Transformer-LT MLPerf BFloat16 Training TensorFlow* Model

Published: 12/09/2020  

Last Updated: 06/15/2022

Download Command



This document has instructions to run a Transformer Language BFloat16 training using Intel® Optimization for TensorFlow*. Detailed information on MLPerf* benchmark can be found in mlperf/training.


Decide the problem you want to run to get the appropriate dataset. We will get the training data of it as an example:

Download dataset for computing BLEU score.

export DATASET_DIR=/home/<user>/transformer_data

For the training dataset, download and untar the model package.

tar -xzf transformer-mlperf-bfloat16-training.tar.gz

export PYTHONPATH=$PYTHONPATH:/home/<user>/transformer-mlperf-bfloat16-training/models/common/tensorflow
export DATASET_DIR=/home/<user>/transformer_data
cd transformer-mlperf-bfloat16-training/models/language_translation/tensorflow/transformer_mlperf/training/bfloat16/transformer
python --data_dir=$DATASET_DIR

Running python --data_dir=$DATASET_DIR assumes you have a Python* environment similar to what the intel/intel-optimized-tensorflow:2.4.0-ubuntu-18.04 container provides. One option would be to run the above within the intel/intel-optimized-tensorflow:2.4.0-ubuntu-18.04 container eg: docker run -u $(id -u):$(id -g) --privileged --entrypoint /bin/bash -v /home/:/home/ -it intel/intel-optimized-tensorflow:2.4.0-ubuntu-18.04

Quick Start Scripts

Transformer Language in MLPerf benchmark can run with full training or fewer training steps. During training we can control if it will do the evaluation or not.

Script name Description
bfloat16_training_demo Runs 100 training steps. The script runs in single-instance mode by default, for multi-instance mode set MPI_NUM_PROCESSES.
bfloat16_training Runs 200 training steps, saves checkpoints and does evaluation. The script runs in single-instance mode by default, for multi-instance mode set MPI_NUM_PROCESSES.

Bare Metal

To run on bare metal, the following prerequisites must be installed in your environment:

After installing the prerequisites, download and untar the model package. Set environment variables for the path to your DATASET_DIR and an OUTPUT_DIR where log files will be written, then run a quickstart script.

DATASET_DIR=<path to the dataset>
OUTPUT_DIR=<directory where log files will be written>

tar -xzf transformer-mlperf-bfloat16-training.tar.gz
cd transformer-mlperf-bfloat16-training

quickstart/<script name>.sh


Documentation and Sources

Get Started
Main GitHub*
Release Notes
Get Started Guide

Code Sources
Report Issue


License Agreement

LEGAL NOTICE: By accessing, downloading or using this software and any required dependent software (the “Software Package”), you agree to the terms and conditions of the software license agreements for the Software Package, which may also include notices, disclaimers, or license terms for third party software included with the Software Package. Please refer to the license file for additional details.

Related Containers and Solutions

Transformer-LT MLPerf BFloat16 Training TensorFlow* Container

View All Containers and Solutions 🡢

Product and Performance Information


Performance varies by use, configuration and other factors. Learn more at