initial llama

3rdparty/Megatron-LM/examples/sc21/SBATCH.sh

+#!/bin/bash
+
+
+sbatch -p ${SLURM_PARTITION} \
+       -A ${SLURM_ACCOUNT} \
+       --job-name=${JOB_NAME} \
+       --nodes=${NNODES} \
+       --export=MEGATRON_CODE_DIR,MEGATRON_PARAMS,DOCKER_MOUNT_DIR SRUN.sh
+
+exit 0
+
+
+

3rdparty/Megatron-LM/examples/sc21/SRUN.sh

+#!/bin/bash
+
+#SBATCH -t 0:30:00 --exclusive --mem=0 --overcommit --ntasks-per-node=8
+
+
+THIS_DIR=`pwd`
+DATETIME=`date +'date_%y-%m-%d_time_%H-%M-%S'`
+mkdir -p ${THIS_DIR}/logs
+
+
+CMD="python -u ${MEGATRON_CODE_DIR}/pretrain_gpt.py ${MEGATRON_PARAMS}"
+
+
+srun -l \
+     --container-image "nvcr.io#nvidia/pytorch:20.12-py3" \
+     --container-mounts "${THIS_DIR}:${THIS_DIR},${MEGATRON_CODE_DIR}:${MEGATRON_CODE_DIR},${DOCKER_MOUNT_DIR}:${DOCKER_MOUNT_DIR}" \
+     --output=${THIS_DIR}/logs/%x_%j_$DATETIME.log sh -c "${CMD}"
+

3rdparty/Megatron-LM/examples/sc21/run_figure_11.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Pipeline-parallel size options = [1, 2, 4, 8].
+PP=1
+
+# Batch size (global batch size) options = [8, 128].
+GBS=8
+
+
+
+
+
+# Set pipeline-parallel size options.
+NLS=$((3*PP))
+NNODES=${PP}
+
+
+# Other params.
+TP=8
+MBS=1
+HS=20480
+NAH=128
+DDP=local
+MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+
+
+# Name of the job.
+export JOB_NAME=results_figure_11_pipeline_parallel_size_${PP}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

3rdparty/Megatron-LM/examples/sc21/run_figure_12.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Interleaved schedule options = [YES, NO].
+INTERLEAVED=YES
+
+# Batch size (global batch size) options = [12, 24, 36, ..., 60].
+GBS=12
+
+
+
+
+
+# Set interleaved schedule options.
+if [ ${INTERLEAVED} == "YES" ]; then
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform --num-layers-per-virtual-pipeline-stage 2 "
+elif [ ${INTERLEAVED} == "NO" ]; then
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+else
+    echo "Invalid configuration"
+    exit 1
+fi
+
+
+# Other params.
+TP=8
+PP=12
+MBS=1
+NLS=96
+HS=12288
+NAH=96
+DDP=local
+NNODES=12
+
+
+# Name of the job.
+export JOB_NAME=results_figure_12_interleaved_${INTERLEAVED}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

3rdparty/Megatron-LM/examples/sc21/run_figure_13.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Pipeline-parallel size options = [2, 4, 8, 16, 32].
+PP=2
+
+# Batch size (global batch size) options = [32, 128].
+GBS=32
+
+
+
+
+
+# Set pipeline-parallel and tensor-parallel size options.
+TP=$((64/PP))
+
+
+# Other params.
+MBS=1
+NLS=32
+HS=20480
+NAH=128
+DDP=local
+MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+NNODES=8
+
+
+# Name of the job.
+export JOB_NAME=results_figure_13_pipeline_parallel_size_${PP}_tensor_parallel_size_${TP}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

3rdparty/Megatron-LM/examples/sc21/run_figure_14.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Pipeline-parallel size options = [2, 4, 8, 16, 32].
+PP=2
+
+# Batch size (global batch size) options = [32, 512].
+GBS=32
+
+
+
+
+
+# Set pipeline-parallel and data-parallel size options.
+DP=$((64/PP))
+
+
+# Other params.
+TP=1
+MBS=1
+NLS=32
+HS=3840
+NAH=32
+DDP=local
+MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+NNODES=8
+
+
+# Name of the job.
+export JOB_NAME=results_figure_14_pipeline_parallel_size_${PP}_data_parallel_size_${DP}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

3rdparty/Megatron-LM/examples/sc21/run_figure_15.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Tensor-parallel size options = [2, 4, 8, 16, 32].
+TP=2
+
+# Batch size (global batch size) options = [32, 128, 512].
+GBS=32
+
+
+
+
+
+# Set tensor-parallel and data-parallel size options.
+DP=$((64/TP))
+
+
+# Other params.
+PP=1
+MBS=1
+NLS=32
+HS=3840
+NAH=32
+DDP=local
+MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+NNODES=8
+
+
+# Name of the job.
+export JOB_NAME=results_figure_15_tensor_parallel_size_${TP}_data_parallel_size_${DP}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

3rdparty/Megatron-LM/examples/sc21/run_figure_16.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Microbatch size options = [1, 2, 4, 8].
+MBS=1
+
+# Batch size (global batch size) options = [128, 512].
+GBS=128
+
+
+
+
+
+# Other params.
+TP=8
+PP=8
+NLS=32
+HS=15360
+NAH=128
+DDP=local
+MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+NNODES=8
+
+
+# Name of the job.
+export JOB_NAME=results_figure_16_microbatch_size_${MBS}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

3rdparty/Megatron-LM/examples/sc21/run_figure_17.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Activation recomputation options = [YES, NO].
+ACTIVATION_RECOMPUTATION=YES
+
+# Batch size (global batch size) options = [1, 2, 4, ..., 256].
+GBS=1
+
+
+
+
+
+# Set activation recomputation.
+if [ ${ACTIVATION_RECOMPUTATION} == "YES" ]; then
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+elif [ ${ACTIVATION_RECOMPUTATION} == "NO" ]; then
+    MEGATRON_EXTRA_PARAMS=""
+else
+    echo "Invalid configuration"
+    exit 1
+fi
+
+
+# Other params.
+TP=8
+PP=16
+MBS=1
+NLS=80
+HS=12288
+NAH=96
+DDP=local
+NNODES=16
+
+
+# Name of the job.
+export JOB_NAME=results_figure_17_activation_recomputation_${ACTIVATION_RECOMPUTATION}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

3rdparty/Megatron-LM/examples/sc21/run_figure_18.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+
+# Scatter-gather communication optimization options = [YES, NO].
+SCATTER_GATHER=YES
+
+# Batch size (global batch size) options = [12, 24, 36, ..., 60].
+GBS=12
+
+
+
+
+
+# Set scatter-gather communication optimization options.
+if [ ${SCATTER_GATHER} == "YES" ]; then
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform --num-layers-per-virtual-pipeline-stage 2 "
+elif [ ${SCATTER_GATHER} == "NO" ]; then
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform --num-layers-per-virtual-pipeline-stage 2 --no-scatter-gather-tensors-in-pipeline "
+else
+    echo "Invalid configuration"
+    exit 1
+fi
+
+
+# Other params.
+TP=8
+PP=12
+MBS=1
+NLS=96
+HS=12288
+NAH=96
+DDP=local
+NNODES=12
+
+
+# Name of the job.
+export JOB_NAME=results_figure_18_scatter_gather_${SCATTER_GATHER}_batch_size_${GBS}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

3rdparty/Megatron-LM/examples/sc21/run_table_1.sh

+#!/bin/bash
+
+# ================================
+# Choose the case to run.
+# ================================
+# model size options = [1.7B, 3.6B, 7.5B, 18B, 39B, 76B, 145B, 310B, 530B, 1T]
+MODEL_SIZE=1.7B
+
+
+
+
+
+
+if [ ${MODEL_SIZE} == "1.7B" ]; then
+    TP=1
+    PP=1
+    MBS=16
+    GBS=512
+    NLS=24
+    HS=2304
+    NAH=24
+    DDP=torch
+    NNODES=4
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+elif [ ${MODEL_SIZE} == "3.6B" ]; then
+    TP=2
+    PP=1
+    MBS=16
+    GBS=512
+    NLS=30
+    HS=3072
+    NAH=32
+    DDP=torch
+    NNODES=8
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+elif [ ${MODEL_SIZE} == "7.5B" ]; then
+    TP=4
+    PP=1
+    MBS=16
+    GBS=512
+    NLS=36
+    HS=4096
+    NAH=32
+    DDP=torch
+    NNODES=16
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+elif [ ${MODEL_SIZE} == "18B" ]; then
+    TP=8
+    PP=1
+    MBS=8
+    GBS=1024
+    NLS=40
+    HS=6144
+    NAH=48
+    DDP=torch
+    NNODES=32
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+elif [ ${MODEL_SIZE} == "39B" ]; then
+    TP=8
+    PP=2
+    MBS=4
+    GBS=1536
+    NLS=48
+    HS=8192
+    NAH=64
+    DDP=local
+    NNODES=64
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+elif [ ${MODEL_SIZE} == "76B" ]; then
+    TP=8
+    PP=4
+    MBS=2
+    GBS=1792
+    NLS=60
+    HS=10240
+    NAH=80
+    DDP=local
+    NNODES=128
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform --num-layers-per-virtual-pipeline-stage 5"
+elif [ ${MODEL_SIZE} == "145B" ]; then
+    TP=8
+    PP=8
+    MBS=2
+    GBS=2304
+    NLS=80
+    HS=12288
+    NAH=96
+    DDP=local
+    NNODES=192
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform --num-layers-per-virtual-pipeline-stage 5 "
+elif [ ${MODEL_SIZE} == "310B" ]; then
+    TP=8
+    PP=16
+    MBS=1
+    GBS=2160
+    NLS=96
+    HS=16384
+    NAH=128
+    DDP=local
+    NNODES=240
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform --num-layers-per-virtual-pipeline-stage 3 "
+elif [ ${MODEL_SIZE} == "530B" ]; then
+    TP=8
+    PP=35
+    MBS=1
+    GBS=2520
+    NLS=105
+    HS=20480
+    NAH=128
+    DDP=local
+    NNODES=315
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform --num-layers-per-virtual-pipeline-stage 1 "
+elif [ ${MODEL_SIZE} == "1T" ]; then
+    TP=8
+    PP=64
+    MBS=1
+    GBS=3072
+    NLS=128
+    HS=25600
+    NAH=160
+    DDP=local
+    NNODES=384
+    MEGATRON_EXTRA_PARAMS="--activations-checkpoint-method uniform "
+else
+    echo "Invalid configuration"
+    exit 1
+fi
+
+
+# Name of the job
+export JOB_NAME=results_table_1_model_size_${MODEL_SIZE}
+
+
+# Import the configs.
+. `pwd`/CONFIG.sh
+
+
+# Submit the job.
+. `pwd`/SBATCH.sh
+
+
+exit 0
+
+
+

3rdparty/Megatron-LM/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 .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 get_signal_handler
+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)

3rdparty/Megatron-LM/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 os
+
+import torch
+
+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_inference_args(parser)
+
+    # Custom arguments.
+    if extra_args_provider is not None:
+        parser = extra_args_provider(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)
+    if args.pipeline_model_parallel_size > 1:
+        if args.pipeline_model_parallel_split_rank is not None:
+            assert args.pipeline_model_parallel_split_rank < \
+                    args.pipeline_model_parallel_size, 'split rank needs'\
+                    ' to be less than pipeline model parallel size ({})'.format(
+                            args.pipeline_model_parallel_size)
+
+    # 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
+    if args.checkpoint_activations:
+        args.activations_checkpoint_method = 'uniform'
+        if args.rank == 0:
+            print('--checkpoint-activations is no longer valid, '
+                  'use --activation-checkpoint-method instead. '
+                  'Defaulting to activation-checkpoint-method=uniform.')
+    del args.checkpoint_activations
+
+    # 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 make sure use-contiguous-buffers-in-local-ddp is not off.
+    if args.accumulate_allreduce_grads_in_fp32:
+        assert args.DDP_impl == 'local'
+        assert args.use_contiguous_buffers_in_local_ddp
+
+    # For torch DDP, we do not use contiguous buffer
+    if args.DDP_impl == 'torch':
+        args.use_contiguous_buffers_in_local_ddp = False
+
+    if args.dataloader_type is None:
+        args.dataloader_type = 'single'
+
+    # Consumed tokens.
+    args.consumed_train_samples = 0
+    args.consumed_valid_samples = 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',
+                     'max_position_embeddings']
+    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.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
+    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.'
+
+    TORCH_MAJOR = int(torch.__version__.split('.')[0])
+    TORCH_MINOR = int(torch.__version__.split('.')[1])
+    # Persistent fused layer norm.
+    if TORCH_MAJOR < 1 or (TORCH_MAJOR == 1 and TORCH_MINOR < 11):
+        args.no_persist_layer_norm = True
+        if args.rank == 0:
+            print('Persistent fused layer norm kernel is supported from '
+                  'pytorch v1.11 (nvidia pytorch container paired with v1.11). '
+                  'Defaulting to no_persist_layer_norm=True')
+
+    # Activation checkpointing.
+    if args.distribute_checkpointed_activations:
+        assert args.tensor_model_parallel_size > 1, 'can distribute ' \
+            'checkpointed activations only across tensor model ' \
+            'parallel groups'
+        assert args.activations_checkpoint_method is not None, \
+            'for distributed checkpoint activations to work you '\
+            'need to use a activation-checkpoint method '
+        assert TORCH_MAJOR >= 1 and TORCH_MINOR >= 10, \
+            'distributed checkpoint activations are supported for pytorch ' \
+            'v1.10 and above (Nvidia Pytorch container >= 21.07). Current ' \
+            'pytorch version is v%s.%s.' % (TORCH_MAJOR, TORCH_MINOR)
+
+    _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)))
+        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_inference_args(parser):
+    group = parser.add_argument_group(title='inference')
+
+    group.add_argument('--inference-batch-times-seqlen-threshold',
+                       type=int, default=512,
+                       help='During inference, if batch-size times '
+                       'sequence-length is smaller than this threshold '
+                       'then we will not use pipelining, otherwise we will.')
+
+    return parser
+
+    
+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('--layernorm-epsilon', type=float, default=1e-5,
+                       help='Layer norm epsilon.')
+    group.add_argument('--apply-residual-connection-post-layernorm',
+                       action='store_true',
+                       help='If set, use original BERT residula connection '
+                       'ordering.')
+    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')
+
+    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.')
+    group.add_argument('--log-memory-to-tensorboard',
+                       action='store_true',
+                       help='Enable memory logging to tensorboard.')
+    group.add_argument('--log-world-size-to-tensorboard',
+                       action='store_true',
+                       help='Enable world size logging 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 incerement> '
+                       '                      <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('--activations-checkpoint-method', type=str, default=None,
+                       choices=['uniform', 'block'],
+                       help='1) uniform: uniformly divide the total number of '
+                       'Transformer layers and checkpoint the input activation of '
+                       'each divided chunk, '
+                       '2) checkpoint the input activations of only a set number of '
+                       'individual Transformer layers per pipeline stage and do the '
+                       'rest without any checkpointing'
+                       'default) do not apply activations checkpoint to any layers')
+    group.add_argument('--activations-checkpoint-num-layers', type=int, default=1,
+                       help='1) uniform: the number of Transformer layers in each '
+                       'uniformly divided checkpoint unit, '
+                       '2) block: the number of individual Transformer layers '
+                       'to checkpoint within each pipeline stage.')
+    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('--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('--exit-signal-handler', action='store_true',
+                       help='Dynamically save the checkpoint and shutdown the '
+                       'training if SIGTERM is received')
+    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('--dataloader-type', type=str, default=None,
+                       choices=['single', 'cyclic'],
+                       help='Single pass vs multiple pass data loader')
+    group.add_argument('--no-async-tensor-model-parallel-allreduce',
+                       action='store_true',
+                       help='Disable asynchronous execution of '
+                       'tensor-model-parallel all-reduce with weight '
+                       'gradient compuation of a column-linear layer.')
+    group.add_argument('--no-persist-layer-norm', action='store_true',
+                       help='Disable using persistent fused layer norm kernel. '
+                       'This kernel supports only a set of hidden sizes. Please '
+                       'check persist_ln_hidden_sizes if your hidden '
+                       'size is supported.')
+    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('--data-parallel-random-init', action='store_true',
+                       help='Enable random initialization of params '
+                       'across data parallel ranks')
+    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-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.')
+
+    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('--pipeline-model-parallel-split-rank',
+                       type=int, default=None,
+                       help='Rank where encoder and decoder should be split.')
+    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('--no-contiguous-buffers-in-local-ddp',
+                       action='store_false', help='If set, dont use '
+                       'contiguous buffer in local DDP.',
+                       dest='use_contiguous_buffers_in_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('--empty-unused-memory-level', default=0, type=int,
+                       choices=[0, 1, 2],
+                       help='Call torch.cuda.empty_cache() each iteration '
+                       '(training and eval), to reduce fragmentation.'
+                       '0=off, 1=moderate, 2=aggressive.')
+    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')
+
+    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='969, 30, 1',
+                       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.')
+    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('--tokenizer-type', type=str,
+                       default=None,
+                       choices=['BertWordPieceLowerCase',
+                                'BertWordPieceCase',
+                                'GPT2BPETokenizer'],
+                       help='What type of tokenizer to use.')
+    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.')
+    group.add_argument('--eod-mask-loss', action='store_true',
+                       help='Mask loss for the end of document tokens.')
+
+    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-h', type=int, default=224,
+                       help='Image height for vision classification task')
+    group.add_argument('--img-w', type=int, default=224,
+                       help='Image height 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')
+    group.add_argument('--classes-fraction', type=float, default=1.0,
+                       help='training with fraction of classes.')
+    group.add_argument('--data-per-class-fraction', type=float, default=1.0,
+                       help='training with fraction of data per class.')
+    group.add_argument('--no-data-sharding', action='store_false',
+                       help='Disable data sharding.',
+                       dest='data_sharding')
+
+    return parser

3rdparty/Megatron-LM/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 import (get_args,
+                      mpu,
+                      print_rank_0,
+                      update_num_microbatches,
+                      utils)
+
+_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')
+    if args.vocab_file:
+        _compare('max_position_embeddings')
+        _compare('make_vocab_size_divisible_by')
+        _compare('padded_vocab_size')
+        _compare('tokenizer_type')
+    if args.data_parallel_random_init:
+        _compare('data_parallel_random_init')
+    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 read_metadata(tracker_filename):
+    # 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)
+
+    # Get the max iteration retrieved across the ranks.
+    iters_cuda = torch.cuda.LongTensor([iteration])
+    torch.distributed.all_reduce(iters_cuda, op=torch.distributed.ReduceOp.MAX)
+    max_iter = iters_cuda[0].item()
+
+    # We should now have all the same iteration.
+    # If not, print a warning and chose the maximum
+    # iteration across all ranks.
+    if iteration != max_iter:
+        print('WARNING: on rank {} found iteration {} in the '
+              'metadata while max iteration across the ranks '
+              'is {}, replacing it with max iteration.'.format(
+                  rank, iteration, max_iter), flush=True)
+    return max_iter, release
+
+
+def get_rng_state():
+    """ collect rng state across data parallel ranks """
+    args = get_args()
+    rng_state = {
+        'random_rng_state': random.getstate(),
+        'np_rng_state': np.random.get_state(),
+        'torch_rng_state': torch.get_rng_state(),
+        'cuda_rng_state': torch.cuda.get_rng_state(),
+        'rng_tracker_states': mpu.get_cuda_rng_tracker().get_states()}
+
+    rng_state_list = None
+    if torch.distributed.is_initialized() and \
+            mpu.get_data_parallel_world_size() > 1 and \
+            args.data_parallel_random_init:
+        rng_state_list = \
+            [None for i in range(mpu.get_data_parallel_world_size())]
+        torch.distributed.all_gather_object(
+            rng_state_list,
+            rng_state,
+            group=mpu.get_data_parallel_group())
+    else:
+        rng_state_list = [rng_state]
+
+    return rng_state_list
+
+
+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.
+    model = utils.unwrap_model(model)
+
+    print_rank_0('saving checkpoint at iteration {:7d} to {}'.format(
+        iteration, args.save))
+
+    # collect rng state across data parallel ranks
+    rng_state = get_rng_state()
+
+    if not torch.distributed.is_initialized() or mpu.get_data_parallel_rank() == 0:
+
+        # Arguments, iteration, and model.
+        state_dict = {}
+        state_dict['args'] = args
+        state_dict['checkpoint_version'] = 3.0
+        state_dict['iteration'] = iteration
+        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["rng_state"] = rng_state
+
+        # Save.
+        checkpoint_name = get_checkpoint_name(args.save, iteration)
+        ensure_directory_exists(checkpoint_name)
+        torch.save(state_dict, checkpoint_name)
+
+    # 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 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()
+
+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)
+
+    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, release = read_metadata(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']
+        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']
+        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)
+    else:
+        print_rank_0('could not find arguments in the checkpoint ...')
+
+    # Model.
+    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 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:
+            if 'rng_state' in state_dict:
+                # access rng_state for data parallel rank
+                if args.data_parallel_random_init:
+
+                    rng_state = state_dict['rng_state'][mpu.get_data_parallel_rank()]
+                else:
+                    rng_state = state_dict['rng_state'][0]
+                random.setstate(rng_state['random_rng_state'])
+                np.random.set_state(rng_state['np_rng_state'])
+                torch.set_rng_state(rng_state['torch_rng_state'])
+                torch.cuda.set_rng_state(rng_state['cuda_rng_state'])
+                # Check for empty states array
+                if not rng_state['rng_tracker_states']:
+                    raise KeyError
+                mpu.get_cuda_rng_tracker().set_states(
+                    rng_state['rng_tracker_states'])
+            else:  # backward compatability
+                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
+

3rdparty/Megatron-LM/megatron/data/Makefile

+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 $@

3rdparty/Megatron-LM/megatron/data/__init__.py

+from . import indexed_dataset

3rdparty/Megatron-LM/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

3rdparty/Megatron-LM/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

3rdparty/Megatron-LM/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