Pre-trained checkpoint and bert config json file

  1. 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.

  2. 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

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

    BERT Wikipedia dataset preparation

  2. 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

  3. 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

  1. Set hosts address in run_multinode.sh

export hosts=('192.168.16.1' '192.168.16.2')

  1. 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.

GitHub

View Github