How to run Megatron BERT using Transformers

Prerequisites

In that guide, we run all the commands from a folder called $MYDIR and defined as (in bash):

export MYDIR=$HOME

Feel free to change the location at your convenience.

To run some of the commands below, you'll have to clone Transformers.

git clone https://github.com/huggingface/transformers.git $MYDIR/transformers

Get the checkpoint from the NVIDIA GPU Cloud

You must create a directory called nvidia/megatron-bert-cased-345m.

mkdir -p $MYDIR/nvidia/megatron-bert-cased-345m

You can download the checkpoint from the NVIDIA GPU Cloud (NGC). For that you have to 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 checkpoint using:

BERT 345M cased

wget --content-disposition https://api.ngc.nvidia.com/v2/models/nvidia/megatron_bert_345m/versions/v0.1_cased/zip -O $MYDIR/nvidia/megatron-bert-cased-345m/checkpoint.zip

Converting the checkpoint

In order to be loaded into Transformers, the checkpoint have to be converted. You should run the following commands for that purpose. Those commands will create config.json and pytorch_model.bin in $MYDIR/nvidia/megatron-bert-cased-345m. You can move those files to different directories if needed.

BERT 345M cased

python3 $MYDIR/transformers/src/transformers/models/megatron_bert/convert_megatron_bert_checkpoint.py $MYDIR/nvidia/megatron-bert-cased-345m/checkpoint.zip

Masked LM

The following code shows how to use the Megatron BERT checkpoint and the Transformers API to perform a Masked LM task.

import os
import torch

from transformers import BertTokenizer, MegatronBertForMaskedLM

# The tokenizer. Megatron was trained with standard tokenizer(s).
tokenizer = BertTokenizer.from_pretrained('nvidia/megatron-bert-cased-345m')
# The path to the config/checkpoint (see the conversion step above).
directory = os.path.join(os.environ['MYDIR'], 'nvidia/megatron-bert-cased-345m')
# Load the model from $MYDIR/nvidia/megatron-bert-cased-345m.
model = MegatronBertForMaskedLM.from_pretrained(directory)

# Copy to the device and use FP16.
assert torch.cuda.is_available()
device = torch.device("cuda")
model.to(device)
model.eval()
model.half()

# Create inputs (from the BERT example page).
input = tokenizer("The capital of France is [MASK]", return_tensors="pt").to(device)
label = tokenizer("The capital of France is Paris",  return_tensors="pt")["input_ids"].to(device)

# Run the model.
with torch.no_grad():
    output = model(**input, labels=label)
    print(output)

Next sentence prediction

The following code shows how to use the Megatron BERT checkpoint and the Transformers API to perform next sentence prediction.

import os
import torch

from transformers import BertTokenizer, MegatronBertForNextSentencePrediction

# The tokenizer. Megatron was trained with standard tokenizer(s).
tokenizer = BertTokenizer.from_pretrained('nvidia/megatron-bert-cased-345m')
# The path to the config/checkpoint (see the conversion step above).
directory = os.path.join(os.environ['MYDIR'], 'nvidia/megatron-bert-cased-345m')
# Load the model from $MYDIR/nvidia/megatron-bert-cased-345m.
model = MegatronBertForNextSentencePrediction.from_pretrained(directory)

# Copy to the device and use FP16.
assert torch.cuda.is_available()
device = torch.device("cuda")
model.to(device)
model.eval()
model.half()

# Create inputs (from the BERT example page).
input = tokenizer('In Italy, pizza served in formal settings is presented unsliced.',
                  'The sky is blue due to the shorter wavelength of blue light.',
                  return_tensors='pt').to(device)
label = torch.LongTensor([1]).to(device)

# Run the model.
with torch.no_grad():
    output = model(**input, labels=label)
    print(output)

Original code

The original code for Megatron can be found here: https://github.com/NVIDIA/Megatron-LM.