maxmustermann/DeepLearningExamples
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity
DeepLearningExamples/TensorFlow/LanguageModeling/BERT/README.md
Krzysztof Kudrynski 49e23b4597 Adding links to performance benchmark page
2021-07-21 14:39:48 +02:00

119 KiB
Raw Blame History

BERT For TensorFlow

This repository provides a script and recipe to train the BERT model for TensorFlow to achieve state-of-the-art accuracy, and is tested and maintained by NVIDIA.

Table Of Contents

Model overview

BERT, or Bidirectional Encoder Representations from Transformers, is a new method of pre-training language representations which obtains state-of-the-art results on a wide array of Natural Language Processing (NLP) tasks. This model is based on the BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding paper. NVIDIA's BERT is an optimized version of Google's official implementation, leveraging mixed precision arithmetic and Tensor Cores on A100, V100 and T4 GPUs for faster training times while maintaining target accuracy.

Other publicly available implementations of BERT include:

  1. NVIDIA PyTorch
  2. Hugging Face
  3. codertimo
  4. gluon-nlp
  5. Google's official implementation

This model is trained with mixed precision using Tensor Cores on NVIDIA Volta, Ampere and Turing GPUs. Therefore, researchers can get results up to 4x faster than training without Tensor Cores, while experiencing the benefits of mixed precision training. This model is tested against each NGC monthly container release to ensure consistent accuracy and performance over time.

Model architecture

BERT's model architecture is a multi-layer bidirectional Transformer encoder. Based on the model size, we have the following two default configurations of BERT:

Model Hidden layers Hidden unit size Attention heads Feedforward filter size Max sequence length Parameters
BERTBASE 12 encoder 768 12 4 x 768 512 110M
BERTLARGE 24 encoder 1024 16 4 x 1024 512 330M

BERT training consists of two steps, pre-training the language model in an unsupervised fashion on vast amounts of unannotated datasets, and then using this pre-trained model for fine-tuning for various NLP tasks, such as question and answer, sentence classification, or sentiment analysis. Fine-tuning typically adds an extra layer or two for the specific task and further trains the model using a task-specific annotated dataset, starting from the pre-trained backbone weights. The end-to-end process in depicted in the following image:

Figure 1: BERT Pipeline

Default configuration

This repository contains scripts to interactively launch data download, training, benchmarking and inference routines in a Docker container for both pre-training and fine tuning for Question Answering. The major differences between the official implementation of the paper and our version of BERT are as follows:

  • Mixed precision support with TensorFlow Automatic Mixed Precision (TF-AMP), which enables mixed precision training without any changes to the code-base by performing automatic graph rewrites and loss scaling controlled by an environmental variable.
  • Scripts to download dataset for:
    • Pre-training - Wikipedia, BookCorpus
    • Fine tuning - SQuAD (Stanford Question Answering Dataset)
    • Fine tuning - GLUE (The General Language Understanding Evaluation benchmark)
    • Pretrained weights from Google
  • Custom fused CUDA kernels for faster computations
  • Multi-GPU/Multi-node support using Horovod

The following performance optimizations were implemented in this model:

  • XLA support (experimental).

These techniques and optimizations improve model performance and reduce training time, allowing you to perform various NLP tasks with no additional effort.

Feature support matrix

The following features are supported by this model.

Feature BERT
Horovod Multi-GPU Yes
Horovod Multi-Node Yes
Automatic mixed precision (AMP) Yes
LAMB Yes

Features

Multi-GPU training with Horovod - Our model uses Horovod to implement efficient multi-GPU training with NCCL. For details, see example sources in this repository or see the TensorFlow tutorial

LAMB stands for Layerwise Adaptive Moments based optimizer, is a large batch optimization technique that helps accelerate training of deep neural networks using large minibatches. It allows using a global batch size of 65536 and 32768 on sequence lengths 128 and 512 respectively, compared to a batch size of 256 for Adam. The optimized implementation accumulates 1024 gradient batches in phase 1 and 4096 steps in phase 2 before updating weights once. This results in 27% training speedup on a single DGX2 node. On multi-node systems, LAMB allows scaling up to 1024 GPUs resulting in training speedups of up to 17x in comparison to Adam. Adam has limitations on the learning rate that can be used since it is applied globally on all parameters whereas LAMB follows a layerwise learning rate strategy.

NVLAMB adds necessary tweaks to LAMB version 1, to ensure correct convergence. A guide to implementating the LAMB optimizer can be found in our article on Medium.com. The algorithm is as follows: NVLAMB

Mixed precision training

Mixed precision is the combined use of different numerical precisions in a computational method. Mixed precision training offers significant computational speedup by performing operations in half-precision format while storing minimal information in single-precision to retain as much information as possible in critical parts of the network. Since the introduction of Tensor Cores in Volta, and following with both the Turing and Ampere architectures, significant training speedups are experienced by switching to mixed precision -- up to 3x overall speedup on the most arithmetically intense model architectures. Using mixed precision training previously required two steps:

  1. Porting the model to use the FP16 data type where appropriate.
  2. Adding loss scaling to preserve small gradient values.

This can now be achieved using Automatic Mixed Precision (AMP) for TensorFlow to enable the full mixed precision methodology in your existing TensorFlow model code. AMP enables mixed precision training on Volta, Turing, and NVIDIA Ampere GPU architectures automatically. The TensorFlow framework code makes all necessary model changes internally.

In TF-AMP, the computational graph is optimized to use as few casts as necessary and maximize the use of FP16, and the loss scaling is automatically applied inside of supported optimizers. AMP can be configured to work with the existing tf.contrib loss scaling manager by disabling the AMP scaling with a single environment variable to perform only the automatic mixed-precision optimization. It accomplishes this by automatically rewriting all computation graphs with the necessary operations to enable mixed precision training and automatic loss scaling.

For information about:

Enabling mixed precision

Mixed precision is enabled in TensorFlow by using the Automatic Mixed Precision (TF-AMP) extension which casts variables to half-precision upon retrieval, while storing variables in single-precision format. Furthermore, to preserve small gradient magnitudes in backpropagation, a loss scaling step must be included when applying gradients. In TensorFlow, loss scaling can be applied statically by using simple multiplication of loss by a constant value or automatically, by TF-AMP. Automatic mixed precision makes all the adjustments internally in TensorFlow, providing two benefits over manual operations. First, programmers need not modify network model code, reducing development and maintenance effort. Second, using AMP maintains forward and backward compatibility with all the APIs for defining and running TensorFlow models.

To enable mixed precision, you can simply add the values to the environmental variables inside your training script:

  • Enable TF-AMP graph rewrite:

    os.environ["TF_ENABLE_AUTO_MIXED_PRECISION_GRAPH_REWRITE"] = "1"

  • Enable Automated Mixed Precision:

    os.environ['TF_ENABLE_AUTO_MIXED_PRECISION'] = '1'

Enabling TF32

TensorFloat-32 (TF32) is the new math mode in NVIDIA A100 GPUs for handling the matrix math also called tensor operations. TF32 running on Tensor Cores in A100 GPUs can provide up to 10x speedups compared to single-precision floating-point math (FP32) on Volta GPUs.

TF32 Tensor Cores can speed up networks using FP32, typically with no loss of accuracy. It is more robust than FP16 for models which require high dynamic range for weights or activations.

For more information, refer to the TensorFloat-32 in the A100 GPU Accelerates AI Training, HPC up to 20x blog post.

TF32 is supported in the NVIDIA Ampere GPU architecture and is enabled by default.

Glossary

Fine-tuning Training an already pretrained model further using a task specific dataset for subject-specific refinements, by adding task-specific layers on top if required.

Language Model Assigns a probability distribution over a sequence of words. Given a sequence of words, it assigns a probability to the whole sequence.

Pre-training Training a model on vast amounts of data on the same (or different) task to build general understandings.

Transformer The paper Attention Is All You Need introduces a novel architecture called Transformer that uses an attention mechanism and transforms one sequence into another.

Setup

The following section lists the requirements in order to start training the BERT model.

Requirements

This repository contains Dockerfile which extends the TensorFlow NGC container and encapsulates some dependencies. Aside from these dependencies, ensure you have the following components:

For more information about how to get started with NGC containers, see the following sections from the NVIDIA GPU Cloud Documentation and the Deep Learning Documentation:

For those unable to use the TensorFlow NGC container, to set up the required environment or create your own container, see the versioned NVIDIA Container Support Matrix.

For multi-node, the sample provided in this repository requires Enroot and Pyxis set up on a SLURM cluster.

More information on how to set up and launch can be found in the Multi-node Documentation.

Quick Start Guide

To pretrain or fine tune your model for Question Answering using mixed precision with Tensor Cores or using FP32/TF32, perform the following steps using the default parameters of the BERT model.

  1. Clone the repository.
git clone https://github.com/NVIDIA/DeepLearningExamples
cd DeepLearningExamples/TensorFlow/LanguageModeling/BERT
  1. Build the BERT TensorFlow NGC container.
bash scripts/docker/build.sh
  1. Download and preprocess the dataset.

This repository provides scripts to download, verify and extract the SQuAD dataset, GLUE dataset and pretrained weights for fine tuning as well as Wikipedia and BookCorpus dataset for pre-training.

To download, verify, and extract the required datasets, run:

bash scripts/data_download.sh

The script launches a Docker container with the current directory mounted and downloads the datasets to a data/ folder on the host.

Note: For fine tuning only, Wikipedia and Bookscorpus dataset download and preprocessing can be skipped by commenting it out.

  • Download Wikipedia only for pretraining

The pretraining dataset is 170GB+ and takes 15+ hours to download. The BookCorpus server most of the times get overloaded and also contain broken links resulting in HTTP 403 and 503 errors. Hence, it is recommended to skip downloading BookCorpus data by running:

bash scripts/data_download.sh wiki_only

  • Download Wikipedia and BookCorpus

Users are welcome to download BookCorpus from other sources to match our accuracy, or repeatedly try our script until the required number of files are downloaded by running the following: bash scripts/data_download.sh wiki_books

Note: Ensure a complete Wikipedia download. If in any case, the download breaks, remove the output file wikicorpus_en.xml.bz2 and start again. If a partially downloaded file exists, the script assumes successful download which causes the extraction to fail. Not using BookCorpus can potentially change final accuracy on a few downstream tasks.

  1. Download the pretrained models from NGC.

We have uploaded checkpoints that have been fine tuned and pre-trained for various configurations on the NGC Model Registry. Our data download scripts, by default download some of them but you can browse and download the relevant checkpoints directly from the NGC model catalog. Download them to the data/download/nvidia_pretrained/ to easily access them in your scripts.

  1. Start an interactive session in the NGC container to run training/inference.

After you build the container image and download the data, you can start an interactive CLI session as follows:

bash scripts/docker/launch.sh

The launch.sh script assumes that the datasets are in the following locations by default after downloading the data.

  • SQuAD v1.1 - data/download/squad/v1.1
  • SQuAD v2.0 - data/download/squad/v2.0
  • GLUE The Corpus of Linguistic Acceptability (CoLA) - data/download/CoLA
  • GLUE Microsoft Research Paraphrase Corpus (MRPC) - data/download/MRPC
  • GLUE The Multi-Genre NLI Corpus (MNLI) - data/download/MNLI
  • BERT Large - data/download/google_pretrained_weights/uncased_L-24_H-1024_A-16
  • BERT Base - data/download/google_pretrained_weights/uncased_L-12_H-768_A-12
  • BERT - data/download/google_pretrained_weights/
  • Wikipedia + BookCorpus TFRecords - data/tfrecords<config>/books_wiki_en_corpus
  1. Start pre-training.

BERT is designed to pre-train deep bidirectional representations for language representations. The following scripts are to replicate pre-training on Wikipedia and BookCorpus from the LAMB paper. These scripts are general and can be used for pre-training language representations on any corpus of choice.

From within the container, you can use the following script to run pre-training using LAMB.

bash scripts/run_pretraining_lamb.sh <train_batch_size_phase1> <train_batch_size_phase2> <eval_batch_size> <learning_rate_phase1> <learning_rate_phase2> <precision> <use_xla> <num_gpus> <warmup_steps_phase1> <warmup_steps_phase2> <train_steps> <save_checkpoint_steps> <num_accumulation_phase1> <num_accumulation_steps_phase2> <bert_model>

For BERT Large FP16 training with XLA using a DGX-1 V100 32GB, run:

bash scripts/run_pretraining_lamb.sh 64 8 8 7.5e-4 5e-4 fp16 true 8 2000 200 7820 100 128 512 large

This repository also contains a number of predefined configurations to run the Lamb pretraining on NVIDIA DGX-1, NVIDIA DGX-2H or NVIDIA DGX A100 nodes in scripts/configs/pretrain_config.sh. For example, to use the default DGX A100 8 gpu config, run:

bash scripts/run_pretraining_lamb.sh $(source scripts/configs/pretrain_config.sh && dgxa100_8gpu_fp16)

Alternatively, to run pre-training with Adam as in the original BERT paper from within the container, run:

bash scripts/run_pretraining_adam.sh <train_batch_size_per_gpu> <eval_batch_size> <learning_rate_per_gpu> <precision> <use_xla> <num_gpus> <warmup_steps> <train_steps> <save_checkpoint_steps>
  1. Start fine tuning.

The above pretrained BERT representations can be fine tuned with just one additional output layer for a state-of-the-art Question Answering system. From within the container, you can use the following script to run fine-training for SQuAD.

bash scripts/run_squad.sh <batch_size_per_gpu> <learning_rate_per_gpu> <precision> <use_xla> <num_gpus> <seq_length> <doc_stride> <bert_model> <squad_version> <checkpoint> <epochs>

For SQuAD 1.1 FP16 training with XLA using a DGX A100 40GB, run:

bash scripts/run_squad.sh 32 5e-6 fp16 true 8 384 128 large 1.1 data/download/google_pretrained_weights/uncased_L-24_H-1024_A-16/bert_model.ckpt 2.0

This repository contains a number of predefined configurations to run the SQuAD fine tuning on NVIDIA DGX-1, NVIDIA DGX-2H or NVIDIA DGX A100 nodes in scripts/configs/squad_config.sh. For example, to use the default DGX A100 8 gpu config, run:

bash scripts/run_squad.sh $(source scripts/configs/squad_config.sh && dgxa100_8gpu_fp16) 1.1 data/download/google_pretrained_weights/uncased_L-24_H-1024_A-16/bert_model.ckpt 2.0

Alternatively, to run fine tuning on GLUE benchmark, run:

bash scripts/run_glue.sh <task_name> <batch_size_per_gpu> <learning_rate_per_gpu> <precision> <use_xla> <num_gpus> <seq_length> <doc_stride> <bert_model> <epochs> <warmup_proportion> <checkpoint>

For MRPC FP16 training with XLA using a DGX A100 40GB, run:

bash scripts/run_glue.sh MRPC 16 3e-6 true 8 128 64 large 3 0.1

The GLUE tasks supported include CoLA, MRPC and MNLI.

  1. Start validation/evaluation.

The run_squad_inference.sh script runs inference on a checkpoint fine tuned for SQuAD and evaluates the validity of predictions on the basis of exact match and F1 score.

bash scripts/run_squad_inference.sh <init_checkpoint> <batch_size> <precision> <use_xla> <seq_length> <doc_stride> <bert_model> <squad_version>

For SQuAD 2.0 FP16 inference with XLA using a DGX-1 V100 32GB using checkpoint at /results/model.ckpt , run:

bash scripts/run_squad_inference.sh /results/model.ckpt 8 fp16 true 384 128 large 2.0

For SQuAD 1.1 FP32 inference without XLA using a DGX A100 40GB using checkpoint at /results/model.ckpt, run:

bash scripts/run_squad_inference.sh /results/model.ckpt 8 fp32 false 384 128 large 1.1

Alternatively, to run inference on GLUE benchmark, run:

bash scripts/run_glue_inference.sh <task_name> <init_checkpoint> <batch_size_per_gpu> <precision> <use_xla> <seq_length> <doc_stride> <bert_model>

Advanced

The following sections provide greater details of the dataset, running training and inference, and the training results.

Scripts and sample code

In the root directory, the most important files are:

  • run_pretraining.py - Serves as entry point for pre-training
  • run_squad.py - Serves as entry point for SQuAD training
  • run_classifier.py - Serves as entry point for GLUE training
  • Dockerfile - Container with the basic set of dependencies to run BERT

The scripts/ folder encapsulates all the one-click scripts required for running various functionalities supported such as:

  • run_squad.sh - Runs SQuAD training and inference using run_squad.py file
  • run_glue.sh - Runs GLUE training and inference using the run_classifier.py file
  • run_pretraining_adam.sh - Runs pre-training with Adam optimizer using the run_pretraining.py file
  • run_pretraining_lamb.sh - Runs pre-training with LAMB optimizer using the run_pretraining.py file in two phases. Phase 1 does 90% of training with sequence length = 128. In phase 2, the remaining 10% of the training is done with sequence length = 512.
  • data_download.sh - Downloads datasets using files in the data/ folder
  • finetune_train_benchmark.sh - Captures performance metrics of training for multiple configurations
  • finetune_inference_benchmark.sh - Captures performance metrics of inference for multiple configurations

Other folders included in the root directory are:

  • data/ - Necessary folders and scripts to download datasets required for fine tuning and pre-training BERT.
  • utils/ - Necessary files for preprocessing data before feeding into BERT and hooks for obtaining performance metrics from BERT.

Parameters

Aside from the options to set hyperparameters, the relevant options to control the behaviour of the run_pretraining.py script are:

  --bert_config_file: The config json file corresponding to the pre-trained BERT model. This specifies the model architecture.
  --init_checkpoint: Initial checkpoint (usually from a pre-trained BERT model).
  --[no]do_eval: Whether to run evaluation on the dev set.(default: 'false')
  --[no]do_train: Whether to run training.(evaluation: 'false')
  --eval_batch_size: Total batch size for eval.(default: '8')(an integer)
  --[no]horovod: Whether to use Horovod for multi-gpu runs(default: 'false')
  --[no]amp: Whether to enable AMP ops. When false, uses TF32 on A100 and FP32 on V100 GPUS.(default: 'True')
  --[no]use_xla: Whether to enable XLA JIT compilation.(default: 'True')
  --input_files_dir: Input TF example files (can be a dir or comma separated).
  --output_dir: The output directory where the model checkpoints will be    written.
  --optimizer_type: Optimizer used for training - LAMB or ADAM
  --num_accumulation_steps: Number of accumulation steps before gradient update. Global batch size = num_accumulation_steps * train_batch_size
  --allreduce_post_accumulation: Whether to all reduce after accumulation of N steps or after each step

Aside from the options to set hyperparameters, some relevant options to control the behaviour of the run_squad.py script are:

  --bert_config_file: The config json file corresponding to the pre-trained BERT model. This specifies the model architecture.
  --output_dir: The output directory where the model checkpoints will be written.
  --[no]do_predict: Whether to run evaluation on the dev set. (default: 'false')
  --[no]do_train: Whether to run training. (default: 'false')
  --learning_rate: The initial learning rate for Adam.(default: '5e-06')(a number)
  --max_answer_length: The maximum length of an answer that can be generated. This is needed because the start and end predictions are not conditioned on one another.(default: '30')(an integer)
  --max_query_length: The maximum number of tokens for the question. Questions longer than this will be truncated to this length.(default: '64')(an integer)
  --max_seq_length: The maximum total input sequence length after WordPiece tokenization. Sequences longer than this will be truncated, and sequences shorter than this will be padded.(default: '384')(an integer)
  --predict_batch_size: Total batch size for predictions.(default: '8')(an integer)
  --train_batch_size: Total batch size for training.(default: '8')(an integer)
  --[no]amp: Whether to enable AMP ops. When false, uses TF32 on A100 and FP32 on V100 GPUS.(default: 'True')
  --[no]use_xla: Whether to enable XLA JIT compilation.(default: 'True')
  --[no]version_2_with_negative: If true, the SQuAD examples contain some that do not have an answer.(default: 'false')

Aside from the options to set hyperparameters, some relevant options to control the behaviour of the run_classifier.py script are:

  --bert_config_file: The config json file corresponding to the pre-trained BERT model. This specifies the model architecture.
  --data_dir: The input data dir. Should contain the .tsv files (or other data files) for the task.
  --[no]do_eval: Whether to run eval on the dev set.
    (default: 'false')
  --[no]do_predict: Whether to run the model in inference mode on the test set.(default: 'false')
  --[no]do_train: Whether to run training.(default: 'false')
  --[no]horovod: Whether to use Horovod for multi-gpu runs(default: 'false')
  --init_checkpoint: Initial checkpoint (usually from a pre-trained BERT model).
  --max_seq_length: The maximum total input sequence length after WordPiece tokenization. Sequences longer than this will be truncated, and sequences shorter than this will be padded.(default: '128')(an integer)
  --num_train_epochs: Total number of training epochs to perform.(default: '3.0')(a number)
  --output_dir: The output directory where the model checkpoints will be written.
  --task_name: The name of the task to train.
  --train_batch_size: Total batch size for training.(default: '32')(an integer)
  --[no]amp: Whether to enable AMP ops. When false, uses TF32 on A100 and FP32 on V100 GPUS.(default: 'True')
  --[no]use_xla: Whether to enable XLA JIT compilation.(default: 'True')
  --vocab_file: The vocabulary file that the BERT model was trained on.
  --warmup_proportion: Proportion of training to perform linear learning rate warmup for. E.g., 0.1 = 10% of training.(default: '0.1')(a number)

Note: When initializing from a checkpoint using --init_checkpoint and a corpus of your choice, keep in mind that bert_config_file and vocab_file should remain unchanged.

Command-line options

To see the full list of available options and their descriptions, use the -h or --help command-line option with the Python file, for example:

python run_pretraining.py --help
python run_squad.py --help
python run_classifier.py --help

Getting the data

For pre-training BERT, we use the concatenation of Wikipedia (2500M words) as well as BookCorpus (800M words). For Wikipedia, we extract only the text passages from here and ignore headers list and tables. It is structured as a document level corpus rather than a shuffled sentence level corpus because it is critical to extract long contiguous sentences.

The next step is to run create_pretraining_data.py with the document level corpus as input, which generates input data and labels for the masked language modeling and next sentence prediction tasks. Pre-training can also be performed on any corpus of your choice. The collection of data generation scripts are intended to be modular to allow modifications for additional preprocessing steps or to use additional data. They can hence easily be modified for an arbitrary corpus.

The preparation of an individual pre-training dataset is described in the create_datasets_from_start.sh script found in the data/ folder. The component steps to prepare the datasets are as follows:

  1. Data download and extract - the dataset is downloaded and extracted.
  2. Clean and format - document tags, etc. are removed from the dataset. The end result of this step is a {dataset_name_one_article_per_line}.txt file that contains the entire corpus. Each line in the text file contains an entire document from the corpus. One file per dataset is created in the formatted_one_article_per_line folder.
  3. Sharding - the sentence segmented corpus file is split into a number of smaller text documents. The sharding is configured so that a document will not be split between two shards. Sentence segmentation is performed at this time using NLTK.
  4. TFRecord file creation - each text file shard is processed by the create_pretraining_data.py script to produce a corresponding TFRecord file. The script generates input data and labels for masked language modeling and sentence prediction tasks for the input text shard.

For fine tuning BERT for the task of Question Answering, we use SQuAD and GLUE. SQuAD v1.1 has 100,000+ question-answer pairs on 500+ articles. SQuAD v2.0 combines v1.1 with an additional 50,000 new unanswerable questions and must not only answer questions but also determine when that is not possible. GLUE consists of single-sentence tasks, similarity and paraphrase tasks and inference tasks. We support one of each: CoLA, MNLI and MRPC.

Dataset guidelines

The procedure to prepare a text corpus for pre-training is described in the previous section. This section provides additional insight into how exactly raw text is processed so that it is ready for pre-training.

First, raw text is tokenized using WordPiece tokenization. A [CLS] token is inserted at the start of every sequence, and the two sentences in the sequence are separated by a [SEP] token.

Note: BERT pre-training looks at pairs of sentences at a time. A sentence embedding token [A] is added to the first sentence and token [B] to the next.

BERT pre-training optimizes for two unsupervised classification tasks. The first is Masked Language Modelling (Masked LM). One training instance of Masked LM is a single modified sentence. Each token in the sentence has a 15% chance of being replaced by a [MASK] token. The chosen token is replaced with [MASK] 80% of the time, 10% with another random token and the remaining 10% with the same token. The task is then to predict the original token.

The second task is next sentence prediction. One training instance of BERT pre-training is two sentences (a sentence pair). A sentence pair may be constructed by simply taking two adjacent sentences from a single document, or by pairing up two random sentences with equal probability. The goal of this task is to predict whether or not the second sentence followed the first in the original document.

The create_pretraining_data.py script takes in raw text and creates training instances for both pre-training tasks.

Multi-dataset

We are able to combine multiple datasets into a single dataset for pre-training on a diverse text corpus. Once TFRecords have been created for each component dataset, you can create a combined dataset by adding the directory to SOURCES in run_pretraining_*.sh. This will feed all matching files to the input pipeline in run_pretraining.py. However, in the training process, only one TFRecord file is consumed at a time, therefore, the training instances of any given training batch will all belong to the same source dataset.

Training process

The training process consists of two steps: pre-training and fine tuning.

Pre-training

Pre-training is performed using the run_pretraining.py script along with parameters defined in the scripts/run_pretraining_lamb.sh.

The run_pretraining_lamb.sh script runs a job on a single node that trains the BERT-large model from scratch using the Wikipedia and BookCorpus datasets as training data. By default, the training script:

  • Runs on 8 GPUs.

  • Has FP16 precision enabled.

  • Is XLA enabled.

  • Creates a log file containing all the output.

  • Saves a checkpoint every 100 iterations (keeps only the latest checkpoint) and at the end of training. All checkpoints, evaluation results and training logs are saved to the /results directory (in the container which can be mounted to a local directory).

  • Evaluates the model at the end of each phase.

  • Phase 1

    • Runs 7038 steps with 2000 warmup steps
    • Sets Maximum sequence length as 128
    • Sets Global Batch size as 64K

  • Phase 2

    • Runs 1564 steps with 200 warm-up steps
    • Sets Maximum sequence length as 512
    • Sets Global Batch size as 32K
    • Starts from Phase1's final checkpoint

These parameters train Wikipedia and BookCorpus with reasonable accuracy on a DGX-1 with 32GB V100 cards.

For example:

scripts/run_pretraining_lamb.sh <train_batch_size_phase1> <train_batch_size_phase2> <eval_batch_size> <learning_rate_phase1> <learning_rate_phase2> <precision> <use_xla> <num_gpus> <warmup_steps_phase1> <warmup_steps_phase2> <train_steps> <save_checkpoint_steps> <num_accumulation_phase1> <num_accumulation_steps_phase2> <bert_model>

Where:

  • <training_batch_size_phase*> is per-GPU batch size used for training in the respective phase. Batch size varies with precision, larger batch sizes run more efficiently, but require more memory.

  • <eval_batch_size> is per-GPU batch size used for evaluation after training.

  • <learning_rate_phase1> is the default rate of 1e-4 is good for global batch size 256.

  • <learning_rate_phase2> is the default rate of 1e-4 is good for global batch size 256.

  • <precision> is the type of math in your model, can be either fp32 or fp16. Specifically:

    • fp32 is 32-bit IEEE single precision floats. Is enabled by default on V100.
    • fp16 is Automatic rewrite of TensorFlow compute graph to take advantage of 16-bit arithmetic whenever it is safe.
    • tf32 uses same 10 bit mantissa as fp16 and 8 bit exponent as fp32. Is enabled by default on A100.

  • <num_gpus> is the number of GPUs to use for training. Must be equal to or smaller than the number of GPUs attached to your node.

  • <warmup_steps_phase*> is the number of warm-up steps at the start of training in the respective phase.

  • <training_steps> is the total number of training steps in both phases combined.

  • <save_checkpoint_steps> controls how often checkpoints are saved. Default is 100 steps.

  • <num_accumulation_phase*> is used to mimic higher batch sizes in the respective phase by accumulating gradients N times before weight update.

  • <bert_model> is used to indicate whether to pretrain BERT Large or BERT Base model

The following sample code trains BERT-large from scratch on a single DGX-2 using FP16 arithmetic. This will take around 4.5 days.

bert_tf/scripts/run_pretraining_lamb.sh 32 8 8 3.75e-4 2.5e-4 fp16 true 16 2000 200 7820 100 128 256 large

Fine tuning

Fine tuning is performed using the run_squad.py script along with parameters defined in scripts/run_squad.sh.

The run_squad.sh script trains a model and performs evaluation on the SQuAD dataset. By default, the training script:

  • Trains for SQuAD v1.1 dataset.
  • Trains on BERT Large Model.
  • Uses 8 GPUs and batch size of 10 on each GPU.
  • Has FP16 precision enabled.
  • Is XLA enabled.
  • Runs for 2 epochs.
  • Saves a checkpoint every 1000 iterations (keeps only the latest checkpoint) and at the end of training. All checkpoints, evaluation results and training logs are saved to the /results directory (in the container which can be mounted to a local directory).
  • Evaluation is done at the end of training. To skip evaluation, modify --do_predict to False.

This script outputs checkpoints to the /results directory, by default, inside the container. Mount point of /results can be changed in the scripts/docker/launch.sh file. The training log contains information about:

  • Loss for the final step
  • Training and evaluation performance
  • F1 and exact match score on the Dev Set of SQuAD after evaluation.

The summary after training is printed in the following format:

I0312 23:10:45.137036 140287431493376 run_squad.py:1332] 0 Total Training Time = 3007.00 Training Time W/O start up overhead = 2855.92 Sentences processed = 175176
I0312 23:10:45.137243 140287431493376 run_squad.py:1333] 0 Training Performance = 61.3378 sentences/sec
I0312 23:14:00.550846 140287431493376 run_squad.py:1396] 0 Total Inference Time = 145.46 Inference Time W/O start up overhead = 131.86 Sentences processed = 10840
I0312 23:14:00.550973 140287431493376 run_squad.py:1397] 0 Inference Performance = 82.2095 sentences/sec
{"exact_match": 83.69914853358561, "f1": 90.8477003317459}

Multi-GPU training is enabled with the Horovod TensorFlow module. The following example runs fine tuning on 8 GPUs:

BERT_DIR=data/download/google_pretrained_weights/uncased_L-24_H-1024_A-16
SQUAD_DIR=data/download/squad/v1.1
mpi_command="mpirun -np 8 -H localhost:8 \
    --allow-run-as-root -bind-to none -map-by slot \
    -x NCCL_DEBUG=INFO \
    -x LD_LIBRARY_PATH \
    -x PATH -mca pml ob1 -mca btl ^openib" \
     python run_squad.py --horovod --vocab_file=$BERT_DIR/vocab.txt \
     --bert_config_file=$BERT_DIR/bert_config.json \
     --output_dir=/results --do_train --train_file=$SQUAD_DIR/train-v1.1.json

Multi-node

Multi-node runs can be launched on a pyxis/enroot Slurm cluster (see Requirements) with the run.sub script with the following command for a 4-node DGX1 example for both phase 1 and phase 2:

BATCHSIZE=16 LEARNING_RATE='1.875e-4' NUM_ACCUMULATION_STEPS=128 PHASE=1 sbatch -N4 --ntasks-per-node=8 run.sub
BATCHSIZE=2 LEARNING_RATE='1.25e-4' NUM_ACCUMULATION_STEPS=512 PHASE=1 sbatch -N4 --ntasks-per-node=8 run.sub

Checkpoint after phase 1 will be saved in checkpointdir specified in run.sub. The checkpoint will be automatically picked up to resume training on phase 2. Note that phase 2 should be run after phase 1.

Variables to re-run the Training performance results are available in the scripts/configs/configurations.yml file.

The batch variables BATCHSIZE, LEARNING_RATE, NUM_ACCUMULATION_STEPS refer to the Python arguments train_batch_size, learning_rate, num_accumulation_steps respectively. The variable PHASE refers to phase specific arguments available in run.sub.

Note that the run.sub script is a starting point that has to be adapted depending on the environment. In particular, variables such as datadir handle the location of the files for each phase.

Refer to the files contents to see the full list of variables to adjust for your system.

Inference process

Inference on a fine tuned Question Answering system is performed using the run_squad.py script along with parameters defined in scripts/run_squad_inference.sh. Inference is supported on a single GPU.

The run_squad_inference.sh script trains a model and performs evaluation on the SQuAD dataset. By default, the inferencing script:

  • Uses SQuAD v1.1 dataset
  • Has FP16 precision enabled
  • Is XLA enabled
  • Evaluates the latest checkpoint present in /results with a batch size of 8

This script outputs predictions file to /results/predictions.json and computes F1 score and exact match score using SQuAD's evaluate file. Mount point of /results can be changed in the scripts/docker/launch.sh file.

The output log contains information about: Inference performance Inference Accuracy (F1 and exact match scores) on the Dev Set of SQuAD after evaluation.

The summary after inference is printed in the following format:

I0312 23:14:00.550846 140287431493376 run_squad.py:1396] 0 Total Inference Time = 145.46 Inference Time W/O start up overhead = 131.86 Sentences processed = 10840
I0312 23:14:00.550973 140287431493376 run_squad.py:1397] 0 Inference Performance = 82.2095 sentences/sec
{"exact_match": 83.69914853358561, "f1": 90.8477003317459}

Inference Process With TensorRT

NVIDIA TensorRT is a platform for high-performance deep learning inference. It includes a deep learning inference optimizer and runtime that delivers low latency and high-throughput for deep learning inference applications. More information on how to perform inference using TensorRT can be found in the subfolder ./trt/README.md

Deploying the BERT model using Triton Inference Server

The NVIDIA Triton Inference Server provides a datacenter and cloud inferencing solution optimized for NVIDIA GPUs. The server provides an inference service via an HTTP or gRPC endpoint, allowing remote clients to request inferencing for any number of GPU or CPU models being managed by the server. More information on how to perform inference using Triton Inference Server can be found in the subfolder ./triton/README.md.

BioBERT

Many works, including BioBERT, SciBERT, NCBI-BERT, ClinicalBERT (MIT), ClinicalBERT (NYU, Princeton), and others at BioNLP’19 workshop, show that pre-training of BERT on large biomedical text corpus such as PubMed results in better performance in biomedical text-mining tasks.

More information on how to download a biomedical corpus and pre-train as well as finetune for biomedical tasks can be found in the subfolder ./biobert/README.md.

Performance

The performance measurements in this document were conducted at the time of publication and may not reflect the performance achieved from NVIDIA’s latest software release. For the most up-to-date performance measurements, go to NVIDIA Data Center Deep Learning Product Performance.

Benchmarking

The following section shows how to run benchmarks measuring the model performance in training and inference modes.

Both of these benchmarking scripts enable you to run a number of epochs, extract performance numbers, and run the BERT model for fine tuning.

Training performance benchmark

Training benchmarking can be performed by running the script:

scripts/finetune_train_benchmark.sh <bert_model> <use_xla> <num_gpu> squad

This script runs 2 epochs by default on the SQuAD v1.1 dataset and extracts performance numbers for various batch sizes and sequence lengths in both FP16 and FP32/TF32. These numbers are saved at /results/squad_train_benchmark_bert_<bert_model>_gpu_<num_gpu>.log.

Inference performance benchmark

Inference benchmarking can be performed by running the script:

scripts/finetune_inference_benchmark.sh squad

This script runs 1024 eval iterations by default on the SQuAD v1.1 dataset and extracts performance and latency numbers for various batch sizes and sequence lengths in both FP16 and FP32/TF32, for base and large models. These numbers are saved at /results/squad_inference_benchmark_bert_<bert_model>.log.

Results

The following sections provide details on how we achieved our performance and accuracy in training and inference for pre-training using LAMB optimizer as well as fine tuning for Question Answering. All results are on BERT-large model unless otherwise mentioned. All fine tuning results are on SQuAD v1.1 using a sequence length of 384 unless otherwise mentioned.

Training accuracy results

Training accuracy
Pre-training accuracy

Our results were obtained by running the scripts/run_pretraining_lamb.sh training script in the TensorFlow 20.06-py3 NGC container.

DGX System Nodes x GPUs Precision Batch Size/GPU: Phase1, Phase2 Accumulation Steps: Phase1, Phase2 Time to Train (Hrs) Final Loss
DGX2H 32 x 16 FP16 64, 8 2, 8 2.63 1.59
DGX2H 32 x 16 FP32 32, 8 4, 8 8.48 1.56
DGXA100 32 x 8 FP16 64, 16 4, 8 3.24 1.56
DGXA100 32 x 8 TF32 64, 8 4, 16 4.58 1.58

Note: Time to train includes upto 16 minutes of start up time for every restart (atleast once for each phase). Experiments were run on clusters with a maximum wall clock time of 8 hours.

Fine-tuning accuracy for SQuAD v1.1: NVIDIA DGX A100 (8x A100 40G)

Our results were obtained by running the scripts/run_squad.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX A100 with 8x A100 40GB GPUs.

GPUs **Batch size / GPU: TF32, FP16 ** Accuracy - TF32 Accuracy - mixed precision Time to Train - TF32 (Hrs) Time to Train - mixed precision (Hrs)
8 16, 24 91.41 91.52 0.26 0.26
Fine-tuning accuracy for GLUE MRPC: NVIDIA DGX A100 (8x A100 40G)

Our results were obtained by running the scripts/run_glue.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX A100 with 8x A100 40GB GPUs for 10 different seeds and picking the maximum accuracy on MRPC dev set.

GPUs Batch size / GPU Accuracy - TF32 Accuracy - mixed precision Time to Train - TF32 (Hrs) Time to Train - mixed precision (Hrs) Throughput - TF32 **Throughput - mixed precision **
8 16 87.99 87.09 0.009 0.009 357.91 230.16
Training stability test
Pre-training SQuAD v1.1 stability test: NVIDIA DGX A100 (256x A100 40GB)

The following tables compare Final Loss scores across 2 different training runs with different seeds, for both FP16 and TF32. The runs showcase consistent convergence on all 2 seeds with very little deviation.

FP16, 256x GPUs seed 1 seed 2 mean std
Final Loss 1.570 1.561 1.565 0.006
TF32, 256x GPUs seed 1 seed 2 mean std
Final Loss 1.583 1.582 1.582 0.0007
Fine-tuning SQuAD v1.1 stability test: NVIDIA DGX A100 (8x A100 40GB)

The following tables compare F1 scores across 5 different training runs with different seeds, for both FP16 and TF32 respectively using (Nvidia's Pretrained Checkpoint)[https://ngc.nvidia.com/catalog/models/nvidia:bert_tf_pretraining_lamb_16n]. The runs showcase consistent convergence on all 5 seeds with very little deviation.

FP16, 8x GPUs seed 1 seed 2 seed 3 seed 4 seed 5 mean std
F1 91.61 91.04 91.59 91.32 91.52 91.41 0.24
TF32, 8x GPUs seed 1 seed 2 seed 3 seed 4 seed 5 mean std
F1 91.50 91.49 91.64 91.29 91.67 91.52 0.15
Fine-tuning GLUE MRPC stability test: NVIDIA DGX A100 (8x A100 40GB)

The following tables compare F1 scores across 10 different training runs with different seeds, for both FP16 and TF32 respectively using (Nvidia's Pretrained Checkpoint)[https://ngc.nvidia.com/catalog/models/nvidia:bert_tf_pretraining_lamb_16n]. The runs showcase consistent convergence on all 10 seeds with very little deviation.

** FP16, 8 GPUs ** ** seed 1 ** ** seed 2 ** ** seed 3 ** ** seed 4 ** ** seed 5 ** ** seed 6 ** ** seed 7 ** ** seed 8 ** ** seed 9 ** ** seed 10 ** ** Mean ** ** Std **
Eval Accuracy 84.31372643 85.78431606 86.76471114 87.00980544 86.27451062 86.27451062 85.5392158 86.51961088 86.27451062 85.2941215 86.00490391 0.795887906
** TF32, 8 GPUs ** ** seed 1 ** ** seed 2 ** ** seed 3 ** ** seed 4 ** ** seed 5 ** ** seed 6 ** ** seed 7 ** ** seed 8 ** ** seed 9 ** ** seed 10 ** ** Mean ** ** Std **
Eval Accuracy 87.00980544 86.27451062 87.99020052 86.27451062 86.02941632 87.00980544 86.27451062 86.51961088 87.74510026 86.02941632 86.7156887 0.7009024515

Training performance results

Training performance: NVIDIA DGX-1 (8x V100 16GB)
Pre-training training performance: single-node on DGX-1 16GB

Our results were obtained by running the scripts/run_pretraining_lamb.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX-1 with 8x V100 16GB GPUs. Performance (in sentences per second) is the steady state throughput.

GPUs Sequence Length Batch size / GPU: mixed precision, FP32 Gradient Accumulation: mixed precision, FP32 Global Batch Size Throughput - mixed precision Throughput - FP32 Throughput speedup (FP32 - mixed precision) Weak scaling - mixed precision Weak scaling - FP32
1 128 16 , 8 4096, 8192 65536 134.34 39.43 3.41 1.00 1.00
4 128 16 , 8 1024, 2048 65536 449.68 152.33 2.95 3.35 3.86
8 128 16 , 8 512, 1024 65536 1001.39 285.79 3.50 7.45 7.25
1 512 4 , 2 8192, 16384 32768 28.72 9.80 2.93 1.00 1.00
4 512 4 , 2 2048, 4096 32768 109.96 35.32 3.11 3.83 3.60
8 512 4 , 2 1024, 2048 32768 190.65 69.53 2.74 6.64 7.09

Note: The respective values for FP32 runs that use a batch size of 16, 4 in sequence lengths 128 and 512 respectively are not available due to out of memory errors that arise.

Fine-tuning training performance for SQuAD v1.1 on DGX-1 16GB

Our results were obtained by running the scripts/run_squad.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX-1 with 8x V100 16GB GPUs. Performance (in sentences per second) is the steady state throughput.

GPUs Batch size / GPU: mixed precision, FP32 Throughput - mixed precision Throughput - FP32 Throughput speedup (FP32 to mixed precision) Weak scaling - FP32 Weak scaling - mixed precision
1 4,2 29.74 7.36 4.04 1.00 1.00
4 4,2 97.28 26.64 3.65 3.27 3.62
8 4,2 189.77 52.39 3.62 6.38 7.12

Note: The respective values for FP32 runs that use a batch size of 4 are not available due to out of memory errors that arise.

To achieve these same results, follow the Quick Start Guide outlined above.

Training performance: NVIDIA DGX-1 (8x V100 32GB)
Pre-training training performance: single-node on DGX-1 32GB

Our results were obtained by running the scripts/run_pretraining_lamb.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX-1 with 8x V100 32GB GPUs. Performance (in sentences per second) is the steady state throughput.

GPUs Sequence Length Batch size / GPU: mixed precision, FP32 Gradient Accumulation: mixed precision, FP32 Global Batch Size Throughput - mixed precision Throughput - FP32 Throughput speedup (FP32 - mixed precision) Weak scaling - mixed precision Weak scaling - FP32
1 128 64 , 32 1024, 2048 65536 168.63 46.78 3.60 1.00 1.00
4 128 64 , 32 256, 512 65536 730.25 179.73 4.06 4.33 3.84
8 128 64 , 32 128, 256 65536 1443.05 357.00 4.04 8.56 7.63
1 512 8 , 8 4096, 4096 32768 31.23 10.67 2.93 1.00 1.00
4 512 8 , 8 1024, 1024 32768 118.84 39.55 3.00 3.81 3.71
8 512 8 , 8 512, 512 32768 255.64 81.42 3.14 8.19 7.63

Note: The respective values for FP32 runs that use a batch size of 64 in sequence lengths 128 are not available due to out of memory errors that arise.

Fine-tuning training performance for SQuAD v1.1 on DGX-1 32GB

Our results were obtained by running the scripts/run_squad.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX-1 with 8x V100 32GB GPUs. Performance (in sentences per second) is the steady state throughput.

GPUs Batch size / GPU: mixed precision, FP32 Throughput - mixed precision Throughput - FP32 Throughput speedup (FP32 to mixed precision) Weak scaling - FP32 Weak scaling - mixed precision
1 24, 10 51.02 10.42 4.90 1.00 1.00
4 24, 10 181.37 39.77 4.56 3.55 3.82
8 24, 10 314.6 79.37 3.96 6.17 7.62

Note: The respective values for FP32 runs that use a batch size of 24 are not available due to out of memory errors that arise.

To achieve these same results, follow the Quick Start Guide outlined above.

Training performance: NVIDIA DGX-2 (16x V100 32GB)
Pre-training training performance: single-node on DGX-2 32GB

Our results were obtained by running the scripts/run_pretraining_lamb.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX-2 with 16x V100 32GB GPUs. Performance (in sentences per second) is the steady state throughput.

GPUs Sequence Length Batch size / GPU: mixed precision, FP32 Gradient Accumulation: mixed precision, FP32 Global Batch Size Throughput - mixed precision Throughput - FP32 Throughput speedup (FP32 - mixed precision) Weak scaling - mixed precision Weak scaling - FP32
1 128 64 , 32 1024 , 8192 65536 188.04 35.32 5.32 1.00 1.00
4 128 64 , 32 256 , 2048 65536 790.89 193.08 4.10 4.21 5.47
8 128 64 , 32 128 , 1024 65536 1556.89 386.89 4.02 8.28 10.95
16 128 64 , 32 64 , 128 65536 3081.69 761.92 4.04 16.39 21.57
1 512 8 , 8 4096 , 4096 32768 35.32 11.67 3.03 1.00 1.00
4 512 8 , 8 1024 , 1024 32768 128.98 42.84 3.01 3.65 3.67
8 512 8 , 8 512 , 512 32768 274.04 86.78 3.16 7.76 7.44
16 512 8 , 8 256 , 256 32768 513.43 173.26 2.96 14.54 14.85

Note: The respective values for FP32 runs that use a batch size of 64 in sequence lengths 128 are not available due to out of memory errors that arise.

Pre-training training performance: multi-node on DGX-2H 32GB

Our results were obtained by running the run.sub training script in the TensorFlow 19.08-py3 NGC container using multiple NVIDIA DGX-2 with 16x V100 32GB GPUs. Performance (in sentences per second) is the steady state throughput.

Num Nodes Sequence Length Batch size / GPU: mixed precision, FP32 Gradient Accumulation: mixed precision, FP32 Global Batch Size Throughput - mixed precision Throughput - FP32 Throughput speedup (FP32 - mixed precision) Weak scaling - mixed precision Weak scaling - FP32
1 128 64 , 32 64 , 128 65536 3081.69 761.92 4.04 1.00 1.00
4 128 64 , 32 16 , 32 65536 13192.00 3389.83 3.89 4.28 4.45
16 128 64 , 32 4 , 8 65536 48223.00 13217.78 3.65 15.65 17.35
32 128 64 , 32 2 , 4 65536 86673.64 25142.26 3.45 28.13 33.00
1 512 8 , 8 256 , 256 32768 577.79 173.26 3.33 1.00 1.00
4 512 8 , 8 64 , 64 32768 2284.23 765.04 2.99 3.95 4.42
16 512 8 , 8 16 , 16 32768 8853.00 3001.43 2.95 15.32 17.32
32 512 8 , 8 8 , 8 32768 17059.00 5893.14 2.89 29.52 34.01

Note: The respective values for FP32 runs that use a batch size of 64 in sequence lengths 128 are not available due to out of memory errors that arise.

Fine-tuning training performance for SQuAD v1.1 on DGX-2 32GB

Our results were obtained by running the scripts/run_squad.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX-2 with 16x V100 32GB GPUs. Performance (in sentences per second) is the steady state throughput.

GPUs Batch size / GPU: mixed precision, FP32 Throughput - mixed precision Throughput - FP32 Throughput speedup (FP32 to mixed precision) Weak scaling - FP32 Weak scaling - mixed precision
1 24, 10 55.28 11.15 4.96 1.00 1.00
4 24, 10 199.53 42.91 4.65 3.61 3.85
8 24, 10 341.55 85.08 4.01 6.18 7.63
16 24, 10 683.37 156.29 4.37 12.36 14.02

Note: The respective values for FP32 runs that use a batch size of 24 are not available due to out of memory errors that arise.

To achieve these same results, follow the Quick Start Guide outlined above.

Training performance: NVIDIA DGX A100 (8x A100 40GB)
Pre-training training performance: single-node on DGX A100 40GB

Our results were obtained by running the scripts/run_pretraining_lamb.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX A100 with 8x A100 40GB GPUs. Performance (in sentences per second) is the steady state throughput.

GPUs Sequence Length Batch size / GPU: mixed precision, TF32 Gradient Accumulation: mixed precision, TF32 Global Batch Size Throughput - mixed precision Throughput - TF32 Throughput speedup (TF32 - mixed precision) Weak scaling - mixed precision Weak scaling -TF32
1 128 64 , 64 1024 , 1024 65536 356.845 238.10 1.50 1.00 1.00
4 128 64 , 64 256 , 256 65536 1422.25 952.39 1.49 3.99 4.00
8 128 64 , 64 128 , 128 65536 2871.89 1889.71 1.52 8.05 7.94
1 512 16 , 8 2048 , 4096 32768 70.856 39.96 1.77 1.00 1.00
4 512 16 , 8 512 , 1024 32768 284.912 160.16 1.78 4.02 4.01
8 512 16 , 8 256 , 512 32768 572.112 316.51 1.81 8.07 7.92

Note: The respective values for TF32 runs that use a batch size of 16 for sequence length 512 are not available due to out of memory errors that arise.

Pre-training training performance: multi-node on DGX A100 40GB

Our results were obtained by running the scripts/run_pretraining_lamb.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX A100 with 8x A100 40GB GPUs. Performance (in sentences per second) is the steady state throughput.

Num Nodes Sequence Length Batch size / GPU: mixed precision, TF32 Gradient Accumulation: mixed precision, TF32 Global Batch Size Throughput - mixed precision Throughput - TF32 Throughput speedup (TF32 - mixed precision) Weak scaling - mixed precision Weak scaling -TF32
1 128 64 , 64 128 , 128 65536 2871.89 1889.71 1.52 1.00 1.00
4 128 64 , 64 32 , 32 65536 11159 7532.00 1.48 3.89 3.99
16 128 64 , 64 8 , 8 65536 41144 28605.62 1.44 14.33 15.14
32 128 64 , 64 4 , 4 65536 77479.87 53585.82 1.45 26.98 28.36
1 512 16 , 8 256 , 512 32768 572.112 316.51 1.81 1.00 1.00
4 512 16 , 8 128 , 128 65536 2197.44 1268.43 1.73 3.84 4.01
16 512 16 , 8 32 , 32 65536 8723.1 4903.39 1.78 15.25 15.49
32 512 16 , 8 16 , 16 65536 16705 9463.80 1.77 29.20 29.90

Note: The respective values for TF32 runs that use a batch size of 16 for sequence length 512 are not available due to out of memory errors that arise.

Fine-tuning training performance for SQuAD v1.1 on DGX A100 40GB

Our results were obtained by running the scripts/run_squad.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX A100 with 8x A100 40GB GPUs. Performance (in sentences per second) is the steady state throughput.

GPUs Batch size / GPU: mixed precision, TF32 Throughput - mixed precision Throughput - TF32 Throughput speedup (TF32 to mixed precision) Weak scaling - TF32 Weak scaling - mixed precision
1 32, 16 102.26 61.364 1.67 1.00 1.00
4 32, 16 366.353 223.187 1.64 3.64 3.58
8 32, 16 767.071 440.47 1.74 7.18 7.50

Note: The respective values for TF32 runs that use a batch size of 32 are not available due to out of memory errors that arise.

To achieve these same results, follow the Quick Start Guide outlined above.

Inference performance results

Inference performance: NVIDIA DGX-1 (1x V100 16GB)
Fine-tuning inference performance for SQuAD v1.1 on 16GB

Our results were obtained by running the scripts/finetune_inference_benchmark.sh script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX-1 with 1x V100 16GB GPUs. Performance numbers (throughput in sentences per second and latency in milliseconds) were averaged from 1024 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.

Model Sequence Length Batch Size Precision Throughput-Average(sent/sec) Latency-Average(ms) Latency-90%(ms) Latency-95%(ms) Latency-99%(ms)
base 128 1 fp16 206.82 7.96 4.98 5.04 5.23
base 128 2 fp16 376.75 8.68 5.42 5.49 5.64
base 128 4 fp16 635 12.31 6.46 6.55 6.83
base 128 8 fp16 962.83 13.64 8.47 8.56 8.75
base 384 1 fp16 167.01 12.77 6.12 6.23 6.52
base 384 2 fp16 252.12 21.05 8.03 8.09 8.61
base 384 4 fp16 341.95 25.09 11.88 11.96 12.52
base 384 8 fp16 421.26 33.16 19.2 19.37 19.91
base 128 1 fp32 174.48 8.17 5.89 5.95 6.12
base 128 2 fp32 263.67 10.33 7.66 7.69 7.92
base 128 4 fp32 349.34 16.31 11.57 11.62 11.87
base 128 8 fp32 422.88 23.27 19.23 19.38 20.38
base 384 1 fp32 99.52 14.99 10.19 10.23 10.78
base 384 2 fp32 118.01 25.98 17.12 17.18 17.78
base 384 4 fp32 128.1 41 31.56 31.7 32.39
base 384 8 fp32 136.1 69.77 59.44 59.66 60.51
large 128 1 fp16 98.63 15.86 10.27 10.31 10.46
large 128 2 fp16 172.59 17.78 11.81 11.86 12.13
large 128 4 fp16 272.86 25.66 14.86 14.94 15.18
large 128 8 fp16 385.64 30.74 20.98 21.1 21.68
large 384 1 fp16 70.74 26.85 14.38 14.47 14.7
large 384 2 fp16 99.9 45.29 20.26 20.43 21.11
large 384 4 fp16 128.42 56.94 31.44 31.71 32.45
large 384 8 fp16 148.57 81.69 54.23 54.54 55.53
large 128 1 fp32 76.75 17.06 13.21 13.27 13.4
large 128 2 fp32 100.82 24.34 20.05 20.13 21.13
large 128 4 fp32 117.59 41.76 34.42 34.55 35.29
large 128 8 fp32 130.42 68.59 62 62.23 62.98
large 384 1 fp32 33.95 37.89 29.82 29.98 30.56
large 384 2 fp32 38.47 68.35 52.56 52.74 53.89
large 384 4 fp32 41.11 114.27 98.19 98.54 99.54
large 384 8 fp32 41.32 213.84 194.92 195.36 196.94

To achieve these same results, follow the Quick Start Guide outlined above.

Inference performance: NVIDIA DGX-1 (1x V100 32GB)
Fine-tuning inference performance for SQuAD v1.1 on 32GB

Our results were obtained by running the scripts/finetune_inference_benchmark.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX-1 with 1x V100 32GB GPUs. Performance numbers (throughput in sentences per second and latency in milliseconds) were averaged from 1024 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.

Model Sequence Length Batch Size Precision Throughput-Average(sent/sec) Latency-Average(ms) Latency-90%(ms) Latency-95%(ms) Latency-99%(ms)
base 128 1 fp16 207.87 7.63 4.94 5.03 5.32
base 128 2 fp16 376.44 8.47 5.44 5.5 5.68
base 128 4 fp16 642.55 11.63 6.3 6.36 6.68
base 128 8 fp16 943.85 13.24 8.56 8.68 8.92
base 384 1 fp16 162.62 12.24 6.31 6.4 6.73
base 384 2 fp16 244.15 20.05 8.34 8.41 8.93
base 384 4 fp16 338.68 23.53 11.88 11.92 12.63
base 384 8 fp16 407.46 32.72 19.84 20.06 20.89
base 128 1 fp32 175.16 8.31 5.85 5.89 6.04
base 128 2 fp32 261.31 10.48 7.75 7.81 8.08
base 128 4 fp32 339.45 16.67 11.95 12.02 12.46
base 128 8 fp32 406.67 24.12 19.86 19.97 20.41
base 384 1 fp32 98.33 15.28 10.27 10.32 10.76
base 384 2 fp32 114.92 26.88 17.55 17.59 18.29
base 384 4 fp32 125.76 41.74 32.06 32.23 33.72
base 384 8 fp32 136.62 69.78 58.95 59.19 60
large 128 1 fp16 96.46 15.56 10.56 10.66 11.02
large 128 2 fp16 168.31 17.42 12.11 12.25 12.57
large 128 4 fp16 267.76 24.76 15.17 15.36 16.68
large 128 8 fp16 378.28 30.34 21.39 21.54 21.97
large 384 1 fp16 68.75 26.02 14.77 14.94 15.3
large 384 2 fp16 95.41 44.01 21.24 21.47 22.01
large 384 4 fp16 124.43 55.14 32.53 32.83 33.58
large 384 8 fp16 143.02 81.37 56.51 56.88 58.05
large 128 1 fp32 75.34 17.5 13.46 13.52 13.7
large 128 2 fp32 99.73 24.7 20.27 20.38 21.45
large 128 4 fp32 116.92 42.1 34.49 34.59 34.98
large 128 8 fp32 130.11 68.95 62.03 62.23 63.3
large 384 1 fp32 33.84 38.15 29.75 29.89 31.23
large 384 2 fp32 38.02 69.31 53.1 53.36 54.42
large 384 4 fp32 41.2 114.34 97.96 98.32 99.55
large 384 8 fp32 42.37 209.16 190.18 190.66 192.77

To achieve these same results, follow the Quick Start Guide outlined above.

Inference performance: NVIDIA DGX-2 (1x V100 32GB)
Fine-tuning inference performance for SQuAD v1.1 on DGX-2 32GB

Our results were obtained by running the scripts/finetune_inference_benchmark.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX-2 with 1x V100 32GB GPUs. Performance numbers (throughput in sentences per second and latency in milliseconds) were averaged from 1024 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.

Model Sequence Length Batch Size Precision Throughput-Average(sent/sec) Latency-Average(ms) Latency-90%(ms) Latency-95%(ms) Latency-99%(ms)
base 128 1 fp16 220.35 7.82 4.7 4.83 5.15
base 128 2 fp16 384.55 8.7 5.49 5.68 6.01
base 128 4 fp16 650.7 36.3 6.35 6.51 6.87
base 128 8 fp16 992.41 13.59 8.22 8.37 8.96
base 384 1 fp16 172.89 12.86 5.94 6.04 6.44
base 384 2 fp16 258.48 20.42 7.89 8.09 9.15
base 384 4 fp16 346.34 24.93 11.97 12.12 12.76
base 384 8 fp16 430.4 33.08 18.75 19.27 20.12
base 128 1 fp32 183.69 7.52 5.86 5.97 6.27
base 128 2 fp32 282.95 9.51 7.31 7.49 7.83
base 128 4 fp32 363.83 15.12 11.35 11.47 11.74
base 128 8 fp32 449.12 21.65 18 18.1 18.6
base 384 1 fp32 104.92 13.8 9.9 9.99 10.48
base 384 2 fp32 123.55 24.21 16.29 16.4 17.61
base 384 4 fp32 139.38 36.69 28.89 29.04 30.01
base 384 8 fp32 146.28 64.69 55.09 55.32 56.3
large 128 1 fp16 98.34 15.85 10.61 10.78 11.5
large 128 2 fp16 172.95 17.8 11.91 12.08 12.42
large 128 4 fp16 278.82 25.18 14.7 14.87 15.65
large 128 8 fp16 402.28 30.45 20.21 20.43 21.24
large 384 1 fp16 71.1 26.55 14.44 14.61 15.32
large 384 2 fp16 100.48 44.04 20.31 20.48 21.6
large 384 4 fp16 131.68 56.19 30.8 31.03 32.3
large 384 8 fp16 151.81 81.53 53.22 53.87 55.34
large 128 1 fp32 77.87 16.33 13.33 13.45 13.77
large 128 2 fp32 105.41 22.77 19.39 19.52 19.86
large 128 4 fp32 124.16 38.61 32.69 32.88 33.9
large 128 8 fp32 137.69 64.61 58.62 58.89 59.94
large 384 1 fp32 36.34 34.94 27.72 27.81 28.21
large 384 2 fp32 41.11 62.54 49.14 49.32 50.25
large 384 4 fp32 43.32 107.53 93.07 93.47 94.27
large 384 8 fp32 44.64 196.28 180.21 180.75 182.41
Inference performance: NVIDIA DGX A100 (1x A100 40GB)
Fine-tuning inference performance for SQuAD v1.1 on DGX A100 40GB

Our results were obtained by running the scripts/finetune_inference_benchmark.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA DGX A100 with 1x A100 40GB GPUs. Performance numbers (throughput in sentences per second and latency in milliseconds) were averaged from 1024 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.

Model Sequence Length Batch Size Precision Throughput-Average(sent/sec) Latency-Average(ms) Latency-90%(ms) Latency-95%(ms) Latency-99%(ms)
base 128 1 fp16 231.37 6.43 4.57 4.68 4.93
base 128 2 fp16 454.54 6.77 4.66 4.77 4.96
base 128 4 fp16 842.34 8.8 4.91 4.98 5.39
base 128 8 fp16 1216.43 10.39 6.77 6.86 7.28
base 384 1 fp16 210.59 9.03 4.83 4.86 5.06
base 384 2 fp16 290.91 14.88 7.09 7.19 7.72
base 384 4 fp16 407.13 18.04 9.93 10.05 10.74
base 384 8 fp16 478.67 26.06 16.92 17.19 17.76
base 128 1 tf32 223.38 6.94 4.73 4.86 5.04
base 128 2 tf32 447.57 7.2 4.68 4.82 5.07
base 128 4 tf32 838.89 9.16 4.88 4.93 5.38
base 128 8 tf32 1201.05 10.81 6.88 6.99 7.21
base 384 1 tf32 206.46 9.74 4.93 4.98 5.25
base 384 2 tf32 287 15.57 7.18 7.27 7.87
base 384 4 tf32 396.59 18.94 10.3 10.41 11.04
base 384 8 tf32 479.04 26.81 16.88 17.25 17.74
base 128 1 fp32 152.92 9.13 6.76 6.91 7.06
base 128 2 fp32 297.42 9.51 6.93 7.07 7.21
base 128 4 fp32 448.57 11.81 9.12 9.25 9.68
base 128 8 fp32 539.94 17.49 15 15.1 15.79
base 384 1 fp32 115.19 13.69 8.89 8.98 9.27
base 384 2 fp32 154.66 18.49 13.06 13.14 13.89
base 384 4 fp32 174.28 28.75 23.11 23.24 24
base 384 8 fp32 191.97 48.05 41.85 42.25 42.8
large 128 1 fp16 127.75 11.18 8.14 8.25 8.53
large 128 2 fp16 219.49 12.76 9.4 9.54 9.89
large 128 4 fp16 315.83 19.01 12.87 12.98 13.37
large 128 8 fp16 495.75 22.21 16.33 16.45 16.79
large 384 1 fp16 96.65 17.46 10.52 10.6 11
large 384 2 fp16 126.07 29.43 16.09 16.22 16.78
large 384 4 fp16 165.21 38.39 24.41 24.61 25.38
large 384 8 fp16 182.13 61.04 44.32 44.61 45.23
large 128 1 tf32 133.24 10.86 7.77 7.87 8.23
large 128 2 tf32 218.13 12.86 9.44 9.56 9.85
large 128 4 tf32 316.25 18.98 12.91 13.01 13.57
large 128 8 tf32 495.21 22.25 16.4 16.51 17.23
large 384 1 tf32 95.43 17.5 10.72 10.83 11.49
large 384 2 tf32 125.99 29.47 16.06 16.15 16.67
large 384 4 tf32 164.28 38.77 24.6 24.83 25.59
large 384 8 tf32 182.46 61 44.2 44.46 45.15
large 128 1 fp32 50.43 23.83 20.11 20.2 20.56
large 128 2 fp32 94.47 25.53 21.36 21.49 21.78
large 128 4 fp32 141.52 32.51 28.44 28.57 28.99
large 128 8 fp32 166.37 52.07 48.3 48.43 49.46
large 384 1 fp32 44.42 30.54 22.67 22.74 23.46
large 384 2 fp32 50.29 48.74 39.95 40.06 40.59
large 384 4 fp32 55.58 81.55 72.31 72.6 73.7
large 384 8 fp32 58.38 147.63 137.43 137.82 138.3

To achieve these same results, follow the Quick Start Guide outlined above.

Inference performance: NVIDIA Tesla T4 (1x T4 16GB)
Fine-tuning inference performance for SQuAD v1.1 on Tesla T4 16GB

Our results were obtained by running the scripts/finetune_inference_benchmark.sh training script in the TensorFlow 20.06-py3 NGC container on NVIDIA Tesla T4 with 1x T4 16GB GPUs. Performance numbers (throughput in sentences per second and latency in milliseconds) were averaged from 1024 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.

Model Sequence Length Batch Size Precision Throughput-Average(sent/sec) Latency-Average(ms) Latency-50%(ms) Latency-90%(ms) Latency-95%(ms) Latency-99%(ms) Latency-100%(ms)
base 128 1 fp16 91.93 13.94 10.93 11.41 11.52 11.94 5491.47
base 128 2 fp16 148.08 16.91 13.65 13.95 14.06 14.74 5757.12
base 128 4 fp16 215.45 24.56 18.68 18.92 19.08 19.84 5894.82
base 128 8 fp16 289.52 33.07 27.77 28.22 28.38 29.16 6074.47
base 384 1 fp16 60.75 23.18 16.6 16.93 17.03 17.45 7006.41
base 384 2 fp16 82.85 37.05 24.26 24.54 24.63 25.67 7529.94
base 384 4 fp16 97.78 54.4 41.02 41.53 41.94 43.91 7995.39
base 384 8 fp16 106.78 89.6 74.98 75.5 76.13 78.02 8461.93
base 128 1 fp32 54.28 20.88 18.52 18.8 18.92 19.29 4401.4
base 128 2 fp32 71.75 30.57 28.08 28.51 28.62 29.12 4573.47
base 128 4 fp32 88.01 50.37 45.61 45.94 46.14 47.04 4992.7
base 128 8 fp32 98.92 85.57 80.98 81.44 81.74 82.75 5408.97
base 384 1 fp32 25.83 43.63 38.75 39.33 39.43 40.02 5148.45
base 384 2 fp32 29.08 77.68 68.89 69.26 69.55 72.08 5462.5
base 384 4 fp32 30.33 141.45 131.86 132.53 133.14 136.7 5975.63
base 384 8 fp32 31.8 262.88 251.62 252.23 253.08 255.56 7124
large 128 1 fp16 40.31 30.61 25.14 25.62 25.87 27.61 10395.87
large 128 2 fp16 63.79 37.43 31.66 32.31 32.66 34.36 10302.2
large 128 4 fp16 87.4 56.5 45.97 46.6 47.01 48.71 10391.17
large 128 8 fp16 107.5 84.29 74.59 75.25 75.64 77.73 10945.1
large 384 1 fp16 23.05 55.73 43.72 44.28 44.74 46.8 12889.23
large 384 2 fp16 29.59 91.61 67.94 68.8 69.45 71.64 13876.35
large 384 4 fp16 34.27 141.56 116.67 118.02 119.1 122.1 14570.73
large 384 8 fp16 38.29 237.85 208.95 210.08 211.33 214.61 16626.02
large 128 1 fp32 21.52 50.46 46.48 47.63 47.94 49.63 7150.38
large 128 2 fp32 25.4 83.3 79.06 79.61 80.06 81.77 7763.11
large 128 4 fp32 28.19 149.49 142.15 143.1 143.65 145.43 7701.38
large 128 8 fp32 30.14 272.84 265.6 266.57 267.21 269.37 8246.3
large 384 1 fp32 8.46 126.81 118.44 119.42 120.31 122.74 9007.96
large 384 2 fp32 9.29 231 215.54 216.64 217.71 220.35 9755.69
large 384 4 fp32 9.55 436.5 418.71 420.05 421.27 424.3 11766.45
large 384 8 fp32 9.75 840.9 820.39 822.19 823.69 827.99 12856.99

To achieve these same results, follow the Quick Start Guide outlined above.

Release notes

Changelog

June 2020

  • Results obtained using 20.06 and on DGX A100 40GB

Janurary 2020

  • Added inference with TensorRT

November 2019

  • Pre-training and Finetuning on BioMedical tasks and corpus

October 2019

  • Disabling Grappler Optimizations for improved performance

September 2019

  • Pre-training using LAMB
  • Multi Node support
  • Fine Tuning support for GLUE (CoLA, MNLI, MRPC)

July 2019

  • Results obtained using 19.06
  • Inference Studies using Triton Inference Server

March 2019

  • Initial release

Known issues

There are no known issues with this model.