This resource is using open-source code maintained in github (see the quick-start-guide section) and available for download from NGC

BERT, or Bidirectional Encoder Representations from Transformers, is a new method of pre-training language representations which obtains state-of-the-art results on a wide array of Natural Language Processing (NLP) tasks. This model is based on the BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding paper. NVIDIA's BERT is an optimized version of Google's official implementation, leveraging mixed precision arithmetic and Tensor Cores on V100 GPUs for faster training times while maintaining target accuracy.

Other publicly available implementations of BERT include:\nNVIDIA PyTorch\nHugging Face\ncodertimo\ngluon-nlp\nGoogle's official implementation

This model is trained with mixed precision using Tensor Cores on Volta, Turing, and the NVIDIA Ampere GPU architectures. Therefore, researchers can get results up to 4x 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

BERT's model architecture is a multi-layer bidirectional transformer encoder. Based on the model size, we have the following two default configurations of BERT:

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n

tr:nth-child(odd)]:bg-surface-sunken\">

Model	Hidden layers	Hidden unit size	Attention heads	Feedforward filter size	Max sequence length	Parameters
BERTBASE	12 encoder	768	12	4 x 768	512	110M
BERTLARGE	24 encoder	1024	16	4 x 1024	512	330M

BERT training consists of two steps, pre-training the language model in an unsupervised fashion on vast amounts of unannotated datasets, and then using this pre-trained model for fine-tuning for various NLP tasks, such as question and answer, sentence classification, or sentiment analysis. Fine-tuning typically adds an extra layer or two for the specific task and further trains the model using a task-specific annotated dataset, starting from the pre-trained backbone weights. The end-to-end process is depicted in the following image:

$\"\"$

Figure 1: BERT Pipeline

Default configuration

This repository contains scripts to interactively launch data download, training, benchmarking, and inference routines in a Docker container for fine tuning Question Answering. The major differences between the official implementation of the paper and our version of BERT are as follows:

Mixed precision support with TensorFlow Automatic Mixed Precision (TF-AMP), which enables mixed precision training without any changes to the code-base by performing automatic graph rewrites and loss scaling controlled by an environmental variable.
Scripts to download dataset and pretrained checkpoints for:\n
- Pre-training - Wikipedia, BookCorpus
- Fine tuning - SQuAD (Stanford Question Answering Dataset)
- Pretrained weights from Google
\n
Custom fused CUDA kernels for faster computations
Multi-GPU/Multi-node support using Horovod

The following performance optimizations were implemented in this model:

XLA support (experimental).

These techniques and optimizations improve model performance and reduce training time, allowing you to perform various NLP tasks with no additional effort.

Feature support matrix

The following features are supported by this model.

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n

tr:nth-child(odd)]:bg-surface-sunken\">

Feature	BERT
Horovod Multi-GPU	Yes
Horovod Multi-node	Yes
Automatic mixed precision (AMP)	Yes
XLA	Yes
LAMB	Yes

Features

Multi-GPU training with Horovod\nOur model uses Horovod to implement efficient multi-GPU training with NCCL. For details, see example sources in this repository or see the TensorFlow tutorial.

XLA support

XLA is a domain-specific compiler for linear algebra that can accelerate TensorFlow models with potentially no source code changes. The results are improvements in speed and memory usage: most internal benchmarks run ~1.1-1.5x faster after XLA is enabled.

Multi-node Training

Supported on a Pyxis/Enroot Slurm cluster.

LAMB

LAMB stands for Layerwise Adaptive Moments based optimizer, is a large batch optimization technique that helps accelerate training of deep neural networks using large minibatches. It allows using a global batch size of 65536 and 32768 on sequence lengths 128 and 512 respectively, compared to a batch size of 256 for Adam. The optimized implementation accumulates 1024 gradient batches in phase 1 and 4096 steps in phase 2 before updating weights once. This results in 27% training speedup on a single DGX2 node. On multi-node systems, LAMB allows scaling up to 1024 GPUs resulting in training speedups of up to 17x in comparison to Adam. Adam has limitations on the learning rate that can be used since it is applied globally on all parameters whereas LAMB follows a layerwise learning rate strategy.

NVLAMB adds necessary tweaks to LAMB version 1, to ensure correct convergence. A guide to implementating the LAMB optimizer can be found in our article on Medium.com. The algorithm is as follows:\n $\"NVLAMB\"$

Mixed precision training

Mixed precision is the combined use of different numerical precision 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 requires 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.

Enabling mixed precision

This implementation exploits the TensorFlow Automatic Mixed Precision feature. To enable AMP, you simply need to supply the --use_fp16 flag to the run_pretraining.py or run_squad.py script. For reference, enabling AMP required us to apply the following changes to the code:

Set the Keras mixed precision policy:

if FLAGS.use_fp16:\n  policy = tf.keras.mixed_precision.experimental.Policy(\"mixed_float16\")\n  tf.keras.mixed_precision.experimental.set_policy(policy)\n

Use the loss scaling wrapper on the optimizer:

if FLAGS.use_fp16:\n   squad_model.optimizer = tf.keras.mixed_precision.LossScaleOptimizer(squad_model.optimizer,\n     dynamic=True)\n

Use scaled loss to calculate the gradients:

if use_float16:\n  scaled_loss = optimizer.get_scaled_loss(loss)\n  scaled_grads = tape.gradient(scaled_loss, training_vars)\n  grads = optimizer.get_unscaled_gradients(scaled_grads)\n

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.

Glossary

Fine-tuning\nTraining an already pretrained model further using a task specific dataset for subject-specific refinements, by adding task-specific layers on top if required.

Language model\nAssigns a probability distribution over a sequence of words. Given a sequence of words, it assigns a probability to the whole sequence.

Pre-training\nTraining a model on vast amounts of data on the same (or different) task to build general understandings.

Transformer\nThe paper Attention Is All You Need introduces a novel architecture called Transformer that uses an attention mechanism and transforms one sequence into another.

","resourceVersions",[235,236,237],"resourceVersionsTotal","hasMoreVersions","versionMetadata",{"_214":29,"_215":-5,"_53":118,"_115":-5,"_119":216,"_121":-5,"_217":132,"_126":18,"_123":35,"_218":-5,"_122":35,"_219":220,"_221":90,"_222":223,"_224":90,"_225":226,"_227":90,"_228":229,"_230":90,"_231":232,"_233":90,"_124":234},"versionNotFound","isLoggedIn","latestVersionPath","/orgs/nvidia/dle/resources/bert_tf2/-","imageReference","cliDownloadCommand","ngc registry resource download-version nvidia/dle/bert_tf2:21.02.1","helmFetchCommand","viewAllVersionsPath","/orgs/nvidia/dle/resources/bert_tf2/21.02.1/version-history","accessCheckResult",{"_213":90},"hasRelatedCollections","context",{"_209":210,"_211":212},1787460896509,[],[],[202,203,204,205,206,207,208],{"_49":63,"_51":73,"_53":65,"_55":66,"_57":74,"_59":60},{"_49":63,"_51":71,"_53":151,"_55":-5,"_57":-5,"_59":60},{"_49":63,"_51":70,"_53":65,"_55":66,"_57":69,"_59":60},{"_49":63,"_51":68,"_53":65,"_55":66,"_57":69,"_59":60},{"_49":63,"_51":64,"_53":65,"_55":66,"_57":67,"_59":60},{"_49":61,"_51":62,"_53":54,"_55":56,"_57":58,"_59":60},{"_49":50,"_51":52,"_53":54,"_55":56,"_57":58,"_59":60},"orgName","nvidia","teamName","dle","allowed","versionId","description",["D",1675203323025],"totalSizeInBytes","malwareScanStatus","releaseNotes","### Changelog\n\nApril 2021\nInitial release\n\n### Known issues\n\nThere are no known issues with this model.","hasReleaseNotes","setup","The following section lists the requirements that you need to meet in order to start training the BERT model.\n\n### Requirements\n\nThis repository contains `Dockerfile` which extends the TensorFlow NGC container and encapsulates some dependencies. Aside from these dependencies, ensure you have the following components:\n- [NVIDIA Docker](https://github.com/NVIDIA/nvidia-docker)\n- [TensorFlow 21.02-py3+](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow) NGC container\n- GPU-based architecture:\n - [NVIDIA Volta](https://www.nvidia.com/en-us/data-center/volta-gpu-architecture/)\n - [NVIDIA Turing](https://www.nvidia.com/en-us/geforce/turing/)\n - [NVIDIA Ampere architecture](https://www.nvidia.com/en-us/data-center/nvidia-ampere-gpu-architecture/)\n\nFor 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:\n- [Getting Started Using NVIDIA GPU Cloud](https://docs.nvidia.com/ngc/ngc-getting-started-guide/index.html)\n- [Accessing And Pulling From The NGC Container Registry](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html#accessing_registry)\n- [Running TensorFlow](https://docs.nvidia.com/deeplearning/frameworks/tensorflow-release-notes/running.html#running)\n\nFor 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](https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html).\n\nFor multi-node, the sample provided in this repository requires [Enroot](https://github.com/NVIDIA/enroot) and [Pyxis](https://github.com/NVIDIA/pyxis) set up on a [SLURM](https://slurm.schedmd.com) cluster.\n\nMore information on how to set up and launch can be found in the [Multi-node Documentation](https://docs.nvidia.com/ngc/multi-node-bert-user-guide).","hasSetup","quickStartGuide","To pretrain or fine tune your model for Question Answering using mixed precision with Tensor Cores or using FP32/TF32, perform the following steps using the default parameters of the BERT model.\n\n1. Clone the repository.\n\n```bash\ngit clone https://github.com/NVIDIA/DeepLearningExamples\ncd DeepLearningExamples/TensorFlow2/LanguageModeling/BERT\n```\n\n2. Build the BERT TensorFlow NGC container.\n\n```bash\nbash scripts/docker/build.sh\n```\n\n3. Download and preprocess the dataset.\n\nThis repository provides scripts to download, verify, extract the SQuAD dataset and pretrained weights for fine tuning as well as Wikipedia and BookCorpus dataset for pre-training.\n\nTo download, verify, and extract the required datasets, run:\n\n```bash\nbash scripts/data_download.sh\n```\n\nThe script launches a Docker container with the current directory mounted and downloads the datasets to a `data/` folder on the host.\n\n- Download datasets for fine tuning and pretraining\n\n`bash scripts/data_download.sh all`\n\n- Download datasets for fine tuning only\n\n`bash scripts/data_download.sh squad`\n\n- Download Wikipedia only for pretraining\n\nThe pretraining dataset is 170GB+ and takes 15+ hours to download. The BookCorpus server most of the times get overloaded and also contains broken links resulting in HTTP 403 and 503 errors. Hence, it is recommended to skip downloading BookCorpus data by running:\n\n`bash scripts/data_download.sh pretrained wiki_only`\n\nNote: Wikipedia dataset is 17GB and takes about one hour to download.\n\n- Download Wikipedia and BookCorpus\n\nUsers can download BookCorpus from other sources to match our accuracy, or repeatedly try our script until the required number of files are downloaded by running the following:\n\n`bash scripts/data_download.sh pretrained wiki_books`\n\nNote: Not using BookCorpus can potentially change final accuracy on a few downstream tasks.\n\n4. Download the pretrained models from NGC.\n\nWe have uploaded checkpoints for fine tuning with various configurations on the NGC Model Registry. You can download them directly from the [NGC model catalog](https://ngc.nvidia.com/catalog/models). Download them to the `results/models/` to easily access them in your scripts.\n\n5. Start an interactive session in the NGC container to run training/inference.\n\nAfter you build the container image and download the data, you can start an interactive CLI session as follows:\n\n```bash\nbash scripts/docker/launch.sh\n```\n\n6. Start pre-training.\n\nBERT is designed to pre-train deep bidirectional representations for language representations. The following scripts are to replicate pre-training on Wikipedia and BookCorpus from the [LAMB paper](https://arxiv.org/pdf/1904.00962.pdf). These scripts are general and can be used for pre-training language representations on any corpus of choice.\n\nFrom within the container, you can use the following script to run pre-training using LAMB.\n```bash\nbash scripts/run_pretraining_lamb.sh \n```\n\nFor BERT-Large FP16 training with XLA using a DGX-2H, run:\n```bash\nbash scripts/run_pretraining_lamb.sh 60 10 8 7.5e-4 5e-4 fp16 true 8 2000 200 7820 100 64 192 large\n```\n\nThis repository also contains a number of predefined configurations to run the LAMB pretraining on NVIDIA DGX-1, NVIDIA DGX-2H or NVIDIA DGX A100 nodes in `scripts/configs/pretrain_config.sh`. For example, to use the default DGX A100 8 GPU config, run:\n\n```bash\nbash scripts/run_pretraining_lamb.sh $(source scripts/configs/pretrain_config.sh && dgxa100_8gpu_fp16)\n```\n\nAlternatively, to run pre-training with Adam as in the original [BERT paper](https://arxiv.org/pdf/1810.04805.pdf) from within the container, run:\n\n```bash\nbash scripts/run_pretraining_adam.sh \n```\n\n7. Start fine tuning.\n\nThe above pretrained BERT representations can be fine tuned with just one additional output layer for a state-of-the-art Question Answering system. From within the container, you can use the following script to run fine-training for SQuAD.\n\n```bash\nbash scripts/run_squad.sh \n```\n\nFor SQuAD 1.1 FP16 training with XLA using a DGX-1 V100 32G, run:\n```bash\nbash scripts/run_squad.sh 8 12 5e-6 fp16 true large 1.1 2\n```\n\nFor SQuAD 2.0 FP32 training without XLA using a DGX-1 V100 32G, run:\n```bash\nbash scripts/run_squad.sh 8 8 5e-6 fp32 false large 2.0 2\n```\n\nThe fine-tuned checkpoint will save to `/results/tf_bert_finetuning_squad_xxxxxx/ctl_step_xxx.ckpt-x`\n\n8. Start validation/evaluation.\n\nThe `run_squad_inference.sh` script runs inference on a checkpoint fine tuned for SQuAD and evaluates the validity of predictions on the basis of exact match and F1 score.\n\n```bash\nbash scripts/run_squad_inference.sh \n```\n\nThe `init_checkpoint` is the fine-tuned checkpoint path. For example, we take the checkpoint from previous step `/results/tf_bert_finetuning_squad_xxxxxx/ctl_step_xxx.ckpt-x` and rename to `/results/model.ckpt`. SQuAD 1.1 and SQuAD 2.0 should be different checkpoints.\n\nFor SQuAD 2.0 FP16 inference with XLA using a DGX-1 V100 32G, run:\n```bash\nbash scripts/run_squad_inference.sh /results/model.ckpt 8 fp16 true large 2.0\n```\n\nFor SQuAD 1.1 FP32 inference without XLA using a DGX-1 V100 32G, run:\n```bash\nbash scripts/run_squad_inference.sh /results/model.ckpt 8 fp32 false large 1.1\n```","hasQuickStartGuide","advanced","The following sections provide greater details of the dataset, running training and inference, and the training results.\n\n### Scripts and sample code\n\nIn the root directory, the most important files are:\n* `run_pretraining.py` - Serves as entry point for pre-training\n* `run_squad.py` - Serves as entry point for SQuAD training\n* `Dockerfile` - Container with the basic set of dependencies to run BERT\n\nThe `scripts/` folder encapsulates all the one-click scripts required for running various functionalities supported such as:\n* `run_squad.sh` - Runs SQuAD training and inference using `run_squad.py` file\n* `run_pretraining_adam.sh` - Runs pre-training with Adam optimizer using the `run_pretraining.py` file\n* `run_pretraining_lamb.sh` - Runs pre-training with LAMB optimizer using the `run_pretraining.py` file in two phases. Phase 1 does 90% of training with sequence length = 128. In phase 2, the remaining 10% of the training is done with sequence length = 512.\n* `data_download.sh` - Downloads datasets using files in the `data/` folder\n* `finetune_train_benchmark.sh` - Captures performance metrics of training for multiple configurations\n* `finetune_inference_benchmark.sh` - Captures performance metrics of inference for multiple configurations\n\nThe `data/` folder contains necessary folders and scripts to download datasets required for fine tuning and pre-training BERT.\n\nAfter downloading the data, the `launch.sh` script assumes that the datasets are in the following locations by default\n\n- SQuAD v1.1 - `data/download/squad/v1.1`\n- SQuAD v2.0 - `data/download/squad/v2.0`\n- BERT-Large - `data/download/google_pretrained_weights/uncased_L-24_H-1024_A-16`\n- BERT-Base - `data/download/google_pretrained_weights/uncased_L-12_H-768_A-12`\n- Wikipedia + BookCorpus TFRecords - `data/tfrecords/books_wiki_en_corpus`\n\nThe `official/` folder contains necessary files of building model architecture and training process.\n\n### Parameters\n\nAside from the options to set hyperparameters, the relevant options to control the behaviour of the `run_pretraining.py` script are:\n\n```\n --bert_config_file: Bert configuration file to define core bert layers.\n --init_checkpoint: Initial checkpoint (usually from a pre-trained BERT model).\n --[no]use_horovod: Whether to use horovod.(default: 'false')\n --[no]use_fp16: Whether to use fp32 or fp16 arithmetic on GPU. When false, uses TF32 on A100 and FP32 on V100 GPUS.(default: 'false')\n --[no]enable_xla: Whether to enable XLA auto jit compilation.(default: 'false')\n --input_files: File path to retrieve training data for pre-training.\n --model_dir: The location of the model checkpoint files.\n --optimizer_type: Optimizer used for training - LAMB or ADAM\n --num_accumulation_steps: Number of accumulation steps before gradient update. Global batch size = num_accumulation_steps * train_batch_size\n```\n\nAside from the options to set hyperparameters, some relevant options to control the behaviour of the `run_squad.py` script are:\n\n```\n --bert_config_file: Bert configuration file to define core bert layers.\n --model_dir: The location of the model checkpoint files.\n --mode: : One of {\"train_and_predict\", \"train\", \"predict\", \"export_only\"}. `train_and_predict`: both train and predict to a json file. `train`: only trains the model. trains the model and evaluates in the meantime. `predict`: predict answers from the squad json file. `export_only`: will take the latest checkpoint inside model_dir and export a `SavedModel`.\n --max_answer_length: The maximum length of an answer that can be generated. (default: '30')(an integer)\n --input_meta_data_path: Path to file that contains meta data about input to be used for training and evaluation.\n --predict_batch_size: Total batch size for predictions.(default: '8')(an integer)\n --train_batch_size: Total batch size for training.(default: '8')(an integer)\n --[no]use_fp16: Whether to use fp32 or fp16 arithmetic on GPU. When false, uses TF32 on A100 and FP32 on V100 GPUS.(default: 'false')\n --[no]enable_xla: Whether to enable XLA auto jit compilation.(default: 'false')\n```\n\n### Command-line options\n\nTo see the full list of available options and their descriptions, use the `-h` or `--helpfull` command-line option with the Python file, for example:\n\n```bash\npython run_pretraining.py --helpfull\npython run_squad.py --helpfull\n```\n\n### Getting the data\n\nFor pre-training BERT, we use the concatenation of Wikipedia (2500M words) as well as BookCorpus (800M words). For Wikipedia, we extract only the text passages from [here](https://dumps.wikimedia.your.org/enwiki/latest/enwiki-latest-pages-articles.xml.bz2) and ignore headers list and tables. It is structured as a document level corpus rather than a shuffled sentence level corpus because it is critical to extract long contiguous sentences.\n\nThe next step is to run `create_pretraining_data.py` with the document level corpus as input, which generates input data and labels for the masked language modeling and next sentence prediction tasks. Pre-training can also be performed on any corpus of your choice. The collection of data generation scripts are intended to be modular to allow modifications for additional preprocessing steps or to use additional data. They can hence easily be modified for an arbitrary corpus.\n\nThe preparation of an individual pre-training dataset is described in the `create_datasets_from_start.sh` script found in the `data/` folder. The component steps to prepare the datasets are as follows:\n\n1. Data download and extract - the dataset is downloaded and extracted.\n2. Clean and format - document tags, etc. are removed from the dataset. The end result of this step is a `{dataset_name_one_article_per_line}.txt` file that contains the entire corpus. Each line in the text file contains an entire document from the corpus. One file per dataset is created in the `formatted_one_article_per_line` folder.\n3. Sharding - the sentence segmented corpus file is split into a number of smaller text documents. The sharding is configured so that a document will not be split between two shards. Sentence segmentation is performed at this time using Natural Language Toolkit (NLTK).\n4. TFRecord file creation - each text file shard is processed by the `create_pretraining_data.py` script to produce a corresponding TFRecord file. The script generates input data and labels for masked language modeling and sentence prediction tasks for the input text shard.\n\nFor fine tuning BERT, for the task of Question Answering, we use SQuAD. SQuAD v1.1 has 100,000+ question-answer pairs on 500+ articles. SQuAD v2.0 combines v1.1 with an additional 50,000 new unanswerable questions and must not only answer questions but also determine when that is not possible.\n\n#### Dataset guidelines\n\nThe procedure to prepare a text corpus for pre-training is described in the previous section. This section provides additional insight into how exactly raw text is processed so that it is ready for pre-training.\n\nFirst, raw text is tokenized using [WordPiece tokenization](https://arxiv.org/pdf/1609.08144.pdf). A [CLS] token is inserted at the start of every sequence, and the two sentences in the sequence are separated by a [SEP] token.\n\nNote: BERT pre-training looks at pairs of sentences at a time. A sentence embedding token [A] is added to the first sentence and token [B] to the next.\n\nBERT pre-training optimizes for two unsupervised classification tasks. The first is Masked Language Modelling (Masked LM). One training instance of Masked LM is a single modified sentence. Each token in the sentence has a 15% chance of being replaced by a [MASK] token. The chosen token is replaced with [MASK] 80% of the time, 10% with another random token, and the remaining 10% with the same token. The task is then to predict the original token.\n\nThe second task is next sentence prediction. One training instance of BERT pre-training is two sentences (a sentence pair). A sentence pair may be constructed by simply taking two adjacent sentences from a single document, or by pairing up two random sentences with equal probability. The goal of this task is to predict whether or not the second sentence followed the first in the original document.\n\nThe `create_pretraining_data.py` script takes in raw text and creates training instances for both pre-training tasks.\n\n#### Multi-dataset\n\nWe are able to combine multiple datasets into a single dataset for pre-training on a diverse text corpus. Once TFRecords have been created for each component dataset, you can create a combined dataset by adding the directory to `SOURCES` in `run_pretraining_*.sh`. This will feed all matching files to the input pipeline in `run_pretraining.py`. However, in the training process, only one TFRecord file is consumed at a time, therefore, the training instances of any given training batch will all belong to the same source dataset.\n\n### Training process\n\n#### Pre-training\n\nPre-training is performed using the `run_pretraining.py` script along with parameters defined in the `scripts/run_pretraining_lamb.sh`.\n\nThe `run_pretraining_lamb.sh` script runs a job on a single node that trains the BERT-Large model from scratch using the Wikipedia and BookCorpus datasets as training data. By default, the training script:\n- Runs on 8 GPUs.\n- Has FP16 precision enabled.\n- Is XLA enabled.\n- Creates a log file containing all the output.\n- Saves a checkpoint every 100 iterations (keeps only the latest checkpoint) at the end of training. All checkpoints, evaluation results, and training logs are saved to the `/results` directory (in the container which can be mounted to a local directory).\n- Evaluates the model at the end of each phase.\n\n- Phase 1\n - Runs 7506 steps with 2133 warmup steps\n - Sets Maximum sequence length to 128\n - Sets Global Batch size to 61K\n\n- Phase 2\n - Runs 1668 steps with 213 warm-up steps\n - Sets Maximum sequence length to 512\n - Sets Global Batch size to 30K\n - Starts from Phase1's final checkpoint\n\nThese parameters train Wikipedia and BookCorpus with reasonable accuracy on a DGX-1 with 32GB V100 cards.\n\nFor example:\n```bash\nscripts/run_pretraining_lamb.sh \n```\n\nWhere:\n- `` is per-GPU batch size used for training in the respective phase. Batch size varies with precision, larger batch sizes run more efficiently, but require more memory.\n\n- `` is per-GPU batch size used for evaluation after training.\n\n- `` is the default rate of 5e-4 is good for global batch size 61K.\n\n- `` is the default rate of 7.5e-4 is good for global batch size 30K.\n\n- `` is the type of math in your model, can be either `fp32` or `fp16`. Specifically:\n\n - `fp32` is 32-bit IEEE single precision floats. Is enabled by default on V100.\n - `fp16` is Automatic rewrite of TensorFlow compute graph to take advantage of 16-bit arithmetic whenever it is safe.\n - `tf32` uses same 10 bit mantissa as FP16 and 8-bit exponent as FP32. Is enabled by default on A100.\n\n- `` is the number of GPUs to use for training. Must be equal to or smaller than the number of GPUs attached to your node.\n\n- `` is the number of warm-up steps at the start of training in the respective phase.\n\n- `` is the total number of training steps in both phases combined.\n\n- `` controls how often checkpoints are saved. Default is 100 steps.\n\n- `` is used to mimic higher batch sizes in the respective phase by accumulating gradients N times before weight update.\n\n- `` is used to indicate whether to pretrain BERT-Large or BERT-Base model\n\nThe following sample code trains BERT-Large from scratch on a single DGX-2 using FP16 arithmetic. This will take around 4.5 days.\n\n```bash\nscripts/run_pretraining_lamb.sh 60 10 8 7.5e-4 5e-4 fp16 true 8 2000 200 7820 100 64 192 large\n```\n\n#### Fine tuning\n\nFine tuning is performed using the `run_squad.py` script along with parameters defined in `scripts/run_squad.sh`.\n\nThe `run_squad.sh` script trains a model and performs evaluation on the SQuAD dataset. By default, the training script:\n\n- Trains for SQuAD v1.1 dataset.\n- Trains on BERT-Large model.\n- Uses 8 GPUs and batch size of 12 on each GPU.\n- Has FP16 precision enabled.\n- Is XLA enabled.\n- Runs for 2 epochs.\n- Saves a checkpoint every 180 seconds (keeps only the latest checkpoint) at the end of training. All checkpoints, evaluation results, and training logs are saved to the `/results` directory (in the container which can be mounted to a local directory).\n- Evaluation is done at the end of training. To skip evaluation, modify `--mode` from `train_and_predict` to `train`.\n\nThis script outputs checkpoints to the `/results` directory, by default, inside the container. Mount point of `/results` can be changed in the `scripts/docker/launch.sh` file. The training log contains information about:\n- Loss for the final step\n- Training and evaluation performance\n- F1 and exact match score on the Dev Set of SQuAD after evaluation.\n\nThe summary after training is printed in the following format:\n```bash\nI0415 18:12:49.376930 140671213582144 model_training_utils.py:82] Training Summary:\n{'total_training_steps': 1846, 'train_loss': 0.6074678301811218}\nI0415 18:12:49.377982 140671213582144 model_training_utils.py:564] -----------------------------\nI0415 18:12:49.377468 140671213582144 model_training_utils.py:558] Batch size = 12\n...\nI0415 18:12:49.379069 140671213582144 model_training_utils.py:568] -----------------------------\n```\n\nMulti-GPU training is enabled with the Horovod TensorFlow module. The following example runs training on 8 GPUs:\n\n```bash\nBERT_DIR=data/download/google_pretrained_weights/uncased_L-24_H-1024_A-16\n\nmpirun -np 8 \\\n --allow-run-as-root -bind-to none -map-by slot \\\n -x NCCL_DEBUG=INFO \\\n -x LD_LIBRARY_PATH \\\n -x PATH -mca pml ob1 -mca btl ^openib \\\n python run_squad.py --use_horovod --vocab_file=$BERT_DIR/vocab.txt \\\n --bert_config_file=$BERT_DIR/bert_config.json \\\n --model_dir=/results\n```\n\n#### Multi-node\n\nMulti-node runs can be launched on a pyxis/enroot Slurm cluster (see [Requirements](#requirements)) with the `run.sub` script with the following command for a 4-node DGX-1 example for both phase 1 and phase 2:\n\n```\nBATCHSIZE=16 LEARNING_RATE='1.875e-4' NUM_ACCUMULATION_STEPS=128 PHASE=1 sbatch -N4 --ntasks-per-node=8 run.sub\nBATCHSIZE=2 LEARNING_RATE='1.25e-4' NUM_ACCUMULATION_STEPS=512 PHASE=1 sbatch -N4 --ntasks-per-node=8 run.sub\n```\n\nCheckpoint after phase 1 will be saved in `model_dir` specified in `run.sub`. The checkpoint will be automatically picked up to resume training on phase 2. Note that phase 2 should be run after phase 1.\n\nVariables to re-run the [Training performance results](#training-performance-results) are available in the `scripts/configs/configurations.yml` file.\n\nThe batch variables `BATCHSIZE`, `LEARNING_RATE`, `NUM_ACCUMULATION_STEPS` refer to the Python arguments `train_batch_size`, `learning_rate`, `num_accumulation_steps` respectively.\nThe variable `PHASE` refers to phase specific arguments available in `run.sub`.\n\nNote that the `run.sub` script is a starting point that has to be adapted depending on the environment. In particular, variables such as `datadir` handle the location of the files for each phase.\n\nRefer to the files contents to see the full list of variables to adjust for your system.\n\n### Inference process\n\nInference on a fine tuned Question Answering system is performed using the `run_squad.py` script along with parameters defined in `scripts/run_squad_inference.sh`. Inference is supported on a single GPU.\n\nThe `run_squad_inference.sh` script trains a model and performs evaluation on the SQuAD dataset. By default, the inferencing script:\n\n- Uses SQuAD v1.1 dataset\n- Has FP16 precision enabled\n- Is XLA enabled\n- Evaluates the latest checkpoint present in `/results` with a batch size of 8\n\nThis script outputs predictions file to `/results/predictions.json` and computes F1 score and exact match score using SQuAD's evaluate file. Mount point of `/results` can be changed in the `scripts/docker/launch.sh` file.\n\nThe output log contains information about:\nInference performance\nInference accuracy (F1 and exact match scores) on the Dev Set of SQuAD after evaluation.\n\nThe summary after inference is printed in the following format:\n```bash\nI0424 23:59:50.030514 139905798453056 run_squad.py:268] -----------------------------\nI0424 23:59:50.030774 139905798453056 run_squad.py:269] Summary Inference Statistics\nI0424 23:59:50.030934 139905798453056 run_squad.py:270] Batch size = 8\nI0424 23:59:50.031085 139905798453056 run_squad.py:271] Sequence Length = 384\nI0424 23:59:50.031238 139905798453056 run_squad.py:272] Precision = fp16\nI0424 23:59:50.031387 139905798453056 run_squad.py:274] Total Inference Time = 88.29 for Sentences = 10840\nI0424 23:59:50.031537 139905798453056 run_squad.py:302] -----------------------------\n{\"exact_match\": 84.08703878902554, \"f1\": 90.87995817872932}\n```","hasAdvanced","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](https://developer.nvidia.com/deep-learning-performance-training-inference).\n\n### Benchmarking\n\nThe following section shows how to run benchmarks measuring the model performance in training and inference modes.\n\nBoth of these benchmarking scripts enable you to run a number of epochs, extract performance numbers, and run the BERT model for fine tuning.\n\n#### Training performance benchmark\n\nTraining benchmarking can be performed by running the script:\n``` bash\nscripts/finetune_train_benchmark.sh \n```\n\nThis script runs 800 steps by default on the SQuAD v1.1 dataset and extracts performance numbers for the given configuration. These numbers are saved at `/results/squad_train_benchmark__gpu_bs.log`.\n\n#### Inference performance benchmark\n\nInference benchmarking can be performed by running the script:\n\n``` bash\nscripts/finetune_inference_benchmark.sh \n```\n\nThis script runs 1000 eval iterations by default on the SQuAD v1.1 dataset and extracts performance and latency numbers for the given configuration. These numbers are saved at `/results/squad_inference_benchmark___bs.log`.\n\n### Results\n\nThe following sections provide details on how we achieved our performance and accuracy in training and inference for fine tuning Question Answering. All results are on BERT-Large model unless otherwise mentioned. All fine tuning results are on SQuAD v1.1 using a sequence length of 384 unless otherwise mentioned.\n\n#### Training accuracy results\n\n##### Pre-training accuracy\n\nOur results were obtained by running the `scripts/run_pretraining_lamb.sh` training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-2 and NVIDIA DGX A100.\n\n| **DGX System** | **Nodes x GPUs** | **Precision** | **Batch Size/GPU: Phase1, Phase2** | **Accumulation Steps: Phase1, Phase2** | **Time to Train (Hrs)** | **Final Loss** |\n|----------------|-----------|---------------|------------------------------------|----------------------------------------|----------------|-------------------------|\n| DGX2H | 32 x 16 | FP16 | 56, 10 | 2, 6 | 2.67 | 1.69 |\n| DGX2H | 32 x 16 | FP32 | 32, 4 | 4, 16 | 8.02 | 1.71 |\n| DGXA100 | 32 x 8 | FP16 | 312, 40 | 1, 3 | 2.02 | 1.68 |\n| DGXA100 | 32 x 8 | TF32 | 176, 22 | 2, 6 | 3.57 | 1.67 |\n\n##### Fine-tuning accuracy for SQuAD v1.1: NVIDIA DGX A100 (8x A100 40GB)\n\nOur results were obtained by running the `scripts/run_squad.sh` training script in the TensorFlow 20.12-py3 NGC container on NVIDIA DGX A100 with 8x A100 80GB GPUs.\n\n| **GPUs** | **Batch size / GPU: TF32, FP16 ** | **Accuracy - TF32** | **Accuracy - mixed precision** | **Time to Train - TF32 (Hrs)** | **Time to Train - mixed precision (Hrs)** |\n|:---:|:----:|:----:|:---:|:----:|:----:|\n| 8 | 38, 76 | 90.88 | 91.12 | 0.16 | 0.11 |\n\n##### Pre-training SQuAD v1.1 stability test: NVIDIA DGX A100 (256x A100 80GB)\n\nThe following tables compare `Final Loss` scores across 3 different training runs with different seeds, for both FP16 and TF32. The runs showcase consistent convergence on all 3 seeds with very little deviation.\n\n| **FP16, 256x GPUs** | **seed 1** | **seed 2** | **seed 3** | **mean** | **std** |\n|:-----------:|:-----:|:-----:|:-----:|:-----:|:-----:|\n|Final Loss |1.657 |1.661 |1.683 |1.667 |0.014 |\n\n| **TF32, 256x GPUs** | **seed 1** | **seed 2** | **seed 3** | **mean** | **std** |\n|:-----------:|:-----:|:-----:|:-----:|:-----:|:-----:|\n|Final Loss |1.67 |1.654 |1.636 |1.653 |0.017 |\n\n##### Fine-tuning SQuAD v1.1 stability test: NVIDIA DGX A100 (8x A100 80GB)\n\nThe following tables compare `F1` scores across 5 different training runs with different seeds, for both FP16 and TF32 respectively using the (NVIDIA Pretrained Checkpoint)[https://ngc.nvidia.com/catalog/models]. The runs showcase consistent convergence on all 5 seeds with very little deviation.\n\n| **FP16, 8x GPUs** | **seed 1** | **seed 2** | **seed 3** | **seed 4** | **seed 5** | **mean** | **std** |\n|:-----------:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|\n|F1 |91.12 |90.80 |90.94 |90.90 |90.94 |90.94 |0.11 |\n\n| **TF32, 8x GPUs** | **seed 1** | **seed 2** | **seed 3** | **seed 4** | **seed 5** | **mean** | **std** |\n|:-----------:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|:-----:|\n|F1 |90.79 |90.88 |90.80 |90.88 |90.83 |90.84 |0.04 |\n\n#### Training performance results\n\n##### Pre-training training performance: Single-node on NVIDIA DGX-2 V100 (16x V100 32GB)\n\nOur results were obtained by running the `scripts/run_pretraining_lamb.sh` training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-2 with 16x V100 32GB GPUs. Performance (in sequences per second) is the steady state throughput.\n\n| **GPUs** | **Sequence Length** | **Batch size / GPU: mixed precision, FP32** | **Gradient Accumulation: mixed precision, FP32** | **Global Batch Size: mixed precision, FP32** | **Throughput - mixed precision** | **Throughput - FP32** | **Throughput speedup (FP32 - mixed precision)** | **Weak scaling - mixed precision** | **Weak scaling - FP32** |\n|:--------:|:-------------------:|:-------------------------------------------:|--------------------------------------------------|:--------------------------------------------:|:--------------------------------:|:---------------------:|-------------------------------------------------|------------------------------------|-------------------------|\n| 1 | 128 | 60 , 32 | 1024 , 2048 | 61440 , 65536 | 206.5 | 49.97 | 4.13 | 1.00\t| 1.00 |\n| 4 | 128 | 60 , 32 | 256 , 512 | 61440 , 65536 | 789.75 | 194.02 | 4.07 | 3.82\t| 3.88 |\n| 8 | 128 | 60 , 32 | 128 , 256 | 61440 , 65536 | 1561.77\t| 367.9 | 4.25 | 7.56\t| 7.36 |\n| 16 | 128 | 60 , 32 | 64 , 128 | 61440 , 65536 | 3077.99\t| 762.22 | 4.04 | 14.9\t| 15.25 |\n| 1 | 512 | 10 , 6 | 3072 , 5120 | 30720 , 30720 | 40.95\t | 11.06 | 3.70 | 1.00\t| 1.00 |\n| 4 | 512 | 10 , 6 | 768 , 1280 | 30720 , 30720 | 158.5\t | 43.05 | 3.68 | 3.87\t| 3.89 |\n| 8 | 512 | 10 , 6 | 384 , 640 | 30720 , 30720 | 312.03\t| 85.51 | 3.65 | 7.62\t| 7.73 |\n| 16 | 512 | 10 , 4 | 192 , 512 | 30720 , 32768 | 614.94\t| 161.38 | 3.81 | 15.02\t| 14.59 |\n\nNote: The respective values for FP32 runs that use a batch size of 60 and 10 in sequence lengths 128 and 512 are not available due to out of memory errors that arise.\n\n##### Pre-training training performance: Multi-node on NVIDIA DGX-2H V100 (16x V100 32GB)\n\nOur results were obtained by running the `run.sub` training script in the TensorFlow 21.02-py3 NGC container using multiple NVIDIA DGX-2 with 16x V100 32GB GPUs. Performance (in sequences per second) is the steady state throughput.\n\n| **Num Nodes** | **Sequence Length** | **Batch size / GPU: mixed precision, FP32** | **Gradient Accumulation: mixed precision, FP32** | **Global Batch Size: mixed precision, FP32** | **Throughput - mixed precision** | **Throughput - FP32** | **Throughput speedup (FP32 - mixed precision)** | **Weak scaling - mixed precision** | **Weak scaling - FP32** |\n|:-------------:|:-------------------:|:-------------------------------------------:|--------------------------------------------------|:--------------------------------------------:|:--------------------------------:|:---------------------:|-------------------------------------------------|------------------------------------|-------------------------|\n| 1 | 128 | 60 , 32 | 64 , 128 | 61440 , 65536 | 3528.51\t| 841.72 | 4.19 | 1.00 | 1.00 |\n| 4 | 128 | 60 , 32 | 16 , 32 | 61440 , 65536 | 13370.21\t| 3060.49 | 4.37 | 3.79\t| 3.64 |\n| 16 | 128 | 60 , 32 | 4 , 8 | 61440 , 65536 | 42697.42\t| 10383.57 | 4.11 | 12.1\t| 12.34 |\n| 32 | 128 | 60 , 32 | 2 , 4 | 61440 , 65536 | 84223.16\t| 20094.14 | 4.19 | 23.87\t| 23.87 |\n| 1 | 512 | 10 , 4 | 192 , 256 | 30720 , 32768 | 678.35\t\t| 180 | 3.77 | 1.00 | 1.00 |\n| 4 | 512 | 10 , 4 | 96 , 64 | 30720 , 32768 | 2678.29\t| 646.76 | 4.14 | 3.95\t| 3.59 |\n| 16 | 512 | 10 , 4 | 24 , 32 | 30720 , 32768 | 7834.72\t| 2204.72 | 3.55 | 11.55\t| 12.25 |\n| 32 | 512 | 10 , 4 | 6 , 16 | 30720 , 32768 | 18786.93\t| 4196.15 | 4.48 | 27.70\t| 23.31 |\n\nNote: The respective values for FP32 runs that use a batch size of 60 and 10 in sequence lengths 128 and 512 are not available due to out of memory errors that arise.\n\n##### Pre-training training performance: Single-node on NVIDIA DGX A100 (8x A100 80GB)\n\nOur results were obtained by running the `scripts/run_pretraining_lamb.sh` training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX A100 with 8x A100 80GB GPUs. Performance (in sequences per second) is the steady state throughput.\n\n| **GPUs** | **Sequence Length** | **Batch size / GPU: mixed precision, TF32** | **Gradient Accumulation: mixed precision, TF32** | **Global Batch Size: mixed precision, FP32** | **Throughput - mixed precision** | **Throughput - TF32** | **Throughput speedup (TF32 - mixed precision)** | **Weak scaling - mixed precision** | **Weak scaling -TF32** |\n|:--------:|:-------------------:|:-------------------------------------------:|--------------------------------------------------|:--------------------------------------------:|:--------------------------------:|:---------------------:|-------------------------------------------------|------------------------------------|------------------------|\n| 1 | 128 | 312 , 176 | 256 , 512 | 79872 , 90112 | 485.59 | 282.98 | 1.72 | 1.00 | 1.00 |\n| 8 | 128 | 312 , 176 | 32 , 64 | 79872 , 90112 | 3799.24 | 1944.77 | 1.95 | 7.82 | 6.87 |\n| 1 | 512 | 40 , 22 | 768 , 1536 | 30720 , 33792 | 96.52 | 54.92 | 1.76 | 1.00 | 1.00 |\n| 8 | 512 | 40 , 22 | 96 , 192 | 30720 , 33792 | 649.69 | 427.39 | 1.52 | 6.73 | 7.78 |\n\nNote: The respective values for TF32 runs that use a batch size of 312 and 40 in sequence lengths 128 and 512 are not available due to out of memory errors that arise.\n\n##### Pre-training training performance: Multi-node on NVIDIA DGX A100 (8x A100 80GB)\n\nOur results were obtained by running the `scripts/run_pretraining_lamb.sh` training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX A100 with 8x A100 40GB GPUs. Performance (in sequences per second) is the steady state throughput.\n\n| **Num Nodes** | **Sequence Length** | **Batch size / GPU: mixed precision, TF32** | **Gradient Accumulation: mixed precision, TF32** | **Global Batch Size: mixed precision, FP32** | **Throughput - mixed precision** | **Throughput - TF32** | **Throughput speedup (TF32 - mixed precision)** | **Weak scaling - mixed precision** | **Weak scaling -TF32** |\n|:-------------:|:-------------------:|:-------------------------------------------:|--------------------------------------------------|:--------------------------------------------:|:--------------------------------:|:---------------------:|-------------------------------------------------|------------------------------------|------------------------|\n| 1 | 128 | 312 , 176 | 32 , 64 | 79872 , 90112 | 3803.82\t| 2062.98 | 1.84 |1.00 | 1.00 |\n| 2 | 128 | 312 , 176 | 16 , 32 | 79872 , 90112 | 7551.37\t| 4084.76 | 1.85 |1.99 | 1.98 |\n| 8 | 128 | 312 , 176 | 4 , 8 | 79872 , 90112 | 29711.11\t| 16134.02 | 1.84 |7.81 | 7.82 |\n| 32 | 128 | 312 , 176 | 1 , 2 | 79872 , 90112 | 110280.73\t| 59569.77 | 1.85 |28.99 | 28.88 |\n| 1 | 512 | 40 , 22 | 96 , 192 | 30720 , 33792 | 749.73\t | 431.89 | 1.74 |1.00 | 1.00 |\n| 2 | 512 | 40 , 22 | 48 , 96 | 30720 , 33792 | 1491.87\t| 739.14 | 2.02 |1.99 | 1.71 |\n| 8 | 512 | 40 , 22 | 12 , 24 | 30720 , 33792 | 5870.83\t| 2926.58 | 2.01 |7.83 | 6.78 |\n| 32 | 512 | 40 , 22 | 3 , 6 | 30720 , 33792 | 22506.23\t| 11240.5 | 2.00 |30.02 | 26.03 |\n\nNote: The respective values for TF32 runs that use a batch size of 312 and 40 in sequence lengths 128 and 512 are not available due to out of memory errors that arise.\n\n##### Fine-tuning training performance for SQuAD v1.1 on NVIDIA DGX-1 V100 (8x V100 16GB)\n\nOur results were obtained by running the `scripts/run_squad.sh` training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-1 with 8x V100 16GB GPUs. Performance (in sequences per second) is the mean throughput from 2 epochs.\n\n| **GPUs** | **Batch size / GPU: mixed precision, FP32** | **Throughput - mixed precision** | **Throughput - FP32** | **Throughput speedup (FP32 to mixed precision)** | **Weak scaling - FP32** | **Weak scaling - mixed precision** |\n|:---:|:---:|:------:|:-----:|:----:|:----:|:----:|\n| 1 | 6,3 | 39.10 | 9.85\t| 3.97 | 1.00 | 1.00 |\n| 4 | 6,3 | 128.48\t| 36.52\t| 3.52 | 3.29 | 3.71 |\n| 8 | 6,3 | 255.36\t| 73.03\t| 3.5 | 6.53 | 7.41 |\n\nNote: The respective values for FP32 runs that use a batch size of 6 are not available due to out of memory errors that arise. Batch size of 6 is only available on using FP16.\n\nTo achieve these same results, follow the [Quick Start Guide](#quick-start-guide) outlined above.\n\n##### Fine-tuning training performance for SQuAD v1.1 on NVIDIA DGX-1 V100 (8x V100 32GB)\n\nOur results were obtained by running the `scripts/run_squad.sh` training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-1 with 8x V100 32GB GPUs. Performance (in sequences per second) is the mean throughput from 2 epochs.\n\n| **GPUs** | **Batch size / GPU: mixed precision, FP32** | **Throughput - mixed precision** | **Throughput - FP32** | **Throughput speedup (FP32 to mixed precision)** | **Weak scaling - FP32** | **Weak scaling - mixed precision** |\n|---|---|-----|------|----|----|----|\n| 1 | 12,8 | 47.06 | 11.11\t| 4.24 | 1.00 | 1.00 |\n| 4 | 12,8 | 165.26\t| 42.84\t| 3.86 | 3.51 | 3.86 |\n| 8 | 12,8 | 330.29\t| 85.91\t| 3.84 | 7.02 | 7.73 |\n\nNote: The respective values for FP32 runs that use a batch size of 12 are not available due to out of memory errors that arise. Batch size of 12 is only available on using FP16.\n\nTo achieve these same results, follow the [Quick Start Guide](#quick-start-guide) outlined above.\n\n##### Fine-tuning training performance for SQuAD v1.1 on NVIDIA DGX A100 (8x A100 80GB)\n\nOur results were obtained by running the `scripts/run_squad.sh` training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-2 with 16x V100 32GB GPUs. Performance (in sequences per second) is the mean throughput from 2 epochs.\n\n| **GPUs** | **Batch size / GPU: mixed precision, TF32** | **Throughput - mixed precision** | **Throughput - FP32** | **Throughput speedup (FP32 to mixed precision)** | **Weak scaling - FP32** | **Weak scaling - mixed precision** |\n|---|---|------|------|----|-----|-----|\n| 1 | 76,38 | 134.22\t| 43.9\t| 3.057 | 1.00 | 1.00 |\n| 8 | 76,38 | 1048.23\t| 341.31\t| 3.071 | 7.81 | 7.77 |\n\nNote: The respective values for TF32 runs that use a batch size of 76 are not available due to out of memory errors that arise. Batch size of 12 is only available on using FP16.\n\nTo achieve these same results, follow the [Quick Start Guide](#quick-start-guide) outlined above.\n\n#### Inference performance results\n\n##### Fine-tuning inference performance for SQuAD v1.1 on NVIDIA DGX-1 V100 (1x V100 16GB)\n\nOur results were obtained by running the `scripts/finetune_inference_benchmark.sh` training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-1 with 1x V100 16GB GPUs. Performance numbers (throughput in sequences per second and latency in milliseconds) were averaged from 1000 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.\n\nBERT-LARGE FP16\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Throughput speedup (FP32 to mixed precision) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|----------------------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 105.04\t| 1.277237354\t| 9.52\t| 9.67\t| 9.77\t| 10.16 |\n| 128 | 2 | 184.9\t | 1.671487977\t| 10.82\t| 11.15\t| 11.27\t| 11.8 |\n| 128 | 4 | 301.9\t | 2.448102498\t| 13.25\t| 13.38\t| 13.45\t| 13.96 |\n| 128 | 8 | 421.98\t| 3.149809659\t| 18.96\t| 19.12\t| 19.2\t| 19.82 |\n| 384 | 1 | 74.99\t | 2.15055922\t| 13.34\t| 13.5\t| 13.58\t| 14.53 |\n| 384 | 2 | 109.84\t| 2.709422792\t| 18.21\t| 18.4\t| 18.6\t| 19.39 |\n| 384 | 4 | 142.58\t| 3.313502208\t| 28.05\t| 28.28\t| 28.48\t| 28.85 |\n| 384 | 8 | 168.34\t| 3.823302294\t| 47.52\t| 47.74\t| 47.86\t| 48.52 |\n\nBERT-Large FP32\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 82.24\t | 12.16\t | 12.28\t| 12.33\t| 12.92 |\n| 128 | 2 | 110.62\t| 18.08\t | 18.22\t| 18.28\t| 18.88 |\n| 128 | 4 | 123.32\t| 32.44\t | 32.72\t| 32.82\t| 32.98 |\n| 128 | 8 | 133.97\t| 59.71\t | 60.29\t| 60.49\t| 60.69 |\n| 384 | 1 | 34.87\t | 28.67\t | 28.92\t| 29.02\t| 29.33 |\n| 384 | 2 | 40.54\t | 49.34\t | 49.74\t| 49.86\t| 50.05 |\n| 384 | 4 | 43.03\t | 92.97\t | 93.59\t| 93.75\t| 94.57 |\n| 384 | 8 | 44.03\t | 181.71\t | 182.34\t| 182.48\t| 183.03 |\n\nBERT-Base FP16\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Throughput speedup (FP32 to mixed precision) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|----------------------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 236.26\t| 1.179589595\t| 4.23\t| 4.37\t| 4.49\t| 4.59 |\n| 128 | 2 | 425.1\t | 1.441554478\t| 4.7\t | 4.84\t| 4.97\t| 5.26 |\n| 128 | 4 | 710.48\t| 1.911691107\t| 5.63\t| 5.78\t| 5.93\t| 6.4 |\n| 128 | 8 | 1081.17\t| 2.523032764\t| 7.4\t | 7.5\t | 7.54\t| 7.73 |\n| 384 | 1 | 190.53\t| 1.757170525\t| 5.25\t| 5.35\t| 5.42\t| 5.8 |\n| 384 | 2 | 289.67\t| 2.248292456\t| 6.9\t | 7.08\t| 7.24\t| 7.57 |\n| 384 | 4 | 404.03\t| 2.946328302\t| 9.9\t | 10\t | 10.03\t| 10.13 |\n| 384 | 8 | 504.24\t| 3.450153951\t| 15.87\t| 15.96\t| 16.01\t| 16.3 |\n\nBERT-Base FP32\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 200.29\t| 4.99\t| 5.08\t| 5.16\t| 5.53 |\n| 128 | 2 | 294.89\t| 6.78\t| 6.89\t| 6.93\t| 7.37 |\n| 128 | 4 | 371.65\t| 10.76\t| 10.89\t| 10.96\t| 11.92 |\n| 128 | 8 | 428.52\t| 18.67\t| 18.89\t| 18.98\t| 19.17 |\n| 384 | 1 | 108.43\t| 9.22\t| 9.26\t| 9.31\t| 10.24 |\n| 384 | 2 | 128.84\t| 15.52\t| 15.6\t| 15.71\t| 16.49 |\n| 384 | 4 | 137.13\t| 29.17\t| 29.4\t| 29.48\t| 29.64 |\n| 384 | 8 | 146.15\t| 54.74\t| 55.19\t| 55.3\t| 55.54 |\n\nTo achieve these same results, follow the [Quick Start Guide](#quick-start-guide) outlined above.\n\n##### Fine-tuning inference performance for SQuAD v1.1 on NVIIDA DGX-1 V100 (1x V100 32GB)\n\nOur results were obtained by running the `scripts/finetune_inference_benchmark.sh` training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-1 with 1x V100 32GB GPUs. Performance numbers (throughput in sequences per second and latency in milliseconds) were averaged from 1000 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.\n\nBERTLarge FP16\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Throughput speedup (FP32 to mixed precision) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|----------------------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 101.58\t| 1.242112986\t| 9.84\t| 9.99\t| 10.06\t| 10.39 |\n| 128 | 2 | 181.89\t| 1.651593571\t| 11\t | 11.14\t| 11.2\t| 11.87 |\n| 128 | 4 | 295.86\t| 2.348840902\t| 13.52\t| 13.67\t| 13.75\t| 14.5 |\n| 128 | 8 | 411.29\t| 3.010246652\t| 19.45\t| 19.62\t| 19.69\t| 20.4 |\n| 384 | 1 | 72.95\t | 2.083690374\t| 13.71\t| 13.93\t| 14.08\t| 14.81 |\n| 384 | 2 | 107.02\t| 2.583775954\t| 18.69\t| 18.8\t| 18.88\t| 19.57 |\n| 384 | 4 | 139.8\t | 3.14652262\t| 28.61\t| 28.75\t| 28.88\t| 29.6 |\n| 384 | 8 | 163.68\t| 3.595782074\t| 48.88\t| 49.09\t| 49.18\t| 49.77 |\n\nBERT-Large FP32\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 81.78\t | 12.23\t| 12.37\t| 12.43\t| 13.2 |\n| 128 | 2 | 110.13\t| 18.16\t| 18.29\t| 18.37\t| 19.27 |\n| 128 | 4 | 125.96\t| 31.76\t| 32.09\t| 32.21\t| 32.42 |\n| 128 | 8 | 136.63\t| 58.55\t| 58.93\t| 59.05\t| 59.4 |\n| 384 | 1 | 35.01\t | 28.56\t| 28.81\t| 28.94\t| 29.16 |\n| 384 | 2 | 41.42\t | 48.29\t| 48.57\t| 48.67\t| 49.02 |\n| 384 | 4 | 44.43\t | 90.03\t| 90.43\t| 90.59\t| 90.89 |\n| 384 | 8 | 45.52\t | 175.76\t| 176.66\t| 176.89\t| 177.33 |\n\nBERT-Base FP16\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Throughput speedup (FP32 to mixed precision) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|----------------------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 234.85\t| 1.217533309\t| 4.26\t| 4.33\t| 4.37\t| 4.62 |\n| 128 | 2 | 415.86\t| 1.435782351\t| 4.81\t| 4.92\t| 5.06\t| 5.55 |\n| 128 | 4 | 680.09\t| 1.84912586\t| 5.88\t| 6.1\t | 6.2\t | 6.53 |\n| 128 | 8 | 1030.03\t| 2.264548752\t| 7.77\t| 7.87\t| 7.95\t| 8.53 |\n| 384 | 1 | 183.18\t| 1.700993593\t| 5.46\t| 5.56\t| 5.61\t| 5.93 |\n| 384 | 2 | 275.77\t| 2.175528558\t| 7.25\t| 7.38\t| 7.44\t| 7.89 |\n| 384 | 4 | 385.61\t| 2.778570399\t| 10.37\t| 10.56\t| 10.63\t| 11.1 |\n| 384 | 8 | 488.45\t| 3.292329469\t| 16.38\t| 16.48\t| 16.52\t| 16.64 |\n\nBERT-Base FP32\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 192.89\t| 5.18\t| 5.3\t | 5.36\t| 5.65 |\n| 128 | 2 | 289.64\t| 6.91\t| 7\t | 7.22\t| 7.83 |\n| 128 | 4 | 367.79\t| 10.88\t| 10.98\t| 11.02\t| 11.59 |\n| 128 | 8 | 454.85\t| 17.59\t| 17.76\t| 17.81\t| 17.92 |\n| 384 | 1 | 107.69\t| 9.29\t| 9.37\t| 9.42\t| 9.88 |\n| 384 | 2 | 126.76\t| 15.78\t| 15.89\t| 15.97\t| 16.72 |\n| 384 | 4 | 138.78\t| 28.82\t| 28.98\t| 29.06\t| 29.88 |\n| 384 | 8 | 148.36\t| 53.92\t| 54.16\t| 54.26\t| 54.58 |\n\nTo achieve these same results, follow the [Quick Start Guide](#quick-start-guide) outlined above.\n\n##### Fine-tuning inference performance for SQuAD v1.1 on NVIDIA DGX A100 (1x A100 80GB)\n\nOur results were obtained by running the `scripts/finetune_inference_benchmark.sh` training script in the TensorFlow 21.02-py3 NGC container on NVIDIA DGX-2 with 1x V100 32GB GPUs. Performance numbers (throughput in sequences per second and latency in milliseconds) were averaged from 1000 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.\n\nBERT-Large FP16\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Throughput speedup (FP32 to mixed precision) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|----------------------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 145.21\t| 0.9435347628\t| 6.89\t| 7.14\t| 7.4\t | 8.35 |\n| 128 | 2 | 272.81\t| 1.093953003\t| 7.33\t| 7.61\t| 7.77\t| 8.35 |\n| 128 | 4 | 468.98\t| 1.273087573\t| 8.53\t| 8.71\t| 8.83\t| 9.85 |\n| 128 | 8 | 705.67\t| 1.191627687\t| 11.34\t| 11.64\t| 11.9\t| 13.1 |\n| 384 | 1 | 118.34\t| 1.042459479\t| 8.45\t| 8.82\t| 8.99\t| 9.52 |\n| 384 | 2 | 197.8\t | 1.231478023\t| 10.11\t| 10.48\t| 10.62\t| 11.4 |\n| 384 | 4 | 275.19\t| 1.268332027\t| 14.54\t| 14.73\t| 14.8\t| 16.8 |\n| 384 | 8 | 342.22\t| 1.416004634\t| 23.38\t| 23.64\t| 23.75\t| 24.1 |\n\nBERT-Large TF32\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 153.9\t | 6.5\t | 6.76\t| 6.86\t| 7.4 |\n| 128 | 2 | 249.38\t| 8.02\t| 8.22\t| 8.34\t| 9.45 |\n| 128 | 4 | 368.38\t| 10.86\t| 11.11\t| 11.24\t| 12.76 |\n| 128 | 8 | 592.19\t| 13.51\t| 13.64\t| 13.77\t| 15.85 |\n| 384 | 1 | 113.52\t| 8.81\t| 9.02\t| 9.16\t| 10.19 |\n| 384 | 2 | 160.62\t| 12.45\t| 12.61\t| 12.68\t| 14.47 |\n| 384 | 4 | 216.97\t| 18.44\t| 18.6\t| 18.7\t| 18.84 |\n| 384 | 8 | 241.68\t| 33.1\t| 33.29\t| 33.36\t| 33.5 |\n\nBERT-Base FP16\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Throughput speedup (FP32 to mixed precision) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|----------------------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 295.01\t| 1.014023992\t| 3.39\t| 3.59\t| 3.65\t| 3.73 |\n| 128 | 2 | 594.81\t| 1.048455898\t| 3.36\t| 3.59\t| 3.68\t| 4.19 |\n| 128 | 4 | 1043.12\t| 1.005145599\t| 3.83\t| 3.97\t| 4.2\t | 4.44 |\n| 128 | 8 | 1786.25\t| 1.198278638\t| 4.48\t| 4.73\t| 4.8\t | 5.19 |\n| 384 | 1 | 278.85\t| 1.103395062\t| 3.59\t| 3.67\t| 3.99\t| 4.15 |\n| 384 | 2 | 464.77\t| 1.252006896\t| 4.3\t | 4.59\t| 4.87\t| 5.29 |\n| 384 | 4 | 675.82\t| 1.264822578\t| 5.92\t| 6.15\t| 6.27\t| 6.94 |\n| 384 | 8 | 846.81\t| 1.31109494\t| 9.45\t| 9.65\t| 9.74\t| 11.03 |\n\nBERT-Base TF32\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 290.93\t| 3.44\t| 3.61\t| 3.73\t| 4.69 |\n| 128 | 2 | 567.32\t| 3.53\t| 3.64\t| 3.96\t| 5.01 |\n| 128 | 4 | 1037.78\t| 3.85\t| 3.95\t| 4.06\t| 4.58 |\n| 128 | 8 | 1490.68\t| 5.37\t| 5.61\t| 5.66\t| 6.19 |\n| 384 | 1 | 252.72\t| 3.96\t| 3.96\t| 4.52\t| 4.66 |\n| 384 | 2 | 371.22\t| 5.39\t| 5.64\t| 5.71\t| 6.38 |\n| 384 | 4 | 534.32\t| 7.49\t| 7.69\t| 7.76\t| 8.56 |\n| 384 | 8 | 645.88\t| 12.39\t| 12.61\t| 12.67\t| 12.77 |\n\nTo achieve these same results, follow the [Quick Start Guide](#quick-start-guide) outlined above.\n\n##### Fine-tuning inference performance for SQuAD v1.1 on NVIDIA Tesla T4 (1x T4 16GB)\n\nOur results were obtained by running the `scripts/finetune_inference_benchmark.sh` training script in the TensorFlow 21.02-py3 NGC container on NVIDIA Tesla T4 with 1x T4 16GB GPUs. Performance numbers (throughput in sequences per second and latency in milliseconds) were averaged from 1000 iterations. Latency is computed as the time taken for a batch to process as they are fed in one after another in the model ie no pipelining.\n\nBERT-Large FP16\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Throughput speedup (FP32 to mixed precision) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|----------------------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 57.6\t | 1.364605544\t| 17.36\t| 18.16\t| 19.02\t| 21.67 |\n| 128 | 2 | 102.76\t| 2.17988969\t| 19.46\t| 20.68\t| 21.27\t| 22.2 |\n| 128 | 4 | 151.11\t| 3.146813828\t| 26.47\t| 26.9\t| 27.06\t| 27.45 |\n| 128 | 8 | 186.99\t| 3.733080455\t| 42.78\t| 43.87\t| 44.18\t| 44.78 |\n| 384 | 1 | 38.88\t | 2.590273151\t| 25.72\t| 26.06\t| 26.16\t| 26.38 |\n| 384 | 2 | 50.53\t | 3.202154626\t| 39.58\t| 39.93\t| 40.35\t| 40.95 |\n| 384 | 4 | 57.69\t | 3.721935484\t| 69.34\t| 70.5\t| 70.77\t| 71.09 |\n| 384 | 8 | 62.99\t | 3.927057357\t| 127\t | 129.18\t| 130.07\t| 131.86 |\n\nBERT-Large FP32\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 42.21\t| 23.69\t| 24.8\t| 25.02\t| 25.48 |\n| 128 | 2 | 47.14\t| 42.42\t| 43.48\t| 43.63\t| 44.32 |\n| 128 | 4 | 48.02\t| 83.29\t| 84.37\t| 84.68\t| 85.14 |\n| 128 | 8 | 50.09\t| 159.72\t| 161.66\t| 161.97\t| 162.52 |\n| 384 | 1 | 15.01\t| 66.63\t| 67.76\t| 68.08\t| 68.66 |\n| 384 | 2 | 15.78\t| 126.78\t| 128.21\t| 128.58\t| 129.08 |\n| 384 | 4 | 15.5\t| 258.1\t| 261.01\t| 261.66\t| 262.55 |\n| 384 | 8 | 16.04\t| 498.61\t| 504.29\t| 504.74\t| 505.55 |\n\nBERT-Base FP16\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Throughput speedup (FP32 to mixed precision) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|----------------------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 116.56\t| 1.039878669\t| 8.58\t| 9.53\t| 10.84\t| 11.74 |\n| 128 | 2 | 238.62\t| 1.675937632\t| 8.38\t| 9.09\t| 9.27\t| 12.33 |\n| 128 | 4 | 402.93\t| 2.440964439\t| 9.93\t| 10.07\t| 10.13\t| 12.17 |\n| 128 | 8 | 532.56\t| 3.052619512\t| 15.02\t| 15.43\t| 15.6\t| 16.52 |\n| 384 | 1 | 102.12\t| 2.035073735\t| 9.79\t| 11.06\t| 11.18\t| 12.07 |\n| 384 | 2 | 149.3\t | 2.910898811\t| 13.4\t| 13.54\t| 13.62\t| 14.36 |\n| 384 | 4 | 177.78\t| 3.563439567\t| 22.5\t| 23.11\t| 23.27\t| 23.59 |\n| 384 | 8 | 192.61\t| 3.752386519\t| 41.53\t| 42.67\t| 42.81\t| 43.31 |\n\nBERT-Base FP32\n\n| Sequence Length | Batch Size | Throughput-Average(seq/sec) | Latency-Average(ms) | Latency-90%(ms) | Latency-95%(ms) | Latency-99%(ms) |\n|-----------------|------------|------------------------------|---------------------|-----------------|-----------------|-----------------|\n| 128 | 1 | 112.09\t| 8.92\t| 9.12\t| 9.49\t| 10.93 |\n| 128 | 2 | 142.38\t| 14.05\t| 14.34\t| 14.48\t| 15.03 |\n| 128 | 4 | 165.07\t| 24.23\t| 24.86\t| 24.92\t| 25.05 |\n| 128 | 8 | 174.46\t| 45.86\t| 46.71\t| 46.8\t| 47.2 |\n| 384 | 1 | 50.18\t | 19.93\t| 20.53\t| 21.04\t| 21.73 |\n| 384 | 2 | 51.29\t | 38.99\t| 39.68\t| 39.93\t| 40.2 |\n| 384 | 4 | 49.89\t | 80.18\t| 81.54\t| 82\t | 82.65 |\n| 384 | 8 | 51.33\t | 155.85\t| 158.11\t| 158.5\t| 159.17 |\n\nTo achieve these same results, follow the [Quick Start Guide](#quick-start-guide) outlined above.","hasPerformance",[],{"_85":240,"_114":241,"_115":-5,"_123":35,"_122":35,"_124":239,"_221":90},{"_85":29,"_114":29,"_115":-5,"_123":35,"_122":35,"_124":239,"_221":90},{"_85":113,"_114":113,"_115":-5,"_123":35,"_122":35,"_124":238,"_221":90},[],[],"-","Latest (21.02.1)","nvidia/dle/bert_tf2","key","nvidia/dle/resources/bert_tf2","resourceId","bert_tf2","shortDescription","BERT is a method of pre-training language representations which obtains state-of-the-art results on a wide array of NLP tasks.","This resource is using open-source code maintained in github (see the quick-start-guide section) and available for download from [NGC](https://catalog.ngc.nvidia.com/orgs/nvidia/teams/dle/resources/bert-tf2)\n\n\nBERT, or Bidirectional Encoder Representations from Transformers, is a new method of pre-training language representations which obtains state-of-the-art results on a wide array of Natural Language Processing (NLP) tasks. This model is based on the [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) paper. NVIDIA's BERT is an optimized version of [Google's official implementation](https://github.com/google-research/bert), leveraging mixed precision arithmetic and Tensor Cores on V100 GPUs for faster training times while maintaining target accuracy.\n\nOther publicly available implementations of BERT include:\n[NVIDIA PyTorch](https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/LanguageModeling/BERT)\n[Hugging Face](https://github.com/huggingface/pytorch-pretrained-BERT)\n[codertimo](https://github.com/codertimo/BERT-pytorch)\n[gluon-nlp](https://github.com/dmlc/gluon-nlp/tree/master/scripts/bert)\n[Google's official implementation](https://github.com/google-research/bert)\n\nThis model is trained with mixed precision using Tensor Cores on Volta, Turing, and the NVIDIA Ampere GPU architectures. Therefore, researchers can get results up to 4x 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.\n\n### Model architecture\n\nBERT's model architecture is a multi-layer bidirectional transformer encoder. Based on the model size, we have the following two default configurations of BERT:\n\n| **Model** | **Hidden layers** | **Hidden unit size** | **Attention heads** | **Feedforward filter size** | **Max sequence length** | **Parameters** |\n|:---------:|:----------:|:----:|:---:|:--------:|:---:|:----:|\n|BERTBASE |12 encoder| 768| 12|4 x 768|512|110M|\n|BERTLARGE|24 encoder|1024| 16|4 x 1024|512|330M|\n\nBERT training consists of two steps, pre-training the language model in an unsupervised fashion on vast amounts of unannotated datasets, and then using this pre-trained model for fine-tuning for various NLP tasks, such as question and answer, sentence classification, or sentiment analysis. Fine-tuning typically adds an extra layer or two for the specific task and further trains the model using a task-specific annotated dataset, starting from the pre-trained backbone weights. The end-to-end process is depicted in the following image:\n\n![](https://raw.githubusercontent.com/NVIDIA/DeepLearningExamples/master/TensorFlow2/LanguageModeling/BERT//data/images/bert_pipeline.png?raw=true)\n\n\nFigure 1: BERT Pipeline\n\n### Default configuration\n\nThis repository contains scripts to interactively launch data download, training, benchmarking, and inference routines in a Docker container for fine tuning Question Answering. The major differences between the official implementation of the paper and our version of BERT are as follows:\n\n- Mixed precision support with TensorFlow Automatic Mixed Precision (TF-AMP), which enables mixed precision training without any changes to the code-base by performing automatic graph rewrites and loss scaling controlled by an environmental variable.\n- Scripts to download dataset and pretrained checkpoints for:\n - Pre-training - [Wikipedia](https://dumps.wikimedia.org/), [BookCorpus](http://yknzhu.wixsite.com/mbweb)\n - Fine tuning - [SQuAD](https://rajpurkar.github.io/SQuAD-explorer/) (Stanford Question Answering Dataset)\n - Pretrained weights from Google\n- Custom fused CUDA kernels for faster computations\n- Multi-GPU/Multi-node support using Horovod\n\nThe following performance optimizations were implemented in this model:\n- [XLA](https://www.tensorflow.org/xla) support (experimental).\n\nThese techniques and optimizations improve model performance and reduce training time, allowing you to perform various NLP tasks with no additional effort.\n\n### Feature support matrix\n\nThe following features are supported by this model.\n\n| **Feature** | **BERT** |\n|:-----------------------:|:--------------------------:|\n| Horovod Multi-GPU | Yes |\n| Horovod Multi-node | Yes |\n| Automatic mixed precision (AMP) | Yes |\n| XLA | Yes |\n| LAMB | Yes |\n\n#### Features\n\n**Multi-GPU training with Horovod**\nOur model uses Horovod to implement efficient multi-GPU training with NCCL. For details, see example sources in this repository or see the [TensorFlow tutorial](https://github.com/horovod/horovod/#usage).\n\n**XLA support**\n\nXLA is a domain-specific compiler for linear algebra that can accelerate TensorFlow models with potentially no source code changes. The results are improvements in speed and memory usage: most internal benchmarks run ~1.1-1.5x faster after XLA is enabled.\n\n**Multi-node Training**\n\nSupported on a Pyxis/Enroot Slurm cluster.\n\n**LAMB**\n\n[LAMB](https://arxiv.org/pdf/1904.00962.pdf) stands for Layerwise Adaptive Moments based optimizer, is a large batch optimization technique that helps accelerate training of deep neural networks using large minibatches. It allows using a global batch size of 65536 and 32768 on sequence lengths 128 and 512 respectively, compared to a batch size of 256 for Adam. The optimized implementation accumulates 1024 gradient batches in phase 1 and 4096 steps in phase 2 before updating weights once. This results in 27% training speedup on a single DGX2 node. On multi-node systems, LAMB allows scaling up to 1024 GPUs resulting in training speedups of up to 17x in comparison to [Adam](https://arxiv.org/pdf/1412.6980.pdf). Adam has limitations on the learning rate that can be used since it is applied globally on all parameters whereas LAMB follows a layerwise learning rate strategy.\n\nNVLAMB adds necessary tweaks to [LAMB version 1](https://arxiv.org/abs/1904.00962v1), to ensure correct convergence. A guide to implementating the LAMB optimizer can be found in our [article](https://medium.com/@NvidiaAI/a-guide-to-optimizer-implementation-for-bert-at-scale-8338cc7f45fd) on Medium.com. The algorithm is as follows:\n ![NVLAMB](https://raw.githubusercontent.com/NVIDIA/DeepLearningExamples/master/TensorFlow2/LanguageModeling/BERT//data/images/images_nvlamb.png)\n\n\n### Mixed precision training\n\nMixed precision is the combined use of different numerical precision in a computational method. [Mixed precision](https://arxiv.org/abs/1710.03740) 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](https://developer.nvidia.com/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 requires two steps:\n1. Porting the model to use the FP16 data type where appropriate.\n2. Adding loss scaling to preserve small gradient values.\n\nThis can now be achieved using Automatic Mixed Precision (AMP) for TensorFlow to enable the full [mixed precision methodology](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html#tensorflow) 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.\n\nIn 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.\n\nFor information about:\n- How to train using mixed precision, see the [Mixed Precision Training](https://arxiv.org/abs/1710.03740) paper and [Training With Mixed Precision](https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html) documentation.\n- Techniques used for mixed precision training, see the [Mixed Precision Training of Deep Neural Networks](https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/) blog.\n- How to access and enable AMP for TensorFlow, see [Using TF-AMP](https://docs.nvidia.com/deeplearning/dgx/tensorflow-user-guide/index.html#tfamp) from the TensorFlow User Guide.\n\n#### Enabling mixed precision\n\nThis implementation exploits the TensorFlow Automatic Mixed Precision feature. To enable AMP, you simply need to supply the `--use_fp16` flag to the `run_pretraining.py` or `run_squad.py` script. For reference, enabling AMP required us to apply the following changes to the code:\n\n1. Set the Keras mixed precision policy:\n ```python\n if FLAGS.use_fp16:\n policy = tf.keras.mixed_precision.experimental.Policy(\"mixed_float16\")\n tf.keras.mixed_precision.experimental.set_policy(policy)\n ```\n\n2. Use the loss scaling wrapper on the optimizer:\n ```python\n if FLAGS.use_fp16:\n squad_model.optimizer = tf.keras.mixed_precision.LossScaleOptimizer(squad_model.optimizer,\n dynamic=True)\n ```\n\n3. Use scaled loss to calculate the gradients:\n ```python\n if use_float16:\n scaled_loss = optimizer.get_scaled_loss(loss)\n scaled_grads = tape.gradient(scaled_loss, training_vars)\n grads = optimizer.get_unscaled_gradients(scaled_grads)\n ```\n\n#### Enabling TF32\n\nTensorFloat-32 (TF32) is the new math mode in [NVIDIA A100](https://www.nvidia.com/en-us/data-center/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.\n\nTF32 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.\n\nFor more information, refer to the [TensorFloat-32 in the A100 GPU Accelerates AI Training, HPC up to 20x](https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/) blog post.\n\nTF32 is supported in the NVIDIA Ampere GPU architecture and is enabled by default.\n\n### Glossary\n\n**Fine-tuning**\nTraining an already pretrained model further using a task specific dataset for subject-specific refinements, by adding task-specific layers on top if required.\n\n**Language model**\nAssigns a probability distribution over a sequence of words. Given a sequence of words, it assigns a probability to the whole sequence.\n\n**Pre-training**\nTraining a model on vast amounts of data on the same (or different) task to build general understandings.\n\n**Transformer**\nThe paper [Attention Is All You Need](https://arxiv.org/abs/1706.03762) introduces a novel architecture called Transformer that uses an attention mechanism and transforms one sequence into another.","displayName","BERT for TensorFlow2",["D",1680631952072],"publisher","NVIDIA Deep Learning Examples","latestVersion","collectionCounts","imageUrl","/api/image/nvidia/dle/resources/bert_tf2","rawImageUrl","https://assets.nvidiagrid.net/ngc/logos/Deep-Learning-Examples.png","canGuestDownload","latestVersionSizeInBytes","licenseTerms","isMultinodeEnabled","isMultiArchitecture","bias","explainability","privacy","safetyAndSecurity","featureLabels",[],"displayLabels",[281],"policyLabels",[],"systemLabels",[],"products",[],"hasNimMetadata",{"_250":282,"_283":284,"_285":286,"_287":87},"Natural Language Understanding","value","uscs_natural_language_understanding","filter",{"_288":289,"_283":284},"category","field","application",{"_156":1269,"_158":1270},"ngcApps",[1222],"csaApiBaseUrl","https://www.nvidia.com/ai-agents/csa-agent/dcsm","datadogRum",{"_1203":1204,"_1205":1206,"_1207":1208,"_1209":1210,"_1211":301,"_109":327,"_1212":1213,"_1214":60,"_1215":90,"_1216":90,"_1217":1218,"_1219":90,"_1220":1221},"amplitude",{"_1200":1201,"_841":-7,"_1202":35},"disableJs","environment","prod","featureFlagsBlob","JTdCJTIydGVzdEZsYWclMjIlM0FmYWxzZSUyQyUyMnNob3dTZWFyY2hMYWJlbFNldEZpbHRlcnMlMjIlM0FmYWxzZSUyQyUyMmVuYWJsZUNzYVdpZGdldCUyMiUzQWZhbHNlJTJDJTIyZW5hYmxlUHJvZHVjdEdyb3VwaW5nQnlTdWJzY3JpcHRpb24lMjIlM0FmYWxzZSUyQyUyMmNvbnRhaW5lclNjYW5uaW5nUmVkZXNpZ25NMiUyMiUzQWZhbHNlJTdE","featureFlagsGuarded","isDevLikeEnv","isPlaywright","isMobileDevice","categorizedOrgs",{"_1194":1195,"_1196":1197,"_1198":1199},"currentOrgContext","currentContextRoles",[],"enterpriseOrgs",[],"manageableExternalUserGroups",[],"ngcApiOrigin","ngcHelmOrigin","origin","https://catalog.ngc.nvidia.com","usesMocks","theme","year",2026,"now",1787460896436,"1.12.0","defaultSearchSuggestions",{"_1081":409,"_1082":1083,"_1084":1085,"_1086":1087},"announcements",[],"consentState",{"_366":1071,"_1072":90,"_1073":35,"_1074":35,"_1075":1076},"consentConfig",{"_348":349,"_350":351,"_352":353,"_354":355,"_356":357,"_358":359,"_360":361,"_362":363,"_364":365,"_366":367},"showContactAdmin",1787460896437,[],[],[341,342,343,344,345,346],{"_49":63,"_51":73,"_53":151,"_55":-5,"_57":-5,"_59":60},{"_49":63,"_51":70,"_53":151,"_55":-5,"_57":-5,"_59":60},{"_49":63,"_51":68,"_53":151,"_55":-5,"_57":-5,"_59":60},{"_49":63,"_51":64,"_53":151,"_55":-5,"_57":-5,"_59":60},{"_49":61,"_51":62,"_53":151,"_55":-5,"_57":-5,"_59":60},{"_49":50,"_51":52,"_53":54,"_55":56,"_57":347,"_59":60},4,"mainInfoText","We and our third-party partners (including social media, advertising, and analytics partners) use cookies and other tracking technologies to collect, store, monitor, and process certain information about you when you visit our website. The information collected might relate to you, your preferences, or your device. We use that information to make the site work, analyze performance and traffic on our website, provide a more personalized web experience, and assist in our marketing efforts.

Under certain privacy laws, you have the right to direct us not to \"sell\" or \"share\" your personal information for targeted advertising. To opt-out of the \"sale\" and \"sharing\" of personal information through cookies, you must opt-out of optional cookies using the toggles below. To opt out of the \"sale\" and \"sharing\" of data collected by other means (e.g., online forms) you must also update your data sharing preferences through the NVIDIA Preference Center.

Click on the different category headings below to find out more and change the settings according to your preference. You cannot opt out of Required Cookies as they are deployed to ensure the proper functioning of our website (such as prompting the cookie banner and remembering your settings, etc.). By clicking \"Save and Accept\" or \"Decline All\" at the bottom, you consent to the use of cookies and other tools as described in our Cookie Policy in accordance with your settings and accept our Terms of Service (which contains important waivers). For more information about our privacy practices, please see our Privacy Policy.","mainText","Cookie Settings","alertNoticeText",{"_1063":1064,"_1065":1066,"_1067":1068,"_1069":1070},"bannerAcceptLabel","Accept All","bannerRejectLabel","Reject Optional","bannerSettingsLabel","Manage Settings","preferenceSaveLabel","Save Choices","preferenceRejectLabel","Decline All","alwaysActiveText","Always Active","groups",[368,369,370,371,372],{"_373":753,"_114":754,"_215":755,"_377":90,"_378":18,"_380":756,"_382":757},{"_373":656,"_114":657,"_215":658,"_377":35,"_378":58,"_380":659,"_382":660},{"_373":564,"_114":565,"_215":566,"_377":35,"_378":347,"_380":567,"_382":568},{"_373":384,"_114":385,"_215":386,"_377":35,"_378":387,"_380":388,"_382":389},{"_373":374,"_114":375,"_215":376,"_377":35,"_378":379,"_380":381,"_382":383},"optanonGroupId","C0005","Social Media Cookies","These cookies are set by a range of social media services that we have added to the site to enable you to share our content with your friends and networks. They are capable of tracking your browser across other sites and building up a profile of your interests. This may impact the content and messages you see on other websites you visit. If you do not allow these cookies you may not be able to use or see these sharing tools.","alwaysActive","order",11,"cookies",[],"hosts",[],"C0004","Advertising Cookies","These cookies record your visit to our websites, the pages you have visited and the links you have followed to influence the advertisements that you see on other websites. These cookies and the information they collect may be managed by other companies, including our advertising partners, and may be used to build a profile of your interests and show you relevant advertising on other sites. We and our advertising partners will use this information to make our websites and the advertising displayed on it, more relevant to your interests.",5,[506,507,508,509,510,511,512,513,514,515,516,517,518,519,520,521,522,523,524,525,526,527],[390,391,392,393,394,395,396,397,398,399,400,401,402,403,404,405,406],{"_407":501,"_250":501,"_215":409,"_380":502},{"_407":497,"_250":497,"_215":409,"_380":498},{"_407":492,"_250":492,"_215":409,"_380":493},{"_407":488,"_250":488,"_215":409,"_380":489},{"_407":483,"_250":483,"_215":409,"_380":484},{"_407":478,"_250":478,"_215":409,"_380":479},{"_407":470,"_250":470,"_215":409,"_380":471},{"_407":462,"_250":462,"_215":409,"_380":463},{"_407":457,"_250":457,"_215":409,"_380":458},{"_407":452,"_250":452,"_215":409,"_380":453},{"_407":447,"_250":447,"_215":409,"_380":448},{"_407":442,"_250":442,"_215":409,"_380":443},{"_407":437,"_250":437,"_215":409,"_380":438},{"_407":430,"_250":430,"_215":409,"_380":431},{"_407":425,"_250":425,"_215":409,"_380":426},{"_407":418,"_250":418,"_215":409,"_380":419},{"_407":408,"_250":408,"_215":409,"_380":410},"hostName","amazon-adsystem.com","",[411,412],{"_114":417,"_414":408,"_415":35,"_215":416},{"_114":413,"_414":408,"_415":35,"_215":416},"ad-privacy","host","isSession","This domain is owned by online retailer Amazon and is used as part of its affiliate marketing programme.","ad-id","doubleclick.net",[420,421],{"_114":424,"_414":418,"_415":35,"_215":423},{"_114":422,"_414":418,"_415":35,"_215":423},"IDE","This domain is owned by Doubleclick (Google). The main business activity is: Doubleclick is Googles real time bidding advertising exchange","test_cookie","bing.com",[427],{"_114":428,"_414":425,"_415":35,"_215":429},"MUID","This domain is owned by Mircosoft - it is the site for the search engine Bing.","assets.adobedtm.com",[432,433],{"_114":436,"_414":430,"_415":35,"_215":435},{"_114":434,"_414":430,"_415":35,"_215":435},"kndctr_F207D74D549850760A4C98C6_AdobeOrg_*","If you know anything about this host and its uses please add your contributions here. ","kndctr_F207D74D549850760A4C98C6_AdobeOrg_consent","tiktok.com",[439],{"_114":440,"_414":437,"_415":35,"_215":441},"_ttp","TikTok is a social media platform owned by ByteDance, known for short-form video sharing. The service is used to create and share videos, and it utilizes cookies for various purposes including advertising and analytics.","t.co",[444],{"_114":445,"_414":442,"_415":35,"_215":446},"muc_ads","This domain is owned by Twitter. It is used as a URL shortening service, which allows Twitter to collect data on how links are shared.","snowplow.com",[449],{"_114":450,"_414":447,"_415":35,"_215":451},"sp","Testing","google.com",[454],{"_114":455,"_414":452,"_415":35,"_215":456},"NID","This domain is owned by Google Inc. Although Google is primarily known as a search engine, the company provides a diverse range of products and services. Its main source of revenue however is advertising. Google tracks users extensively both through its own products and sites, and the numerous technologies embedded into many millions of websites around the world. It uses the data gathered from most of these services to profile the interests of web users and sell advertising space to organisations based on such interest profiles as well as aligning adverts to the content on the pages where its customer's adverts appear.","linkedin.com",[459],{"_114":460,"_414":457,"_415":35,"_215":461},"bcookie","This domain is owned by LinkedIn, the business networking platform. It typically acts as a third party host where website owners have placed one of its content sharing buttons in their pages, although its content and services can be embedded in other ways. Although such buttons add functionality to the website they are on, cookies are set regardless of whether or not the visitor has an active Linkedin profile, or agreed to their terms and conditions. For this reason it is classified as a primarily tracking/targeting domain.","twitter.com",[464,465],{"_114":468,"_414":462,"_415":35,"_215":469},{"_114":466,"_414":462,"_415":35,"_215":467},"personalization_id","This domain is owned by Twitter. The main business activity is: Social Networking Services. Where twitter acts as a third party host, it collects data through a range of plug-ins and integrations, that is primarily used for tracking and targeting.","guest_*","This domain is owned by X (formerly Twitter). The main business activity is: Social Networking Services. Where X (formerly Twitter) acts as a third party host, it collects data through a range of plug-ins and integrations, that is primarily used for tracking and targeting.","youtube.com",[472,473],{"_114":476,"_414":470,"_415":35,"_215":477},{"_114":474,"_414":470,"_415":35,"_215":475},"VISITOR_INFO1_LIVE","This cookie is used as a unique identifier to track viewing of videos","__Secure-*","YouTube is a Google owned platform for hosting and sharing videos. YouTube collects user data through videos embedded in websites, which is aggregated with profile data from other Google services in order to display targeted advertising to web visitors across a broad range of their own and other websites.","demdex.net",[480],{"_114":481,"_414":478,"_415":35,"_215":482},"demdex","This cookie helps Adobe Audience Manger perform basic functions such as visitor identification, ID synchronization, segmentation, modeling, reporting, etc.","im-apps.net",[485],{"_114":486,"_414":483,"_415":35,"_215":487},"receive-cookie-deprecation","This domain is owned by im-apps.net which provides information and resources about various instant messaging (IM) applications to help users make informed decisions based on their communication needs.","influ2.com",[490],{"_114":491,"_414":488,"_415":35,"_215":451},"R","6sc.co",[494],{"_114":495,"_414":492,"_415":35,"_215":496},"6suuid","This domain is associated with the company 6Sense Insights, Inc., which provides predictive analytics and account engagement services. It is primarily used for tracking and targeting in marketing automation and advertising.","hushly.com",[499],{"_114":500,"_414":497,"_415":35,"_215":451},"_hly_gvid","t.contentsquare.net",[503],{"_114":504,"_414":501,"_415":90,"_215":505},"_cs_*","This is a session cookie. This is a pattern type cookie where a random 13-digit number is appended to the prefix _cs_.",{"_114":562,"_414":529,"_415":35,"_215":563},{"_114":560,"_414":529,"_415":35,"_215":561},{"_114":558,"_414":529,"_415":35,"_215":559},{"_114":440,"_414":529,"_415":35,"_215":557},{"_114":534,"_414":529,"_415":90,"_215":536},{"_114":556,"_414":529,"_415":35,"_215":409},{"_114":555,"_414":532,"_415":35,"_215":409},{"_114":553,"_414":529,"_415":35,"_215":554},{"_114":551,"_414":529,"_415":35,"_215":552},{"_114":550,"_414":529,"_415":35,"_215":409},{"_114":548,"_414":529,"_415":35,"_215":549},{"_114":534,"_414":547,"_415":90,"_215":536},{"_114":545,"_414":532,"_415":35,"_215":546},{"_114":545,"_414":535,"_415":35,"_215":546},{"_114":504,"_414":529,"_415":35,"_215":505},{"_114":543,"_414":529,"_415":35,"_215":544},{"_114":450,"_414":541,"_415":35,"_215":542},{"_114":539,"_414":529,"_415":35,"_215":540},{"_114":537,"_414":529,"_415":35,"_215":538},{"_114":534,"_414":535,"_415":90,"_215":536},{"_114":531,"_414":532,"_415":35,"_215":533},{"_114":528,"_414":529,"_415":90,"_215":530},"AMCVS_xxxxxAdobeOrg","nvidia.com","This is a pattern type cookie name associated with Adobe Marketing Cloud. It stores a unique visitor identifier, and uses an organisation identifier to allow a company to track users across their domains and services.This is a pattern type cookie name associated with Adobe Marketing Cloud. It stores a unique visitor identifier, and uses an organisation identifier.","_rdt_pn","www.nvidia.com","This cookie name is associated with www.mind.org.uk which is used for marketing purposes.","AMP_*","build.nvidia.com","Supports Amplitude product analytics.","_gcl_au","Used by Google AdSense for experimenting with advertisement efficiency across websites using their services","_mkto_trk","This cookie is associated with an email marketing service provided by Marketo. This tracking cookie allows a website to link visitor behaviour to the recipient of an email marketing campaign, to measure campaign effectiveness.","investor.nvidia.com","This cookie is associated with Eventbrite and is used to deliver content tailored to the end user's interests and improve content creation. This cookie is also used for event-booking purposes.","_gat_gtag_UA_140825383_2","Google Analytics","_rdt_uuid","This cookie is set by Reddit and is used for tracking the views of embedded videos and performance of the advertisements.","catalog.ngc.nvidia.com","_uetsid","This cookie is used by Bing to determine what ads should be shown that may be relevant to the end user perusing the site. ","kndctr_F207D74D549850760A4C98C6_AdobeOrg_bc_session_id","ttcsid_XXXXXXXXXXXXXXXXXXXXCategory","This cookie name is associated with TikTok which is used to track your activity on the platform, including how you interact with their embedded services on other websites.This cookie is associated with TikTok. It helps track your activity on the platform, including how you interact with their embedded services on other websites.","_fbp","Used by Facebook to deliver a series of advertisement products such as real time bidding from third party advertisers","_zitok","_twpid","This cookie is typically associated with the use of Taptap Digital, a company specializing in digital advertising. It is used for tracking and ad personalization purposes.","_tt_enable_cookie","This cookie name is associated with TikTok which is used to collect data about behavior and purchases on the website and to measure the effect of advertising. It is also used to track website activity to help optimize advertising efforts.","AMCV_xxxxxAdobeOrg","This cookie is used to identify a unique visitor for tracking purposes by Adobe Experience Cloud.This is a pattern type cookie name associated with Adobe Marketing Cloud. It stores a unique visitor identifier, and uses an organisation identifier to allow a company to track users across their domains and services.","ttcsid","This cookie name is associated with TikTok which is used to enable TikTok Ads session tracking by assigning a unique session ID to users who click on TikTok ads, allowing for the attribution of user behavior and conversions within the same session.This cookie is associated with TikTok. It helps track your activity on the platform, including how you interact with their embedded services on other websites.","C0003","Personalization Cookies","These cookies collect data about how you have interacted with our website to help us improve your web experience, such as which pages you have visited. These cookies may store a unique ID so that our system will remember you when you return. They may be set by us or by third party providers whose services we have added to our pages. These cookies enable us to provide enhanced website functionality and personalization as well as make the marketing messages we send to you more relevant to your interests. If you do not allow these cookies, then some or all of these services may not function properly.",[598,599,600,601,602,603,604,605,606,607,608,609,610,611,612,613,614,615,616,617,618,619,620,621,622],[569,570,571,572,573],{"_407":483,"_250":483,"_215":409,"_380":593},{"_407":457,"_250":457,"_215":409,"_380":588},{"_407":583,"_250":583,"_215":409,"_380":584},{"_407":578,"_250":578,"_215":409,"_380":579},{"_407":574,"_250":574,"_215":409,"_380":575},"nvidia.partners",[576],{"_114":577,"_414":574,"_415":35,"_215":451},"edge","app.hushly.com",[580],{"_114":581,"_414":578,"_415":35,"_215":582},"*CORS","This domain is associated with Hushly, a platform that provides lead conversion solutions for content marketing by using personalization and content engagement strategies.","github.com",[585],{"_114":586,"_414":583,"_415":35,"_215":587},"_octo","This domain is owned by Github. The main business activity is: Source code repository and collaboration platform",[589,590],{"_114":592,"_414":457,"_415":35,"_215":461},{"_114":591,"_414":457,"_415":35,"_215":461},"lidc","li_gc",[594,595],{"_114":597,"_414":483,"_415":35,"_215":487},{"_114":596,"_414":483,"_415":35,"_215":487},"imid_*","imid",{"_114":623,"_414":532,"_415":35,"_215":625},{"_114":654,"_414":529,"_415":35,"_215":655},{"_114":623,"_414":653,"_415":35,"_215":625},{"_114":651,"_414":529,"_415":90,"_215":652},{"_114":649,"_414":529,"_415":35,"_215":650},{"_114":648,"_414":631,"_415":90,"_215":409},{"_114":623,"_414":627,"_415":35,"_215":625},{"_114":646,"_414":529,"_415":35,"_215":647},{"_114":623,"_414":535,"_415":35,"_215":625},{"_114":634,"_414":532,"_415":35,"_215":409},{"_114":623,"_414":645,"_415":35,"_215":625},{"_114":643,"_414":529,"_415":35,"_215":644},{"_114":640,"_414":641,"_415":35,"_215":642},{"_114":638,"_414":639,"_415":35,"_215":409},{"_114":623,"_414":635,"_415":35,"_215":625},{"_114":637,"_414":529,"_415":35,"_215":409},{"_114":623,"_414":636,"_415":35,"_215":625},{"_114":634,"_414":635,"_415":35,"_215":409},{"_114":633,"_414":631,"_415":90,"_215":409},{"_114":623,"_414":632,"_415":35,"_215":625},{"_114":623,"_414":631,"_415":35,"_215":625},{"_114":623,"_414":630,"_415":35,"_215":625},{"_114":629,"_414":529,"_415":90,"_215":625},{"_114":626,"_414":627,"_415":90,"_215":628},{"_114":623,"_414":624,"_415":35,"_215":625},"_gd_*","info.nvidia.com","This cookie contains no identifying information and is used for diagnostic purposes by Impact Radius, a service we use to track signup commissions for our affiliate program. We have categorized it under Functionality because it is necessary for the operation of our affiliate program.","nv_country_code","developer.nvidia.com","This cookie is likely used to store the country code of the visitor to provide localized content and possibly for legal compliance with regional data privacy laws.","_gdxxxxxxxxxxxxx","nvidianews.nvidia.com","docs.api.nvidia.com","research.nvidia.com","readme_library","wcmmode","ai-podcast.nvidia.com","resources.nvidia.com","geo_locator","openedx-language-preference","learn.nvidia.com","doxygen_width","raytracing-docs.nvidia.com","This cookie is likely used to remember the width preferences of a sidebar or interface element within a documentation platform like Doxygen.","_sfid_eb78","This cookie is categorized as Functionality Cookie.","crowdsource.nvidia.com","mbox","Cookie set by Adobe SiteCatalyst analytics software - used for measuring the performance of page content using A/B split testing","readme_language","vid","This cookie name 'vid' is used for tracking the search request made by the user on the 'Find a practitioner' page.","at_check","Used by Adobe Target to check if cookies are enabled/supported on the browser","blogs.nvidia.com","nvHeadless","This cookie is likely used to track if a user is using the site in a 'headless' non-GUI mode, which could be for accessibility reasons or for automated processes.","C0002","Performance Cookies","These cookies are used to provide quantitative measures of our website visitors, such as the number of times you visit, time on page, your mouse movements, scrolling, clicks and keystroke activity on the websites; other browsing, search, or product research behavior; and what brought you to our site. These cookies may store a unique ID so that our system will remember you when you return. Information collected with these cookies is used to measure and find ways to improve website performance.",[681,682,683,684,685,686,687,688,689,690,691,692,693,694,695,696,697,698,699,700,701,702,703,704,705,706,707,708,709],[661,662,663,664],{"_407":676,"_250":676,"_215":409,"_380":677},{"_407":671,"_250":671,"_215":409,"_380":672},{"_407":470,"_250":470,"_215":409,"_380":668},{"_407":583,"_250":583,"_215":409,"_380":665},[666],{"_114":667,"_414":583,"_415":90,"_215":587},"_gh_sess",[669],{"_114":670,"_414":470,"_415":90,"_215":477},"YSC","bizible.com",[673],{"_114":674,"_414":671,"_415":35,"_215":675},"_BUID","Bizible is a marketing attribution software that tracks the performance of marketing efforts and connects them to revenue, primarily used for B2B environments.","a.usbrowserspeed.com",[678],{"_114":679,"_414":676,"_415":35,"_215":680},"tuid","This domain is owned by Ookla which provides internet speed testing services through its Speedtest application.",{"_114":751,"_414":529,"_415":35,"_215":752},{"_114":736,"_414":750,"_415":90,"_215":737},{"_114":719,"_414":532,"_415":35,"_215":644},{"_114":719,"_414":632,"_415":35,"_215":644},{"_114":748,"_414":529,"_415":35,"_215":749},{"_114":736,"_414":535,"_415":90,"_215":737},{"_114":746,"_414":529,"_415":35,"_215":747},{"_114":736,"_414":532,"_415":90,"_215":737},{"_114":744,"_414":631,"_415":35,"_215":745},{"_114":736,"_414":653,"_415":90,"_215":737},{"_114":716,"_414":532,"_415":35,"_215":718},{"_114":742,"_414":529,"_415":90,"_215":743},{"_114":740,"_414":529,"_415":35,"_215":741},{"_114":738,"_414":529,"_415":90,"_215":739},{"_114":736,"_414":627,"_415":90,"_215":737},{"_114":734,"_414":529,"_415":35,"_215":735},{"_114":732,"_414":529,"_415":35,"_215":733},{"_114":730,"_414":529,"_415":90,"_215":731},{"_114":728,"_414":636,"_415":35,"_215":729},{"_114":716,"_414":627,"_415":35,"_215":718},{"_114":726,"_414":529,"_415":35,"_215":727},{"_114":724,"_414":529,"_415":90,"_215":725},{"_114":722,"_414":529,"_415":35,"_215":723},{"_114":720,"_414":529,"_415":35,"_215":721},{"_114":719,"_414":529,"_415":35,"_215":644},{"_114":716,"_414":717,"_415":35,"_215":718},{"_114":714,"_414":529,"_415":35,"_215":715},{"_114":712,"_414":529,"_415":90,"_215":713},{"_114":710,"_414":529,"_415":35,"_215":711},"_ga_xxxxxxxxxx","Used by Google Analytics to identify and track an individual session with your device.","s_ppvl","This cookie is set by Adobe Analytics to store information on the percentage of the page displayed to the end user.","coveo_visitorId","This cookie name is associated with Coveo which is used to store a unique user ID for marketing purposes and to track the number of distinct users who submitted a query or clicked an item, counting them only once even if they did the same query or clicked the same item over the course of many visits.","_dd_s","marketplace.nvidia.com","Datadog Cookie used to group all events generated from a unique user session across multiple pages. It contains the current session ID, whether the session is excluded due to sampling, and the expiration date of the session. The cookie is extended for an extra 15 minutes every time the user interacts with the website, up to the maximum user session duration (4 hours).","nvweb_A","_ga","This cookie name is associated with Google Universal Analytics - which is a significant update to Google's more commonly used analytics service. This cookie is used to distinguish unique users by assigning a randomly generated number as a client identifier. It is included in each page request in a site and used to calculate visitor, session and campaign data for the sites analytics reports. By default it is set to expire after 2 years, although this is customisable by website owners.","_uetvid","This is a cookie utilised by Microsoft Bing Ads and is a tracking cookie. It allows us to engage with a user that has previously visited our website.","tp","This cookie is associated with Adobe Analytics site tracking. It registers if the website's scroll-depth detection is active. This function remembers how far the user has scrolled on the current session across the website's sub-pages.","_hly_vid","This cookie _hly_vid is used to record website and server performance with each visit to the site. The data is aggregated and analyzed; therefore, it remains anonymous upon use.","_session_id","This cookie creates an interim session ID used as an in-session user ID.\n","s_cc","This cookie is associated with the Adobe Site Catalyst. It determines whether cookies are enabled in the web browser.","s_ecid","Adobe Analytics cookie containing the the Adobe Experience Cloud 1st party visitor id if the Analytics AMCV visitor id is not found.","_hjSessionUser_xxxxxx","Hotjar cookie that is set when a user first lands on a page with the Hotjar script. It is used to persist the Hotjar User ID, unique to that site on the browser. This ensures that behavior in subsequent visits to the same site will be attributed to the same user ID.","_hjHasCachedUserAttributes","This cookie name is associated with Hotjar which is used to know whether the data set in _hjUserAttributes Local Storage item is up to date or not.","s_ppv","This cookie is associated with Adobe Analytics site tracking and is used to measure a user's scroll activity to see how much of a page they view before moving to another page.","_gid","This cookie name is associated with Google Universal Analytics. This appears to be a new cookie and as of Spring 2017 no information is available from Google. It appears to store and update a unique value for each page visited.","_hjTLDTest","When the Hotjar script executes we try to determine the most generic cookie path we should use, instead of the page hostname. This is done so that cookies can be shared across subdomains (where applicable). To determine this, we try to store the _hjTLDTest cookie for different URL substring alternatives until it fails. After this check, the cookie is removed.","ekfls-2025-03-27","This cookie name is associated with readme.io which is used to help maintain data for the API reference support page.","s_getNewRepeat","This cookie is used to determine whether a visitor is a new visitor or a repeat visitor to the site, and is used for analytical purposes.","_hly_sid","This cookie is likely associated with website analytics, as the prefix '_hly_' suggests a connection to Holistics, a business intelligence and analytics platform.","docs.nvidia.com","_hjSession_xxxxxx","A cookie that holds the current session data. This ensues that subsequent requests within the session window will be attributed to the same Hotjar session.","C0001","Required Cookies","These cookies enable core functionality such as security, network management, and accessibility. These cookies are required for the site to function and cannot be turned off.",[898,899,900,901,902,903,904,905,906,907,908,909,910,911,912,913,914,915,916,917,918,919,920,921,922,923,924,925,926,927,928,929,930,931,932,933,934,935,936,937,938,939,940,941,942,943,944,945,946,947,948,949,950,951,952,953,954,955,956,957,958,959,960,961,962,963,964,965,966,967,968,969,970],[758,759,760,761,762,763,764,765,766,767,768,769,770,771,772,773,774,775,776,777,778,779,780,781,782],{"_407":895,"_250":895,"_215":409,"_380":896},{"_407":889,"_250":889,"_215":409,"_380":890},{"_407":886,"_250":886,"_215":409,"_380":887},{"_407":874,"_250":874,"_215":409,"_380":875},{"_407":871,"_250":871,"_215":409,"_380":872},{"_407":868,"_250":868,"_215":409,"_380":869},{"_407":863,"_250":863,"_215":409,"_380":864},{"_407":470,"_250":470,"_215":409,"_380":860},{"_407":857,"_250":857,"_215":409,"_380":858},{"_407":462,"_250":462,"_215":409,"_380":855},{"_407":457,"_250":457,"_215":409,"_380":853},{"_407":848,"_250":848,"_215":409,"_380":849},{"_407":843,"_250":843,"_215":409,"_380":844},{"_407":838,"_250":838,"_215":409,"_380":839},{"_407":835,"_250":835,"_215":409,"_380":836},{"_407":583,"_250":583,"_215":409,"_380":832},{"_407":442,"_250":442,"_215":409,"_380":829},{"_407":578,"_250":578,"_215":409,"_380":822},{"_407":574,"_250":574,"_215":409,"_380":816},{"_407":810,"_250":810,"_215":409,"_380":811},{"_407":805,"_250":805,"_215":409,"_380":806},{"_407":800,"_250":800,"_215":409,"_380":801},{"_407":796,"_250":796,"_215":409,"_380":797},{"_407":791,"_250":791,"_215":409,"_380":792},{"_407":783,"_250":783,"_215":409,"_380":784},"polyfill.io",[785,786],{"_114":789,"_414":783,"_415":35,"_215":790},{"_114":787,"_414":783,"_415":35,"_215":788},"cf_ob_info","This cookie is associated with sites using CloudFlare. It is used to improve page load times and to override any security restrictions based on the visitors IP address. It does not contain any user identification information.","cf_use_ob","This cookie is associated with sites using CloudFlare. It is used to improve page load times and to override any security restrictions based on the visitors IP address. It does not contain any user identification information. ","www.recaptcha.net",[793],{"_114":794,"_414":791,"_415":35,"_215":795},"_GRECAPTCHA","This domain is used by Google reCAPTCHA, which provides spam protection services to websites by detecting if a user is a human or a bot.","api2.hcaptcha.com",[798],{"_114":799,"_414":796,"_415":35,"_215":451},"__cflb","instagram.com",[802],{"_114":803,"_414":800,"_415":35,"_215":804},"csrftoken","Instagram is a social networking service owned by Facebook, Inc. It's designed for sharing photos and videos from a smartphone. The service also uses cookies for various purposes, including analytics, personalization, and advertising.","readme.io",[807],{"_114":808,"_414":805,"_415":35,"_215":809},"__cf_bm","The __cf_bm cookie is a cookie necessary to support Cloudflare Bot Management, currently in private beta. As part of our bot management service, this cookie helps manage incoming traffic that matches criteria associated with bots.","nvidiacorp.us-5.evergage.com",[812,813],{"_114":815,"_414":810,"_415":35,"_215":451},{"_114":814,"_414":810,"_415":35,"_215":451},"AWSALBTG","AWSALBTGCORS",[817,818],{"_114":820,"_414":574,"_415":35,"_215":821},{"_114":819,"_414":574,"_415":35,"_215":451},"ak_bmsc","bm_mi","This domain is likely associated with NVIDIA's partner program, which could involve marketing, sales, and support services for partners.",[823,824,825],{"_114":828,"_414":578,"_415":35,"_215":451},{"_114":827,"_414":578,"_415":90,"_215":451},{"_114":826,"_414":578,"_415":35,"_215":451},"AWSALBCORS","JSESSIONID","AWSALB",[830],{"_114":808,"_414":442,"_415":35,"_215":831},"This is a CloudFoundry cookie",[833],{"_114":834,"_414":583,"_415":35,"_215":587},"logged_in","hcaptcha.com",[837],{"_114":808,"_414":835,"_415":35,"_215":831},"tracking.intentsify.io",[840],{"_114":841,"_414":838,"_415":35,"_215":842},"userId","Used to track usage of the embedded video player.","static.reo.dev",[845],{"_114":846,"_414":843,"_415":35,"_215":847},"cookietest","Common cookie name could have a number of different origins. Where this is first party and a session cookie, its most likely to do with checking to see if the browser is set to block or allow cookies.","www.datadoghq-browser-agent.com",[850],{"_114":851,"_414":848,"_415":35,"_215":852},"dd_cookie_test_","dd_cookie_test",[854],{"_114":808,"_414":457,"_415":35,"_215":831},[856],{"_114":808,"_414":462,"_415":35,"_215":831},"supabase.co",[859],{"_114":808,"_414":857,"_415":35,"_215":831},[861],{"_114":862,"_414":470,"_415":35,"_215":477},"VISITOR_PRIVACY_METADATA","dosa-prd.azurewebsites.net",[865],{"_114":866,"_414":863,"_415":35,"_215":867},"access_token","This cookie is used to obtain the permissions for users of the portal to allow for API access in order to retrieve information required to perform portal functions.","w.hcaptcha.com",[870],{"_114":808,"_414":868,"_415":35,"_215":831},"cdn.readme.io",[873],{"_114":808,"_414":871,"_415":35,"_215":831},"queue-it.net",[876,877,878,879,880],{"_114":885,"_414":874,"_415":35,"_215":409},{"_114":884,"_414":874,"_415":35,"_215":409},{"_114":883,"_414":874,"_415":35,"_215":409},{"_114":882,"_414":874,"_415":35,"_215":409},{"_114":881,"_414":874,"_415":35,"_215":409},"Queue-it-{queueId}","Queue-it-{dmonvidia }________{eventId}","Queue-it-token","queueitsoftblock_{queueId}","Queue-it-{nvidia}________{eventId}","x.com",[888],{"_114":808,"_414":886,"_415":35,"_215":809},"api.hcaptcha.com",[891,892],{"_114":894,"_414":889,"_415":35,"_215":451},{"_114":799,"_414":889,"_415":35,"_215":893},"This domain is associated with hCaptcha, a service that provides CAPTCHA and reCAPTCHA alternatives for user verification and bot protection.","hmt_id","vimeo.com",[897],{"_114":808,"_414":895,"_415":35,"_215":831},{"_114":1031,"_414":529,"_415":35,"_215":1062},{"_114":1060,"_414":624,"_415":90,"_215":1061},{"_114":1053,"_414":1022,"_415":90,"_215":1054},{"_114":1059,"_414":717,"_415":90,"_215":409},{"_114":1057,"_414":529,"_415":35,"_215":1058},{"_114":815,"_414":1043,"_415":35,"_215":982},{"_114":990,"_414":1052,"_415":90,"_215":992},{"_114":814,"_414":981,"_415":35,"_215":982},{"_114":993,"_414":1052,"_415":35,"_215":994},{"_114":1055,"_414":529,"_415":90,"_215":1056},{"_114":1053,"_414":1029,"_415":90,"_215":1054},{"_114":987,"_414":1052,"_415":35,"_215":989},{"_114":1050,"_414":529,"_415":35,"_215":1051},{"_114":814,"_414":973,"_415":35,"_215":982},{"_114":978,"_414":1049,"_415":35,"_215":980},{"_114":826,"_414":979,"_415":35,"_215":995},{"_114":828,"_414":979,"_415":35,"_215":1046},{"_114":993,"_414":988,"_415":35,"_215":994},{"_114":1047,"_414":627,"_415":35,"_215":1048},{"_114":828,"_414":653,"_415":35,"_215":1046},{"_114":1044,"_414":529,"_415":35,"_215":1045},{"_114":820,"_414":653,"_415":35,"_215":1039},{"_114":1016,"_414":632,"_415":90,"_215":409},{"_114":814,"_414":1043,"_415":35,"_215":982},{"_114":1041,"_414":981,"_415":90,"_215":1042},{"_114":1005,"_414":529,"_415":35,"_215":996},{"_114":978,"_414":529,"_415":35,"_215":980},{"_114":1040,"_414":541,"_415":35,"_215":409},{"_114":820,"_414":529,"_415":35,"_215":1039},{"_114":828,"_414":973,"_415":35,"_215":1038},{"_114":1036,"_414":636,"_415":90,"_215":1037},{"_114":1034,"_414":529,"_415":90,"_215":1035},{"_114":1033,"_414":529,"_415":35,"_215":409},{"_114":1031,"_414":979,"_415":35,"_215":1032},{"_114":1030,"_414":541,"_415":35,"_215":409},{"_114":1021,"_414":1029,"_415":90,"_215":1023},{"_114":1027,"_414":632,"_415":35,"_215":1028},{"_114":1025,"_414":717,"_415":35,"_215":1026},{"_114":1024,"_414":535,"_415":35,"_215":409},{"_114":1021,"_414":1022,"_415":90,"_215":1023},{"_114":1019,"_414":988,"_415":90,"_215":1020},{"_114":1017,"_414":529,"_415":35,"_215":1018},{"_114":1016,"_414":532,"_415":90,"_215":409},{"_114":987,"_414":991,"_415":35,"_215":989},{"_114":819,"_414":1006,"_415":35,"_215":996},{"_114":1014,"_414":532,"_415":35,"_215":1015},{"_114":1012,"_414":529,"_415":35,"_215":1013},{"_114":1010,"_414":529,"_415":35,"_215":1011},{"_114":1009,"_414":529,"_415":35,"_215":409},{"_114":851,"_414":532,"_415":35,"_215":1008},{"_114":1005,"_414":653,"_415":35,"_215":996},{"_114":808,"_414":624,"_415":35,"_215":1007},{"_114":1005,"_414":1006,"_415":35,"_215":996},{"_114":1004,"_414":541,"_415":35,"_215":409},{"_114":1003,"_414":529,"_415":35,"_215":976},{"_114":1001,"_414":988,"_415":35,"_215":1002},{"_114":814,"_414":999,"_415":35,"_215":1000},{"_114":998,"_414":541,"_415":35,"_215":409},{"_114":997,"_414":529,"_415":35,"_215":409},{"_114":990,"_414":988,"_415":90,"_215":992},{"_114":819,"_414":653,"_415":35,"_215":996},{"_114":826,"_414":653,"_415":35,"_215":995},{"_114":993,"_414":991,"_415":35,"_215":994},{"_114":990,"_414":991,"_415":90,"_215":992},{"_114":987,"_414":988,"_415":35,"_215":989},{"_114":985,"_414":981,"_415":90,"_215":986},{"_114":983,"_414":639,"_415":35,"_215":984},{"_114":815,"_414":981,"_415":35,"_215":982},{"_114":978,"_414":979,"_415":35,"_215":980},{"_114":808,"_414":631,"_415":35,"_215":977},{"_114":975,"_414":529,"_415":35,"_215":976},{"_114":826,"_414":973,"_415":35,"_215":974},{"_114":971,"_414":529,"_415":35,"_215":972},"LoginHints","Keeps users logged in across NVIDIA domains.","api.ngc.nvidia.com","This cookie is managed by AWS and is used for load balancing.","bm_so","This cookie name is used to load the site as quickly as possible and to recognize bots using Akamai.","The __cf_bm cookie is a cookie necessary to support Cloudflare Bot Management, currently in private beta. As part of our bot management service, this cookie helps manage incoming traffic that matches criteria associated with bots.\nThis is a CloudFoundry cookie","OptanonConsent","nvid.nvidia.com","This cookie is set by the cookie compliance solution from OneTrust. It stores information about the categories of cookies the site uses and whether visitors have given or withdrawn consent for the use of each category. This enables site owners to prevent cookies in each category from being set in the users browser, when consent is not given. The cookie has a normal lifespan of one year, so that returning visitors to the site will have their preferences remembered. It contains no information that can identify the site visitor.","events.nvidia.com","This cookie is associated with the Amazon Web Services Elastic Load Balancing functionality for routing client request on the server.Used to attribute commission to affiliates when you arrive at the website from an affiliate referral link. It is set when you click on one of our links and used to let the advertiser and us know the website from which you came.","sessionid","This is a very generic cookie name that may have different purposes on different sites, but generally it will be some kind of anonymous session identifier.","_csrf","This cookie is used to prevent Cross-Site Request Forgery (CSRF) attacks, ensuring that each request made to the server is from the user's browser and is intentional.","LSKey-c$CookieConsentPolicy","enterprise-support.nvidia.com","Used to apply end-user cookie consent preferences set by our client-side utility.Custom- Short Duration. Unique ID of an Enquiry- used to validate access to case, and load exisiting data.\n","renderCtx","academicgrants.nvidia.com","Used to store site parameters in the session for reuse across requests by a single client for functionality and performance reasons.","CookieConsentPolicy","Used to apply end-user cookie consent preferences set by our client-side utility.","This cookie is managed by AWS and is used for load balancing.Used to attribute commission to affiliates when you arrive at the website from an affiliate referral link. It is set when you click on one of our links and used to let the advertiser and us know the website from which you came.","This cookie is associated with Akamai and is used to differentiate between traffic from humans and bots.","nv_co","__uxq412__ses.88c1","learn.learn.nvidia.com","This cookie is associated with the Amazon Web Services Elastic Load Balancing functionality for routing client request on the server.","pctrk","Used to count page views by unauthenticated users against license usage.","bm_s","bpazaws52gukakzc__ctrl0_ctl36_uccaptcha","bm_sv","store.nvidia.com","The __cf_bm cookie is a cookie necessary to support Cloudflare Bot Management, currently in private beta. As part of our bot management service, this cookie helps manage incoming traffic that matches criteria associated with bots.\n","This cookie name is used by DataDog to determine if cookies are supported on the website.dd_cookie_test","AMP_TLDTEST_xxxxxxxx","SESSlsd","This cookie is likely to be a session identifier used to maintain user session state across page requests.","mboxEdgeCluster","This cookie name is associated with Adobe Target which is used to communicate the correct edge cluster to at.js so that Target profiles can stay in sync as users navigate across a site.","_rdt_em","This cookie name is associated with Reddit which is used to manage session state and user interactions.","megaNavCollapsed","_abck","The _abck cookie is associated with Akamai's Bot Manager, which is used to prevent fraud and abuse by safeguarding online interactions and detecting irregularities that could suggest malicious activity, such as bot-driven attacks or hacking attempts.","idccsrf","This cookie name is associated with Salesforce which is used to track CrossSiteRequestForgery validation for certain single sign-on (SSO) flows.","ApplicationGatewayAffinity","partners.nvidia.com","This cookie is used to maintain a user session by the server in a load balanced environment, to ensure the user has a consistent experience.","AMP_TEST_xxxxxxxx","bm_lso","This cookie name is associated with Akamai which is used to load the site as quickly as possible and to recognize bots.","SSESS###########","This cookie is likely to be a session cookie used to maintain an anonymized user session by the server in a secure environment, often seen in sites built with the Drupal content management system.","nvdeveloper.nvidia.com","__uxq412__id.88c1","OptanonAlertBoxClosed","This cookie is set by websites using certain versions of the cookie law compliance solution from OneTrust. It is set after visitors have seen a cookie information notice and in some cases only when they actively close the notice down. It enables the website not to show the message more than once to a user. The cookie has a normal lifespan of one year and contains no personal information.","QueueITAccepted-SDFrts345E-V3_{eventId}","TEST_AMCV_COOKIE_WRITE","This cookie is used to test whether cookies are enabled in the browser. It is a temporary cookie used to determine if the browser accepts cookies.","XSRF-TOKEN","This cookie is written to help with site security in preventing Cross-Site Request Forgery attacks.This cookie helps with site security in preventing Cross-Site Request Forgery attacks.","AWS ELB application load balancer","This cookie is associated with Akamai. It is used to analyse traffic to determine if it is automated traffic generated by bots or a human user.","bpazaws52gukakzc__ctrl0_ctl78_list_ctl09_systemcaptcha","PHPSESSID","PHP session cookie associated with embedded content from this domain.\n","base.manager.learn.nvidia.com","bm_ss","This cookie is used to help protect websites from automated bot traffic.","AWS ELB application load balancerUsed to attribute commission to affiliates when you arrive at the website from an affiliate referral link. It is set when you click on one of our links and used to let the advertiser and us know the website from which you came.","_devzone_session","This cookie is likely used to maintain a user session on the website by the server. It is common for session identifiers to be used to keep users logged in and track their interaction with the site during a session.","docs.ngc.nvidia.com","bm_sz","This cookie is used by Akamai to optimize site performance and security. It is typically associated with the delivery of web content and services.","vcalliance.nvidia.com","ASP.NET_SessionId","General purpose platform session cookie, used by sites written with Microsoft .NET based technologies. Usually used to maintain an anonymised user session by the server.","com.adobe.alloy.getTld","This cookie name is associated with the Adobe WebSDK which is used to determine the top-most domain that can accept cookies.","SESSls","This cookie is likely to be a session identifier, used to maintain a user's session state across page requests.","c_code","BIGipServerx","This cookie name is associated with the BIG-IP product suite from company F5. Usually associated with managing sessions on load balanced servers, to ensure user requests are routed consistently to the correct server. The common root is BIGipServer most commonly followed by a domain name, usually the one that it is hosted on, but not always.","This cookie is set by websites using certain versions of the cookie law compliance solution from OneTrust. It is set after visitors have seen a cookie information notice and in some cases only when they actively close the notice down. It enables the website not to show the message more than once to a user. The cookie has a one year lifespan and contains no personal information.","default","

NVIDIA uses cookies to improve your experience on our web site. We and our third-party partners also use cookies and other tools to collect and record information you provide as well as information about your interactions with our websites for performance improvement, analytics, and to assist in marketing efforts. By clicking \"Accept All\", you consent to our use of cookies and other tools as described in our Cookie Policy. You can manage your cookie settings by clicking on \"Manage Settings.\" By continuing to use this site or by clicking one of the buttons below, you agree to our Terms of Service (which contains important waivers). Please see our Privacy Policy for more information on our privacy practices.

","gpc","

We have detected the Global Privacy Control (GPC) signal and recorded your rejection of all optional cookies on this site for this browser. You can manage your cookie settings by clicking on \"Manage Settings\". Please see our Cookie Policy for more information. To opt out of non-cookie personal information \"sales\" / \"sharing\" for targeted advertising purposes, please visit the NVIDIA Preference Center. Please see our Privacy Policy for more information on our privacy practices.

","gpcLoggedIn","

We have detected the Global Privacy Control Signal (GPC) and have opted you out of all optional cookies on this browser. You can manage your cookie settings by clicking on \"Manage Settings\". Please see our Cookie Policy for more information. We have also opted you out of \"sharing\"/\"sales\" of personal information outside of cookies. You can manage these settings in the NVIDIA NVIDIA Preference Center. Please see our Privacy Policy for more information.

","gpcLoggedInOverridden","

We have detected the Global Privacy Control Signal (GPC) and have opted you out of all optional cookies on this browser. You can manage your cookie settings by clicking on \"Manage Settings\". Please see our Cookie Policy for more information. We have also opted you out of \"sharing\"/\"sales\" of personal information outside of cookies which overrides at least one of your previous settings. You can manage them in the NVIDIA Preference Center. Please see our Privacy Policy for more information.

",{"_753":18,"_656":60,"_564":60,"_384":60,"_374":60},"hasConsented","gpcApplied","showBanner","geolocation",{"_1077":1078,"_1079":1080},"country","US","state","OR","query","categories",[1100,1101,1102,1103,1104],"optionUrls",[1088,1089,1090,1091,1092,1093,1094,1095,1096,1097,1098,1099],"total",12,"/orgs/nim/deepseek-ai/containers/deepseek-r1/-","/orgs/nim/nvidia/containers/llama-3.1-nemotron-70b-instruct/-","/orgs/nvidia/-/containers/pytorch/-","/orgs/nvidia/omniverse/collections/kit/-","/orgs/nvidia/-/collections/deepstream_sdk/-","/orgs/nvidia/omniverse/collections/kit-appstreaming-collection/-","/orgs/nvidia/research/models/stylegan3/-","/orgs/nvidia/tao/models/peoplenet/-","/orgs/nvidia/tao/models/trafficcamnet/-","/orgs/nvidia/riva/resources/riva_quickstart/-","/orgs/nvidia/-/helm-charts/gpu-operator/-","/orgs/nvidia/-/helm-charts/k8s-nim-operator/-",{"_21":1178,"_1106":1179,"_1108":60},{"_21":1143,"_1106":1144,"_1108":58},{"_21":1126,"_1106":1127,"_1108":1128},{"_21":22,"_1106":1119,"_1108":1120},{"_21":1105,"_1106":1107,"_1108":20},"helm-charts","items",[1109,1110],"startIndex",{"_21":1105,"_250":1116,"_209":210,"_211":-7,"_245":1117,"_253":1113,"_278":1118,"_1115":1098},{"_21":1105,"_250":1111,"_209":210,"_211":-7,"_245":1112,"_253":1113,"_278":1114,"_1115":1099},"NVIDIA NIM Operator","k8s-nim-operator","NVIDIA",[],"url","GPU Operator","gpu-operator",[],[1121],9,{"_21":22,"_250":1122,"_209":210,"_211":1123,"_245":1124,"_253":1113,"_278":1125,"_1115":1097},"Riva Skills Quick Start","riva","riva_quickstart",[],"models",[1129,1130,1131],6,{"_21":1126,"_250":1139,"_209":210,"_211":1140,"_245":1141,"_253":1113,"_278":1142,"_1115":1094},{"_21":1126,"_250":1136,"_209":210,"_211":1133,"_245":1137,"_253":1113,"_278":1138,"_1115":1095},{"_21":1126,"_250":1132,"_209":210,"_211":1133,"_245":1134,"_253":1113,"_278":1135,"_1115":1096},"TrafficCamNet","tao","trafficcamnet",[],"PeopleNet","peoplenet",[],"StyleGAN3 pretrained models","research","stylegan3",[],"collections",[1145,1146,1147],{"_21":1143,"_250":1175,"_209":210,"_211":1149,"_245":1176,"_253":1113,"_278":1177,"_1115":1091},{"_21":1143,"_250":1172,"_209":210,"_211":-7,"_245":1173,"_253":1113,"_278":1174,"_1115":1092},{"_21":1143,"_250":1148,"_209":210,"_211":1149,"_245":1150,"_253":1113,"_278":1151,"_1115":1093},"Omniverse Kit App Streaming","omniverse","kit-appstreaming-collection",[1152,1153,1154],{"_114":1166,"_250":1167,"_247":1168,"_257":1169,"_1159":1170},{"_114":1149,"_250":1162,"_247":1163,"_257":1158,"_1159":1164},{"_114":1155,"_250":1156,"_247":1157,"_257":1158,"_1159":1160},"omniverse-dev","NVIDIA Developer Program","Join the Developer Program for access to free tools, support, and tech resources.","https://store-images.s-microsoft.com/image/apps.33936.6f1bcf42-55d0-45d1-b4b6-f08c68a50bbe.e91ca707-db64-49d6-9761-4781590a9685.7ba64fc5-7c44-4864-a5e6-7bde05126935","supportedFeatures",[1161],"UNIFIED_CATALOG","NVIDIA Omniverse","NVIDIA Omniverse is a platform of APIs, services, and software development kits (SDKs) that enable developers to build generative AI-enabled tools and applications.",[1161,1165],"SHAREABLE_TO_USER_GROUPS","nv-ai-enterprise","NVIDIA AI Enterprise","Accelerate your AI agent development","https://assets.ngc.nvidia.com/products/catalog/images/product-NVAIE.jpg",[1161,1165,1171],"NGC_CONSOLE","DeepStream SDK","deepstream_sdk",[],"Omniverse Kit (FB)","kit",[1152,1153,1154],"containers",[1180,1181,1182],{"_21":1178,"_250":1190,"_209":1187,"_211":1191,"_245":1192,"_253":1113,"_278":1193,"_1115":1088},{"_21":1178,"_250":1186,"_209":1187,"_211":210,"_245":1188,"_253":1113,"_278":1189,"_1115":1089},{"_21":1178,"_250":1183,"_209":210,"_211":-7,"_245":1184,"_253":1113,"_278":1185,"_1115":1090},"PyTorch","pytorch",[],"Llama-3.1-Nemotron-70B-Instruct","nim","llama-3.1-nemotron-70b-instruct",[],"DeepSeek-R1","deepseek-ai","deepseek-r1",[],"ownerOrgs",[],"memberOrgs",[],"externalOrgs",[],"apiKey","2b44f86fea50e90ffd2db29ee2028419","consentGranted","applicationId","4c857e4f-5ffa-45fa-b75e-af773c22f0ad","clientToken","pubb6927db33dae60ba5031a186cde0242a","site","datadoghq.com","service","catalog-ui","env","sessionSampleRate",100,"sessionReplaySampleRate","compressIntakeRequests","storeContextsAcrossPages","defaultPrivacyLevel","mask-user-input","trackViewsManually","trackingConsent","not-granted",{"_1223":1224,"_1225":320,"_1226":35,"_1227":1228,"_1106":1229,"_1230":1231,"_1232":1224,"_1233":35,"_1234":35,"_1235":35,"_1236":1237},"appName","ngc-catalog","appOrigin","hideOnPrivateOrg","icon","ObjectsModules2",[1238,1239,1240,1241,1242,1243,1244],"menuIcon","editor-swatches","productName","requiresAllScopes","requiresLogin","requiresOrgLevelScopesOnly","title","Catalog",{"_1236":1266,"_51":1267,"_1230":1268},{"_1259":1260,"_1236":1261,"_51":1262,"_1232":1166},{"_1236":1257,"_51":1258,"_1230":1231},{"_1236":1254,"_51":1255,"_1230":1256},{"_1236":1251,"_51":1252,"_1230":1253},{"_1236":1248,"_51":1249,"_1230":1250},{"_1236":1245,"_51":1246,"_1230":1247},"Resources","/resources","misc-library","Models","/models","misc-neural-network","Helm Charts","/helm-charts","misc-ship-wheel","Containers","/containers","3d-cube","Collections","/collections","access",[1263],"Enterprise Catalog","/enterprise",[1264,1265],"NVAIE","READ","Explore Catalog","/","common-home",{"_161":162,"_163":164,"_165":166,"_167":168,"_169":170,"_171":172},{"_160":35}]

	Files	Compressed Size		Download
21.02.1Selected	1	73.55 KB	January 31, 2023 UTC
21.02.0	1	73.57 KB	September 22, 2022 UTC