GNMT v2 For TensorFlow

This repository provides a script and recipe to train the GNMT v2 model to achieve state-of-the-art accuracy and is tested and maintained by NVIDIA. GNMT model for TensorFlow1 is no longer maintained and will soon become unavailable, please consider PyTorch or TensorFlow2 models as a substitute for your requirements.

Table Of Contents

Model overview
Setup
- Requirements
Quick Start Guide
Advanced
Performance
- Benchmarking
  - Training performance benchmark
  - Inference performance benchmark
- Results
Release notes
- Changelog
- Known issues

Model overview

The GNMT v2 model is similar to the one discussed in the Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation paper.

The most important difference between the two models is in the attention mechanism. In our model, the output from the first LSTM layer of the decoder goes into the attention module, then the re-weighted context is concatenated with inputs to all subsequent LSTM layers in the decoder at the current timestep.

The same attention mechanism is also implemented in the default GNMT-like models from TensorFlow Neural Machine Translation Tutorial and NVIDIA OpenSeq2Seq Toolkit.

This model is trained with mixed precision using Tensor Cores on Volta, Turing, and the NVIDIA Ampere GPU architectures. Therefore, researchers can get results 2x 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

The following image shows the GNMT model architecture:

Default configuration

The following features were implemented in this model:

general:
- encoder and decoder are using shared embeddings
- data-parallel multi-GPU training
- dynamic loss scaling with backoff for Tensor Cores (mixed precision) training
- trained with label smoothing loss (smoothing factor 0.1)
encoder:
- 4-layer LSTM, hidden size 1024, first layer is bidirectional, the rest are unidirectional
- with residual connections starting from 3rd layer
- dropout is applied on input to all LSTM layers, probability of dropout is set to 0.2
- hidden state of LSTM layers is initialized with zeros
- weights and bias of LSTM layers is initialized with uniform (-0.1, 0.1) Distribution
decoder:
- 4-layer unidirectional LSTM with hidden size 1024 and fully-connected classifier
- with residual connections starting from 3rd layer
- dropout is applied on input to all LSTM layers, probability of dropout is set to 0.2
- hidden state of LSTM layers is initialized with the last hidden state from encoder
- weights and bias of LSTM layers is initialized with uniform (-0.1, 0.1) distribution
- weights and bias of fully-connected classifier is initialized with uniform (-0.1, 0.1) distribution
attention:
- normalized Bahdanau attention
- output from first LSTM layer of decoder goes into attention, then re-weighted context is concatenated with the input to all subsequent LSTM layers of the decoder at the current timestep
- linear transform of keys and queries is initialized with uniform (-0.1, 0.1), normalization scalar is initialized with 1.0 / sqrt(1024), normalization bias is initialized with zero
inference:
- beam search with default beam size of 5
- with coverage penalty and length normalization, coverage penalty factor is set to 0.1, length normalization factor is set to 0.6 and length normalization constant is set to 5.0
- de-tokenized BLEU computed by SacreBLEU
- motivation for choosing SacreBLEU

When comparing the BLEU score, there are various tokenization approaches and BLEU calculation methodologies; therefore, ensure you align similar metrics.

Code from this repository can be used to train a larger, 8-layer GNMT v2 model. Our experiments show that a 4-layer model is significantly faster to train and yields comparable accuracy on the public WMT16 English-German dataset. The number of LSTM layers is controlled by the --num_layers parameter in the nmt.py script.

Feature support matrix

The following features are supported by this model.

Feature	GNMT TF
Automatic Mixed Precision	yes

Features

The following features are supported by this model.

Automatic Mixed Precision (AMP) - Computation graphs can be modified by TensorFlow on runtime to support mixed precision training. Detailed explanation of mixed precision can be found in the next section.

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:

Porting the model to use the FP16 data type where appropriate.
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:

How to train using mixed precision, see the Mixed Precision Training paper and Training With Mixed Precision documentation.
Techniques used for mixed precision training, see the Mixed-Precision Training of Deep Neural Networks blog.
How to access and enable AMP for TensorFlow, see Using TF-AMP from the TensorFlow User Guide.
APEX tools for mixed precision training, see the NVIDIA Apex: Tools for Easy Mixed-Precision Training in PyTorch.

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.

Setup

The following section lists the requirements that you need to meet in order to start training the GNMT v2 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.

Quick Start Guide

To train your model using mixed or TF32 precision with Tensor Cores or using FP32, perform the following steps using the default parameters of the GNMT v2 model on the WMT16 English German dataset. For the specifics concerning training and inference, see the Advanced section.

1. Clone the repository.

git clone https://github.com/NVIDIA/DeepLearningExamples
cd DeepLearningExamples/TensorFlow/Translation/GNMT

2. Build the GNMT v2 TensorFlow container.

bash scripts/docker/build.sh

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

bash scripts/docker/interactive.sh

4. Download and preprocess the dataset.

Data will be downloaded to the data directory (on the host). The data directory is mounted to the /workspace/gnmt/data location in the Docker container.

bash scripts/wmt16_en_de.sh

5. Start training.

All results and logs are saved to the results directory (on the host) or to the /workspace/gnmt/results directory (in the container). The training script saves the checkpoint after every training epoch and after every 2000 training steps within each epoch. You can modify the results directory using the --output_dir argument.

To launch mixed precision training on 1 GPU, run:

python nmt.py --output_dir=results --batch_size=128 --learning_rate=5e-4 --amp

To launch mixed precision training on 8 GPUs, run:

python nmt.py --output_dir=results --batch_size=1024 --num_gpus=8 --learning_rate=2e-3 --amp

To launch FP32 (TF32 on NVIDIA Ampere GPUs) training on 1 GPU, run:

python nmt.py --output_dir=results --batch_size=128 --learning_rate=5e-4

To launch FP32 (TF32 on NVIDIA Ampere GPUs) training on 8 GPUs, run:

python nmt.py --output_dir=results --batch_size=1024 --num_gpus=8 --learning_rate=2e-3

6. Start evaluation.

The training process automatically runs evaluation and outputs the BLEU score after each training epoch. Additionally, after the training is done, you can manually run inference on test dataset with the checkpoint saved during the training.

To launch mixed precision inference on 1 GPU, run:

python nmt.py --output_dir=results --infer_batch_size=128 --mode=infer --amp

To launch FP32 (TF32 on NVIDIA Ampere GPUs) inference on 1 GPU, run:

python nmt.py --output_dir=results --infer_batch_size=128 --mode=infer

7. Start translation.

After the training is done, you can translate custom sentences with the checkpoint saved during the training.

echo "The quick brown fox jumps over the lazy dog" >file.txt
python nmt.py --output_dir=results --mode=translate --translate-file=file.txt
cat file.txt.trans

Der schnelle braune Fuchs springt über den faulen Hund

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:

nmt.py: serves as the entry point to launch the training
Dockerfile: container with the basic set of dependencies to run GNMT v2
requirements.txt: set of extra requirements for running GNMT v2
attention_wrapper.py, gnmt_model.py, model.py: model definition
estimator.py: functions for training and inference

In the script directory, the most important files are:

translate.py: wrapped on nmt.py for benchmarking and running inference
parse_log.py: script for retrieving information in JSON format from the training log
wmt16_en_de.sh: script for downloading and preprocessing the dataset

In the script/docker directory, the files are:

build.sh: script for building the GNMT container
interactive.sh: script for running the GNMT container interactively

Parameters

The most useful arguments are as follows:

  --learning_rate LEARNING_RATE
                        Learning rate.
  --warmup_steps WARMUP_STEPS
                        How many steps we inverse-decay learning.
  --max_train_epochs MAX_TRAIN_EPOCHS
                        Max number of epochs.
  --target_bleu TARGET_BLEU
                        Target bleu.
  --data_dir DATA_DIR   Training/eval data directory.
  --translate_file TRANSLATE_FILE
                        File to translate, works only with translate mode
  --output_dir OUTPUT_DIR
                        Store log/model files.
  --batch_size BATCH_SIZE
                        Total batch size.
  --log_step_count_steps LOG_STEP_COUNT_STEPS
                        The frequency, in number of global steps, that the
                        global step and the loss will be logged during training
  --num_gpus NUM_GPUS   Number of gpus in each worker.
  --random_seed RANDOM_SEED
                        Random seed (>0, set a specific seed).
  --ckpt CKPT           Checkpoint file to load a model for inference.
                        (defaults to newest checkpoint)
  --infer_batch_size INFER_BATCH_SIZE
                        Batch size for inference mode.
  --beam_width BEAM_WIDTH
                        beam width when using beam search decoder. If 0, use
                        standard decoder with greedy helper.
  --amp                 use amp for training and inference
  --mode {train_and_eval,infer,translate}

Command-line options

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

python nmt.py --help

Getting the data

The GNMT v2 model was trained on the WMT16 English-German dataset and newstest2014 is used as a testing dataset.

This repository contains the scripts/wmt16_en_de.sh download script which automatically downloads and preprocesses the training and test datasets. By default, data is downloaded to the data directory.

Our download script is very similar to the wmt16_en_de.sh script from the tensorflow/nmt repository. Our download script contains an extra preprocessing step, which discards all pairs of sentences which can't be decoded by latin-1 encoder. The scripts/wmt16_en_de.sh script uses the subword-nmt package to segment text into subword units (Byte Pair Encodings - BPE). By default, the script builds the shared vocabulary of 32,000 tokens.

In order to test with other datasets, the scripts need to be customized accordingly.

Dataset guidelines

The process of downloading and preprocessing the data can be found in the scripts/wmt16_en_de.sh script.

Initially, data is downloaded from www.statmt.org. Then, europarl-v7, commoncrawl and news-commentary corpora are concatenated to form the training dataset, similarly newstest2015 and newstest2016 are concatenated to form the validation dataset. Raw data is preprocessed with Moses, first by launching Moses tokenizer (tokenizer breaks up text into individual words), then by launching clean-corpus-n.perl which removes invalid sentences and does initial filtering by sequence length.

Second stage of preprocessing is done by launching the scripts/filter_dataset.py script, which discards all pairs of sentences that can't be decoded by latin-1 encoder.

Third state of preprocessing uses the subword-nmt package. First it builds shared byte pair encoding vocabulary with 32,000 merge operations (command subword-nmt learn-bpe), then it applies generated vocabulary to training, validation and test corpora (command subword-nmt apply-bpe).

Training process

The training configuration can be launched by running the nmt.py script. By default, the training script saves the checkpoint after every training epoch and after every 2000 training steps within each epoch. Results are stored in the results directory.

The training script launches data-parallel training on multiple GPUs. We have tested reliance on up to 8 GPUs on a single node.

After each training epoch, the script runs an evaluation and outputs a BLEU score on the test dataset (newstest2014). BLEU is computed by the SacreBLEU package. Logs from the training and evaluation are saved to the results directory.

The training script automatically runs testing after each training epoch. The results from the testing are printed to the standard output and saved to the log files.

The summary after each training epoch is printed in the following format:

training time for epoch 1: 29.37 mins (2918.36 sent/sec, 139640.48 tokens/sec)
[...]
bleu is 20.50000
eval time for epoch 1: 1.57 mins (78.48 sent/sec, 4283.88 tokens/sec)

The BLEU score is computed on the test dataset. Performance is reported in total sentences per second and in total tokens per second. The performance result is averaged over an entire training epoch and summed over all GPUs participating in the training.

To view all available options for training, run python nmt.py --help.

Inference process

Validation and translation can be run by launching the nmt.py script, although, it requires a pre-trained model checkpoint and tokenized input (for validation) and non-tokenized input (for translation).

Validation process

The nmt.py script supports batched validation (--mode=infer flag). By default, it launches beam search with beam size of 5, coverage penalty term and length normalization term. Greedy decoding can be enabled by setting the --beam_width=1 flag for the nmt.py inference script. To control the batch size use the --infer_batch_size flag.

To view all available options for validation, run python nmt.py --help.

Translation process

The nmt.py script supports batched translation (--mode=translate flag). By default, it launches beam search with beam size of 5, coverage penalty term and length normalization term. Greedy decoding can be enabled by setting the --beam_width=1 flag for the nmt.py prediction script. To control the batch size use the --infer_batch_size flag.

The input file may contain many sentences, each on a new line. The file can be specified by the --translate_file <file> flag. This script will create a new file called <file>.trans, with translation of the input file.

To view all available options for translation, run python nmt.py --help.

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.

Training performance benchmark

To benchmark training performance, run:

python nmt.py --output_dir=results --max_train_epochs=1 --num_gpus <num GPUs> --batch_size <total batch size> --amp for mixed precision
python nmt.py --output_dir=results --max_train_epochs=1 --num_gpus <num GPUs> --batch_size <total batch size> for FP32/TF32

The log file will contain training performance in the following format:

training time for epoch 1: 25.75 mins (3625.19 sent/sec, 173461.27 tokens/sec)

Inference performance benchmark

To benchmark inference performance, run the scripts/translate.py script:

For FP32/TF32: python scripts/translate.py --output_dir=/path/to/trained/model --beam_width <comma separated beam widths> --infer_batch_size <comma separated batch sizes>
For mixed precision python scripts/translate.py --output_dir=/path/to/trained/model --amp --beam_width <comma separated beam widths> --infer_batch_size <comma separated batch sizes>

The benchmark requires a checkpoint from a fully trained model.

Results

The following sections provide details on how we achieved our performance and accuracy in training and inference.

Training accuracy results

Training accuracy: NVIDIA DGX A100 (8x A100 40GB)

Our results were obtained by running the examples/DGXA100_{TF32,AMP}_8GPU.sh training script in the tensorflow-20.06-tf1-py3 NGC container on NVIDIA DGX A100 (8x A100 40GB) GPUs.

GPUs	Batch size / GPU	Accuracy - mixed precision (BLEU)	Accuracy - TF32 (BLEU)	Time to train - mixed precision	Time to train - TF32	Time to train speedup (TF32 to mixed precision)
8	128	25.1	24.31	96 min	139 min	1.45

Training accuracy: NVIDIA DGX-1 (8x V100 16GB)

Our results were obtained by running the nmt.py script in the tensorflow-19.07-py3 NGC container on NVIDIA DGX-1 with (8x V100 16GB) GPUs.

GPUs	Batch size / GPU	Accuracy - mixed precision (BLEU)	Accuracy - FP32 (BLEU)	Time to train - mixed precision	Time to train - FP32	Time to train speedup (FP32 to mixed precision)
1	128	24.90	24.84	763 min	1237 min	1.62
8	128	24.33	24.34	168 min	237 min	1.41

In the following plot, the BLEU scores after each training epoch for different configurations are displayed.

Training stability test

The GNMT v2 model was trained for 6 epochs, starting from 6 different initial random seeds. After each training epoch, the model was evaluated on the test dataset and the BLEU score was recorded. The training was performed in the tensorflow-20.06-tf1-py3 NGC container.

In the following tables, the BLEU scores after each training epoch for different initial random seeds are displayed.

NVIDIA DGX A100 with 8 Ampere A100 40GB GPUs with TF32.

Epoch	Average	Standard deviation	Minimum	Median	Maximum
1	20.272	0.165	19.760	20.295	20.480
2	21.911	0.145	21.650	21.910	22.230
3	22.731	0.140	22.490	22.725	23.020
4	23.142	0.164	22.930	23.090	23.440
5	23.967	0.137	23.760	23.940	24.260
6	24.358	0.143	24.120	24.360	24.610

NVIDIA DGX-1 with 8 Tesla V100 16GB GPUs with FP32.

Epoch	Average	Standard deviation	Minimum	Median	Maximum
1	20.259	0.225	19.820	20.300	20.590
2	21.954	0.194	21.540	21.955	22.370
3	22.729	0.150	22.480	22.695	23.110
4	23.218	0.210	22.820	23.225	23.470
5	23.921	0.114	23.680	23.910	24.080
6	24.381	0.131	24.160	24.375	24.590

Inference accuracy results

Inference accuracy: NVIDIA DGX-1 (8x V100 16GB)

Our results were obtained by running the scripts/translate.py script in the tensorflow-19.07-py3 NGC container on NVIDIA DGX-1 8x V100 16GB GPUs.

For mixed precision: python scripts/translate.py --output_dir=/path/to/trained/model --beam_width 1,2,5 --infer_batch_size 128 --amp
For FP32: python scripts/translate.py --output_dir=/path/to/trained/model --beam_width 1,2,5 --infer_batch_size 128

Batch size	Beam size	Mixed precision BLEU	FP32 BLEU
128	1	23.80	23.80
128	2	24.58	24.59
128	5	25.10	25.09

Training performance results

Training performance: NVIDIA DGX A100 (8x A100 40GB)

Our results were obtained by running the examples/DGXA100_{TF32,AMP}_{1,8}GPU.sh training script in the tensorflow-20.06-tf1-py3 NGC container on NVIDIA DGX A100 (8x A100 40GB) GPUs. Performance numbers (in items/images per second) were averaged over an entire training epoch.

GPUs	Batch size / GPU	Throughput - mixed precision (tokens/s)	Throughput - TF32 (tokens/s)	Throughput speedup (TF32 - mixed precision)	Weak scaling - mixed precision	Weak scaling - TF32
1	128	29 911	31 110	0.96	1.00	1.00
8	128	181 384	175 292	1.03	6.06	5.63

To achieve these same results, follow the steps in the Quick Start Guide.

Training performance: NVIDIA DGX-1 (8x V100 16GB)

Our results were obtained by running the nmt.py script in the tensorflow-19.07-py3 NGC container on NVIDIA DGX-1 with 8x V100 16G GPUs. Performance numbers (in tokens per second) were averaged over an entire training epoch.

GPUs	Batch size / GPU	Throughput - mixed precision (tokens/s)	Throughput - FP32 (tokens/s)	Throughput speedup (FP32 - mixed precision)	Weak scaling - mixed precision	Weak scaling - FP32
1	128	23 011	14 106	1.63	1.00	1.00
8	128	138 106	93 688	1.47	6.00	6.64

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

Inference performance results

The benchmark requires a checkpoint from a fully trained model.

To launch the inference benchmark in mixed precision on 1 GPU, run:

python scripts/translate.py --output_dir=/path/to/trained/model --beam_width 1,2,5 --infer_batch_size 1,2,4,8,32,128,512 --amp

To launch the inference benchmark in FP32/TF32 on 1 GPU, run:

python scripts/translate.py --output_dir=/path/to/trained/model --beam_width 1,2,5 --infer_batch_size 1,2,4,8,32,128,512

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

Inference performance: NVIDIA DGX A100 (1x A100 40GB)

Our results were obtained by running the python scripts/translate.py --infer_batch_size 1,2,4,8,32,128,512 --beam_width 1,2,5 {--amp} inferencing benchmarking script in the tensorflow-20.06-tf1-py3 NGC container on NVIDIA DGX A100 (1x A100 40GB) GPU.

FP16

Batch size	Beam width	Bleu	Sentences/sec	Tokens/sec	Latency Avg	Latency 50%	Latency 90%	Latency 95%	Latency 99%	Latency 100%
1	1	23.80	13.67	737.89	73.15	67.69	121.98	137.20	162.74	201.06
1	2	24.58	13.40	721.18	74.65	69.12	123.99	138.82	169.58	198.49
1	5	25.10	12.12	647.78	82.53	76.53	136.35	152.59	196.09	216.55
2	1	23.80	21.55	1163.16	92.82	88.15	139.88	152.49	185.18	208.35
2	2	24.58	21.07	1134.42	94.91	89.62	142.08	158.12	188.00	205.08
2	5	25.10	19.59	1047.21	102.10	96.20	152.36	172.46	211.96	219.87
4	1	23.80	36.98	1996.27	108.16	105.07	150.42	161.56	200.99	205.87
4	2	24.57	34.92	1880.48	114.53	111.42	160.29	177.14	205.32	211.80
4	5	25.10	31.56	1687.34	126.74	122.06	179.68	201.38	225.08	229.14
8	1	23.80	64.52	3482.81	123.99	122.89	159.89	174.66	201.12	205.59
8	2	24.57	59.04	3178.17	135.50	135.23	180.50	191.66	214.95	216.84
8	5	25.09	55.51	2967.82	144.11	141.98	198.39	218.88	223.55	225.61
32	1	23.80	193.54	10447.04	165.34	163.56	211.67	215.37	221.07	221.14
32	2	24.57	182.00	9798.09	175.82	176.04	220.33	224.25	226.45	227.05
32	5	25.10	141.63	7572.02	225.94	225.59	278.38	279.56	281.61	282.13
128	1	23.80	556.57	30042.59	229.98	226.81	259.05	260.26	260.74	260.85
128	2	24.57	400.02	21535.38	319.98	328.23	351.31	352.82	353.01	353.06
128	5	25.10	235.14	12570.95	544.35	576.62	581.95	582.64	583.61	583.85
512	1	23.80	903.83	48786.58	566.48	570.44	579.74	580.66	581.39	581.57
512	2	24.58	588.63	31689.07	869.81	894.90	902.65	902.85	903.00	903.04
512	5	25.10	285.86	15283.40	1791.06	1835.19	1844.29	1845.59	1846.63	1846.89

TF32

Batch size	Beam width	Bleu	Sentences/sec	Tokens/sec	Latency Avg	Latency 50%	Latency 90%	Latency 95%	Latency 99%	Latency 100%
1	1	23.82	13.25	715.47	75.45	69.81	125.63	141.89	169.70	209.78
1	2	24.59	13.21	711.16	75.72	70.06	124.75	140.20	173.23	201.39
1	5	25.08	12.38	661.99	80.76	74.90	131.93	148.91	187.05	208.39
2	1	23.82	21.61	1166.56	92.55	87.25	139.54	151.77	180.24	209.05
2	2	24.59	21.24	1143.63	94.17	88.78	139.70	156.61	189.09	205.06
2	5	25.10	19.49	1042.17	102.62	96.14	153.38	172.89	213.99	219.54
4	1	23.81	35.84	1934.49	111.62	108.73	154.52	165.42	207.88	211.29
4	2	24.58	34.71	1869.20	115.24	111.24	161.24	177.73	208.12	212.74
4	5	25.09	32.24	1723.86	124.07	119.35	177.54	196.69	221.10	223.52
8	1	23.80	64.08	3459.74	124.84	123.61	161.92	177.06	205.47	206.47
8	2	24.61	59.31	3193.52	134.89	133.44	182.92	192.71	216.04	218.78
8	5	25.10	56.60	3026.29	141.35	138.61	194.52	213.65	220.24	221.45
32	1	23.80	195.31	10544.22	163.85	162.80	212.71	215.41	216.92	217.34
32	2	24.61	185.66	9996.59	172.36	171.07	216.46	221.64	223.68	225.25
32	5	25.11	147.24	7872.61	217.34	214.97	269.75	270.71	271.44	272.87
128	1	23.81	576.54	31123.19	222.02	219.25	249.44	249.75	249.88	249.91
128	2	24.57	419.87	22609.82	304.86	314.47	332.18	334.13	336.22	336.74
128	5	25.10	245.76	13138.84	520.83	552.68	558.89	559.09	559.13	559.13
512	1	23.80	966.24	52156.34	529.89	534.82	558.30	559.33	560.16	560.36
512	2	24.58	642.41	34590.81	797.00	812.40	824.23	825.92	827.27	827.61
512	5	25.10	289.33	15468.09	1769.61	1817.19	1849.83	1855.17	1859.45	1860.51

To achieve these same results, follow the steps in the Quick Start Guide.

Inference performance: NVIDIA DGX-1 (1x V100 16GB)

Our results were obtained by running the python scripts/translate.py --infer_batch_size 1,2,4,8,32,128,512 --beam_width 1,2,5 {--amp} inferencing benchmarking script in the tensorflow-20.06-tf1-py3 NGC container on NVIDIA DGX-1 with (1x V100 16GB) GPU.

FP16

Batch size	Sequence length	Throughput Avg	Latency Avg	Latency 90%	Latency 95%	Latency 99%
1	1	23.78	9.06	489.00	110.41	102.80
1	2	24.58	8.68	467.35	115.22	107.17
1	5	25.09	8.39	448.32	119.25	109.79
2	1	23.82	14.59	787.70	137.04	129.38
2	2	24.57	14.44	777.60	138.51	131.07
2	5	25.11	13.78	736.99	145.11	136.76
4	1	23.82	23.79	1284.24	168.14	164.13
4	2	24.59	22.67	1220.66	176.45	171.40
4	5	25.08	22.33	1194.00	179.12	174.04
8	1	23.81	43.33	2338.68	184.63	183.25
8	2	24.60	39.12	2106.44	204.49	200.96
8	5	25.10	37.16	1987.05	215.26	210.92
32	1	23.82	129.52	6992.15	247.06	245.81
32	2	24.55	123.28	6637.86	259.57	261.07
32	5	25.05	88.74	4744.33	360.61	359.27
128	1	23.80	332.81	17964.83	384.60	382.14
128	2	24.59	262.87	14153.59	486.93	506.45
128	5	25.08	143.91	7695.36	889.42	932.93
512	1	23.80	613.57	33126.42	834.46	848.06
512	2	24.59	387.72	20879.62	1320.54	1343.05
512	5	25.10	199.48	10664.34	2566.67	2628.50

FP32

Batch size	Sequence length	Throughput Avg	Latency Avg	Latency 90%	Latency 95%	Latency 99%
1	1	23.80	8.37	451.86	119.46	111.26
1	2	24.59	8.83	475.11	113.31	104.54
1	5	25.09	7.74	413.92	129.15	119.44
2	1	23.80	13.96	753.79	143.22	135.73
2	2	24.59	12.96	697.63	154.33	145.01
2	5	25.09	12.67	677.23	157.88	148.24
4	1	23.80	22.42	1209.97	178.44	172.70
4	2	24.59	20.55	1106.07	194.68	188.83
4	5	25.09	21.19	1132.58	188.81	182.77
8	1	23.80	39.32	2122.26	203.48	201.89
8	2	24.59	37.51	2019.43	213.26	211.55
8	5	25.09	31.69	1694.02	252.46	245.33
32	1	23.80	118.51	6396.93	270.02	269.22
32	2	24.59	100.23	5395.33	319.28	318.89
32	5	25.09	68.59	3666.77	466.55	466.84
128	1	23.80	256.49	13845.09	499.04	492.36
128	2	24.59	176.83	9519.12	723.86	754.89
128	5	25.09	96.21	5143.17	1330.48	1420.94
512	1	23.80	366.07	19759.97	1398.63	1421.81
512	2	24.59	225.48	12137.77	2270.75	2323.62
512	5	25.09	106.02	5667.78	4829.31	4946.65

To achieve these same results, follow the steps in the Quick Start Guide.

Inference performance: NVIDIA T4

Our results were obtained by running the scripts/translate.py script in the tensorflow-19.07-py3 NGC container on NVIDIA T4.

Reported mixed precision speedups are relative to FP32 numbers for corresponding configuration.

Batch size	Beam size	Mixed precision tokens/s	Speedup	Mixed precision average latency (ms)	Average latency speedup	Mixed precision latency 50% (ms)	Latency 50% speedup	Mixed precision latency 90% (ms)	Latency 90% speedup	Mixed precision latency 95% (ms)	Latency 95% speedup	Mixed precision latency 99% (ms)	Latency 99% speedup	Mixed precision latency 100% (ms)	Latency 100% speedup
1	1	643	1.278	84	1.278	78	1.279	138	1.309	154	1.312	180	1.304	220	1.296
1	2	584	1.693	92	1.692	86	1.686	150	1.743	168	1.737	201	1.770	236	1.742
1	5	552	1.702	97	1.701	90	1.696	158	1.746	176	1.738	218	1.769	244	1.742
2	1	948	1.776	114	1.776	108	1.769	170	1.803	184	1.807	218	1.783	241	1.794
2	2	912	1.761	118	1.760	112	1.763	175	1.776	192	1.781	226	1.770	246	1.776
2	5	832	1.900	128	1.900	121	1.910	192	1.912	214	1.922	258	1.922	266	1.905
4	1	1596	1.792	135	1.792	132	1.791	187	1.799	197	1.815	241	1.784	245	1.796
4	2	1495	1.928	144	1.927	141	1.926	201	1.927	216	1.936	250	1.956	264	1.890
4	5	1308	1.702	164	1.702	159	1.702	230	1.722	251	1.742	283	1.708	288	1.699
8	1	2720	1.981	159	1.981	158	1.992	204	1.975	219	1.986	249	1.987	252	1.966
8	2	2554	1.809	169	1.808	168	1.829	224	1.797	237	1.783	260	1.807	262	1.802
8	5	1979	1.768	216	1.768	213	1.780	292	1.797	319	1.793	334	1.760	336	1.769
32	1	7449	1.775	232	1.774	231	1.777	292	1.789	300	1.760	301	1.768	301	1.768
32	2	5569	1.670	309	1.669	311	1.672	389	1.652	392	1.665	401	1.651	404	1.644
32	5	3079	1.867	556	1.867	555	1.865	692	1.858	695	1.860	702	1.847	703	1.847
128	1	12986	1.662	532	1.662	529	1.667	607	1.643	608	1.645	609	1.647	609	1.647
128	2	7856	1.734	878	1.734	911	1.755	966	1.742	967	1.741	968	1.744	968	1.744
128	5	3361	1.683	2036	1.682	2186	1.678	2210	1.673	2210	1.674	2211	1.674	2211	1.674
512	1	14932	1.825	1851	1.825	1889	1.808	1927	1.801	1928	1.800	1929	1.800	1930	1.799
512	2	8109	1.786	3400	1.786	3505	1.783	3520	1.782	3523	1.781	3525	1.781	3525	1.781
512	5	3370	1.802	8123	1.801	8376	1.798	8391	1.804	8394	1.804	8396	1.805	8397	1.805

Release notes

Changelog

Mar 18, 2019

Initial release

June, 2019

Performance improvements

June, 2020

Updated performance tables to include A100 results

April 2023

Ceased maintenance of this model in TensorFlow1

Known issues

There are no known issues in this release.