np, tokentypes_np, labels_np, padding_mask_np, loss_mask_np
+
+
+def build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
+                                    train_valid_test_num_samples,
+                                    max_seq_length,
+                                    masked_lm_prob, short_seq_prob, seed,
+                                    skip_warmup, binary_head=False,
+                                    max_seq_length_dec=None,
+                                    dataset_type='standard_bert'):
+    if len(data_prefix) == 1:
+        return _build_train_valid_test_datasets(data_prefix[0],
+                                                data_impl, splits_string,
+                                                train_valid_test_num_samples,
+                                                max_seq_length, masked_lm_prob,
+                                                short_seq_prob, seed,
+                                                skip_warmup,
+                                                binary_head,
+                                                max_seq_length_dec,
+                                                dataset_type=dataset_type)
+    # Blending dataset.
+    # Parse the values.
+    output = get_datasets_weights_and_num_samples(data_prefix,
+                                                  train_valid_test_num_samples)
+    prefixes, weights, datasets_train_valid_test_num_samples = output
+
+    # Build individual datasets.
+    train_datasets = []
+    valid_datasets = []
+    test_datasets = []
+    for i in range(len(prefixes)):
+        train_ds, valid_ds, test_ds = _build_train_valid_test_datasets(
+            prefixes[i], data_impl, splits_string,
+            datasets_train_valid_test_num_samples[i],
+            max_seq_length, masked_lm_prob, short_seq_prob,
+            seed, skip_warmup, binary_head, dataset_type=dataset_type)
+        if train_ds:
+            train_datasets.append(train_ds)
+        if valid_ds:
+            valid_datasets.append(valid_ds)
+        if test_ds:
+            test_datasets.append(test_ds)
+
+        # Blend.
+    blending_train_dataset = None
+    if train_datasets:
+        blending_train_dataset = BlendableDataset(train_datasets, weights)
+    blending_valid_dataset = None
+    if valid_datasets:
+        blending_valid_dataset = BlendableDataset(valid_datasets, weights)
+    blending_test_dataset = None
+    if test_datasets:
+        blending_test_dataset = BlendableDataset(test_datasets, weights)
+
+    return (blending_train_dataset, blending_valid_dataset,
+            blending_test_dataset)
+
+
+def _build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
+                                     train_valid_test_num_samples,
+                                     max_seq_length,
+                                     masked_lm_prob, short_seq_prob, seed,
+                                     skip_warmup, binary_head,
+                                     max_seq_length_dec,
+                                     dataset_type='standard_bert'):
+
+    if dataset_type not in DSET_TYPES:
+        raise ValueError("Invalid dataset_type: ", dataset_type)
+
+    # Indexed dataset.
+    indexed_dataset = get_indexed_dataset_(data_prefix,
+                                           data_impl,
+                                           skip_warmup)
+
+    if dataset_type == DSET_TYPE_ICT:
+        args = get_args()
+        title_dataset = get_indexed_dataset_(args.titles_data_path,
+                                             data_impl,
+                                             skip_warmup)
+
+    # Get start and end indices of train/valid/train into doc-idx
+    # Note that doc-idx is desinged to be num-docs + 1 so we can
+    # easily iterate over it.
+    total_num_of_documents = indexed_dataset.doc_idx.shape[0] - 1
+    splits = get_train_valid_test_split_(splits_string, total_num_of_documents)
+
+    # Print stats about the splits.
+    print_rank_0(' > dataset split:')
+
+    def print_split_stats(name, index):
+        print_rank_0('    {}:'.format(name))
+        print_rank_0('     document indices in [{}, {}) total of {} '
+                     'documents'.format(splits[index], splits[index + 1],
+                                        splits[index + 1] - splits[index]))
+        start_index = indexed_dataset.doc_idx[splits[index]]
+        end_index = indexed_dataset.doc_idx[splits[index + 1]]
+        print_rank_0('     sentence indices in [{}, {}) total of {} '
+                     'sentences'.format(start_index, end_index,
+                                        end_index - start_index))
+    print_split_stats('train', 0)
+    print_split_stats('validation', 1)
+    print_split_stats('test', 2)
+
+    def build_dataset(index, name):
+        from megatron.data.bert_dataset import BertDataset
+        from megatron.data.ict_dataset import ICTDataset
+        from megatron.data.t5_dataset import T5Dataset
+        dataset = None
+        if splits[index + 1] > splits[index]:
+            # Get the pointer to the original doc-idx so we can set it later.
+            doc_idx_ptr = indexed_dataset.get_doc_idx()
+            # Slice the doc-idx
+            start_index = splits[index]
+            # Add +1 so we can index into the dataset to get the upper bound.
+            end_index = splits[index + 1] + 1
+            # New doc_idx view.
+            indexed_dataset.set_doc_idx(doc_idx_ptr[start_index:end_index])
+            # Build the dataset accordingly.
+            kwargs = dict(
+                name=name,
+                data_prefix=data_prefix,
+                num_epochs=None,
+                max_num_samples=train_valid_test_num_samples[index],
+                max_seq_length=max_seq_length,
+                seed=seed,
+            )
+
+            if dataset_type == DSET_TYPE_ICT:
+                args = get_args()
+                dataset = ICTDataset(
+                    block_dataset=indexed_dataset,
+                    title_dataset=title_dataset,
+                    query_in_block_prob=args.query_in_block_prob,
+                    use_one_sent_docs=args.use_one_sent_docs,
+                    binary_head=binary_head,
+                    **kwargs
+                )
+            elif dataset_type == DSET_TYPE_T5:
+                dataset = T5Dataset(
+                    indexed_dataset=indexed_dataset,
+                    masked_lm_prob=masked_lm_prob,
+                    max_seq_length_dec=max_seq_length_dec,
+                    short_seq_prob=short_seq_prob,
+                    **kwargs
+                )
+            elif dataset_type == DSET_TYPE_BERT:
+                dataset = BertDataset(
+                    indexed_dataset=indexed_dataset,
+                    masked_lm_prob=masked_lm_prob,
+                    short_seq_prob=short_seq_prob,
+                    binary_head=binary_head,
+                    **kwargs
+                )
+            else:
+                raise NotImplementedError("Dataset type not fully implemented.")
+
+            # Set the original pointer so dataset remains the main dataset.
+            indexed_dataset.set_doc_idx(doc_idx_ptr)
+            # Checks.
+            assert indexed_dataset.doc_idx[0] == 0
+            assert indexed_dataset.doc_idx.shape[0] == \
+                (total_num_of_documents + 1)
+        return dataset
+
+    train_dataset = build_dataset(0, 'train')
+    valid_dataset = build_dataset(1, 'valid')
+    test_dataset = build_dataset(2, 'test')
+
+    return (train_dataset, valid_dataset, test_dataset)
+
+
+def get_indexed_dataset_(data_prefix, data_impl, skip_warmup):
+
+    print_rank_0(' > building dataset index ...')
+
+    start_time = time.time()
+    indexed_dataset = make_indexed_dataset(data_prefix,
+                                           data_impl,
+                                           skip_warmup)
+    assert indexed_dataset.sizes.shape[0] == indexed_dataset.doc_idx[-1]
+    print_rank_0(' > finished creating indexed dataset in {:4f} '
+                 'seconds'.format(time.time() - start_time))
+
+    print_rank_0(' > indexed dataset stats:')
+    print_rank_0('    number of documents: {}'.format(
+        indexed_dataset.doc_idx.shape[0] - 1))
+    print_rank_0('    number of sentences: {}'.format(
+        indexed_dataset.sizes.shape[0]))
+
+    return indexed_dataset
+
+
+def get_split_by_range_(range_string, size):
+    """ Get dataset splits based on a range:
+    range_string is in the form  START%:END%  for e.g. 0.2:0.8
+    outputs an array of two values [start_index, end_index]
+    """
+    # some checks that range is given in the correct form
+    splits = [float(i) for i in range_string.split(":")]
+    assert len(splits) == 2, "splits should be passed as start:end"
+    assert splits[0] <= 1 and splits[1] <= 1
+    splits_sum = sum(splits)
+    assert splits_sum > 0.0
+    splits_index = [round(s * float(size)) for s in splits]
+    assert len(splits_index) == 2
+    return splits_index
+
+def get_train_valid_test_split_(splits_string, size):
+    """ Get dataset splits from comma or '/' separated string list."""
+    splits = []
+    if splits_string.find(',') != -1:
+        splits = [float(s) for s in splits_string.split(',')]
+    elif splits_string.find('/') != -1:
+        splits = [float(s) for s in splits_string.split('/')]
+    else:
+        splits = [float(splits_string)]
+    while len(splits) < 3:
+        splits.append(0.)
+    splits = splits[:3]
+    splits_sum = sum(splits)
+    assert splits_sum > 0.0
+    splits = [split / splits_sum for split in splits]
+    splits_index = [0]
+    for index, split in enumerate(splits):
+        splits_index.append(splits_index[index] +
+                            int(round(split * float(size))))
+    diff = splits_index[-1] - size
+    for index in range(1, len(splits_index)):
+        splits_index[index] -= diff
+    assert len(splits_index) == 4
+    assert splits_index[-1] == size
+    return splits_index
+
+def get_samples_mapping(indexed_dataset,
+                        data_prefix,
+                        num_epochs,
+                        max_num_samples,
+                        max_seq_length,
+                        short_seq_prob,
+                        seed,
+                        name,
+                        binary_head):
+    """Get a list that maps a sample index to a starting sentence index, end sentence index, and length"""
+
+    if not num_epochs:
+        if not max_num_samples:
+            raise ValueError("Need to specify either max_num_samples "
+                             "or num_epochs")
+        num_epochs = np.iinfo(np.int32).max - 1
+    if not max_num_samples:
+        max_num_samples = np.iinfo(np.int64).max - 1
+
+    # Filename of the index mapping
+    indexmap_filename = data_prefix
+    indexmap_filename += '_{}_indexmap'.format(name)
+    if num_epochs != (np.iinfo(np.int32).max - 1):
+        indexmap_filename += '_{}ep'.format(num_epochs)
+    if max_num_samples != (np.iinfo(np.int64).max - 1):
+        indexmap_filename += '_{}mns'.format(max_num_samples)
+    indexmap_filename += '_{}msl'.format(max_seq_length)
+    indexmap_filename += '_{:0.2f}ssp'.format(short_seq_prob)
+    indexmap_filename += '_{}s'.format(seed)
+    indexmap_filename += '.npy'
+
+    # Build the indexed mapping if not exist.
+    if torch.distributed.get_rank() == 0 and \
+       not os.path.isfile(indexmap_filename):
+        print(' > WARNING: could not find index map file {}, building '
+              'the indices on rank 0 ...'.format(indexmap_filename))
+
+        # Make sure the types match the helpers input types.
+        assert indexed_dataset.doc_idx.dtype == np.int64
+        assert indexed_dataset.sizes.dtype == np.int32
+
+        # Build samples mapping
+        verbose = torch.distributed.get_rank() == 0
+        start_time = time.time()
+        print_rank_0(' > building sapmles index mapping for {} ...'.format(
+            name))
+        # First compile and then import.
+        from megatron.data import helpers
+        samples_mapping = helpers.build_mapping(
+            indexed_dataset.doc_idx,
+            indexed_dataset.sizes,
+            num_epochs,
+            max_num_samples,
+            max_seq_length,
+            short_seq_prob,
+            seed,
+            verbose,
+            2 if binary_head else 1)
+        print_rank_0(' > done building sapmles index maping')
+        np.save(indexmap_filename, samples_mapping, allow_pickle=True)
+        print_rank_0(' > saved the index mapping in {}'.format(
+            indexmap_filename))
+        # Make sure all the ranks have built the mapping
+        print_rank_0(' > elasped time to build and save samples mapping '
+                     '(seconds): {:4f}'.format(
+                         time.time() - start_time))
+    # This should be a barrier but nccl barrier assumes
+    # device_index=rank which is not the case for model
+    # parallel case
+    counts = torch.cuda.LongTensor([1])
+    torch.distributed.all_reduce(counts, group=mpu.get_data_parallel_group())
+    torch.distributed.all_reduce(counts, group=mpu.get_pipeline_model_parallel_group())
+    assert counts[0].item() == (
+        torch.distributed.get_world_size() //
+        torch.distributed.get_world_size(group=mpu.get_tensor_model_parallel_group()))
+
+    # Load indexed dataset.
+    print_rank_0(' > loading indexed mapping from {}'.format(
+        indexmap_filename))
+    start_time = time.time()
+    samples_mapping = np.load(indexmap_filename, allow_pickle=True, mmap_mode='r')
+    print_rank_0('    loaded indexed file in {:3.3f} seconds'.format(
+        time.time() - start_time))
+    print_rank_0('    total number of samples: {}'.format(
+        samples_mapping.shape[0]))
+
+    return samples_mapping

megatron-deepspeed_dtk22.10/megatron/data/decoder_packed_mtf_dataset.py

+import os
+import time
+
+import numpy as np
+import torch
+
+from megatron import print_rank_0, mpu, logging
+from megatron.data.blendable_dataset import BlendableDataset
+from megatron.data.dataset_utils import get_datasets_weights_and_num_samples, get_split_by_range_, \
+    get_train_valid_test_split_
+from megatron.data.mtf_dataset import MTFDataset
+from megatron.data.indexed_dataset import make_dataset as make_indexed_dataset
+
+logger = logging.get_logger(__name__)
+
+def build_train_valid_test_datasets(
+    data_prefix,
+    data_impl,
+    splits_string,
+    seq_length: int,
+    pad_token: int,
+    eos_token: int,
+    train_valid_test_num_samples,
+    seed,
+    skip_warmup
+):
+    """Build train, valid, and test datasets."""
+
+    # Single dataset.
+    if len(data_prefix) == 1:
+        all_train_datasets, all_valid_datasets, all_test_datasets = _build_train_valid_test_datasets(
+            data_prefix=data_prefix[0],
+            data_impl=data_impl,
+            splits_string=splits_string,
+            seq_length=seq_length,
+            pad_token=pad_token,
+            eos_token=eos_token,
+            train_valid_test_num_samples=train_valid_test_num_samples,
+            seed=seed,
+            skip_warmup=skip_warmup
+        )
+    # Blending dataset.
+    else:
+
+        output = get_datasets_weights_and_num_samples(data_prefix=data_prefix, train_valid_test_num_samples=train_valid_test_num_samples)
+        prefixes, weights, datasets_train_valid_test_num_samples = output
+
+        # Build individual datasets.
+        train_datasets = []
+        valid_datasets = []
+        test_datasets = []
+        for i in range(len(prefixes)):
+            train_ds, valid_ds, test_ds = _build_train_valid_test_datasets(
+                data_prefix=prefixes[i],
+                data_impl=data_impl,
+                splits_string=splits_string,
+                seq_length=seq_length,
+                pad_token=pad_token,
+                eos_token=eos_token,
+                train_valid_test_num_samples=datasets_train_valid_test_num_samples[i],
+                seed=seed,
+                skip_warmup=skip_warmup
+            )
+            if train_ds:
+                train_datasets.append(train_ds)
+            if valid_ds:
+                valid_datasets.append(valid_ds)
+            if test_ds:
+                test_datasets.append(test_ds)
+
+        all_train_datasets = BlendableDataset(train_datasets, weights) \
+                            if train_datasets else None
+        all_valid_datasets = BlendableDataset(valid_datasets, weights) \
+                            if valid_datasets else None
+        all_test_datasets = BlendableDataset(test_datasets, weights) \
+                            if test_datasets else None
+
+    return all_train_datasets, all_valid_datasets, all_test_datasets
+
+
+def build_dataset_group(
+    dataset_group_name,
+    paths,
+    weights,
+    splits,
+    data_impl,
+    seq_length: int,
+    pad_token: int,
+    eos_token: int,
+    train_valid_test_num_samples,
+    seed,
+    skip_warmup,
+    train_valid_test
+):
+    '''
+    Build a single dataset group corresponding to Option 2 of data loading see arguments.py
+    a dataset group is passed in the following form
+    GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2, WEIGHT2 START:END PATH2
+    or alternatively
+    GIVEN_NAME PATH1    # for a single dataset to be used fully
+    '''
+
+    assert train_valid_test in ["train","valid","test"]
+
+    # Single dataset.
+    if len(paths) == 1:
+        dataset = _build_single_datasets(
+            data_prefix=paths[0],
+            range_string=splits[0],
+            data_impl=data_impl,
+            seq_length=seq_length,
+            pad_token=pad_token,
+            eos_token=eos_token,
+            train_valid_test_num_samples=train_valid_test_num_samples,
+            seed=seed,
+            skip_warmup=skip_warmup,
+            dataset_group_name=dataset_group_name,
+            train_valid_test=train_valid_test
+        )
+        return dataset
+    # Blending dataset.
+    else:
+
+        data_prefix = []
+        # data_prefix is of the shape:
+        # ["WEIGHT1", "PATH1", "WEIGHT2", "PATH2", "WEIGHT3", "PATH3"]
+        for w,p in zip(weights, paths):
+            data_prefix += [w,p]
+
+        output = get_datasets_weights_and_num_samples(data_prefix,
+                                                    train_valid_test_num_samples)
+        prefixes, weights, datasets_train_valid_test_num_samples = output
+
+        # Build individual datasets.
+        datasets = []
+        for i in range(len(prefixes)):
+            ds = _build_single_datasets(
+                data_prefix=prefixes[i],
+                range_string=splits[i],
+                data_impl=data_impl,
+                seq_length=seq_length,
+                pad_token=pad_token,
+                eos_token=eos_token,
+                train_valid_test_num_samples=datasets_train_valid_test_num_samples[i],
+                seed=seed,
+                skip_warmup=skip_warmup,
+                dataset_group_name=dataset_group_name,
+                train_valid_test=train_valid_test
+            )
+
+            datasets.append(ds)
+        all_datasets = BlendableDataset(datasets, weights)
+
+        return all_datasets
+
+def _build_single_datasets(
+    data_prefix,
+    range_string,
+    data_impl,
+    seq_length: int,
+    pad_token: int,
+    eos_token: int,
+    train_valid_test_num_samples,
+    seed,
+    skip_warmup,
+    dataset_group_name,
+    train_valid_test
+):
+    """Build a single dataset"""
+
+    assert train_valid_test in ["train","valid","test"]
+    index = ["train","valid","test"].index(train_valid_test)
+
+    # Target indexed dataset.
+    target_indexed_dataset = get_indexed_dataset(
+        data_prefix=data_prefix,
+        is_input=False,
+        data_impl=data_impl,
+        skip_warmup=skip_warmup
+    )
+
+    total_num_of_documents = target_indexed_dataset.sizes.shape[0]
+    # this corresponds to option2 for data loading on the form
+    # WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2, WEIGHT3 START:END PATH3
+    # splits here is an array of size 2  [start_index, end_index]
+    splits = get_split_by_range_(range_string=range_string, size=total_num_of_documents)
+
+    # Print stats about the splits.
+    print_rank_0(' > dataset split:')
+
+    print_rank_0('    {}:'.format(dataset_group_name))
+    print_rank_0('     document indices in [{}, {}) total of {} '
+                     'documents'.format(splits[0], splits[1],
+                                        splits[1] - splits[0]))
+
+    def build_dataset(name):
+        dataset = None
+        if splits[1] > splits[0]:
+            documents = np.arange(start=splits[0], stop=splits[1],
+                                  step=1, dtype=np.int32)
+            dataset = DecoderPackedMTFDataset(
+                name=name,
+                data_prefix=data_prefix,
+                data_impl=data_impl,
+                skip_warmup=skip_warmup,
+                documents=documents,
+                seq_length=seq_length,
+                pad_token=pad_token,
+                eos_token=eos_token,
+                num_samples=train_valid_test_num_samples[index],
+                seed=seed
+            )
+        return dataset
+
+    dataset = build_dataset(dataset_group_name)
+
+    return dataset
+
+
+def _build_train_valid_test_datasets(
+    data_prefix,
+    data_impl,
+    splits_string,
+    seq_length: int,
+    pad_token: int,
+    eos_token: int,
+    train_valid_test_num_samples,
+    seed,
+    skip_warmup
+):
+    """Build train, valid, and test datasets."""
+
+    # Target indexed dataset.
+    target_indexed_dataset = get_indexed_dataset(data_prefix, is_input=False, data_impl=data_impl, skip_warmup=skip_warmup)
+
+    total_num_of_documents = target_indexed_dataset.sizes.shape[0]
+    # splits here is an array of size 4  [train_start_index, valid_start_index, test_start_index, test_end_index]
+    splits = get_train_valid_test_split_(splits_string, total_num_of_documents)
+    # Print stats about the splits.
+    print_rank_0(' > dataset split:')
+
+    def print_split_stats(name, index):
+        print_rank_0('    {}:'.format(name))
+        print_rank_0('     document indices in [{}, {}) total of {} '
+                     'documents'.format(splits[index], splits[index + 1],
+                                        splits[index + 1] - splits[index]))
+    print_split_stats('train', 0)
+    print_split_stats('validation', 1)
+    print_split_stats('test', 2)
+
+    def build_dataset(index, name):
+        dataset = None
+        if splits[index + 1] > splits[index]:
+            documents = np.arange(start=splits[index], stop=splits[index + 1],
+                                  step=1, dtype=np.int32)
+            dataset = DecoderPackedMTFDataset(
+                name=name,
+                data_prefix=data_prefix,
+                data_impl=data_impl,
+                skip_warmup=skip_warmup,
+                documents=documents,
+                seq_length=seq_length,
+                pad_token=pad_token,
+                eos_token=eos_token,
+                num_samples=train_valid_test_num_samples[index],
+                seed=seed
+            )
+        return dataset
+
+    train_dataset = build_dataset(0, 'train')
+    valid_dataset = build_dataset(1, 'valid')
+    test_dataset = build_dataset(2, 'test')
+
+    return (train_dataset, valid_dataset, test_dataset)
+
+
+class DecoderPackedMTFDataset(torch.utils.data.Dataset):
+
+    def __init__(
+        self,
+        name,
+        data_prefix,
+        data_impl,
+        skip_warmup,
+        documents,
+        num_samples,
+        seq_length: int,
+        pad_token: int,
+        eos_token: int,
+        seed,
+    ):
+        self.mtf_dataset = MTFDataset(name=name, data_prefix=data_prefix, data_impl=data_impl, skip_warmup=skip_warmup, documents=documents)
+
+        self.pad_token = pad_token
+        self.seq_length = seq_length
+
+        self.sample_index, self.shuffle_index = _build_index_mappings(name=name, data_prefix=data_prefix, nb_documents=len(documents), mtf_dataset=self.mtf_dataset, num_samples=num_samples, seq_length=seq_length, seed=seed)
+
+    def __len__(self):
+        return len(self.sample_index)
+
+    def __getitem__(self, idx):
+        # Get the shuffled index.
+        start, end = self.sample_index[idx]
+        mtf_samples_indices = self.shuffle_index[start: end]
+        # TODO @thomasw21 build a dataset that generates an entire batch instead of a row (allows for more optimization)
+        items = [self.mtf_dataset[sample_id] for sample_id in mtf_samples_indices]
+
+        return self.pack_samples(items)
+
+    def pack_samples(self, items):
+        """
+        Greedily packs samples.
+
+        Items:
+            [
+                {
+                    'input_tokens': array([6, 7]),
+                    'target_tokens': array([8])
+                },
+                {
+                    'input_tokens': array([3, 4]),
+                    'target_tokens': array([5])
+                }
+            ]
+
+        Output:
+            decoder_tokens = [[6, 7, 8, 3, 4, 5, <pad>]]: Concatenation of tokens followed with padding tokens.
+            decoder_segment_ids = [[1, 1, 1, 2, 2, 2, 0]]: Segment ids determine original documents.
+            decoder_is_inputs = [[1, 1, 0, 1, 1, 0, 0]]: `1` depicts inputs, `0` depicts target.
+        """
+
+        decoder_tokens = np.full((self.seq_length,), self.pad_token, dtype=np.int64)
+        decoder_segment_ids = np.zeros((self.seq_length,), dtype=np.int64)
+        decoder_is_inputs = np.full((self.seq_length,), False, dtype=bool)
+
+        # `0` is reserved for padding
+        item_num = 1
+        cur_len = 0
+
+        assert len(items) > 0
+
+        for token_dict in items:
+            input_token_len = len(token_dict["input_tokens"])
+            target_token_len = len(token_dict["target_tokens"])
+
+            total_len = input_token_len + target_token_len
+
+            if cur_len + total_len > self.seq_length:
+                # This should not happen at the indexing should only allow the correct number of items
+                raise ValueError(f"""Items to be packed do not fit inside a single sample.
+                    current length: {cur_len}
+                    input tokens length: {input_token_len}
+                    target token length: {target_token_len}
+                    expected sequence length: {self.seq_length}
+                """)
+
+            decoder_tokens[cur_len: cur_len + input_token_len] = token_dict["input_tokens"]
+            decoder_tokens[cur_len + input_token_len: cur_len + total_len] = token_dict["target_tokens"]
+            decoder_segment_ids[cur_len: cur_len + total_len] = item_num
+            decoder_is_inputs[cur_len: cur_len + input_token_len] = 1  # inputs
+            # targets are already 0 at init, no need to update `decoder_is_inputs`
+
+            item_num += 1
+            cur_len += total_len
+            assert cur_len <= self.seq_length
+
+        return {
+            "decoder_token_ids": decoder_tokens,
+            "decoder_segment_ids": decoder_segment_ids,
+            "decoder_is_inputs": decoder_is_inputs,
+        }
+
+
+def _build_index_mappings(
+    name,
+    data_prefix,
+    nb_documents,
+    mtf_dataset,
+    num_samples: int,
+    seq_length: int,
+    seed,
+):
+    """
+    - `shuffle_index` is [num_epoch * len(self.mtf)]
+    - `sample_index` is [num_sample, 2] (storing the start and end of the sample). We query the sample via `self.shuffle_index[start:end]`
+
+    TODO @thomas21 Instead of loading individually samples, we save the packing one and for all
+    """
+    # rng state
+    np_rng = np.random.RandomState(seed=seed)
+
+    # Filename of the index mappings.
+    _filename = data_prefix
+    _filename += '_{}_indexmap'.format(name)
+    _filename += '_{}ns'.format(num_samples)
+    _filename += '_{}s'.format(seed)
+    sample_idx_filename = _filename + '_decoder_packed_batch_idx.npy'
+    shuffle_idx_filename = _filename + '_decoder_packed_shuffle_idx.npy'
+
+    # Build the indexed mapping if not exist.
+    if torch.distributed.get_rank() == 0:
+        if (not os.path.isfile(sample_idx_filename)) or \
+           (not os.path.isfile(shuffle_idx_filename)):
+
+            print_rank_0(' > WARNING: could not find index map files, building '
+                         'the indices on rank 0 ...')
+
+            # iteratively add the entire dataset for every epoch and see if it's enough given current packing strategy
+            start_time = time.time()
+            row_offset = 0
+            old_sample_start = 0
+            epoch = 0
+            shuffle_idx = []
+            sample_idx = []
+            while len(sample_idx) <= num_samples:
+                new_document_ids = _build_shuffle_idx(nb_documents=nb_documents, np_rng=np_rng)
+                # Generate a shuffling of the entire dataset
+                shuffle_idx.append(new_document_ids)
+                # Packs them into a single sample
+                new_samples, row_offset, old_sample_start = _build_sample_idx(
+                    mtf_dataset=mtf_dataset,
+                    document_ids=new_document_ids,
+                    seq_length=seq_length,
+                    row_offset=row_offset,
+                    old_sample_start=old_sample_start,
+                    epoch=epoch
+                )
+                sample_idx.extend(new_samples)
+                epoch += 1
+
+            shuffle_idx = np.concatenate(shuffle_idx, axis=0)
+            sample_idx = np.stack(sample_idx, axis=0)
+
+            np.save(shuffle_idx_filename, shuffle_idx, allow_pickle=True)
+            np.save(sample_idx_filename, sample_idx, allow_pickle=True)
+            print_rank_0(' > elasped time to build and save shuffle-idx and sample-idx mapping'
+                         ' (seconds): {:4f}'.format(time.time() - start_time))
+
+    # This should be a barrier but nccl barrier assumes
+    # device_index=rank which is not the case for model
+    # parallel case
+    counts = torch.cuda.LongTensor([1])
+    torch.distributed.all_reduce(counts, group=mpu.get_data_parallel_group())
+    torch.distributed.all_reduce(counts, group=mpu.get_pipeline_model_parallel_group())
+    assert counts[0].item() == (
+        torch.distributed.get_world_size() //
+        torch.distributed.get_world_size(group=mpu.get_tensor_model_parallel_group()))
+
+    # Load mappings.
+    start_time = time.time()
+    print_rank_0(' > loading doc-idx mapping from {}'.format(
+        sample_idx_filename))
+    sample_idx = np.load(sample_idx_filename, allow_pickle=True, mmap_mode='r')
+    print_rank_0(' > loading shuffle-idx mapping from {}'.format(
+        shuffle_idx_filename))
+    shuffle_idx = np.load(shuffle_idx_filename, allow_pickle=True, mmap_mode='r')
+    print_rank_0('    loaded indexed file in {:3.3f} seconds'.format(
+        time.time() - start_time))
+
+    return sample_idx, shuffle_idx
+
+def _build_sample_idx(mtf_dataset, document_ids, seq_length, row_offset, old_sample_start, epoch):
+    """Build start and off index of each `full` batch, return that list of batch + start of the unfinished batch"""
+    row_length = row_offset
+
+    full_samples = []
+    current_sample_start = old_sample_start
+    epoch_offset = epoch * len(document_ids)
+
+    assert epoch_offset >= current_sample_start
+    for current_sample_end, document_id in enumerate(document_ids):
+        current_sample_end = epoch_offset + current_sample_end
+        sample_sizes = mtf_dataset.size(document_id)
+
+        # TODO @thomasw21 figure out if we add <eos> tokens
+        tok_len = sample_sizes["input_tokens"] + sample_sizes["target_tokens"]
+
+        row_length = row_length + tok_len
+        if row_length > seq_length:
+            # current sample can't be added and requires to be added in the next one
+            if current_sample_end > current_sample_start:
+                full_samples.append(np.asarray([current_sample_start, current_sample_end]))
+            current_sample_start = current_sample_end
+            row_length = tok_len
+
+            if tok_len > seq_length:
+                # TODO @thomasw21 handle the case where a single sample cannot fit inside a row. We can
+                #   - silently skip that value [currently implemented]
+                #   - truncate to `seq_length`, and keep the right part
+                logger.warning(f"Skipping sample id={document_id}. Maximum sequence length: {seq_length}, sample length: {tok_len}")
+                current_sample_start = current_sample_end + 1  # skipping
+                row_length = 0
+                continue
+
+    return full_samples, row_length, current_sample_start
+
+def _build_shuffle_idx(nb_documents: int, np_rng):
+    """Build the range [0, dataset_size) and shuffle."""
+    dtype_ = np.int64
+
+    result = np.arange(start=0, stop=nb_documents, step=1, dtype=dtype_)
+
+    # in-place shuffling
+    np_rng.shuffle(result)
+
+    return result
+
+
+def get_indexed_dataset(data_prefix: str, is_input: bool, data_impl: str, skip_warmup: bool):
+    if is_input:
+        field = "inputs"
+    else:
+        field = "targets"
+
+    return get_indexed_dataset_(f"{data_prefix}_{field}_document", data_impl, skip_warmup)
+
+
+def get_indexed_dataset_(path, data_impl, skip_warmup):
+    """Build indexed dataset."""
+    print_rank_0(' > building dataset index ...')
+    start_time = time.time()
+    indexed_dataset = make_indexed_dataset(path,
+                                           data_impl,
+                                           skip_warmup)
+    print_rank_0(' > finished creating indexed dataset in {:4f} '
+                 'seconds'.format(time.time() - start_time))
+    print_rank_0('    number of documents: {}'.format(
+        indexed_dataset.sizes.shape[0]))
+
+    return indexed_dataset

megatron-deepspeed_dtk22.10/megatron/data/distdata.py

+import os
+import numpy as np
+
+import torch
+import torch.nn.functional as F
+import torch.distributed as dist
+
+class DistDataError(Exception):
+    """Defines an empty exception to throw when some other rank hit a real exception."""
+    pass
+
+class DistData(object):
+    def __init__(self, backend='gloo'):
+        assert backend in ['gloo', 'mpi'], f"torch.distributed backend '{backend}' is not supported, valid options are 'gloo' or 'mpi'"
+
+        dist.init_process_group(backend, init_method="env://")
+
+        # lookup our process rank and the group size
+        self.rank = dist.get_rank()
+        self.numranks = dist.get_world_size()
+
+    def allassert(self, cond, msg):
+        """Check that cond is True on all ranks, assert with msg everywhere if not.
+
+        To prevent deadlocks in cases where an assertion might only fail on one rank,
+        this executes an allreduce to ensure that if any rank finds that an assertion
+        has been violated, all ranks fail an assertion check.
+        The condition must be true on all ranks for this not to assert.
+        """
+        alltrue = self.alltrue(cond)
+        assert alltrue, msg
+
+    def allraise_if(self, err):
+        """Raise exception if err is not None on any rank.
+
+        Similarly to allassert, this raises an exception on all ranks if err
+        is set to an exception on any rank.  Rank(s) where err is not None
+        re-raise err as exception, and ranks where err is None raise DistDataError.
+        Thus all ranks raise an exception if any rank has an active exception,
+        which helps avoid deadlocks in cases where an exception may be raised
+        on a subset of ranks.
+        """
+        alltrue = self.alltrue(err is None)
+        if not alltrue:
+            # At least one rank raised an exception.
+            # Re-raise the actual exception if this rank threw one.
+            if err is not None:
+                raise err
+
+            # TODO: is there a better exception to use here?
+            # On other ranks, raise an "empty" exception to indicate
+            # that we're only failing because someone else did.
+            raise DistDataError
+
+    def barrier(self):
+        """Globally synchronize all processes"""
+        dist.barrier()
+
+    def bcast(self, val, root):
+        """Broadcast a scalar value from root to all ranks"""
+        vals = [val]
+        dist.broadcast_object_list(vals, src=root)
+        return vals[0]
+
+    def scatterv_(self, invals: np.array, counts: list, root:int=0):
+        """Scatter int64 values from invals according to counts array, return received portion in a new tensor"""
+
+        self.allassert(len(counts) == self.numranks,
+            f"Length of counts list {len(counts)} does not match number of ranks {self.numranks}")
+
+        # Define list of tensors to scatter on the root.
+        # torch.distributed.scatter requires each tensor to be the same shape,
+        # so find the max size across all count values and pad.
+        max_size = max(counts)
+        scatterlist = None
+        if self.rank == root:
+            slices = list(torch.split(torch.from_numpy(invals), counts))
+            scatterlist = [F.pad(s, (0, max_size - len(s))) for s in slices]
+
+        # Receive a tensor of the max count size from the root,
+        # then copy values into output numpy array, which may be smaller.
+        recvtensor = torch.zeros(max_size, dtype=torch.int64)
+        dist.scatter(recvtensor, scatterlist, src=root)
+        return recvtensor[:counts[self.rank]]
+
+    def alltrue(self, val):
+        """Returns True if all procs input True, False otherwise"""
+        # torch.dist does not support reductions with bool types
+        # so we cast to int and cast the result back to bool
+        tensor = torch.tensor([int(val)], dtype=torch.int32)
+        dist.all_reduce(tensor, op=dist.ReduceOp.BAND)
+        return bool(tensor[0])
+
+    def sum(self, val):
+        """Compute sum of a scalar val, and return total on all ranks."""
+        tensor = torch.tensor([val])
+        dist.all_reduce(tensor, op=dist.ReduceOp.SUM)
+        return tensor[0]
+
+    def exscan(self, val: int):
+        """Compute prefix sum (exclusive scan) of int64 val, and return offset of each rank."""
+        # torch.distributed doesn't have a scan, so fallback to allreduce
+        tensor = torch.zeros(self.numranks, dtype=torch.int64)
+        tensor[self.rank:] = val
+        dist.all_reduce(tensor, op=dist.ReduceOp.SUM)
+        return int(tensor[self.rank]) - val
+
+    def min(self, val):
+        """Return minimum of scalar val to all ranks."""
+        tensor = torch.tensor([val])
+        dist.all_reduce(tensor, op=dist.ReduceOp.MIN)
+        return tensor[0]
+
+    def minrank(self, cond):
+        """Find first rank whose condition is True, return that rank if any, None otherwise."""
+        minrank = self.numranks
+        if cond:
+            minrank = self.rank
+        minrank = self.min(minrank)
+
+        if minrank < self.numranks:
+            return minrank
+        return None
+
+    def bcast_first(self, val):
+        """Broadcast val from first rank where it is not None, return val if any, None otherwise"""
+        # Find the first rank with a valid value.
+        minrank = self.minrank(val is not None)
+
+        # If there is no rank with a valid value, return None
+        if minrank is None:
+            return None
+
+        # Otherwise broadcast the value from the first valid rank.
+        val = self.bcast(val, root=minrank)
+        return val
+
+    def all_sum_(self, vals: np.array):
+        """Sums values in numpy array vals element-wise and update vals in place with final result on all ranks"""
+        # Builds torch.tensor with from_numpy to use same underlying memory as numpy array.
+        tensor = torch.from_numpy(vals)
+        dist.all_reduce(tensor, op=dist.ReduceOp.SUM)
+
+    def open(self, filename, truncate=None):
+        """Create, truncate, and open a file shared by all ranks."""
+
+        # Don't truncate existing file until all ranks reach this point
+        self.barrier()
+
+        # We'll capture any exception in this variable
+        err = None
+
+        # Rank 0 creates and truncates file.
+        if self.rank == 0:
+            try:
+                f = open(filename, 'wb')
+
+                # Some file systems like GPFS deliver faster write speed
+                # if the file size is known before data is written to the file.
+                if truncate is not None:
+                    f.truncate(truncate)
+
+            except Exception as e:
+                err = e
+
+        # Verify that rank 0 created the file
+        self.allraise_if(err)
+
+        # Wait for rank 0 to open (and truncate) file,
+        # then have all ranks open file for writing.
+        if self.rank != 0:
+            try:
+                f = open(filename, 'r+b')
+            except Exception as e:
+                err = e
+
+        # Verify that all ranks successfully opened the file
+        self.allraise_if(err)
+
+        return f
+
+    def remove(self, filename):
+        """Remove a shared file."""
+
+        # Don't remove the file until all are ready
+        self.barrier()
+
+        # We'll capture any exception in this variable
+        err = None
+
+        # Rank 0 removes the file if it exists.
+        if self.rank == 0:
+            try:
+                if os.path.exists(filename):
+                    os.remove(filename)
+            except Exception as e:
+                err = e
+
+        # Verify that rank 0 successfully removed the file.
+        self.allraise_if(err)
+
+    def rename(self, srcfile, destfile):
+        """Rename a shared file."""
+
+        # Don't rename until all are ready
+        self.barrier()
+
+        # We'll capture any exception in this variable
+        err = None
+
+        # Rank 0 renames the file.
+        if self.rank == 0:
+            try:
+                if os.path.exists(srcfile):
+                    os.rename(srcfile, destfile)
+            except Exception as e:
+                err = e
+
+        # Verify that the rename succeeded
+        self.allraise_if(err)

megatron-deepspeed_dtk22.10/megatron/data/gpt_dataset.py

+# coding=utf-8
+# Copyright (c) 2020, 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.
+
+"""GPT style dataset."""
+
+import os
+import time
+
+import numpy as np
+import torch
+
+from megatron import mpu, print_rank_0
+from megatron.data.blendable_dataset import BlendableDataset
+from megatron.data.dataset_utils import get_datasets_weights_and_num_samples
+from megatron.data.dataset_utils import get_train_valid_test_split_, get_split_by_range_
+from megatron.data.indexed_dataset import make_dataset as make_indexed_dataset
+
+
+def build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
+                                    train_valid_test_num_samples,
+                                    seq_length, seed, skip_warmup):
+    """Build train, valid, and test datasets."""
+
+    # Single dataset.
+    if len(data_prefix) == 1:
+        all_train_datasets, all_valid_datasets, all_test_datasets = _build_train_valid_test_datasets(data_prefix[0],
+                                                data_impl, splits_string,
+                                                train_valid_test_num_samples,
+                                                seq_length, seed, skip_warmup)
+    # Blending dataset.
+    else:
+
+        output = get_datasets_weights_and_num_samples(data_prefix,
+                                                    train_valid_test_num_samples)
+        prefixes, weights, datasets_train_valid_test_num_samples = output
+
+        # Build individual datasets.
+        train_datasets = []
+        valid_datasets = []
+        test_datasets = []
+        for i in range(len(prefixes)):
+            train_ds, valid_ds, test_ds = _build_train_valid_test_datasets(
+                                            prefixes[i], data_impl, splits_string,
+                                            datasets_train_valid_test_num_samples[i],
+                                            seq_length, seed, skip_warmup)
+            if train_ds:
+                train_datasets.append(train_ds)
+            if valid_ds:
+                valid_datasets.append(valid_ds)
+            if test_ds:
+                test_datasets.append(test_ds)
+
+        all_train_datasets = BlendableDataset(train_datasets, weights) \
+                            if train_datasets else None
+        all_valid_datasets = BlendableDataset(valid_datasets, weights) \
+                            if valid_datasets else None
+        all_test_datasets = BlendableDataset(test_datasets, weights) \
+                            if test_datasets else None
+
+    return all_train_datasets, all_valid_datasets, all_test_datasets
+
+
+def build_dataset_group(dataset_group_name, paths, weights, splits, data_impl,
+                        train_valid_test_num_samples,
+                        seq_length, seed, skip_warmup, train_valid_test):
+    '''
+    Build a single dataset group corresponding to Option 2 of data loading see arguments.py
+    a dataset group is passed on the following form
+    GIVEN_NAME WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2, WEIGHT2 START:END PATH2
+    or alternatively
+    GIVEN_NAME PATH1    # for a single dataset to be used fully
+    '''
+
+    assert train_valid_test in ["train","valid","test"]
+
+    # Single dataset.
+    if len(paths) == 1:
+        dataset =  _build_single_datasets(paths[0],
+                                          splits[0],
+                                          data_impl,
+                                          train_valid_test_num_samples,
+                                          seq_length, seed, skip_warmup,
+                                          dataset_group_name, train_valid_test)
+        return dataset
+    # Blending dataset.
+    else:
+
+        data_prefix = []
+        # data_prefix is on the shape:
+        # ["WEIGHT1", "PATH1", "WEIGHT2", "PATH2", "WEIGHT3", "PATH3"]
+        for w,p in zip(weights, paths):
+            data_prefix += [w,p]
+
+        output = get_datasets_weights_and_num_samples(data_prefix,
+                                                    train_valid_test_num_samples)
+        prefixes, weights, datasets_train_valid_test_num_samples = output
+
+        # Build individual datasets.
+        datasets = []
+        for i in range(len(prefixes)):
+            ds = _build_single_datasets(prefixes[i],
+                                        splits[i],
+                                        data_impl,
+                                        datasets_train_valid_test_num_samples[i],
+                                        seq_length,
+                                        seed, skip_warmup,
+                                        dataset_group_name, train_valid_test)
+
+            datasets.append(ds)
+        all_datasets = BlendableDataset(datasets, weights)
+
+        return all_datasets
+
+def _build_single_datasets(data_prefix, range_string, data_impl, train_valid_test_num_samples,
+                            seq_length, seed, skip_warmup, dataset_group_name, train_valid_test):
+    """Build a single dataset"""
+
+    assert train_valid_test in ["train","valid","test"]
+    index = ["train","valid","test"].index(train_valid_test)
+
+    # Indexed dataset.
+    indexed_dataset = get_indexed_dataset_(data_prefix,
+                                           data_impl,
+                                           skip_warmup)
+
+    total_num_of_documents = indexed_dataset.sizes.shape[0]
+    # this corresponds to option2 for data loading on the form
+    # WEIGHT1 START:END PATH1, WEIGHT2 START:END PATH2, WEIGHT3 START:END PATH3
+    # splits here is an array of size 2  [start_index, end_index]
+    splits = get_split_by_range_(range_string=range_string, size=total_num_of_documents)
+
+    # Print stats about the splits.
+    print_rank_0(' > dataset split:')
+
+    print_rank_0('    {}:'.format(dataset_group_name))
+    print_rank_0('     document indices in [{}, {}) total of {} '
+                     'documents'.format(splits[0], splits[1],
+                                        splits[1] - splits[0]))
+
+    def build_dataset(name):
+        dataset = None
+        if splits[1] > splits[0]:
+            documents = np.arange(start=splits[0], stop=splits[1],
+                                  step=1, dtype=np.int32)
+            dataset = GPTDataset(name, data_prefix,
+                                  documents, indexed_dataset,
+                                  train_valid_test_num_samples[index],
+                                  seq_length, seed)
+        return dataset
+
+    dataset = build_dataset(dataset_group_name)
+
+    return dataset
+
+
+def _build_train_valid_test_datasets(data_prefix, data_impl, splits_string,
+                                     train_valid_test_num_samples,
+                                     seq_length, seed, skip_warmup):
+    """Build train, valid, and test datasets."""
+
+
+    # Indexed dataset.
+    indexed_dataset = get_indexed_dataset_(data_prefix,
+                                           data_impl,
+                                           skip_warmup)
+
+    total_num_of_documents = indexed_dataset.sizes.shape[0]
+    # splits here is an array of size 4  [train_start_index, valid_start_index, test_start_index, test_end_index]
+    splits = get_train_valid_test_split_(splits_string, total_num_of_documents)
+    # Print stats about the splits.
+    print_rank_0(' > dataset split:')
+
+    def print_split_stats(name, index):
+        print_rank_0('    {}:'.format(name))
+        print_rank_0('     document indices in [{}, {}) total of {} '
+                     'documents'.format(splits[index], splits[index + 1],
+                                        splits[index + 1] - splits[index]))
+    print_split_stats('train', 0)
+    print_split_stats('validation', 1)
+    print_split_stats('test', 2)
+
+    def build_dataset(index, name):
+        dataset = None
+        if splits[index + 1] > splits[index]:
+            documents = np.arange(start=splits[index], stop=splits[index + 1],
+                                  step=1, dtype=np.int32)
+            dataset = GPTDataset(name, data_prefix,
+                                  documents, indexed_dataset,
+                                  train_valid_test_num_samples[index],
+                                  seq_length, seed)
+        return dataset
+
+    train_dataset = build_dataset(0, 'train')
+    valid_dataset = build_dataset(1, 'valid')
+    test_dataset = build_dataset(2, 'test')
+
+    return (train_dataset, valid_dataset, test_dataset)
+
+
+def get_indexed_dataset_(path, data_impl, skip_warmup):
+    """Build indexed dataset."""
+    print_rank_0(' > building dataset index ...')
+    start_time = time.time()
+    indexed_dataset = make_indexed_dataset(path,
+                                           data_impl,
+                                           skip_warmup)
+    print_rank_0(' > finished creating indexed dataset in {:4f} '
+                 'seconds'.format(time.time() - start_time))
+    print_rank_0('    number of documents: {}'.format(
+        indexed_dataset.sizes.shape[0]))
+
+    return indexed_dataset
+
+
+class GPTDataset(torch.utils.data.Dataset):
+
+    def __init__(self, name, data_prefix, documents, indexed_dataset,
+                 num_samples, seq_length, seed):
+
+        self.name = name
+        self.indexed_dataset = indexed_dataset
+
+        # Checks
+        assert np.min(documents) >= 0
+        assert np.max(documents) < indexed_dataset.sizes.shape[0]
+
+        # Build index mappings.
+        self.doc_idx, self.sample_idx, self.shuffle_idx = _build_index_mappings(
+            self.name, data_prefix, documents, self.indexed_dataset.sizes,
+            num_samples, seq_length, seed)
+
+    def __len__(self):
+        # -1 is due to data structure used to retieve the index:
+        #    sample i --> [sample_idx[i], sample_idx[i+1])
+        return self.sample_idx.shape[0] - 1
+
+    def __getitem__(self, idx):
+        # Get the shuffled index.
+        idx = self.shuffle_idx[idx]
+        # Start and end documents and offsets.
+        doc_index_f = self.sample_idx[idx][0]
+        doc_index_l = self.sample_idx[idx + 1][0]
+        offset_f = self.sample_idx[idx][1]
+        offset_l = self.sample_idx[idx + 1][1]
+        # If we are within the same document, just extract the chunk.
+        if doc_index_f == doc_index_l:
+            sample = self.indexed_dataset.get(self.doc_idx[doc_index_f],
+                                              offset=offset_f,
+                                              length=offset_l - offset_f + 1)
+        else:
+            # Otherwise, get the rest of the initial document.
+            sample_list = [self.indexed_dataset.get(self.doc_idx[doc_index_f],
+                                                    offset=offset_f)]
+            # Loop over all in between documents and add the entire document.
+            for i in range(doc_index_f + 1, doc_index_l):
+                sample_list.append(self.indexed_dataset.get(self.doc_idx[i]))
+            # And finally add the relevant portion of last document.
+            sample_list.append(self.indexed_dataset.get(
+                self.doc_idx[doc_index_l],
+                length=offset_l + 1))
+            sample = np.concatenate(sample_list)
+
+        return {'text': np.array(sample, dtype=np.int64)}
+
+
+def _build_index_mappings(name, data_prefix, documents, sizes,
+                          num_samples, seq_length, seed, cutoff_last_epoch=0.95):
+    """Build doc-idx, sample-idx, and shuffle-idx.
+    doc-idx: is an array (ordered) of documents to be used in training.
+    sample-idx: is the start document index and document offset for each
+       training sample.
+    shuffle-idx: maps the sample index into a random index into sample-idx.
+    """
+    # Number of tokens in each epoch and number of required epochs.
+    tokens_per_epoch = _num_tokens(documents, sizes)
+    num_epochs = _num_epochs(tokens_per_epoch, seq_length, num_samples)
+    # rng state
+    np_rng = np.random.RandomState(seed=seed)
+
+    # Filename of the index mappings.
+    _filename = data_prefix
+    _filename += '_{}_indexmap'.format(name)
+    _filename += '_{}ns'.format(num_samples)
+    _filename += '_{}sl'.format(seq_length)
+    _filename += '_{}s'.format(seed)
+    doc_idx_filename = _filename + '_doc_idx.npy'
+    sample_idx_filename = _filename + '_sample_idx.npy'
+    shuffle_idx_filename = _filename + '_shuffle_idx.npy'
+
+    # Build the indexed mapping if not exist.
+    if torch.distributed.get_rank() == 0:
+        if (not os.path.isfile(doc_idx_filename)) or \
+           (not os.path.isfile(sample_idx_filename)) or \
+           (not os.path.isfile(shuffle_idx_filename)):
+
+            print_rank_0(' > WARNING: could not find index map files, building '
+                         'the indices on rank 0 ...')
+
+            # For the last epoch, decide whether include the entire epoch
+            # in the global shuffle or not.
+
+            # If we need only one epoch, then separating last epoch  does
+            # not mean anything.
+            if num_epochs == 1:
+                separate_last_epoch = False
+                print(' > only one epoch required, setting '
+                      'separate_last_epoch to False', flush=True)
+
+            else:
+                # Get the number of samples for the last epoch
+                num_samples_from_epochs_minus_one = (
+                    (num_epochs - 1) * tokens_per_epoch - 1) // seq_length
+                last_epoch_num_samples = num_samples - \
+                                         num_samples_from_epochs_minus_one
+                assert last_epoch_num_samples >= 0, \
+                    f'last epoch number of samples {last_epoch_num_samples} should be non-negative.'
+                num_samples_per_epoch = (tokens_per_epoch - 1) // seq_length
+                assert last_epoch_num_samples <= num_samples_per_epoch, \
+                    f'last epoch number of samples {last_epoch_num_samples} exceeded max value {num_samples_per_epoch}.'
+                # If we have less than cutoff_last_epoch * samples_per_epoch of the samples for the last epoch,
+                # seperate out the epoch and treat it differently.
+                separate_last_epoch = (last_epoch_num_samples <
+                                       int(cutoff_last_epoch * num_samples_per_epoch))
+                if separate_last_epoch:
+                    string = ' > last epoch number of samples ({}) is smaller '\
+                             'than {}% of number of samples per epoch ({}), '\
+                             'setting separate_last_epoch to True'
+                else:
+                    string = ' > last epoch number of samples ({}) is larger '\
+                             'than {}% of number of samples per epoch ({}), '\
+                             'setting separate_last_epoch to False'
+                print(string.format(last_epoch_num_samples, cutoff_last_epoch * 100,
+                                    num_samples_per_epoch), flush=True)
+
+            # doc-idx.
+            start_time = time.time()
+            doc_idx = _build_doc_idx(documents, num_epochs, np_rng,
+                                     separate_last_epoch)
+            np.save(doc_idx_filename, doc_idx, allow_pickle=True)
+            print_rank_0(' > elasped time to build and save doc-idx mapping '
+                         '(seconds): {:4f}'.format(time.time() - start_time))
+            # sample-idx.
+            start_time = time.time()
+            # Use C++ implementation for speed.
+            # First compile and then import.
+            from megatron.data import helpers
+            assert doc_idx.dtype == np.int32
+            assert sizes.dtype == np.int32
+            sample_idx = helpers.build_sample_idx(sizes, doc_idx, seq_length,
+                                                  num_epochs, tokens_per_epoch)
+            # sample_idx = _build_sample_idx(sizes, doc_idx, seq_length,
+            #                               num_epochs, tokens_per_epoch)
+            np.save(sample_idx_filename, sample_idx, allow_pickle=True)
+            print_rank_0(' > elasped time to build and save sample-idx mapping '
+                         '(seconds): {:4f}'.format(time.time() - start_time))
+            # shuffle-idx.
+            start_time = time.time()
+            # -1 is due to data structure used to retieve the index:
+            #    sample i --> [sample_idx[i], sample_idx[i+1])
+            if separate_last_epoch:
+                num_samples_ = num_samples_from_epochs_minus_one
+            else:
+                num_samples_ = sample_idx.shape[0] - 1
+            shuffle_idx = _build_shuffle_idx(num_samples_,
+                                             sample_idx.shape[0] - 1, np_rng)
+            np.save(shuffle_idx_filename, shuffle_idx, allow_pickle=True)
+            print_rank_0(' > elasped time to build and save shuffle-idx mapping'
+                         ' (seconds): {:4f}'.format(time.time() - start_time))
+
+    # This should be a barrier but nccl barrier assumes
+    # device_index=rank which is not the case for model
+    # parallel case
+    counts = torch.cuda.LongTensor([1])
+    torch.distributed.all_reduce(counts, group=mpu.get_data_parallel_group())
+    torch.distributed.all_reduce(counts, group=mpu.get_pipeline_model_parallel_group())
+    assert counts[0].item() == (
+        torch.distributed.get_world_size() //
+        torch.distributed.get_world_size(group=mpu.get_tensor_model_parallel_group()))
+
+    # Load mappings.
+    start_time = time.time()
+    print_rank_0(' > loading doc-idx mapping from {}'.format(
+        doc_idx_filename))
+    doc_idx = np.load(doc_idx_filename, allow_pickle=True, mmap_mode='r')
+    print_rank_0(' > loading sample-idx mapping from {}'.format(
+        sample_idx_filename))
+    sample_idx = np.load(sample_idx_filename, allow_pickle=True, mmap_mode='r')
+    print_rank_0(' > loading shuffle-idx mapping from {}'.format(
+        shuffle_idx_filename))
+    shuffle_idx = np.load(shuffle_idx_filename, allow_pickle=True, mmap_mode='r')
+    print_rank_0('    loaded indexed file in {:3.3f} seconds'.format(
+        time.time() - start_time))
+    print_rank_0('    total number of samples: {}'.format(
+        sample_idx.shape[0]))
+    print_rank_0('    total number of epochs: {}'.format(num_epochs))
+
+    return doc_idx, sample_idx, shuffle_idx
+
+
+def _num_tokens(documents, sizes):
+    """Total number of tokens in the dataset."""
+    return np.sum(sizes[documents])
+
+
+def _num_epochs(tokens_per_epoch, seq_length, num_samples):
+    """Based on number of samples and sequence lenght, calculate how many
+    epochs will be needed."""
+    num_epochs = 0
+    total_tokens = 0
+    while True:
+        num_epochs += 1
+        total_tokens += tokens_per_epoch
+        # -1 is because we need to retrieve seq_length + 1 token each time
+        # but the last token will overlap with the first token of the next
+        # sample except for the last sample.
+        if ((total_tokens - 1) // seq_length) >= num_samples:
+            return num_epochs
+
+
+def _build_doc_idx(documents, num_epochs, np_rng, separate_last_epoch):
+    """Build an array with length = number-of-epochs * number-of-dcuments.
+    Each index is mapped to a corresponding document."""
+    if not separate_last_epoch or num_epochs == 1:
+        doc_idx = np.mgrid[0:num_epochs, 0:len(documents)][1]
+        doc_idx[:] = documents
+        doc_idx = doc_idx.reshape(-1)
+        doc_idx = doc_idx.astype(np.int32)
+        np_rng.shuffle(doc_idx)
+        return doc_idx
+
+    doc_idx_first = _build_doc_idx(documents, num_epochs-1, np_rng, False)
+    doc_idx_last = _build_doc_idx(documents, 1, np_rng, False)
+    return np.concatenate((doc_idx_first, doc_idx_last))
+
+
+def _build_sample_idx(sizes, doc_idx, seq_length,
+                      num_epochs, tokens_per_epoch):
+    """Sample index mapping is a 2D array with sizes
+    [number-of-samples + 1, 2] where [..., 0] contains
+    the index into `doc_idx` and [..., 1] is the
+    starting offset in that document."""
+
+    # Total number of samples. For -1 see comments in `_num_epochs`.
+    num_samples = (num_epochs * tokens_per_epoch - 1) // seq_length
+    sample_idx = np.zeros([num_samples + 1, 2], dtype=np.int32)
+
+    # Index into sample_idx.
+    sample_index = 0
+    # Index into doc_idx.
+    doc_idx_index = 0
+    # Begining offset for each document.
+    doc_offset = 0
+    # Start with first document and no offset.
+    sample_idx[sample_index][0] = doc_idx_index
+    sample_idx[sample_index][1] = doc_offset
+    sample_index += 1
+    while sample_index <= num_samples:
+        # Start with a fresh sequence.
+        remaining_seq_length = seq_length + 1
+        while remaining_seq_length != 0:
+            # Get the document length.
+            doc_id = doc_idx[doc_idx_index]
+            doc_length = sizes[doc_id] - doc_offset
+            # And add it to the current sequence.
+            remaining_seq_length -= doc_length
+            # If we have more than a full sequence, adjust offset and set
+            # remaining length to zero so we return from the while loop.
+            # Note that -1 here is for the same reason we have -1 in
+            # `_num_epochs` calculations.
+            if remaining_seq_length <= 0:
+                doc_offset += (remaining_seq_length + doc_length - 1)
+                remaining_seq_length = 0
+            else:
+                # Otherwise, start from the begining of the next document.
+                doc_idx_index += 1
+                doc_offset = 0
+        # Record the sequence.
+        sample_idx[sample_index][0] = doc_idx_index
+        sample_idx[sample_index][1] = doc_offset
+        sample_index += 1
+
+    return sample_idx
+
+
+def _build_shuffle_idx(num_samples, total_size, np_rng):
+    """Build the range [0, size) and shuffle."""
+    print(' > building shuffle index with split [0, {}) and [{}, {}) '
+          '...'.format(num_samples, num_samples, total_size), flush=True)
+
+    dtype_ = np.uint32
+    if total_size >= (np.iinfo(np.uint32).max - 1):
+        dtype_ = np.int64
+
+    shuffle_idx_first = np.arange(start=0, stop=num_samples,
+                                  step=1, dtype=dtype_)
+    np_rng.shuffle(shuffle_idx_first)
+    if num_samples == total_size:
+        return shuffle_idx_first
+
+    shuffle_idx_last = np.arange(start=num_samples, stop=total_size,
+                                 step=1, dtype=dtype_)
+    np_rng.shuffle(shuffle_idx_last)
+
+    return np.concatenate((shuffle_idx_first, shuffle_idx_last))

megatron-deepspeed_dtk22.10/megatron/data/helpers.cpp

+/*
+ coding=utf-8
+ Copyright (c) 2020, 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.
+ */
+
+
+/* Helper methods for fast index mapping builds */
+
+#include <algorithm>
+#include <iostream>
+#include <limits>
+#include <math.h>
+#include <stdexcept>
+#include <pybind11/pybind11.h>
+#include <pybind11/numpy.h>
+#include <random>
+
+namespace py = pybind11;
+using namespace std;
+
+const int32_t LONG_SENTENCE_LEN = 512;
+
+
+void build_blending_indices(py::array_t<uint8_t>& dataset_index,
+			    py::array_t<int64_t>& dataset_sample_index,
+			    const py::array_t<double>& weights,
+			    const int32_t num_datasets,
+			    const int64_t size, const bool verbose) {
+  /* Given multiple datasets and a weighting array, build samples
+   such that it follows those wieghts.*/
+
+  if (verbose) {
+    std::cout << "> building indices for blendable datasets ..." << std::endl;
+  }
+
+  // Get the pointer access without the checks.
+  auto dataset_index_ptr = dataset_index.mutable_unchecked<1>();
+  auto dataset_sample_index_ptr = dataset_sample_index.mutable_unchecked<1>();
+  auto weights_ptr = weights.unchecked<1>();
+
+  // Initialize buffer for number of samples used for each dataset.
+  int64_t current_samples[num_datasets];
+  for(int64_t i = 0; i < num_datasets; ++i) {
+    current_samples[i] = 0;
+  }
+
+  // For each sample:
+  for(int64_t sample_idx = 0; sample_idx < size; ++sample_idx) {
+
+    // Determine where the max error in sampling is happening.
+    auto sample_idx_double = std::max(static_cast<double>(sample_idx), 1.0);
+    int64_t max_error_index = 0;
+    double max_error = weights_ptr[0] * sample_idx_double -
+      static_cast<double>(current_samples[0]);
+    for (int64_t dataset_idx = 1; dataset_idx < num_datasets; ++dataset_idx) {
+      double error = weights_ptr[dataset_idx] * sample_idx_double -
+	static_cast<double>(current_samples[dataset_idx]);
+      if (error > max_error) {
+	max_error = error;
+	max_error_index = dataset_idx;
+      }
+    }
+
+    // Populate the indices.
+    dataset_index_ptr[sample_idx] = static_cast<uint8_t>(max_error_index);
+    dataset_sample_index_ptr[sample_idx] = current_samples[max_error_index];
+
+    // Update the total samples.
+    current_samples[max_error_index] += 1;
+    
+  }
+
+  // print info
+  if (verbose) {
+    std::cout << " > sample ratios:" << std::endl;
+    for (int64_t dataset_idx = 0; dataset_idx < num_datasets; ++dataset_idx) {
+      auto ratio = static_cast<double>(current_samples[dataset_idx]) /
+	static_cast<double>(size);
+      std::cout << "   dataset " << dataset_idx << ", input: " <<
+	weights_ptr[dataset_idx] << ", achieved: " << ratio << std::endl; 
+    }
+  }
+
+}
+
+
+py::array build_sample_idx(const py::array_t<int32_t>& sizes_,
+			   const py::array_t<int32_t>& doc_idx_,
+			   const int32_t seq_length,
+			   const int32_t num_epochs,
+			   const int64_t tokens_per_epoch) {
+    /* Sample index (sample_idx) is used for gpt2 like dataset for which
+       the documents are flattened and the samples are built based on this
+       1-D flatten array. It is a 2D array with sizes [number-of-samples + 1, 2]
+       where [..., 0] contains the index into `doc_idx` and [..., 1] is the
+       starting offset in that document.*/
+
+    // Consistency checks.
+    assert(seq_length > 1);
+    assert(num_epochs > 0);
+    assert(tokens_per_epoch > 1);
+
+    // Remove bound checks.
+    auto sizes = sizes_.unchecked<1>();
+    auto doc_idx = doc_idx_.unchecked<1>();
+
+    // Mapping and it's length (1D).
+    int64_t num_samples = (num_epochs * tokens_per_epoch - 1) / seq_length;
+    int32_t* sample_idx = new int32_t[2*(num_samples+1)];
+
+    cout << "    using:" << endl << std::flush;
+    cout << "     number of documents:       " <<
+      doc_idx_.shape(0) / num_epochs << endl << std::flush;
+    cout << "     number of epochs:          " << num_epochs <<
+      endl << std::flush;
+    cout << "     sequence length:           " << seq_length <<
+      endl << std::flush;
+    cout << "     total number of samples:   " << num_samples <<
+      endl << std::flush;
+
+    // Index into sample_idx.
+    int64_t sample_index = 0;
+    // Index into doc_idx.
+    int64_t doc_idx_index = 0;
+    // Begining offset for each document.
+    int32_t doc_offset = 0;
+    // Start with first document and no offset.
+    sample_idx[2 * sample_index] = doc_idx_index;
+    sample_idx[2 * sample_index + 1] = doc_offset;
+    ++sample_index;
+
+    while (sample_index <= num_samples) {
+        // Start with a fresh sequence.
+      int32_t remaining_seq_length = seq_length + 1;
+      while (remaining_seq_length != 0) {
+            // Get the document length.
+	auto doc_id = doc_idx[doc_idx_index];
+	auto doc_length = sizes[doc_id] - doc_offset;
+	// And add it to the current sequence.
+	remaining_seq_length -= doc_length;
+	// If we have more than a full sequence, adjust offset and set
+	// remaining length to zero so we return from the while loop.
+	// Note that -1 here is for the same reason we have -1 in
+	// `_num_epochs` calculations.
+	if (remaining_seq_length <= 0) {
+	  doc_offset += (remaining_seq_length + doc_length - 1);
+	  remaining_seq_length = 0;
+	} else {
+	  // Otherwise, start from the begining of the next document.
+	  ++doc_idx_index;
+	  doc_offset = 0;
+	}
+      }
+      // Record the sequence.
+      sample_idx[2 * sample_index] = doc_idx_index;
+      sample_idx[2 * sample_index + 1] = doc_offset;
+      ++sample_index;
+    }
+
+    // Method to deallocate memory.
+    py::capsule free_when_done(sample_idx, [](void *mem_) {
+	int32_t *mem = reinterpret_cast<int32_t*>(mem_);
+	delete[] mem;
+      });
+
+    // Return the numpy array.
+    const auto byte_size = sizeof(int32_t);
+    return py::array(std::vector<int64_t>{num_samples+1, 2}, // shape
+                     {2*byte_size, byte_size}, // C-style contiguous strides
+                     sample_idx, // the data pointer
+                     free_when_done); // numpy array references
+    
+}
+
+
+inline int32_t get_target_sample_len(const int32_t short_seq_ratio,
+				     const int32_t max_length,
+				     std::mt19937& rand32_gen) {
+    /* Training sample length. */
+    if (short_seq_ratio == 0) {
+      return max_length;
+    }
+    const auto random_number = rand32_gen();
+    if ((random_number % short_seq_ratio) == 0) {
+      return 2 + random_number % (max_length - 1);
+    }
+    return max_length;
+}
+
+
+template<typename DocIdx>
+py::array build_mapping_impl(const py::array_t<int64_t>& docs_,
+                             const py::array_t<int32_t>& sizes_,
+                             const int32_t num_epochs,
+                             const uint64_t max_num_samples,
+                             const int32_t max_seq_length,
+                             const double short_seq_prob,
+                             const int32_t seed,
+			     const bool verbose,
+			     const int32_t min_num_sent) {
+    /* Build a mapping of (start-index, end-index, sequence-length) where
+       start and end index are the indices of the sentences in the sample
+       and sequence-length is the target sequence length.
+    */
+
+    // Consistency checks.
+    assert(num_epochs > 0);
+    assert(max_seq_length > 1);
+    assert(short_seq_prob >= 0.0);
+    assert(short_seq_prob <= 1.0);
+    assert(seed > 0);
+
+    // Remove bound checks.
+    auto docs = docs_.unchecked<1>();
+    auto sizes = sizes_.unchecked<1>();
+
+    // For efficiency, convert probability to ratio. Note: rand() generates int.
+    int32_t short_seq_ratio = 0;
+    if (short_seq_prob > 0) {
+      short_seq_ratio = static_cast<int32_t>(round(1.0 / short_seq_prob));
+    }
+
+    if (verbose) {
+        const auto sent_start_index = docs[0];
+	const auto sent_end_index = docs[docs_.shape(0) - 1];
+	const auto num_sentences = sent_end_index - sent_start_index;
+	cout << "    using:" << endl << std::flush;
+	cout << "     number of documents:            " << docs_.shape(0) - 1 <<
+	  endl << std::flush;
+	cout << "     sentences range:                [" << sent_start_index <<
+	", " << sent_end_index << ")" << endl << std::flush;
+	cout << "     total number of sentences:      " << num_sentences <<
+	  endl << std::flush;
+	cout << "     number of epochs:               " << num_epochs <<
+	  endl << std::flush;
+	cout << "     maximum number of samples:      " << max_num_samples <<
+	  endl << std::flush;
+	cout << "     maximum sequence length:        " << max_seq_length <<
+	  endl << std::flush;
+	cout << "     short sequence probability:     " << short_seq_prob <<
+	endl << std::flush;
+	cout << "     short sequence ration (1/prob): " << short_seq_ratio <<
+	  endl << std::flush;
+	cout << "     seed:                           " << seed << endl <<
+	  std::flush;
+    }
+
+    // Mapping and it's length (1D).
+    int64_t num_samples = -1;
+    DocIdx* maps = NULL;
+
+    // Perform two iterations, in the first iteration get the size
+    // and allocate memory and in the second iteration populate the map.
+    bool second = false;
+    for (int32_t iteration=0; iteration<2; ++iteration) {
+
+        // Set the seed so both iterations produce the same results.
+        std::mt19937 rand32_gen(seed);
+
+        // Set the flag on second iteration.
+        second = (iteration == 1);
+
+        // Counters:
+        uint64_t empty_docs = 0;
+        uint64_t one_sent_docs = 0;
+	uint64_t long_sent_docs = 0;
+
+        // Current map index.
+        uint64_t map_index = 0;
+
+        // For each epoch:
+        for (int32_t epoch=0; epoch<num_epochs; ++epoch) {
+            if (map_index >= max_num_samples) {
+	        if (verbose && (!second)) {
+		  cout << "    reached " << max_num_samples << " samples after "
+		       << epoch << " epochs ..." << endl << std::flush;
+		}
+                break;
+            }
+            // For each document:
+            for (int32_t doc=0; doc<(docs.shape(0) - 1); ++doc) {
+
+                // Document sentences are in [sent_index_first, sent_index_last)
+                const auto sent_index_first = docs[doc];
+                const auto sent_index_last = docs[doc + 1];
+
+                // At the begining of the document previous index is the
+		// start index.
+                auto prev_start_index = sent_index_first;
+
+                // Remaining documents.
+                auto num_remain_sent = sent_index_last - sent_index_first;
+
+                // Some bookkeeping
+                if ((epoch == 0) && (!second)) {
+                    if (num_remain_sent == 0) {
+		        ++empty_docs;
+                    }
+                    if (num_remain_sent == 1) {
+		        ++one_sent_docs;
+                    }
+                }
+
+		// Detect documents with long sentences.
+		bool contains_long_sentence = false;
+		if (num_remain_sent > 1) {
+		    for (auto sent_index=sent_index_first;
+			 sent_index < sent_index_last; ++sent_index) {
+		        if (sizes[sent_index] > LONG_SENTENCE_LEN){
+			    if ((epoch == 0) && (!second)) {
+			        ++long_sent_docs;
+			    }
+			    contains_long_sentence = true;
+			    break;
+			}
+		    }
+		}
+
+                // If we have more than two sentences.
+                if ((num_remain_sent >= min_num_sent) && (!contains_long_sentence)) {
+
+                    // Set values.
+                    auto seq_len = int32_t{0};
+                    auto num_sent = int32_t{0};
+                    auto target_seq_len = get_target_sample_len(short_seq_ratio,
+								max_seq_length,
+								rand32_gen);
+
+                    // Loop through sentences.
+                    for (auto sent_index=sent_index_first;
+                         sent_index < sent_index_last; ++sent_index) {
+
+		        // Add the size and number of sentences.
+		        seq_len += sizes[sent_index];
+		        ++num_sent;
+			--num_remain_sent;
+
+			// If we have reached the target length.
+			// and if not only one sentence is left in the document.
+			// and if we have at least two sentneces.
+			// and if we have reached end of the document.
+			if (((seq_len >= target_seq_len) &&
+			     (num_remain_sent > 1) &&
+			     (num_sent >= min_num_sent) ) || (num_remain_sent == 0)) {
+
+			    // Check for overflow.
+			    if ((3 * map_index + 2) >
+				std::numeric_limits<int64_t>::max()) {
+			        cout << "number of samples exceeded maximum "
+				     << "allowed by type int64: "
+				     << std::numeric_limits<int64_t>::max()
+				     << endl;
+				throw std::overflow_error("Number of samples");
+			    }
+
+			    // Populate the map.
+			    if (second) {
+			        const auto map_index_0 = 3 * map_index;
+				maps[map_index_0] = static_cast<DocIdx>(prev_start_index);
+				maps[map_index_0 + 1] = static_cast<DocIdx>(sent_index + 1);
+				maps[map_index_0 + 2] = static_cast<DocIdx>(target_seq_len);
+			    }
+
+			    // Update indices / counters.
+			    ++map_index;
+			    prev_start_index = sent_index + 1;
+			    target_seq_len = get_target_sample_len(short_seq_ratio,
+								   max_seq_length,
+								   rand32_gen);
+			    seq_len = 0;
+			    num_sent = 0;
+			}
+
+                    } // for (auto sent_index=sent_index_first; ...
+                } // if (num_remain_sent > 1) {
+            } // for (int doc=0; doc < num_docs; ++doc) {
+        } // for (int epoch=0; epoch < num_epochs; ++epoch) {
+
+        if (!second) {
+	    if (verbose) {
+	        cout << "   number of empty documents: " << empty_docs <<
+		  endl << std::flush;
+		cout << "   number of documents with one sentence: " <<
+		  one_sent_docs << endl << std::flush;
+		cout << "   number of documents with long sentences: " <<
+		  long_sent_docs << endl << std::flush;
+		cout << "   will create mapping for " << map_index <<
+		  " samples" << endl << std::flush;
+	    }
+	    assert(maps == NULL);
+	    assert(num_samples < 0);
+            maps = new DocIdx[3*map_index];
+            num_samples = static_cast<int64_t>(map_index);
+        }
+
+    } // for (int iteration=0; iteration < 2; ++iteration) {
+
+    // Shuffle.
+    // We need a 64 bit random number generator as we might have more
+    // than 2 billion samples.
+    std::mt19937_64 rand64_gen(seed + 1);
+    for (auto i=(num_samples - 1); i > 0; --i) {
+      const auto j = static_cast<int64_t>(rand64_gen() % (i + 1));
+      const auto i0 = 3 * i;
+      const auto j0 = 3 * j;
+      // Swap values.
+      swap(maps[i0], maps[j0]);
+      swap(maps[i0 + 1], maps[j0 + 1]);
+      swap(maps[i0 + 2], maps[j0 + 2]);
+    }
+
+    // Method to deallocate memory.
+    py::capsule free_when_done(maps, [](void *mem_) {
+            DocIdx *mem = reinterpret_cast<DocIdx*>(mem_);
+	    delete[] mem;
+        });
+
+    // Return the numpy array.
+    const auto byte_size = sizeof(DocIdx);
+    return py::array(std::vector<int64_t>{num_samples, 3}, // shape
+                     {3*byte_size, byte_size}, // C-style contiguous strides
+                     maps, // the data pointer
+                     free_when_done); // numpy array references
+
+}
+
+
+py::array build_mapping(const py::array_t<int64_t>& docs_,
+                        const py::array_t<int>& sizes_,
+                        const int num_epochs,
+                        const uint64_t max_num_samples,
+                        const int max_seq_length,
+                        const double short_seq_prob,
+                        const int seed,
+			const bool verbose,
+			const int32_t min_num_sent) {
+
+    if (sizes_.size() > std::numeric_limits<uint32_t>::max()) {
+        if (verbose) {
+	   cout << "    using uint64 for data mapping..." << endl << std::flush;
+	}
+	return build_mapping_impl<uint64_t>(docs_, sizes_, num_epochs,
+					    max_num_samples, max_seq_length,
+					    short_seq_prob, seed, verbose,
+					    min_num_sent);
+    } else {
+       if (verbose) {
+	   cout << "    using uint32 for data mapping..." << endl << std::flush;
+       }
+       return build_mapping_impl<uint32_t>(docs_, sizes_, num_epochs,
+					   max_num_samples, max_seq_length,
+					   short_seq_prob, seed, verbose,
+					   min_num_sent);
+    }
+}
+
+template<typename DocIdx>
+py::array build_blocks_mapping_impl(const py::array_t<int64_t>& docs_,
+                                    const py::array_t<int32_t>& sizes_,
+                                    const py::array_t<int32_t>& titles_sizes_,
+                                    const int32_t num_epochs,
+                                    const uint64_t max_num_samples,
+                                    const int32_t max_seq_length,
+                                    const int32_t seed,
+                                    const bool verbose,
+                                    const bool use_one_sent_blocks) {
+    /* Build a mapping of (start-index, end-index, sequence-length) where
+       start and end index are the indices of the sentences in the sample
+       and sequence-length is the target sequence length.
+    */
+
+    // Consistency checks.
+    assert(num_epochs > 0);
+    assert(max_seq_length > 1);
+    assert(seed > 0);
+
+    // Remove bound checks.
+    auto docs = docs_.unchecked<1>();
+    auto sizes = sizes_.unchecked<1>();
+    auto titles_sizes = titles_sizes_.unchecked<1>();
+
+    if (verbose) {
+        const auto sent_start_index = docs[0];
+        const auto sent_end_index = docs[docs_.shape(0) - 1];
+        const auto num_sentences = sent_end_index - sent_start_index;
+        cout << "    using:" << endl << std::flush;
+        cout << "     number of documents:            " << docs_.shape(0) - 1 <<
+          endl << std::flush;
+        cout << "     sentences range:                [" << sent_start_index <<
+        ", " << sent_end_index << ")" << endl << std::flush;
+        cout << "     total number of sentences:      " << num_sentences <<
+          endl << std::flush;
+        cout << "     number of epochs:               " << num_epochs <<
+          endl << std::flush;
+        cout << "     maximum number of samples:      " << max_num_samples <<
+          endl << std::flush;
+        cout << "     maximum sequence length:        " << max_seq_length <<
+          endl << std::flush;
+        cout << "     seed:                           " << seed << endl <<
+          std::flush;
+    }
+
+    // Mapping and its length (1D).
+    int64_t num_samples = -1;
+    DocIdx* maps = NULL;
+
+    // Acceptable number of sentences per block.
+    int min_num_sent = 2;
+    if (use_one_sent_blocks) {
+        min_num_sent = 1;
+    }
+
+    // Perform two iterations, in the first iteration get the size
+    // and allocate memory and in the second iteration populate the map.
+    bool second = false;
+    for (int32_t iteration=0; iteration<2; ++iteration) {
+
+        // Set the flag on second iteration.
+        second = (iteration == 1);
+
+        // Current map index.
+        uint64_t map_index = 0;
+
+        uint64_t empty_docs = 0;
+        uint64_t one_sent_docs = 0;
+        uint64_t long_sent_docs = 0;
+        // For each epoch:
+        for (int32_t epoch=0; epoch<num_epochs; ++epoch) {
+            // assign every block a unique id
+            int32_t block_id = 0;
+
+            if (map_index >= max_num_samples) {
+                if (verbose && (!second)) {
+                cout << "    reached " << max_num_samples << " samples after "
+                     << epoch << " epochs ..." << endl << std::flush;
+                }
+                break;
+            }
+            // For each document:
+            for (int32_t doc=0; doc<(docs.shape(0) - 1); ++doc) {
+
+                // Document sentences are in [sent_index_first, sent_index_last)
+                const auto sent_index_first = docs[doc];
+                const auto sent_index_last = docs[doc + 1];
+                const auto target_seq_len = max_seq_length - titles_sizes[doc];
+
+                // At the begining of the document previous index is the
+                // start index.
+                auto prev_start_index = sent_index_first;
+
+                // Remaining documents.
+                auto num_remain_sent = sent_index_last - sent_index_first;
+
+                // Some bookkeeping
+                if ((epoch == 0) && (!second)) {
+                    if (num_remain_sent == 0) {
+		                ++empty_docs;
+                    }
+                    if (num_remain_sent == 1) {
+		                ++one_sent_docs;
+                    }
+                }
+                // Detect documents with long sentences.
+                bool contains_long_sentence = false;
+                if (num_remain_sent >= min_num_sent) {
+                    for (auto sent_index=sent_index_first;
+                    sent_index < sent_index_last; ++sent_index) {
+                        if (sizes[sent_index] > LONG_SENTENCE_LEN){
+                            if ((epoch == 0) && (!second)) {
+                                ++long_sent_docs;
+                            }
+                            contains_long_sentence = true;
+                            break;
+                        }
+                    }
+                }
+                // If we have enough sentences and no long sentences.
+                if ((num_remain_sent >= min_num_sent) && (!contains_long_sentence)) {
+
+                    // Set values.
+                    auto seq_len = int32_t{0};
+                    auto num_sent = int32_t{0};
+
+                    // Loop through sentences.
+                    for (auto sent_index=sent_index_first;
+                         sent_index < sent_index_last; ++sent_index) {
+
+                            // Add the size and number of sentences.
+                            seq_len += sizes[sent_index];
+                            ++num_sent;
+                            --num_remain_sent;
+
+                        // If we have reached the target length.
+                        // and there are an acceptable number of sentences left
+                        // and if we have at least the minimum number of sentences.
+                        // or if we have reached end of the document.
+                        if (((seq_len >= target_seq_len) &&
+                             (num_remain_sent >= min_num_sent) &&
+                             (num_sent >= min_num_sent) ) || (num_remain_sent == 0)) {
+
+                            // Populate the map.
+                            if (second) {
+                                const auto map_index_0 = 4 * map_index;
+                                // Each sample has 4 items: the starting sentence index, ending sentence index,
+                                // the index of the document from which the block comes (used for fetching titles)
+                                // and the unique id of the block (used for creating block indexes)
+
+                                maps[map_index_0] = static_cast<DocIdx>(prev_start_index);
+                                maps[map_index_0 + 1] = static_cast<DocIdx>(sent_index + 1);
+                                maps[map_index_0 + 2] = static_cast<DocIdx>(doc);
+                                maps[map_index_0 + 3] = static_cast<DocIdx>(block_id);
+                            }
+
+                            // Update indices / counters.
+                            ++map_index;
+                            ++block_id;
+                            prev_start_index = sent_index + 1;
+                            seq_len = 0;
+                            num_sent = 0;
+                        }
+                    } // for (auto sent_index=sent_index_first; ...
+                } // if (num_remain_sent > 1) {
+            } // for (int doc=0; doc < num_docs; ++doc) {
+        } // for (int epoch=0; epoch < num_epochs; ++epoch) {
+
+        if (!second) {
+            if (verbose) {
+	        cout << "   number of empty documents: " << empty_docs <<
+              endl << std::flush;
+            cout << "   number of documents with one sentence: " <<
+              one_sent_docs << endl << std::flush;
+            cout << "   number of documents with long sentences: " <<
+              long_sent_docs << endl << std::flush;
+            cout << "   will create mapping for " << map_index <<
+              " samples" << endl << std::flush;
+            }
+            assert(maps == NULL);
+            assert(num_samples < 0);
+            maps = new DocIdx[4*map_index];
+            num_samples = static_cast<int64_t>(map_index);
+        }
+
+    } // for (int iteration=0; iteration < 2; ++iteration) {
+
+    // Shuffle.
+    // We need a 64 bit random number generator as we might have more
+    // than 2 billion samples.
+    std::mt19937_64 rand64_gen(seed + 1);
+    for (auto i=(num_samples - 1); i > 0; --i) {
+        const auto j = static_cast<int64_t>(rand64_gen() % (i + 1));
+        const auto i0 = 4 * i;
+        const auto j0 = 4 * j;
+        // Swap values.
+        swap(maps[i0], maps[j0]);
+        swap(maps[i0 + 1], maps[j0 + 1]);
+        swap(maps[i0 + 2], maps[j0 + 2]);
+        swap(maps[i0 + 3], maps[j0 + 3]);
+    }
+
+    // Method to deallocate memory.
+    py::capsule free_when_done(maps, [](void *mem_) {
+            DocIdx *mem = reinterpret_cast<DocIdx*>(mem_);
+	    delete[] mem;
+        });
+
+    // Return the numpy array.
+    const auto byte_size = sizeof(DocIdx);
+    return py::array(std::vector<int64_t>{num_samples, 4}, // shape
+                     {4*byte_size, byte_size}, // C-style contiguous strides
+                     maps, // the data pointer
+                     free_when_done); // numpy array references
+
+}
+
+py::array build_blocks_mapping(const py::array_t<int64_t>& docs_,
+                               const py::array_t<int>& sizes_,
+                               const py::array_t<int>& titles_sizes_,
+                               const int num_epochs,
+                               const uint64_t max_num_samples,
+                               const int max_seq_length,
+                               const int seed,
+                    const bool verbose,
+                    const bool use_one_sent_blocks) {
+
+    if (sizes_.size() > std::numeric_limits<uint32_t>::max()) {
+        if (verbose) {
+	   cout << "    using uint64 for data mapping..." << endl << std::flush;
+	}
+	return build_blocks_mapping_impl<uint64_t>(docs_, sizes_, titles_sizes_,
+	                    num_epochs, max_num_samples, max_seq_length, seed, verbose, use_one_sent_blocks);
+    } else {
+       if (verbose) {
+	   cout << "    using uint32 for data mapping..." << endl << std::flush;
+       }
+       return build_blocks_mapping_impl<uint32_t>(docs_, sizes_, titles_sizes_,
+                        num_epochs, max_num_samples, max_seq_length, seed, verbose, use_one_sent_blocks);
+    }
+}
+
+PYBIND11_MODULE(helpers, m) {
+    m.def("build_mapping", &build_mapping);
+    m.def("build_blocks_mapping", &build_blocks_mapping);
+    m.def("build_sample_idx", &build_sample_idx);
+    m.def("build_blending_indices", &build_blending_indices);
+}