Megatron (1 and 2) is a large, powerful transformer developed by the Applied Deep Learning Research team at NVIDIA. This repository is for ongoing research on training large transformer language models at scale. We developed efficient, model-parallel (tensor and pipeline), and multi-node pre-training oftransformer based models such as GPT, BERT, and T5 using mixed precision.

Below are some of the projects where we have directly used Megatron: * BERT and GPT Studies Using Megatron * BioMegatron: Larger Biomedical Domain Language Model * End-to-End Training of Neural Retrievers for Open-Domain Question Answering * Large Scale Multi-Actor Generative Dialog Modeling * Local Knowledge Powered Conversational Agents * MEGATRON-CNTRL: Controllable Story Generation with External Knowledge Using Large-Scale Language Models * RACE Reading Comprehension Dataset Leaderboard * Scaling Language Model Training to a Trillion Parameters Using Megatron * Training Question Answering Models From Synthetic Data

Our codebase is capable of efficiently training very large (hundreds of billions of parameters) language models with both model and data parallelism. To demonstrate how the code scales with multiple GPUs and model sizes, we consider GPT models from 1 billion all the way to 1 trillion parameters. All models use a vocabulary size of 51,200 and a sequence length of 2048. We vary hidden size, number of attention heads, and number of layers to arrive at a specifc model size. As the model size increases, we also modestly increase the batch size. We leverage NVIDIA's Selene supercomputer to perform scaling studies and use up to 3072 A100 GPUs for the largest model. The table below shows the model configurations along with the achieved FLOPs (both per GPU and aggregate over all GPUs). Note that the FLOPs are measured for end-to-end training, i.e., includes all operations including data loading, optimization, and even logging.

All the cases from 1 billion to 1 trillion parameters achieve more than 43% half precision utilization, which is high for an end-to-end application. We observe that initially the utilization remains constant but as hidden size increases for larger models, utilization starts increasing and reaches 52% for the largest model. We also note that achieved aggregate petaFLOPs across all GPUs increases almost linearly with number of GPUs, demonstrating good weak scaling.

Contents

• Contents
• Setup
  • Downloading Checkpoints
• Usage
• Training
  • Data Preprocessing
  • BERT Pretraining
  • GPT Pretraining
  • GPT Pretraining
  • T5 Pretraining
  • Distributed Pretraining
  • GPT-3 Example
• Evaluation and Tasks
  • GPT Text Generation
  • GPT Evaluation
    • WikiText Perplexity Evaluation
    • LAMBADA Cloze Accuracy
  • BERT Task Evaluation
    • RACE Evaluation
    • MNLI Evaluation
• Datasets
  • Collecting Wikipedia Training Data
  • Collecting GPT Webtext Data

Setup

We have tested Megatron with NGC's PyTorch container version 20.12, which uses python 3.8, pytorch 1.8, cuda 11.1, and nccl 2.8.3.

To use this repository, please install the latest supported versions of PyTorch with GPU support (python 3.8, pytorch 1.8, cuda 11.1, and nccl 2.8.3 and above) and NVIDIA APEX. We strongly recommend using one of NGC's recent PyTorch containers (the latest compatible version at time of publication can be pulled with docker pull nvcr.io/nvidia/pytorch:20.12-py3). Data preprocessing requires NLTK, though this is not required for training, evaluation, or downstream tasks.

Downloading Checkpoints

We have provided pretrained BERT-345M and GPT-345M checkpoints for use to evaluate or finetuning downstream tasks. To access these checkpoints, first sign up for and setup the NVIDIA GPU Cloud (NGC) Registry CLI. Further documentation for downloading models can be found in the NGC documentation.

Alternatively, you can directly download the checkpoints using:

BERT-345M-uncased: wget --content-disposition https://api.ngc.nvidia.com/v2/models/nvidia/megatron_bert_345m/versions/v0.1_uncased/zip -O megatron_bert_345m_v0.1_uncased.zip
BERT-345M-cased: wget --content-disposition https://api.ngc.nvidia.com/v2/models/nvidia/megatron_bert_345m/versions/v0.1_cased/zip -O megatron_bert_345m_v0.1_cased.zip
GPT-345M: wget --content-disposition https://api.ngc.nvidia.com/v2/models/nvidia/megatron_lm_345m/versions/v0.0/zip -O megatron_lm_345m_v0.0.zip

The models require vocabulary files to run. The BERT WordPiece vocab file can be extracted from Google's pretrained BERT models: uncased, cased. The GPT vocab file and merge table can be downloaded directly.

Usage

After installation, there are several possible workflows. The most comprehensive is: 1. Data preprocessing 2. Pretraining 3. Finetuning (Optional for zero-shot tasks) 4. Downstream task evaluation or text generation

However, steps 1 and 2 can be replaced by using one of the pretrained models mentioned above.

We've provided several scripts for pretraining both BERT and GPT in examples directory, as well as scripts for both zero-shot and fine-tuned downstream tasks including MNLI, RACE, WikiText103, and LAMBADA evaluation. There is also a script for GPT interactive text generation.

Training

Data Preprocessing

The training data requires preprocessing. First, place your training data in a loose json format, with one json containing a text sample per line. For example:

{"src": "www.nvidia.com", "text": "The quick brown fox", "type": "Eng", "id": "0", "title": "First Part"}
{"src": "The Internet", "text": "jumps over the lazy dog", "type": "Eng", "id": "42", "title": "Second Part"}

The name of the text field of the json can be changed by using the --json-key flag in preprocess_data.py The other metadata are optional and are not used in training.

The loose json is then processed into a binary format for training. To convert the json into mmap, cached index file, or the lazy loader format use preprocess_data.py. Set the --dataset-impl flag to mmap, cached, or lazy, respectively (default is mmap). An example script to prepare data for BERT training is:

python tools/preprocess_data.py \
       --input my-corpus.json \
       --output-prefix my-bert \
       --vocab bert-vocab.txt \
       --dataset-impl mmap \
       --tokenizer-type BertWordPieceLowerCase \
       --split-sentences

The output will be two files named, in this case, my-bert_text_sentence.bin and my-bert_text_sentence.idx. The --data-path specified in later BERT training is the full path and new filename, but without the file extension.

Some minor modifications are required for GPT data preprocessing, namely, the addition of a merge table, an end-of-document token, removal of sentence splitting, and a change to the tokenizer type:

python tools/preprocess_data.py \
       --input my-corpus.json \
       --output-prefix my-gpt2 \
       --vocab gpt2-vocab.json \
       --dataset-impl mmap \
       --tokenizer-type GPT2BPETokenizer \
       --merge-file gpt2-merges.txt \
       --append-eod

Here the output files are named my-gpt2_text_document.bin and my-gpt2_text_document.idx. As before, in GPT training, use the longer name without the extension as --data-path.

Further command line arguments are described in the source file preprocess_data.py.

BERT Pretraining

The examples/pretrain_bert.sh script runs single GPU 345M parameter BERT pretraining. Debugging is the primary use for single GPU training, as the code base and command line arguments are optimized for highly distributed training. Most of the arguments are fairly self-explanatory. By default, the learning rate decays linearly over the training iterations starting at --lr to a minimum set by --min-lr over --lr-decay-iters iterations. The fraction of training iterations used for warmup is set by --lr-warmup-fraction. While this is single GPU training, the batch size specified by --micro-batch-size is a single forward-backward path batch-size and the code will perform gradient accumulation steps until it reaches global-batch-size whcih is the batch size per iteration. The data is partitioned into a 949:50:1 ratio for training/validation/test sets (default is 969:30:1). This partitioning happens on the fly, but is consistent across runs with the same random seed (1234 by default, or specified manually with --seed). We use train-iters as the training iterations requested. Alternatively, one can provide --train-samples which is total number of samples to train on. If this option is present, then instead of providing --lr-decay-iters, one will need to provide --lr-decay-samples.

The logging, checkpoint-saving, and evaluation intervals are specified. Checkpointing the activations facilitates the training of larger models and/or batches. Note that the --data-path now includes the additional _text_sentence suffix added in preprocessing, but does not include the file extensions.

CHECKPOINT_PATH=checkpoints/bert_345m
VOCAB_FILE=bert-vocab.txt
DATA_PATH=my-bert_text_sentence

BERT_ARGS="--num-layers 24 \
           --hidden-size 1024 \
           --num-attention-heads 16 \
           --seq-length 512 \
           --max-position-embeddings 512 \
           --lr 0.0001 \
           --lr-decay-iters 990000 \
           --train-iters 2000000 \
           --min-lr 0.00001 \
           --lr-warmup-fraction 0.01 \
       --micro-batch-size 4 \
           --global-batch-size 8 \
           --vocab-file $VOCAB_FILE \
           --split 949,50,1 \
           --fp16"

OUTPUT_ARGS="--log-interval 10 \
             --save-interval 500 \
             --eval-interval 100 \
             --eval-iters 10 \
             --checkpoint-activations"

python pretrain_bert.py \
       $BERT_ARGS \
       $OUTPUT_ARGS \
       --save $CHECKPOINT_PATH \
       --load $CHECKPOINT_PATH \
       --data-path $DATA_PATH

Further command line arguments are described in the source file arguments.py.

GPT Pretraining

The examples/pretrain_gpt.sh script runs single GPU 345M parameter GPT pretraining. As mentioned above, single GPU training is primarily intended for debugging purposes, as the code is optimized for distributed training.

It follows largely the same format as the previous BERT script with a few notable differences: the tokenization scheme used is BPE (which requires a merge table and a json vocabulary file) instead of WordPiece, the model architecture allows for longer sequences (note that the max position embedding must be greater than or equal to the maximum sequence length), and the --lr-decay-style has been set to cosine decay. Note that the --data-path now includes the additional _text_document suffix added in preprocessing, but does not include the file extensions.

CHECKPOINT_PATH=checkpoints/gpt2_345m
VOCAB_FILE=gpt2-vocab.json
MERGE_FILE=gpt2-merges.txt
DATA_PATH=my-gpt2_text_document

GPT_ARGS="--num-layers 24 \
          --hidden-size 1024 \
          --num-attention-heads 16 \
          --seq-length 1024 \
          --max-position-embeddings 1024 \
          --micro-batch-size 4 \
          --global-batch-size 8 \
          --lr 0.00015 \
          --train-iters 500000 \
          --lr-decay-iters 320000 \
          --lr-decay-style cosine \
          --vocab-file $VOCAB_FILE \
          --merge-file $MERGE_FILE \
          --lr-warmup-fraction .01 \
          --fp16"

OUTPUT_ARGS=<same as those in BERT pretraining above>

python