BERT model training impelmentation using 1024 A100 GPUs for MLPerf Training v1.1

Last update: Apr 27, 2022

Related tags

Overview

Pre-trained checkpoint and bert config json file

Location of checkpoint and bert config json file

This MLCommons members Google Drive location contains these files.
- TensorFlow checkpoint (tf1_ckpt) containing the pre-trained weights.
- Config file (bert_config.json) which specifies the hyperparameters of the model.
Checkpoint conversion

python convert_tf_checkpoint.py --tf_checkpoint <path/to/checkpointdir_phase1/model.ckpt-28252.index> --bert_config_path <path/to/checkpointdir_phase1/bert_config.json> --output_checkpoint model.ckpt-28252.pt

Download and preprocess datasets

Download dataset and generate the TFRecords for training data and eval data

BERT Wikipedia dataset preparation

Convert training data and eval data from TFRecords to HDF5

TF_INPUT_DIR=<path/to/tfrecord_input_dir> HDF5_OUTPUT_DIR=<path/to/hdf5_output_dir> ./run_trans_tfrecord_to_hdf5.sh

4bins training data

We split dataset to enable data-load balacning and it can reduce communication overhead.

Based on the sequence length distribution, split HDF5 training data into 4 part:

part 1: 0 < sequence length <= 128

part 2: 128 < sequence length <= 256

part 3: 256 < sequence length <= 384

part 4: 384 < sequence length <= 512

The output_dir contains 4 sub-directories 128, 256, 384 and 512.

cd cleanup_scripts
python run_split_and_chop_hdf5_files.py --input_dir=<path/to/hdf5_datadir> --output_dir=<path/to/4bins_training_datadir>

Prepare the environment

Create a virtualenv and install the required packages:

virtualenv venv -p python3.8.7
source venv/bin/activate
pip install -r requirements.txt

# Install mlperf-logging Python package
git clone https://github.com/mlperf/logging.git mlperf-logging
pip install -e mlperf-logging

# Install apex
git clone https://github.com/NVIDIA/apex.git
cd apex
git reset --hard d06404fecab73f152c6cbb89ac2c2e9b7fc24124
git submodule update --init --recursive
git apply ../patch_for_mlperf_trining_v1.1_by_samsung.patch
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" --global-option="--distributed_adam" --global-option="--distributed_lamb" --global-option="--bnp" --global-option="--xentropy" --global-option="--fast_layer_norm" --global-option="--deprecated_fused_adam"  --global-option="--fmha"  --global-option="--fast_multihead_attn" ./

# Compile mhalib
cd mhalib
python setup.py build
cp build/lib*/mhalib* ../

Other software requirements

Softeware	Version
python	3.8.7
pytorch	1.9.1
NCCL	2.9.9
CUDA	11.3.0
cudnn	8.2.1.32
cublas	11.4.2
nvidia driver	470.57.02
mofed version	5.4-1.0.3

Run the model

Set hosts address in run_multinode.sh

export hosts=('192.168.16.1' '192.168.16.2')

Launch the training

Use the following command to run the config_Samsung_Supercomputer21_DGXA100_128x8x16x1.sh in python virtual environment.

PYTHON=<path/to/python> DGXSYSTEM=Samsung_Supercomputer21_DGXA100_128x8x16x1 INPUT_DIR=<path/to/4bins_training_datadir> EVAL_DIR=<path/to/eval_datadir> CHECKPOINTDIR_PHASE1=<path/to/checkpointdir_phase1> NEXP=10 ./run_multinode.sh

Appendix

Our source code is based on MLPerf BERT v0.7, and all the files newly added and modified are as follows.

File Name	Status	Description
config_Samsung_Supercomputer21_DGXA100_128x8x16x1.sh	Newly added	The file contains configurations used for 1024 GPUs experiment.
run_split_and_chop_hdf5_files.py	Newly added	The file is used for generating 4-bin training data.
mhalib/setup.py	Modified	The file is modified since CUDA upgraded.
optim/init.py	Newly added	The file is used as the entrance of "optim" module.
optim/acclip.py	Newly added	The file implements ACClip optimizer for trial.
optim/madgrad.py	Newly added	The file implements MADGRAD optimizer for trial.
bind_launch.py	Newly added	The file is added for BERT training on python environment.
bind_pyt.py	Modified	The file is modified for the following items. (1) Log compliance; (2) Add new NUMA binding.
fmha.py	Newly added	The file is used for adding FMHA operator (refer to MLPerf v1.0).
mlperf_logger.py	Modified	The file is modified for log compliance.
modeling.py	Modified	The file is modified for adding FMHA (refer to MLPerf v1.0).
padding.py	Modified	The file is modified for adding FMHA (refer to MLPerf v1.0).
README.md	Modified	It is modified to run Samsung optimized implematation.
requirements.txt	Modified	The file shows required software version.
run_multinode.sh	Newly added	The file is startup script about how to run BERT training on python environment
run_pretraining.py	Modified	The file is modified for the following items. (1) Load splitting training data; (2) Add exchange padding feature (refer to MLPerf v1.0); (3) Add NCCL warmup (refer to MLPerf v1.0); (4) Add SAIT local/group exchange padding; (5) Add NCCL warmup for group exchange padding; (6) Add per-device local gradient clipping before all-reduce; (7) Add pytorch DDP.
schedulers.py	Modified	The file is modified for optimizing learning rate scheduler
utils.py	Modified	The file is modified for the following items. (1) Add get_optimzer() interface; (2) Add a batch sampler (SplitRandomSampler) for 4-bin splitting training data.

BERT model training impelmentation using 1024 A100 GPUs for MLPerf Training v1.1

Related tags

Overview

Pre-trained checkpoint and bert config json file

Download and preprocess datasets

Prepare the environment

Run the model

Appendix

Owner

SAIT (Samsung Advanced Institute of Technology)

Paper Title: Heterogeneous Knowledge Distillation for Simultaneous Infrared-Visible Image Fusion and Super-Resolution

TFOD-MASKRCNN - Tensorflow MaskRCNN With Python

A high-level Python library for Quantum Natural Language Processing

Official PyTorch implementation of the paper "Deep Constrained Least Squares for Blind Image Super-Resolution", CVPR 2022.

Online Pseudo Label Generation by Hierarchical Cluster Dynamics for Adaptive Person Re-identification

BigDetection: A Large-scale Benchmark for Improved Object Detector Pre-training

Multi-modal Text Recognition Networks: Interactive Enhancements between Visual and Semantic Features

The project is associated with the recently-launched ICASSP 2022 Multi-channel Multi-party Meeting Transcription Challenge (M2MeT) to provide participants with baseline systems for speech recognition and speaker diarization in conference scenario.

Code used to generate the results appearing in "Train longer, generalize better: closing the generalization gap in large batch training of neural networks"

(CVPR2021) Kaleido-BERT: Vision-Language Pre-training on Fashion Domain

System Design course at HSE (2021)

Asymmetric Bilateral Motion Estimation for Video Frame Interpolation, ICCV2021

This project provides an unsupervised framework for mining and tagging quality phrases on text corpora with pretrained language models (KDD'21).

RATE: Overcoming Noise and Sparsity of Textual Features in Real-Time Location Estimation (CIKM'17)

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.

Unofficial implementation of Pix2SEQ

Open CV - Convert a picture to look like a cartoon sketch in python

yolov5目标检测模型的知识蒸馏（基于响应的蒸馏）

Stacked Hourglass Network with a Multi-level Attention Mechanism: Where to Look for Intervertebral Disc Labeling

Evaluation suite for large-scale language models.