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

This repository provides an implementation of the Transformer-XL model in\nPyTorch from the paper Transformer-XL: Attentive\nLanguage Models Beyond a Fixed-Length\nContext. Transformer-XL is a\ntransformer-based language model with a segment-level recurrence and a novel\nrelative positional encoding. Enhancements introduced in Transformer-XL help\ncapture better long-term dependencies by attending to tokens from multiple\nprevious segments.

Our implementation is based on the\ncodebase published by the\nauthors of the Transformer-XL paper.\nOur implementation uses a modified model architecture. Our\nmodifications were made to achieve better hardware utilization and to take\nadvantage of Tensor Cores. Similar modifications were also proposed in an\nimplementation available from\ngithub.com/cybertronai/transformer-xl.\nRefer to the Model architecture section for more\ndetails.

This model is trained with mixed precision using Tensor Cores on NVIDIA Volta\nand the NVIDIA Ampere GPU architectures and evaluated on Volta, Turing and the\nNVIDIA Ampere GPU architectures.\nTherefore, researchers can get results up to 2.5x faster than training without\nTensor Cores, while experiencing the benefits of mixed precision training. This\nmodel is tested against each NGC monthly container release to ensure consistent\naccuracy and performance over time.

Model architecture

The Transformer-XL \"base\" model for WikiText-103 dataset available in this\nrepository was modified to use the following hyperparameter values:

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

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

Hyperparameter	Description	Original setting for the base model	Our modification for the base model
`d_model`	hidden size	410	512
`n_head`	number of attention heads	10	8
`d_head`	size of each attention head	41	64
`d_inner`	hidden size in fully-connected layers	2100	2048
`tgt_len`	number of tokens to predict during training	150	192
`mem_len`	number of tokens cached from previous iterations during training	150	192

Changes described above were made to align certain hyperparameters with powers\nof two, with this modification, the model is able to achieve better hardware\nutilization, and therefore higher training throughput.

The Transformer-XL \"large\" model for WikiText-103 dataset available in this\nrepository uses the original hyperparameters from the reference\nimplementation.

The following table lists the hyperparameters for the large and the base\nTransformer-XL models for WikiText-103 dataset available in this repository.

\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\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\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\">

Hyperparameter	Description	Base model	Large model
`n_layer`	number of layers	16	18
`d_model`	hidden size	512	1024
`n_head`	number of attention heads	8	16
`d_head`	size of each attention head	64	64
`d_inner`	inner hidden size in fully-connected layers	2048	4096
`dropout`	dropout	0.1	0.2
`dropatt`	dropout after softmax in the attention	0.0	0.2
`lr`	base learning rate	0.01	0.01
`eta_min`	minimum learning rate (for cosine decay)	0.001	0.0001
`max_step`	number of training steps	40,000	100,000
`warmup_step`	number of learning rate warmup steps	1,000	16,000
`batch_size`	training batch size	256	128
`tgt_len`	number of tokens to predict during training	192	384
`mem_len`	number of tokens cached from previous iterations during training	192	384

The Transformer-XL model addresses the limitations of vanilla transformer-based\nlanguage models, which are only able to use relatively short context, bounded\nby the segment length. The Transformer-XL introduces a recurrence mechanism,\nwhich is able to use a cached hidden state from previous segments. During\ntraining, the context consists of a concatenation of current segment's hidden\nstate and cached states from previous iterations. Gradients are backpropagated\nonly through the current segment, although the model is able to take advantage\nof the extra information stored in the cache and therefore is able to model\nlong-term dependencies.

An illustration of the recurrence mechanism taken from the Transformer-XL\npaper is shown below.\n $\"model\"$

Default configuration

The following features were implemented in this model:

general\n
- single-node or multi-node, data-parallel multi-GPU training
- training and inference with mixed precision using Tensor Cores
- mixed precision training implemented using\nApex AMP, with O2 optimization\nlevel and with a dynamic loss scaling
\n
model\n
- 16-layer base Transformer-XL model with hidden size 512, 8 attention heads,\neach head with hidden size 64
- 18-layer large Transformer-XL model with hidden size 1024, 16 attention\nheads, each head with hidden size 64
- the model trained on\nWikiText-103\ndataset, using word-level vocabulary and\nadaptive softmax
- embedding weights are tied with weights in the classifier
\n
training\n
- training with LAMB optimizer, the\nimplementation of the optimizer uses\nTorchScript, which enables\nthe fusion of elementwise operations and accelerates the training
- support for training with a gradient accumulation
- base model:\n
  - linear learning rate warmup for 1,000 iterations, followed by the cosine\nlearning rate schedule, the initial learning rate is set to 0.01, and the final\nlearning rate is set to 0.001
  - training for 40,000 steps, using a batch size of 256
  \n
- large model:\n
  - single node:\n
    - linear learning rate warmup for 16,000 iterations, followed by the cosine\nlearning rate schedule, the initial learning rate is set to 0.01, and the final\nlearning rate is set to 0.0001
    - training for 100,000 steps, using a batch size of 128
    \n
  - multi node:\n
    - linear learning rate warmup for 16,000 iterations, followed by the cosine\nlearning rate schedule, the initial learning rate is set to 0.02, and the final\nlearning rate is set to 0.0002
    - training for 25,000 steps, using a batch size of 512
    \n
  \n
\n
inference\n
- support for multi-gpu inference
- support for TorchScript and\npure Python inference
- each token is using the same size of the context from previous time steps.
- base model:\n
  - target length is set to 64, length of memory is set to 640
  - positional embeddings are clamped after 400 time steps
  \n
- large model:\n
  - target length is set to 128, length of memory is set to 1,600
  - positional embeddings are clamped after 1,000 time steps
  \n
\n

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	Transformer-XL
Apex AMP	Yes
PyTorch DistributedDataParallel	Yes
LAMB	Yes
Inference with TorchScript	Yes
Multi-node training	Yes

Features

Apex AMP - a tool that enables Tensor\nCore-accelerated training. Refer to the Enabling mixed\nprecision section for more details.

PyTorch\nDistributedDataParallel - a module\nwrapper that enables easy multiprocess distributed data-parallel\ntraining.

LAMB - stands for Layerwise Adaptive\nMoments Based optimizer, is a large batch optimization technique that helps\naccelerate training of deep neural networks using large minibatches.

TorchScript - is a way to create\nserializable and optimizable models from PyTorch code. Any TorchScript program\ncan be saved from a Python process and loaded in a process where there is no\nPython dependency.

Mixed precision training

Mixed precision is the combined use of different numerical precisions in a\ncomputational method.\nMixed precision training offers significant\ncomputational speedup by performing operations in half-precision format while\nstoring minimal information in single-precision to retain as much information\nas possible in critical parts of the network. Since the introduction of Tensor\nCores in Volta, and following with\nboth the Turing and Ampere architectures, significant training speedups are\nexperienced by switching to\nmixed precision -- up to 3x overall speedup on the most arithmetically intense\nmodel architectures. Using mixed precision training previously required two\nsteps:

Porting the model to use the FP16 data type where appropriate.
Manually adding loss scaling to preserve small gradient values.

The ability to train deep learning networks with lower precision was introduced\nin the Pascal architecture and first supported in CUDA\n8 in the NVIDIA Deep\nLearning SDK.

For information about:

How to train using mixed precision, see the Mixed Precision\nTraining paper and Training With Mixed\nPrecision\ndocumentation.
Techniques used for mixed precision training, see the Mixed-Precision\nTraining of Deep Neural\nNetworks\nblog.
APEX tools for mixed precision training, see the NVIDIA Apex: Tools for Easy\nMixed-Precision Training in\nPyTorch\n.

Enabling mixed precision

The pytorch/train.py training script launches mixed precision training\nwith Tensor Cores if the flag --fp16 is set.

Mixed precision is enabled in PyTorch by using the Automatic Mixed Precision\n(AMP), library from APEX that casts variables\nto half-precision upon retrieval, while storing variables in single-precision\nformat. Furthermore, to preserve small gradient magnitudes in backpropagation,\na loss\nscaling\nstep must be included when applying gradients. In PyTorch, loss scaling can be\neasily applied by using scale_loss() method provided by AMP. The scaling\nvalue to be used can be\ndynamic or fixed.

For an in-depth walk through on AMP, check out sample usage\nhere.\nAPEX is a PyTorch extension that contains\nutility libraries, such as AMP, which require minimal network code changes to\nleverage Tensor Cores performance.

The following steps were needed to enable mixed precision training in\nTransformer-XL:

Import AMP from APEX:

from apex import amp\n

Initialize AMP and wrap the model and the optimizer before starting the\ntraining:

model, optimizer = amp.initialize(\n    model,\n    optimizer,\n    opt_level='O2',\n    )\n

Apply scale_loss context manager:

with amp.scale_loss(loss, optimizer) as scaled_loss:\n    scaled_loss.backward()\n

Apply gradient clipping on single precision master weights:

torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.clip)\n

Enabling TF32

TensorFloat-32 (TF32) is the new math mode in NVIDIA\nA100 GPUs for handling the\nmatrix math also called tensor operations. TF32 running on Tensor Cores in A100\nGPUs can provide up to 10x speedups compared to single-precision floating-point\nmath (FP32) on Volta GPUs.

TF32 Tensor Cores can speed up networks using FP32, typically with no loss of\naccuracy. It is more robust than FP16 for models which require high dynamic\nrange for weights or activations.

For more information, refer to the TensorFloat-32 in the A100 GPU Accelerates\nAI Training, HPC up to\n20x\nblog post.

TF32 is supported in the NVIDIA Ampere GPU architecture and is enabled by\ndefault.

","resourceVersions",[231,232],"resourceVersionsTotal","hasMoreVersions","versionMetadata",{"_210":28,"_211":-5,"_52":116,"_113":-5,"_117":212,"_119":-5,"_213":115,"_124":16,"_121":34,"_214":-5,"_120":34,"_215":216,"_217":90,"_218":219,"_220":90,"_221":222,"_223":90,"_224":225,"_226":90,"_227":228,"_229":90,"_122":230},"versionNotFound","isLoggedIn","latestVersionPath","/orgs/nvidia/dle/resources/transformerxl_pyt/-","-","imageReference","cliDownloadCommand","ngc registry resource download-version nvidia/dle/transformerxl_pyt:20.06.0","helmFetchCommand","viewAllVersionsPath","/orgs/nvidia/dle/resources/transformerxl_pyt/-/version-history","accessCheckResult",{"_209":90},"hasRelatedCollections","context",{"_205":206,"_207":208},1785955791229,[],[],[198,199,200,201,202,203,204],{"_48":63,"_50":74,"_52":65,"_54":66,"_56":73,"_58":59},{"_48":63,"_50":72,"_52":146,"_54":-5,"_56":-5,"_58":59},{"_48":63,"_50":70,"_52":65,"_54":66,"_56":71,"_58":59},{"_48":63,"_50":68,"_52":65,"_54":66,"_56":69,"_58":59},{"_48":63,"_50":64,"_52":65,"_54":66,"_56":67,"_58":59},{"_48":60,"_50":61,"_52":53,"_54":55,"_56":62,"_58":59},{"_48":49,"_50":51,"_52":53,"_54":55,"_56":57,"_58":59},"orgName","nvidia","teamName","dle","allowed","versionId","description",["D",1663857816874],"totalSizeInBytes","malwareScanStatus","releaseNotes","### Changelog\n\n* June 2020\n * Added support for NVIDIA DGX A100\n * Updated default NGC container to pytorch-20.06-py3\n* December 2019\n * Added support for the large Transformer-XL model trained on WikiText-103\n dataset, the large model was trained on NVIDIA DGX-1, NVIDIA DGX-2 and on\n 8x NVIDIA DGX-2H (multi-node training)\n * Updated default NGC container to pytorch-19.11-py3\n * Added support for inference with TorchScript\n* October 2019\n * Initial release\n * Support for FP32 and mixed precision training on NVIDIA\n DGX-1, NVIDIA DGX-2, and inference on NVIDIA Tesla V100 16GB\n and NVIDIA T4\n\n### Known issues\nThere are no known issues with this model.","hasReleaseNotes","setup","The following section lists the requirements that you need to meet in order to\nstart training the Transformer-XL model.\n\n### Requirements\n\nThis repository contains `Dockerfile` which extends the PyTorch NGC container\nand encapsulates some dependencies. Aside from these dependencies, ensure you\nhave the following components:\n\n* [NVIDIA Docker](https://github.com/NVIDIA/nvidia-docker)\n* [PyTorch 20.06-py3 NGC container](https://ngc.nvidia.com/registry/nvidia-pytorch)\n* GPU 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\nfollowing sections from the NVIDIA GPU Cloud Documentation and the Deep\nLearning DGX Documentation:\n\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/dgx/user-guide/index.html#accessing_registry),\n* [Running PyTorch](https://docs.nvidia.com/deeplearning/dgx/pytorch-release-notes/running.html#running).\n\nFor those unable to use the Pytorch NGC container, to set up the required\nenvironment or create your own container, see the versioned [NVIDIA Container\nSupport\nMatrix](https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html).\n\nFor multi-node, the sample provided in this repository requires\n[Enroot](https://github.com/NVIDIA/enroot) and\n[Pyxis](https://github.com/NVIDIA/pyxis) set up on a\n[SLURM](https://slurm.schedmd.com) cluster.","hasSetup","quickStartGuide","To train your model using mixed or TF32 precision with Tensor Cores or using\nFP32, perform the following steps using the default parameters of the\nTransformer-XL base model on the\n[WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/)\ndataset. \n\nFor the specifics concerning training\nand inference, see the [Advanced](#advanced) section.\n\n1. Clone the repository.\n\n```\ngit clone https://github.com/NVIDIA/DeepLearningExamples\ncd DeepLearningExamples/PyTorch/LanguageModeling/Transformer-XL\n```\n\n2. Download and preprocess the dataset.\n\n```\nbash getdata.sh\n```\n\n3. Build the Transformer-XL PyTorch NGC container.\n\n```\nbash pytorch/scripts/docker/build.sh\n```\n\n4. Start an interactive session in the NGC container to run training/inference.\n\n```\nbash pytorch/scripts/docker/interactive.sh\n```\n\n5. Start training.\n\nThis repository contains a number of predefined configurations to run the\ntraining on NVIDIA DGX-1, NVIDIA DGX-2H or NVIDIA DGX A100 nodes.\n\nTo start the training on NVIDIA DGX-1 or NVIDIA DGX-2H, run:\n\n```\nbash run_wt103_{base,large}.sh train <#GPUs> --config {dgx1,dgx2}_<#GPUs>gpu_{fp16,fp32}\n```\n\nTo start the training on NVIDIA DGX A100, run:\n\n```\nbash run_wt103_{base,large}.sh train <#GPUs> --config dgxa100_<#GPUs>gpu_{fp16,tf32}\n```\n\n* use the `run_wt103_base.sh` script to train the base model, and use the\n `run_wt103_large.sh` script to train the large model\n* the training is executed on `<#GPUs>` GPUs, supported values for `<#GPUs>`\n for NVIDIA DGX-1 and NVIDIA DGX A100 are: 1, 2, 4, 8 and for NVIDIA DGX-2H:\n 1, 2, 4, 8, 16\n* use configs with the `dgx1` prefix to run on a NVIDIA DGX-1, configs with the\n `dgx2` prefix to run on a NVIDIA DGX-2H and configs with the `dgxa100` prefix\n to run on a NVIDIA DGX A100\n* configs with the `fp16` suffix are launching mixed precision training,\n configs with the `fp32` suffix are launching FP32 training, configs with the\n `tf32` suffix are launching TF32 training\n\nExamples:\n\nTo launch TF32 training of the base Transformer-XL model on a NVIDIA DGX A100\nusing 8 GPUs, run:\n\n```\nbash run_wt103_base.sh train 8 --config dgxa100_8gpu_tf32\n```\n\nTo launch FP32 training of the base Transformer-XL model on a NVIDIA DGX-1\nusing 8 GPUs, run:\n```\nbash run_wt103_base.sh train 8 --config dgx1_8gpu_fp32\n```\n\nTo launch mixed precision training of the large Transformer-XL model on a\nNVIDIA DGX-2H using 16 GPUs, run:\n```\nbash run_wt103_large.sh train 16 --config dgx2_16gpu_fp16\n```\n\nTo launch mixed precision training of the large Transformer-XL model on a\nNVIDIA DGX A100 using 8 GPUs, run:\n\n```\nbash run_wt103_large.sh train 8 --config dgxa100_8gpu_fp16\n```\n\nTo run on multiple nodes, see the [Multi-node](#multi-node) section. \n\nFor more information on the available options, and for an explanation of what\nhappens at the end of training, refer to the [Training\nprocess](#training-process) section.\n\n6. Start evaluation.\n\nTo start inference on the test set using `<#GPUs>` GPUs, run:\n\n```\nbash run_wt103_{base,large}.sh eval <#GPUs> [--fp16] [--type {pytorch, torchscript}]\n```\n\nSelect `run_wt103_base.sh` for the base Transformer-XL model and\n`run_wt103_large.sh` for the large Transformer-XL model.\nThe `--fp16` flag is optional, however, if it's specified, then the script\nlaunches mixed precision inference with Tensor Cores. If the flag is not\npresent, then the script launches FP32 inference on NVIDIA Volta and NVIDIA\nTuring GPUs and TF32 inference on NVIDIA Ampere GPUs.\n\nBy default, the script is loading the checkpoint from\n`LM-TFM/checkpoint_best.pt`, which contains the model corresponding to the\nlowest value of the validation loss from the previous training run. Path to the\ncheckpoint can be customized by setting the `--model` flag.\n\nInference can use pure Python execution or TorchScript from using the `--type`\nflag.\n\nSupported values for `<#GPUs>` are: 1, 2, 4, 8 for NVIDIA DGX-1 and NVIDIA DGX\nA100 and 1, 2, 4, 8, 16 for NVIDIA DGX-2H.\n\nAdditionally, one can pass the input text directly from the command-line using\nthe `--manual` flag. This mode of operation supports only 1 GPU and batch size\nof 1. The script outputs average loss and perplexity for the provided input\ntext.\n\nExamples:\n\n```\nbash run_wt103_base.sh eval 1 \\\n --model LM-TFM/checkpoint_best.pt \\\n --fp16 \\\n --manual \"recognize speech\"\n\n===============================================================================\n\n| test loss 6.20 | test ppl 494.291\n===============================================================================\n```\n\n```\nbash run_wt103_base.sh eval 1 \\\n --model LM-TFM/checkpoint_best.pt \\\n --fp16 \\\n --manual \"wreck a nice beach\"\n\n===============================================================================\n\n| test loss 8.04 | test ppl 3099.706\n===============================================================================\n```\n\nFor more information on the available options, refer to the [Inference\nprocess](#inference-process) section.","hasQuickStartGuide","advanced","The following sections provide greater details of the dataset, running training\nand inference, and the training results.\n\n### Scripts and sample code\n\nIn the root directory, the most important files are:\n\n* `Dockerfile`: container with the basic set of dependencies to run\n Transformer-XL\n* `requirements.txt`: set of extra requirements for running Transformer-XL\n* `getdata.sh`: script for downloading datasets\n\nIn the `pytorch` directory, the most important files are:\n\n* `data_utils.py`: data loading utilities\n* `eval.py`: serves as the entry point to launch the evaluation and inference\n* `lamb.py`: implementation of [LAMB](https://arxiv.org/abs/1904.00962)\n optimizer\n* `mem_transformer.py`: implementation of the Transformer-XL model\n* `train.py`: serves as the entry point to launch the training\n* `run.sub`: Slurm batch script for launching multi-node training\n\nThe `pytorch/utils` directory contains the following additional modules:\n\n* `adaptive_softmax.py`: implementation of adaptive softmax\n* `data_parallel.py`: implementation of `BalancedDataParallel` class\n* `distributed.py`: utility functions for running distributed training\n* `exp_utils.py`: utility functions for running training and benchmarking\n* `log_uniform_sampler.py`: implementation of log-uniform sampler\n* `proj_adaptive_softmax.py`: implementation of projected adaptive softmax\n* `vocabulary.py`: implementation of word-level vocabulary and BPE-based\n vocabulary\n\nThe `pytorch/inference` directory contains modules optimized for running\ninference with TorchScript:\n* `mem_transformer_jit.py`: implementation of TorchScript-compatible\n Transformer-XL model\n* `proj_adaptive_softmax_jit.py`: implementation of TorchScript-compatible\n projected adaptive softmax\n\n### Parameters\n\n**Training**\n\nThe complete list of available parameters for the `pytorch/train.py` training\nscript contains:\n\n```\ngeneral setup:\n --work_dir WORK_DIR Directory for the results\n --append_dataset Automatically append dataset name to work_dir\n --append_time Automatically append current time to work_dir\n --cuda Run training on a GPU using CUDA\n --fp16 Run training in fp16/mixed precision\n --restart RESTART Restart training from the saved checkpoint\n --debug Run in debug mode (do not create exp dir)\n --log_all_ranks Enable logging from all distributed ranks\n --dllog_file DLLOG_FILE\n Name of the DLLogger output file\n --txtlog_file TXTLOG_FILE\n Name of the txt log file\n --save_all Save all checkpoints\n --no_env Do not print info on execution env\n --no_eval Disable model evaluation\n --log_interval LOG_INTERVAL\n Report interval\n --target_throughput TARGET_THROUGHPUT\n Target training throughput (for benchmarking)\n --target_perplexity TARGET_PERPLEXITY\n Target validation perplexity (for benchmarking)\n --amp_mode {O0,O1,O2,O3}\n Optimization level for apex amp\n\ndataset setup:\n --data DATA Location of the data corpus\n --dataset {wt103,lm1b,enwik8,text8}\n Dataset name\n --vocab {word,bpe} Type of vocabulary\n\nmodel setup:\n --n_layer N_LAYER Number of total layers\n --n_head N_HEAD Number of heads\n --d_head D_HEAD Head dimension\n --d_embed D_EMBED Embedding dimension\n --d_model D_MODEL Model dimension\n --d_inner D_INNER Inner dimension in feedforward layer\n --dropout DROPOUT Global dropout rate\n --dropatt DROPATT Attention probability dropout rate\n --pre_lnorm Apply LayerNorm to the input instead of the output\n --attn_type ATTN_TYPE\n Attention type. 0 for ours, 1 for Shaw et al,2 for\n Vaswani et al, 3 for Al Rfou et al.\n --not_tied Do not tie the word embedding and softmax weights\n --clamp_len CLAMP_LEN\n Use the same pos embeddings after clamp_len\n --adaptive Use adaptive softmax\n --div_val DIV_VAL Dividend value for adaptive input and softmax\n --sample_softmax SAMPLE_SOFTMAX\n Number of samples in sampled softmax\n --init INIT Parameter initializer to use\n --emb_init EMB_INIT Parameter initializer to use\n --init_range INIT_RANGE\n Parameters initialized by U(-init_range, init_range)\n --emb_init_range EMB_INIT_RANGE\n Parameters initialized by U(-init_range, init_range)\n --init_std INIT_STD Parameters initialized by N(0, init_std)\n --proj_init_std PROJ_INIT_STD\n Parameters initialized by N(0, init_std)\n\noptimizer setup:\n --optim {adam,sgd,adagrad,lamb,jitlamb}\n Optimizer to use\n --lr LR Initial learning rate\n --mom MOM Momentum for sgd\n --scheduler {cosine,inv_sqrt,dev_perf,constant}\n LR scheduler to use\n --max_step_scheduler MAX_STEP_SCHEDULER\n Max number of training steps for LR scheduler\n --warmup_step WARMUP_STEP\n Number of iterations for LR warmup\n --decay_rate DECAY_RATE\n Decay factor when ReduceLROnPlateau is used\n --lr_min LR_MIN Minimum learning rate during annealing\n --clip CLIP Gradient clipping\n --weight_decay WEIGHT_DECAY\n Weight decay for adam|lamb\n --clip_nonemb Only clip the gradient of non-embedding params\n --patience PATIENCE Patience\n --eta_min ETA_MIN Min learning rate for cosine scheduler\n\ntraining setup:\n --max_step MAX_STEP Max number of training steps\n --batch_size BATCH_SIZE\n Global batch size\n --local_batch_size LOCAL_BATCH_SIZE\n Local (per-device) batch size, this setting overrides\n global --batch_size and sets batch_size to\n local_batch_size * world_size\n --batch_chunk BATCH_CHUNK\n Split batch into chunks and train with gradient\n accumulation\n --roll Enable random shifts within each data stream\n --tgt_len TGT_LEN Number of tokens to predict\n --ext_len EXT_LEN Length of the extended context\n --mem_len MEM_LEN Length of the retained previous heads\n --seed SEED Random seed\n --multi_gpu {ddp,dp} Use multiple GPU\n --gpu0_bsz GPU0_BSZ Batch size on gpu 0 (for \"dp\" backend)\n --same_length Use the same attn length for all tokens\n --varlen Use variable length\n\nvalidation setup:\n --eval_tgt_len EVAL_TGT_LEN\n Number of tokens to predict for evaluation\n --eval_batch_size EVAL_BATCH_SIZE\n Eval batch size\n --eval_max_steps EVAL_MAX_STEPS\n Max eval steps\n --eval_interval EVAL_INTERVAL\n Evaluation interval\n```\n\n**Inference**\n\nThe complete list of available parameters for the `eval.py` inference\nscript contains:\n\n```\n --work_dir WORK_DIR experiment directory\n --debug run in debug mode (do not create exp dir)\n --data DATA location of the data corpus\n --manual MANUAL [MANUAL ...]\n run model on raw input data\n --dataset {wt103,lm1b,enwik8,text8}\n dataset name\n --split {all,valid,test}\n which split to evaluate\n --type {pytorch,torchscript}\n type of runtime to use\n --batch_size BATCH_SIZE\n batch size\n --tgt_len TGT_LEN number of tokens to predict\n --ext_len EXT_LEN length of the extended context\n --mem_len MEM_LEN length of the retained previous heads\n --seed SEED Random seed\n --clamp_len CLAMP_LEN\n max positional embedding index\n --cuda Run evaluation on a GPU using CUDA\n --model MODEL path to the checkpoint\n --manual_config MANUAL_CONFIG\n Manually specify config for the model\n --manual_vocab {word,bpe}\n Manually specify type of vocabulary\n --fp16 Run training in fp16/mixed precision\n --log_all_ranks Enable logging for all distributed ranks\n --dllog_file DLLOG_FILE\n Name of the DLLogger output file\n --same_length set same length attention with masking\n --no_env Do not print info on execution env\n --log_interval LOG_INTERVAL\n Report interval\n --target_perplexity TARGET_PERPLEXITY\n target perplexity\n --target_throughput TARGET_THROUGHPUT\n target throughput\n --save_data save latency and throughput data to a file\n --repeat REPEAT loop over the dataset REPEAT times\n --max_size MAX_SIZE run inference on up to MAX_SIZE batches\n --percentiles PERCENTILES [PERCENTILES ...]\n percentiles for latency confidence intervals\n --save_torchscript SAVE_TORCHSCRIPT\n save torchscript model to a file\n --load_torchscript LOAD_TORCHSCRIPT\n load torchscript model from a file\n```\n\n### Command-line options\n\nTo see the full list of available options and their descriptions, use the `-h`\nor `--help` command-line option. For example, for training:\n\n```\npython3 train.py --help\n\nusage: train.py [-h] [--work_dir WORK_DIR] [--append_dataset] [--append_time]\n [--cuda] [--fp16] [--restart RESTART] [--debug]\n [--log_all_ranks] [--dllog_file DLLOG_FILE]\n [--txtlog_file TXTLOG_FILE] [--save_all] [--no_env]\n [--no_eval] [--log_interval LOG_INTERVAL]\n [--target_throughput TARGET_THROUGHPUT]\n [--target_perplexity TARGET_PERPLEXITY]\n [--amp_mode {O0,O1,O2,O3}] [--data DATA]\n [--dataset {wt103,lm1b,enwik8,text8}] [--vocab {word,bpe}]\n [--n_layer N_LAYER] [--n_head N_HEAD] [--d_head D_HEAD]\n [--d_embed D_EMBED] [--d_model D_MODEL] [--d_inner D_INNER]\n [--dropout DROPOUT] [--dropatt DROPATT] [--pre_lnorm]\n [--attn_type ATTN_TYPE] [--not_tied] [--clamp_len CLAMP_LEN]\n [--adaptive] [--div_val DIV_VAL]\n [--sample_softmax SAMPLE_SOFTMAX] [--init INIT]\n [--emb_init EMB_INIT] [--init_range INIT_RANGE]\n [--emb_init_range EMB_INIT_RANGE] [--init_std INIT_STD]\n [--proj_init_std PROJ_INIT_STD]\n [--optim {adam,sgd,adagrad,lamb,jitlamb}] [--lr LR]\n [--mom MOM] [--scheduler {cosine,inv_sqrt,dev_perf,constant}]\n [--max_step_scheduler MAX_STEP_SCHEDULER]\n [--warmup_step WARMUP_STEP] [--decay_rate DECAY_RATE]\n [--lr_min LR_MIN] [--clip CLIP] [--weight_decay WEIGHT_DECAY]\n [--clip_nonemb] [--patience PATIENCE] [--eta_min ETA_MIN]\n [--max_step MAX_STEP] [--batch_size BATCH_SIZE]\n [--local_batch_size LOCAL_BATCH_SIZE]\n [--batch_chunk BATCH_CHUNK] [--roll] [--tgt_len TGT_LEN]\n [--ext_len EXT_LEN] [--mem_len MEM_LEN] [--seed SEED]\n [--multi_gpu {ddp,dp}] [--gpu0_bsz GPU0_BSZ] [--same_length]\n [--varlen] [--eval_tgt_len EVAL_TGT_LEN]\n [--eval_batch_size EVAL_BATCH_SIZE]\n [--eval_max_steps EVAL_MAX_STEPS]\n [--eval_interval EVAL_INTERVAL] [--local_rank LOCAL_RANK]\n```\n\nFor example, for inference:\n\n```\npython3 eval.py --help\n\nusage: eval.py [-h] [--work_dir WORK_DIR] [--debug] [--data DATA]\n [--manual MANUAL [MANUAL ...]]\n [--dataset {wt103,lm1b,enwik8,text8}]\n [--split {all,valid,test}] [--type {pytorch,torchscript}]\n [--batch_size BATCH_SIZE] [--tgt_len TGT_LEN]\n [--ext_len EXT_LEN] [--mem_len MEM_LEN] [--seed SEED]\n [--clamp_len CLAMP_LEN] [--cuda] [--model MODEL]\n [--manual_config MANUAL_CONFIG] [--manual_vocab {word,bpe}]\n [--fp16] [--log_all_ranks] [--dllog_file DLLOG_FILE]\n [--same_length] [--no_env] [--log_interval LOG_INTERVAL]\n [--target_perplexity TARGET_PERPLEXITY]\n [--target_throughput TARGET_THROUGHPUT] [--save_data]\n [--repeat REPEAT] [--max_size MAX_SIZE]\n [--percentiles PERCENTILES [PERCENTILES ...]]\n [--save_torchscript SAVE_TORCHSCRIPT]\n [--load_torchscript LOAD_TORCHSCRIPT] [--local_rank LOCAL_RANK]\n```\n\n### Getting the data\n\nThe Transformer-XL model was trained on the\n[WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/)\ndataset. The WikiText-103 dataset is a collection of over 100 million tokens\nextracted from the set of verified\n[Good](https://en.wikipedia.org/wiki/Wikipedia:Good_articles) and\n[Featured](https://en.wikipedia.org/wiki/Wikipedia:Featured_articles) articles\non Wikipedia.\n\nThis repository contains the `getdata.sh` download script which\nautomatically downloads and extracts the training, validation and test\ndatasets. By default, data is downloaded to the `data` directory.\n\nIn order to test with other datasets, the script needs to be customized\naccordingly.\n\n#### Dataset guidelines\n\nThe WikiText-103 dataset was already pre-tokenized with word-level tokens. The\ndataset features a large vocabulary of 267,735 tokens and retains the original\ncase, punctuation and numbers.\n\nThe `getdata.sh` script downloads the data, extracts the archive and renames\nthe training, validation, and test set to `train.txt`, `valid.txt`, `test.txt`\nrespectively.\n\n#### Multi-dataset\n\nUsing other datasets requires changes in the following files:\n\n* `pytorch/train.py`:\n * the name of the new dataset should be added to the `dataset` argument in\n the `parse_args()` function\n * desired values of cutoffs for adaptive softmax should be added in the\n `main()` function, after the section which builds train/valid/test data\n iterators\n* `pytorch/data_utils.py`:\n * the support for the new dataset needs to be added to the `Corpus` class:\n names of files containing training, validation and test data, options for\n the tokenizer, and dataset iterator\n\nThe current codebase supports training with word-level vocabulary\n(automatically generated based on the provided dataset) and with BPE vocabulary\n(using pre-built vocabulary from pretrained GPT2 model imported from\n[github.com/huggingface/transformers](https://github.com/huggingface/transformers).\n\nAdditionally, using other datasets may require changes in some hyperparameters\n(for example, batch size, learning rate, number of training steps,\nand the configuration of learning rate scheduler). \n\n### Training process\n\nThe default training configuration can be launched by running the\n`run_wt103_base.sh` or the `run_wt103_large.sh` script with the first argument\nset to `train`. By default, the training results are saved to the `LM-TFM`\ndirectory; this can be customized by setting the `--work_dir` parameter.\n\nThe training script launches a single-node data-parallel training with a fixed\nglobal batch size of 256, optionally with gradient accumulation to allow\ntraining on configurations with less than 8 GPUs. Logs from the training are\nautomatically saved to the `LM-TFM/train_log.log` file.\n\n**Command-line**\n\nYou can launch training of the Transformer-XL base/large model on the\nWikiText-103 dataset with the word-based vocabulary and adaptive softmax using\n`<#GPUs>` GPUs. For example:\n\n```\nbash run_wt103_base.sh train <#GPUs> [--fp16] [--batch_chunk CHUNK]\n```\n\nand\n\n```\nbash run_wt103_large.sh train <#GPUs> [--fp16] [--batch_chunk CHUNK]\n```\n\nThe `--fp16` flag is optional, however, if it's specified, then the script\nlaunches mixed precision training with Tensor Cores; if the flag is not\npresent, then the script launches FP32 training on NVIDIA Volta GPUs and TF32\ntraining on NVIDIA Ampere GPUs.\n\nThe `--batch_chunk CHUNK` parameter controls gradient accumulation. With\ngradient accumulation, the batch size is split into `CHUNK` chunks of equal\nsize, the training script executes the forward and backward pass using each\nchunk and then executes the optimizer using accumulated gradients.\n\n**Examples**\n\nYou can launch mixed precision training of the Transformer-XL base model on the\nWikiText-103 dataset using 16 GPUs. For example:\n\n```\nbash run_wt103_base.sh train 16 --fp16 --batch_chunk 1\n```\n\nThe batch size per GPU is equal to the default global batch size of 256 divided\nby the product of the number of GPUs times the number of chunks, in this case\nbatch size per GPU is equal to `256 / (16 * 1) = 16`.\n\nYou can launch FP32 training using 8 GPUs; the batch size per GPU is equal to\n16 (`--batch_chunk` was set to `2` because a local batch size of 32 runs out of\nmemory on a NVIDIA DGX-1 with Tesla V100 16GB in FP32 training). For example:\n\n```\nbash run_wt103_base.sh train 8 --batch_chunk 2\n```\n\nA progress summary of the training progress is printed after every 10 training\niterations; this can be customized by setting the `--log_interval` parameter.\nThe summary is printed in the following format:\n\n```\n\n| epoch 18 step 36000 | batches 283 / 2101 | lr 1.220e-03 | ms/batch 185.1 | tok/s 265585 | loss 3.12 | ppl 22.71\n```\n\nwhich contains information about a current training epoch, current training\nstep, number of batches processed within the current epoch, current learning\nrate, execution time in milliseconds per batch, throughput in tokens per\nsecond, current training loss and training perplexity.\n\nThe script saves two checkpoints: `checkpoint_best.pt` which contains the model\ncorresponding to the lowest value of the validation loss and\n`checkpoint_last.pt` which contains the model corresponding to the last\nexecution of the validation step. By default, the validation is executed every\n5000 training steps, this can be customized by setting the `--eval_interval`\nparameter. The summary of results on the validation dataset is printed in the\nfollowing format:\n\n```\n\n| Eval 7 at step 35000 | time: 1.37s | valid loss 3.14 | valid ppl 23.132\n```\n\nwhich contains information about the current epoch, current training step, time\nneeded to execute the validation, current validation loss, and validation\nperplexity.\n\nAt the end of the training, the training script automatically runs evaluation\non the test dataset. This automatic evaluation is executed with values of\n`mem_len` and `tgt_len` hyperparameters inherited from the training setup.\nEvaluation (inference) benefits from longer attention sequences, therefore to\nreproduce perplexity values reported in the [Transformer-XL\npaper](https://arxiv.org/abs/1901.02860), it's necessary to run the final\nevaluation with a dedicated inference script. Refer to the [Inference\nprocess](#inference-process) section for more details.\n\n#### Multi-node\n\nMulti-node runs can be launched on a pyxis/enroot Slurm cluster (see\n[Requirements](#requirements)). To launch a multi-node run, issue the\n`run.sub` script with the following command for an 8-node DGX-2H training, for\nexample:\n\n```\nsbatch run.sub all\n```\n\nThis repository contains a number of predefined configurations to run the\nmulti-node training on DGX-2H nodes. By default, `run.sub` launches 8-node\ntraining. \n\nTo launch multi-node training on `` DGX-2H nodes, run:\n\n```\nCONFIG=dgx2_16gpu_{fp16,fp32} sbatch -N run.sub all\n```\n\n* supported values for `` parameter are: 1, 2, 4, 8\n* configs with `fp16` suffix launch mixed precision training, configs with\n `fp32` suffix launch FP32 training\n\nExamples:\n\nTo launch 4-node mixed-precision training, run:\n\n```\nCONFIG=4dgx2_16gpu_fp16 sbatch -N 4 run.sub all\n```\n\nTo launch 2-node FP32 training, run:\n\n```\nCONFIG=2dgx2_16gpu_fp32 sbatch -N 2 run.sub all\n```\n\nNote that the `run.sub` script is a starting point that has to be adapted\ndepending on the environment. In particular, variables such as `WORK_DIR`\nhandle the location of the workspace in the file system. The variable `CONT`\nshould point to the location of the Transformer-XL Docker container. It's\nassumed that the Docker container built with the `scripts/docker/build.sh`\nscript was pushed to a Docker registry accessible from all compute nodes.\n\nRefer to the contents of the file to see the full list of variables to adjust\nfor your system.\n\n### Inference process\n\nInference can be run by launching the `run_wt103_base.sh` or the\n`run_wt103_large.sh` script with the first argument set to `eval`. Running\ninference requires a pre-trained model checkpoint.\n\nThe script supports single-node multi-GPU inference, each batch is split\nequally among all GPUs running the inference and the loss is averaged over the\nglobal batch. Logs from the inference are automatically saved to the\n`LM-TFM/eval_log.log` file.\n\n**Command-line**\n\nYou can launch inference of the Transformer-XL base/large model on the\nWikiText-103 dataset with the word-based vocabulary and adaptive softmax using\n`<#GPUs>` GPUs. For example:\n\n```\nbash run_wt103_base.sh eval <#GPUs> --model [--fp16] [--type {pytorch, torchscript}]\n```\n\nand\n\n```\nbash run_wt103_large.sh eval <#GPUs> --model [--fp16] [--type {pytorch, torchscript}]\n```\n\nThe `--fp16` flag is optional, however, if it's specified, then the script\nlaunches inference with Tensor Cores; if the flag is not present, then the\nscript launches FP32 inference on NVIDIA Volta and NVIDIA Turing GPUs and TF32\ninference on NVIDIA Ampere GPUs.\n\nThe `--type` flag selects between pure Python PyTorch execution and TorchScript\nexecution.\n\nSupported values for `<#GPUs>` are: 1, 2, 4, 8 for NVIDIA DGX-1 and NVIDIA DGX\nA100 and 1, 2, 4, 8, 16 for NVIDIA DGX-2H.\n\n**Examples**\n\nTo launch TorchScript mixed precision inference on 8 GPUs using a checkpoint\nloaded from `LM-TFM/checkpoint_best.pt`, run:\n```\nbash run_wt103_base.sh eval 8 --model LM-TFM/checkpoint_best.pt --fp16 --type torchscript\n```\n\nTo launch pure Python TF32/FP32 inference on a single GPU using a checkpoint loaded\nfrom `LM-TFM/checkpoint_best.pt`, run:\n\n```\nbash run_wt103_base.sh eval 1 --model LM-TFM/checkpoint_best.pt --type pytorch\n```\n\nAfter the execution, the script prints a summary in the following format:\n\n```\nEvaluating with math fp16 type torchscript bsz 16 tgt_len 64 ext_len 0 mem_len 640 clamp_len 400\nTime : 5.29s, 22.05ms/segment\n====================================================================================================\n\n| test loss 3.15 | test ppl 23.304\n====================================================================================================\n```\n\nwhich contains information about runtime parameters, execution time, loss and\nperplexity on the test dataset.","hasAdvanced","performance","The performance measurements in this document were conducted at the time of\npublication and may not reflect the performance achieved from NVIDIA's latest\nsoftware release. For the most up-to-date performance measurements, go to\n[NVIDIA Data Center Deep Learning Product\nPerformance](https://developer.nvidia.com/deep-learning-performance-training-inference).\n\n### Benchmarking\n\nThe following section shows how to run benchmarks measuring the model\nperformance in training and inference modes.\n\n#### Training performance benchmark\n\nTo benchmark the training performance for a specific local (per-gpu) batch size\n``, with a specific number of GPUs `<#GPUs>` for a specific number of\ntraining iterations ``, run:\n\n```\nbash run_wt103_{base,large}.sh train <#GPUs> --config trainbench --local_batch_size --max_step [--fp16]\n```\n\n* use the `run_wt103_base.sh` script to run the benchmark for the base model,\n and use the `run_wt103_large.sh` script to run the benchmark for the large\n model\n* it's recommended to launch at least 500 training steps to get a reliable\n estimate of training performace.\n* the `--fp16` flag is optional, however, if it's specified, then the script\n launches mixed precision training with Tensor Cores. If the flag is not\n present, then the script launches FP32 training on NVIDIA Volta GPUs and TF32\n training on NVIDIA Ampere GPUs.\n\nFor more information about the available options, refer to the [Training\nprocess](#training-process) section.\n\nThe training script prints information in the following format:\n\n```\n(...)\n\n| epoch 1 step 499 | batches 499 / 16802 | lr 4.990e-03 | ms/batch 219.9 | tok/s 27947 | loss 6.43 | ppl 620.80\n| epoch 1 step 500 | batches 500 / 16802 | lr 5.000e-03 | ms/batch 221.4 | tok/s 27747 | loss 6.42 | ppl 611.70\n-------------------------------------------------------------------------------\n(...)\nTraining time: 1.81 minutes\nTraining throughput: 28508.91 tok/s\n```\n\nThe last two lines contain information on the total training time and on the\naverage training throughput measured in tokens per second.\n\n##### Training performance benchmark for multi-node\n\nTo benchmark the multi-node training performance of the large model on a\nspecific number of DGX-2H nodes `` and a specific local batch size\n``, run:\n\nFor mixed precision:\n```\nFP16=1 LOCAL_BATCH_SIZE= CONFIG=trainbench_multinode sbatch -N run.sub train\n```\n\nFor FP32:\n\n```\nLOCAL_BATCH_SIZE= CONFIG=trainbench_multinode sbatch -N run.sub train\n```\n\n#### Inference performance benchmark\n\nThe inference performance and accuracy benchmarks require a checkpoint from a\ntrained model.\n\nTo benchmark the inference performance on a specific global batch size ``\nwith a specific number of GPUs `<#GPUs>`, run:\n\nFor the base model:\n\n```\nbash run_wt103_base.sh eval <#GPUs> --model --batch_size --save_data [--fp16] [--type {pytorch, torchscript}]\n```\n\nFor the large model:\n\n```\nbash run_wt103_large.sh eval <#GPUs> --model --batch_size --save_data [--fp16] [--type {pytorch, torchscript}]\n```\n\nThe inference script prints information in the following format:\n\n```\nEvaluating with math fp16 type torchscript bsz 16 tgt_len 64 ext_len 0 mem_len 640 clamp_len 400\nTime : 5.25s, 21.88ms/segment\n====================================================================================================\n\n| test loss 3.15 | test ppl 23.304\n====================================================================================================\nThroughput Avg: 46316.64 tok/s\nLatency Avg: 22.09 ms\nLatency 90%: 22.22 ms\nLatency 95%: 22.25 ms\nLatency 99%: 22.37 ms\n====================================================================================================\n```\n\nThe output contains information on the achieved test loss and test perplexity,\naverage inference throughput (measured in tokens per second), average inference\nlatency and latency at 90%, 95% and 99% confidence intervals (measured in\nmilliseconds).\n\nThe `scripts/inference_benchmark.sh` benchmarking script is provided for\nconvenience, it automatically launches TF32/FP32 and FP16 inference for various\nbatch sizes.\n\n### Results\n\nThe following sections provide details on how we achieved our performance and\naccuracy in training and inference.\n\n#### Training accuracy results\n\n##### Training accuracy: NVIDIA DGX A100 (8x A100 40GB)\n\n###### Base model\n\nOur results were obtained by running the `pytorch/run_wt103_base.sh`\ntraining script in the pytorch-20.06-py3 NGC container on NVIDIA DGX A100\nwith 8x A100 40GB GPUs.\n\n|**GPUs**|**Batch Size / GPU**|**Accuracy - TF32 (perplexity)**|**Accuracy - Mixed precision (perplexity)**|**Time to Train - TF32 (minutes)**|**Time to Train - Mixed precision (minutes)**|**Time to Train Speedup (TF32 to Mixed precision)**|\n|-------:|-------------------:|-------------------------------:|------------------------------------------:|---------------------------------:|--------------------------------------------:|--------------------------------------------------:|\n| 8 | 32 | 23.24 | 23.24 | 110 | 76 | 1.45 |\n\n###### Large model\n\nOur results were obtained by running the `pytorch/run_wt103_large.sh`\ntraining script in the pytorch-20.06-py3 NGC container on NVIDIA DGX A100\nwith 8x A100 40GB GPUs.\n\n|**GPUs**|**Batch Size / GPU**|**Accuracy - TF32 (perplexity)**|**Accuracy - Mixed precision (perplexity)**|**Time to Train - TF32 (minutes)**|**Time to Train - Mixed precision (minutes)**|**Time to Train Speedup (TF32 to Mixed precision)**|\n|-------:|-------------------:|-------------------------------:|------------------------------------------:|---------------------------------:|--------------------------------------------:|--------------------------------------------------:|\n| 8 | 8 | 18.18 | 18.18 | 735 | 477 | 1.54 |\n| 8 | 16 | N/A | 18.19 | N/A | 430 | 1.71 |\n\n##### Training accuracy: NVIDIA DGX-1 (8x V100 16GB)\n\n###### Base model\n\nOur results were obtained by running the `pytorch/run_wt103_base.sh`\ntraining script in the pytorch-20.06-py3 NGC container on NVIDIA DGX-1\nwith 8x V100 16GB GPUs.\n\n|**GPUs**|**Batch Size / GPU**|**Accuracy - FP32 (perplexity)**|**Accuracy - Mixed precision (perplexity)**|**Time to Train - FP32 (minutes)**|**Time to Train - Mixed precision (minutes)**|**Time to Train Speedup (FP32 to Mixed precision)**|\n|-------:|-------------------:|-------------------------------:|------------------------------------------:|---------------------------------:|--------------------------------------------:|--------------------------------------------------:|\n| 1 | 16 | 23.12 | 23.13 | 2146 | 960 | 2.24 |\n| 8 | 16 | 23.17 | 23.14 | 316 | 167 | 1.89 |\n| 1 | 32 | N/A | 23.15 | N/A | 766 | 2.80 |\n| 8 | 32 | N/A | 23.18 | N/A | 121 | 2.61 |\n\n###### Large model\n\nOur results were obtained by running the `pytorch/run_wt103_large.sh`\ntraining script in the pytorch-20.06-py3 NGC container on NVIDIA DGX-1\nwith 8x V100 16GB GPUs.\n\n|**GPUs**|**Batch Size / GPU**|**Accuracy - FP32 (perplexity)**|**Accuracy - Mixed precision (perplexity)**|**Time to Train - FP32 (minutes)**|**Time to Train - Mixed precision (minutes)**|**Time to Train Speedup (FP32 to Mixed precision)**|\n|-------:|-------------------:|-------------------------------:|------------------------------------------:|---------------------------------:|--------------------------------------------:|--------------------------------------------------:|\n| 8 | 2 | 18.22 | 18.20 | 2983 | 1480 | 2.01 |\n| 8 | 4 | N/A | 18.17 | N/A | 984 | 3.03 |\n\n##### Training accuracy: NVIDIA DGX-2H (16x V100 32GB)\n\n###### Base model\n\nOur results were obtained by running the `pytorch/run_wt103_base.sh`\ntraining script in the pytorch-20.06-py3 NGC container on NVIDIA DGX-2H\nwith 16x V100 32GB GPUs.\n\n|**GPUs**|**Batch Size / GPU**|**Accuracy - FP32 (perplexity)**|**Accuracy - Mixed precision (perplexity)**|**Time to Train - FP32 (minutes)**|**Time to Train - Mixed precision (minutes)**|**Time to Train Speedup (FP32 to Mixed precision)**|\n|-------:|-------------------:|-------------------------------:|------------------------------------------:|---------------------------------:|--------------------------------------------:|--------------------------------------------------:|\n| 16 | 16 | 23.22 | 23.22 | 149 | 80 | 1.86 |\n\n###### Large model\n\nOur results were obtained by running the `pytorch/run_wt103_large.sh`\ntraining script in the pytorch-20.06-py3 NGC container on NVIDIA DGX-2H\nwith 16x V100 32GB GPUs.\n\n|**GPUs**|**Batch Size / GPU**|**Accuracy - FP32 (perplexity)**|**Accuracy - Mixed precision (perplexity)**|**Time to Train - FP32 (minutes)**|**Time to Train - Mixed precision (minutes)**|**Time to Train Speedup (FP32 to Mixed precision)**|\n|-------:|-------------------:|-------------------------------:|------------------------------------------:|---------------------------------:|--------------------------------------------:|--------------------------------------------------:|\n| 16 | 8 | 18.21 | 18.20 | 1075 | 394 | 2.73 |\n\n##### Training accuracy: 8x NVIDIA DGX-2H (16x V100 32GB)\n\n###### Large model\n\nOur results were obtained by running the `pytorch/run.sub`\ntraining script in the pytorch-20.06-py3 NGC container on 8x NVIDIA DGX-2H\nwith 16x V100 32GB GPUs.\n\n|**DGX System**|**Nodes**|**Batch Size / GPU**|**Accuracy - FP32 (perplexity)**|**Accuracy - Mixed precision (perplexity)**|**Time to Train - FP32 (minutes)**|**Time to Train - Mixed precision (minutes)**|**Time to Train Speedup (FP32 to Mixed precision)**|\n|-------------:|--------:|-------------------:|-------------------------------:|------------------------------------------:|---------------------------------:|--------------------------------------------:|--------------------------------------------------:|\n| DGX-2H | 8 | 4 | 18.27 | 18.28 | 156 | 74 | 2.11 |\n\n##### Training accuracy plots\n\n###### Base model\n\n![TrainingLossBase](https://raw.githubusercontent.com/NVIDIA/DeepLearningExamples/master/PyTorch/LanguageModeling/Transformer-XL//pytorch/img/training_loss_base.png)\n\n\n###### Large model (single-node)\n\n![TrainingLossLarge](https://raw.githubusercontent.com/NVIDIA/DeepLearningExamples/master/PyTorch/LanguageModeling/Transformer-XL//pytorch/img/training_loss_large.png)\n\n\n###### Large model (multi-node)\n\n![TrainingLossLargeMultiNode](https://raw.githubusercontent.com/NVIDIA/DeepLearningExamples/master/PyTorch/LanguageModeling/Transformer-XL//pytorch/img/training_loss_large_multinode.png)\n\n\n##### Training stability test\n\n###### Base model\n\nThe Transformer-XL base model was trained for 40,000 training steps, starting\nfrom 16 different initial random seeds. After every 5,000 training steps, the\nmodel was evaluated on the validation dataset and validation perplexity was\nrecorded. The training was performed in the pytorch-20.06-py3 NGC container on\nNVIDIA DGX A100 with 8x A100 40GB GPUs. The following table summarizes the\nperplexity of our validation dataset.\n\n|**Training step**|**Average perplexity**|**Standard deviation**|**Minimum**|**Maximum**|**Median**|\n|----------------:|----------:|---------------------:|----------:|----------:|---------:|\n| 5000 | 42.62 | 0.27311 | 42.01 | 43.09 | 42.67 |\n| 10000 | 32.31 | 0.12814 | 32.10 | 32.59 | 32.31 |\n| 15000 | 28.38 | 0.10764 | 28.23 | 28.57 | 28.35 |\n| 20000 | 26.14 | 0.10218 | 25.96 | 26.36 | 26.14 |\n| 25000 | 24.59 | 0.09060 | 24.42 | 24.81 | 24.60 |\n| 30000 | 23.71 | 0.07259 | 23.61 | 23.84 | 23.71 |\n| 35000 | 23.15 | 0.04781 | 23.05 | 23.26 | 23.15 |\n| 40000 | 22.93 | 0.05593 | 22.83 | 23.04 | 22.94 |\n\nAfter training, the models were evaluated on the test dataset. The following\ntable summarizes the final perplexity on the test set.\n\n|**Average perplexity**|**Standard deviation**|**Minimum**|**Maximum**|**Median**|\n|----------:|---------------------:|----------:|----------:|---------:|\n| 23.24| 0.07794| 23.11| 23.38| 23.25|\n\n###### Large model (single-node)\n\nThe Transformer-XL large model was trained for 100,000 training steps, starting\nfrom 16 different initial random seeds. After every 10,000 training steps, the\nmodel was evaluated on the validation dataset and validation perplexity was\nrecorded. The training was performed in the pytorch-20.06-py3 NGC container on\nNVIDIA DGX A100 with 8x A100 40GB GPUs. The following table summarizes the\nperplexity of our validation dataset.\n\n|**Training step**|**Average perplexity**|**Standard deviation**|**Minimum**|**Maximum**|**Median**|\n|----------------:|----------:|---------------------:|----------:|----------:|---------:|\n| 10000 | 32.63 | 0.20432 | 32.34 | 33.05 | 32.62 |\n| 20000 | 24.08 | 0.10980 | 23.90 | 24.28 | 24.10 |\n| 30000 | 21.52 | 0.09069 | 21.36 | 21.66 | 21.52 |\n| 40000 | 20.17 | 0.06922 | 20.06 | 20.27 | 20.17 |\n| 50000 | 19.23 | 0.05975 | 19.11 | 19.33 | 19.24 |\n| 60000 | 18.57 | 0.06008 | 18.47 | 18.72 | 18.56 |\n| 70000 | 18.17 | 0.06473 | 18.08 | 18.32 | 18.15 |\n| 80000 | 17.95 | 0.06506 | 17.82 | 18.08 | 17.94 |\n| 90000 | 17.80 | 0.04350 | 17.71 | 17.90 | 17.80 |\n| 100000 | 17.80 | 0.03592 | 17.74 | 17.86 | 17.81 |\n\nAfter training, the models were evaluated on the test dataset. The following\ntable summarizes the final perplexity on the test set.\n\n|**Average perplexity**|**Standard deviation**|**Minimum**|**Maximum**|**Median**|\n|---------------------:|---------------------:|----------:|----------:|---------:|\n| 18.17 | 0.04016 | 18.09 | 18.24 | 18.17 |\n\n###### Large model (multi-node)\n\nThe Transformer-XL large model was trained for 25,000 training steps, starting\nfrom 10 different initial random seeds. After every 1,000 training steps, the\nmodel was evaluated on the validation dataset and validation perplexity was\nrecorded. The training was performed in the pytorch-20.06-py3 NGC container on\n8x NVIDIA DGX-2H with 16x V100 32GB GPUs. The following table summarizes the\nperplexity of our validation dataset.\n\n|**Training step**|**Average perplexity**|**Standard deviation**|**Minimum**|**Maximum**|**Median**|\n|----------------:|----------:|---------------------:|----------:|----------:|---------:|\n| 1000 | 608.09 | 3.80116 | 600.65 | 613.73 | 609.40 |\n| 2000 | 142.75 | 0.94452 | 141.21 | 143.84 | 143.07 |\n| 3000 | 62.19 | 0.44544 | 61.38 | 63.01 | 62.18 |\n| 4000 | 40.22 | 0.16397 | 39.93 | 40.54 | 40.20 |\n| 5000 | 32.00 | 0.15850 | 31.61 | 32.19 | 32.02 |\n| 6000 | 28.05 | 0.17854 | 27.81 | 28.41 | 28.05 |\n| 7000 | 25.65 | 0.10946 | 25.51 | 25.87 | 25.65 |\n| 8000 | 24.20 | 0.11385 | 23.98 | 24.36 | 24.20 |\n| 9000 | 23.18 | 0.14936 | 22.84 | 23.37 | 23.20 |\n| 10000 | 22.88 | 0.22752 | 22.54 | 23.33 | 22.94 |\n| 11000 | 21.99 | 0.16232 | 21.73 | 22.29 | 21.97 |\n| 12000 | 21.69 | 0.10824 | 21.46 | 21.81 | 21.73 |\n| 13000 | 21.42 | 0.09154 | 21.25 | 21.57 | 21.44 |\n| 14000 | 21.33 | 0.13821 | 21.15 | 21.55 | 21.27 |\n| 15000 | 21.24 | 0.15526 | 20.95 | 21.57 | 21.20 |\n| 16000 | 21.19 | 0.10521 | 21.01 | 21.44 | 21.18 |\n| 17000 | 20.89 | 0.18239 | 20.69 | 21.18 | 20.82 |\n| 18000 | 20.36 | 0.10715 | 20.21 | 20.53 | 20.34 |\n| 19000 | 19.74 | 0.12803 | 19.45 | 19.92 | 19.75 |\n| 20000 | 19.18 | 0.10020 | 19.05 | 19.39 | 19.15 |\n| 21000 | 18.49 | 0.06319 | 18.36 | 18.60 | 18.49 |\n| 22000 | 18.17 | 0.03674 | 18.11 | 18.22 | 18.16 |\n| 23000 | 17.98 | 0.03682 | 17.90 | 18.04 | 17.99 |\n| 24000 | 17.88 | 0.02880 | 17.84 | 17.92 | 17.89 |\n| 25000 | 17.85 | 0.02793 | 17.80 | 17.90 | 17.86 |\n\nAfter training, the models were evaluated on the test dataset. The following\ntable summarizes the final perplexity on the test set.\n\n|**Average perplexity**|**Standard deviation**|**Minimum**|**Maximum**|**Median**|\n|----------:|---------------------:|----------:|----------:|---------:|\n| 18.30 | 0.02747 | 18.24 | 18.33 | 18.30 |\n\n#### Training performance results\n\n##### Training performance: NVIDIA DGX A100 (8x A100 40GB)\n\n###### Base model\n\nOur results were obtained by running the `pytorch/run_wt103_base.sh` training\nscript in the pytorch-20.06-py3 NGC container on NVIDIA DGX A100 with 8x A100\n40GB GPUs. Performance numbers (in tokens per second) were averaged over 500\ntraining iterations.\n\n|**GPUs**|**Batch Size / GPU**|**Throughput - TF32 (tok/s)**|**Throughput - Mixed precision (tok/s)**|**Throughput speedup (TF32 to Mixed precision)**|**Weak Scaling - TF32**|**Weak Scaling - Mixed precision**|\n|-------:|-------------------:|----------------------------:|---------------------------------------:|-----------------------------------------------:|----------------------:|---------------------------------:|\n| 1 | 32 | 41,527 | 59,961 | 1.444 | 1.000 | 1.000 |\n| 2 | 32 | 77,625 | 113,238 | 1.459 | 1.869 | 1.889 |\n| 4 | 32 | 153,945 | 225,609 | 1.466 | 3.707 | 3.763 |\n| 8 | 32 | 305,933 | 449,890 | 1.471 | 7.367 | 7.503 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and setup \nthe container, and then proceed to the \n[Training performance benchmark](#training-performance-benchmark) section for \ninstruction on how to launch the benchmark.\n\n###### Large model\n\nOur results were obtained by running the `pytorch/run_wt103_large.sh` training\nscript in the pytorch-20.06-py3 NGC container on NVIDIA DGX A100 with 8x A100\n40GB GPUs. Performance numbers (in tokens per second) were averaged over 500\ntraining iterations.\n\n|**GPUs**|**Batch Size / GPU**|**Throughput - TF32 (tok/s)**|**Throughput - Mixed precision (tok/s)**|**Throughput speedup (TF32 to Mixed precision)**|**Weak Scaling - TF32**|**Weak Scaling - Mixed precision**|\n|-------:|-------------------:|----------------------------:|---------------------------------------:|-----------------------------------------------:|----------------------:|---------------------------------:|\n| 1 | 8 | 14,497 | 21,554 | 1.487 | 1.000 | 1.000 |\n| 2 | 8 | 27,304 | 40,222 | 1.473 | 1.883 | 1.866 |\n| 4 | 8 | 53,756 | 80,226 | 1.492 | 3.708 | 3.722 |\n| 8 | 8 | 106,651 | 159,185 | 1.493 | 7.357 | 7.385 |\n| 1 | 16 | N/A | 25,084 | 1.730 | N/A | 1.000 |\n| 2 | 16 | N/A | 48,562 | 1.779 | N/A | 1.936 |\n| 4 | 16 | N/A | 95,997 | 1.786 | N/A | 3.827 |\n| 8 | 16 | N/A | 191,148 | 1.792 | N/A | 7.620 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and setup \nthe container, and then proceed to the \n[Training performance benchmark](#training-performance-benchmark) section for \ninstruction on how to launch the benchmark.\n\n##### Training performance: NVIDIA DGX-1 (8x V100 16GB)\n\n###### Base model\n\nOur results were obtained by running the `pytorch/run_wt103_base.sh` training\nscript in the pytorch-20.06-py3 NGC container on NVIDIA DGX-1 with 8x V100 16GB\nGPUs. Performance numbers (in tokens per second) were averaged over 500\ntraining iterations.\n\n|**GPUs**|**Batch Size / GPU**|**Throughput - FP32 (tok/s)**|**Throughput - Mixed precision (tok/s)**|**Throughput speedup (FP32 to Mixed precision)**|**Weak Scaling - FP32**|**Weak Scaling - Mixed precision**|\n|-------:|-------------------:|----------------------------:|---------------------------------------:|-----------------------------------------------:|----------------------:|---------------------------------:|\n| 1 | 16 | 13,981 | 26,639 | 1.905 | 1.000 | 1.000 |\n| 2 | 16 | 23,163 | 45,299 | 1.956 | 1.657 | 1.700 |\n| 4 | 16 | 48,893 | 92,618 | 1.894 | 3.497 | 3.477 |\n| 8 | 16 | 97,005 | 170,532 | 1.758 | 6.938 | 6.402 |\n| 1 | 32 | N/A | 36,692 | 2.624 | N/A | 1.000 |\n| 2 | 32 | N/A | 65,889 | 2.845 | N/A | 1.796 |\n| 4 | 32 | N/A | 133,838 | 2.737 | N/A | 3.648 |\n| 8 | 32 | N/A | 258,648 | 2.666 | N/A | 7.049 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and setup \nthe container, and then proceed to the \n[Training performance benchmark](#training-performance-benchmark) section for \ninstruction on how to launch the benchmark.\n\n###### Large model\n\nOur results were obtained by running the `pytorch/run_wt103_large.sh` training\nscript in the pytorch-20.06-py3 NGC container on NVIDIA DGX-1 with 8x V100 16GB\nGPUs. Performance numbers (in tokens per second) were averaged over 500\ntraining iterations.\n\n|**GPUs**|**Batch Size / GPU**|**Throughput - FP32 (tok/s)**|**Throughput - Mixed precision (tok/s)**|**Throughput speedup (FP32 to Mixed precision)**|**Weak Scaling - FP32**|**Weak Scaling - Mixed precision**|\n|-------:|-------------------:|----------------------------:|---------------------------------------:|-----------------------------------------------:|----------------------:|---------------------------------:|\n| 1 | 2 | 3,558 | 6,907 | 1.941 | 1.000 | 1.000 |\n| 2 | 2 | 6,153 | 11,272 | 1.832 | 1.729 | 1.632 |\n| 4 | 2 | 12,492 | 22,530 | 1.804 | 3.511 | 3.262 |\n| 8 | 2 | 24,595 | 40,920 | 1.664 | 6.913 | 5.925 |\n| 1 | 4 | N/A | 10,210 | 2.870 | N/A | 1.000 |\n| 2 | 4 | N/A | 17,984 | 2.923 | N/A | 1.761 |\n| 4 | 4 | N/A | 36,340 | 2.909 | N/A | 3.559 |\n| 8 | 4 | N/A | 66,716 | 2.713 | N/A | 6.535 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and setup \nthe container, and then proceed to the \n[Training performance benchmark](#training-performance-benchmark) section for \ninstruction on how to launch the benchmark.\n\n##### Training performance: NVIDIA DGX-2H (16x V100 32GB)\n\n###### Base model\n\nOur results were obtained by running the `pytorch/run_wt103_base.sh` training\nscript in the pytorch-20.06-py3 NGC container on NVIDIA DGX-2H with 16x V100\n32GB GPUs. Performance numbers (in tokens per second) were averaged over 500\ntraining iterations.\n\n|**GPUs**|**Batch Size / GPU**|**Throughput - FP32 (tok/s)**|**Throughput - Mixed precision (tok/s)**|**Throughput speedup (FP32 to Mixed precision)**|**Weak Scaling - FP32**|**Weak Scaling - Mixed precision**|\n|-------:|-------------------:|----------------------------:|---------------------------------------:|-----------------------------------------------:|----------------------:|---------------------------------:|\n| 1 | 16 | 16,150 | 32,875 | 2.036 | 1.000 | 1.000 |\n| 2 | 16 | 29,712 | 59,058 | 1.988 | 1.840 | 1.796 |\n| 4 | 16 | 58,011 | 113,985 | 1.965 | 3.592 | 3.467 |\n| 8 | 16 | 114,655 | 223,907 | 1.953 | 7.099 | 6.811 |\n| 16 | 16 | 222,920 | 414,994 | 1.862 | 13.803 | 12.623 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and setup \nthe container, and then proceed to the \n[Training performance benchmark](#training-performance-benchmark) section for \ninstruction on how to launch the benchmark.\n\n###### Large model\n\nOur results were obtained by running the `pytorch/run_wt103_large.sh` training\nscript in the pytorch-20.06-py3 NGC container on NVIDIA DGX-2H with 16x V100\n32GB GPUs. Performance numbers (in tokens per second) were averaged over 500\ntraining iterations.\n\n|**GPUs**|**Batch Size / GPU**|**Throughput - FP32 (tok/s)**|**Throughput - Mixed precision (tok/s)**|**Throughput speedup (FP32 to Mixed precision)**|**Weak Scaling - FP32**|**Weak Scaling - Mixed precision**|\n|-------:|-------------------:|----------------------------:|---------------------------------------:|-----------------------------------------------:|----------------------:|---------------------------------:|\n| 1 | 8 | 5,169 | 14,787 | 2.861 | 1.000 | 1.000 |\n| 2 | 8 | 9,977 | 27,710 | 2.777 | 1.930 | 1.874 |\n| 4 | 8 | 19,691 | 54,207 | 2.753 | 3.810 | 3.666 |\n| 8 | 8 | 39,157 | 107,073 | 2.734 | 7.576 | 7.241 |\n| 16 | 8 | 77,568 | 211,387 | 2.725 | 15.008 | 14.296 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and setup \nthe container, and then proceed to the \n[Training performance benchmark](#training-performance-benchmark) section for \ninstruction on how to launch the benchmark.\n\n##### Training performance: 8x NVIDIA DGX-2H (16x V100 32GB)\n\nOur results were obtained by running the `pytorch/run.sub` training script in\nthe pytorch-20.06-py3 NGC container. Performance numbers (in tokens per second)\nwere averaged over 500 training iterations.\n\n###### Large model\n\n|**DGX System**|**Nodes**|**Batch Size / GPU**|**Throughput - FP32 (tok/s)**|**Throughput - Mixed precision (tok/s)**|**Throughput speedup (FP32 to Mixed precision)**|**Weak Scaling - FP32**|**Weak scaling - Mixed precision**|\n|-------------:|--------:|-------------------:|-------------------------------:|------------------------------------------:|---------------------------------:|--------------------------------------------:|--------------------------------------------------:|\n| DGX-2H | 1 | 4 | 69,070 | 154,950 | 2.24 | 1.00 | 1.00 |\n| DGX-2H | 2 | 4 | 136,960 | 307,520 | 2.25 | 1.98 | 1.98 |\n| DGX-2H | 4 | 4 | 270,120 | 605,530 | 2.24 | 3.91 | 3.91 |\n| DGX-2H | 8 | 4 | 514,500 | 1,189,700 | 2.31 | 7.45 | 7.68 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and then\nproceed to the\n[Training performance benchmark for\nmulti-node](#training-performance-benchmark-for-multi-node) section for\ninstruction on how to launch the multi-node performance benchmark. The numbers\npresented above were obtained with `LOCAL_BATCH_SIZE=4`.\n\n#### Inference performance results\n\n##### Inference performance: NVIDIA DGX A100 (1x A100 40GB)\n\n###### Base model\n\nOur results were obtained by running the\n`pytorch/scripts/inference_benchmark.sh` inferencing benchmarking script in the\npytorch-20.06-py3 NGC container on NVIDIA DGX A100 with 1x A100 40GB GPU.\n\nThe command to launch the inference performance benchmark is provided in the\n[Inference performance benchmark](#inference-performance-benchmark) section.\n\n**FP16, pure Python**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 64 | 640 | 4,163.7 | 15.38 | 15.58 | 15.66 | 16.12 |\n| 2 | 64 | 640 | 7,915.4 | 16.17 | 16.36 | 16.42 | 17.19 |\n| 4 | 64 | 640 | 15,710.2 | 16.29 | 16.45 | 16.49 | 17.38 |\n| 8 | 64 | 640 | 32,712.1 | 15.64 | 15.77 | 15.82 | 16.65 |\n| 16 | 64 | 640 | 59,378.6 | 17.23 | 17.32 | 17.36 | 18.39 |\n| 32 | 64 | 640 | 91,654.2 | 22.33 | 22.39 | 22.53 | 23.63 |\n\n**FP16, TorchScript**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 64 | 640 | 6,935.9 | 9.231 | 9.388 | 9.445 | 9.534 |\n| 2 | 64 | 640 | 12,649.4 | 10.120 | 10.253 | 10.294 | 10.945 |\n| 4 | 64 | 640 | 25,029.5 | 10.223 | 10.346 | 10.381 | 10.475 |\n| 8 | 64 | 640 | 52,666.3 | 9.716 | 9.808 | 9.851 | 10.540 |\n| 16 | 64 | 640 | 90,767.8 | 11.274 | 11.321 | 11.334 | 11.800 |\n| 32 | 64 | 640 | 107,082.4 | 19.109 | 19.138 | 19.162 | 19.608 |\n\n**TF32, pure Python**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 64 | 640 | 4,003.8 | 15.99 | 16.26 | 16.36 | 16.58 |\n| 2 | 64 | 640 | 7,499.2 | 17.07 | 17.32 | 17.39 | 17.86 |\n| 4 | 64 | 640 | 14,835.4 | 17.25 | 17.46 | 17.50 | 18.34 |\n| 8 | 64 | 640 | 30,001.5 | 17.06 | 17.22 | 17.28 | 18.40 |\n| 16 | 64 | 640 | 50,189.7 | 20.39 | 20.48 | 20.52 | 21.41 |\n| 32 | 64 | 640 | 63,660.5 | 32.14 | 32.17 | 32.29 | 33.19 |\n\n**TF32, TorchScript**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 64 | 640 | 6,084.5 | 10.52 | 10.74 | 10.84 | 10.95 |\n| 2 | 64 | 640 | 11,680.6 | 10.96 | 11.17 | 11.22 | 11.76 |\n| 4 | 64 | 640 | 22,867.3 | 11.19 | 11.35 | 11.40 | 12.07 |\n| 8 | 64 | 640 | 45,165.5 | 11.33 | 11.46 | 11.49 | 12.03 |\n| 16 | 64 | 640 | 61,042.0 | 16.76 | 16.84 | 16.86 | 17.13 |\n| 32 | 64 | 640 | 71,124.1 | 28.77 | 28.81 | 28.84 | 28.86 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and setup \nthe container, and then proceed to the \n[Inference performance benchmark](#inference-performance-benchmark) section for \ninstruction on how to launch the benchmark.\n\n###### Large model\n\nOur results were obtained by running the\n`pytorch/scripts/inference_benchmark.sh` inferencing benchmarking script in the\npytorch-20.06-py3 NGC container on NVIDIA DGX A100 with 1x A100 40GB GPU.\n\nThe command to launch the inference performance benchmark is provided in the\n[Inference performance benchmark](#inference-performance-benchmark) section.\n\n**FP16, pure Python**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 128 | 1,600 | 7,033.0 | 18.20 | 18.57 | 18.64 | 18.93 |\n| 2 | 128 | 1,600 | 12,832.5 | 19.94 | 20.23 | 20.29 | 21.07 |\n| 4 | 128 | 1,600 | 21,500.2 | 23.80 | 23.99 | 24.07 | 25.09 |\n| 8 | 128 | 1,600 | 25,797.1 | 39.66 | 39.74 | 39.91 | 41.00 |\n| 16 | 128 | 1,600 | 28,143.5 | 72.71 | 72.74 | 73.12 | 74.00 |\n| 32 | 128 | 1,600 | 28,533.6 | 143.44 | 143.30 | 143.48 | 149.07 |\n\n**FP16, TorchScript**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 128 | 1,600 | 11,068.2 | 11.57 | 11.83 | 11.88 | 12.42 |\n| 2 | 128 | 1,600 | 19,847.0 | 12.89 | 13.09 | 13.11 | 13.27 |\n| 4 | 128 | 1,600 | 24,450.7 | 20.92 | 21.08 | 21.10 | 21.15 |\n| 8 | 128 | 1,600 | 27,938.4 | 36.62 | 36.72 | 36.75 | 36.86 |\n| 16 | 128 | 1,600 | 30,783.0 | 66.48 | 66.54 | 66.59 | 66.98 |\n| 32 | 128 | 1,600 | 32,161.6 | 127.26 | 127.19 | 127.34 | 131.64 |\n\n**TF32, pure Python**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 128 | 1,600 | 6,558.8 | 19.52 | 19.87 | 19.95 | 20.44 |\n| 2 | 128 | 1,600 | 10,658.4 | 24.00 | 24.28 | 24.36 | 25.17 |\n| 4 | 128 | 1,600 | 14,769.6 | 34.64 | 34.82 | 34.89 | 35.74 |\n| 8 | 128 | 1,600 | 16,852.6 | 60.71 | 60.82 | 61.05 | 62.17 |\n| 16 | 128 | 1,600 | 18,071.8 | 113.23 | 113.28 | 113.37 | 114.64 |\n| 32 | 128 | 1,600 | 17,619.2 | 234.04 | 229.98 | 239.30 | 328.15 |\n\n**TF32, TorchScript**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 128 | 1,600 | 9,084.4 | 14.09 | 14.37 | 14.40 | 14.46 |\n| 2 | 128 | 1,600 | 12,839.4 | 19.92 | 20.15 | 20.17 | 20.25 |\n| 4 | 128 | 1,600 | 15,582.4 | 32.83 | 33.00 | 33.02 | 33.28 |\n| 8 | 128 | 1,600 | 17,825.0 | 57.40 | 57.55 | 57.59 | 57.94 |\n| 16 | 128 | 1,600 | 19,419.2 | 105.38 | 105.49 | 105.54 | 105.91 |\n| 32 | 128 | 1,600 | 20,079.4 | 203.81 | 203.77 | 203.84 | 207.47 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and setup \nthe container, and then proceed to the \n[Inference performance benchmark](#inference-performance-benchmark) section for \ninstruction on how to launch the benchmark.\n\n##### Inference performance: NVIDIA DGX-1 (1x V100 16GB)\n\n###### Base model\n\nOur results were obtained by running the\n`pytorch/scripts/inference_benchmark.sh` inferencing benchmarking script in the\npytorch-20.06-py3 NGC container on NVIDIA DGX-1 with 1x V100 16GB GPU.\n\nThe command to launch the inference performance benchmark is provided in the\n[Inference performance benchmark](#inference-performance-benchmark) section.\n\n**FP16, pure Python**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 64 | 640 | 2,999.6 | 21.36 | 21.72 | 21.90 | 24.86 |\n| 2 | 64 | 640 | 5,738.5 | 22.32 | 22.64 | 22.89 | 25.97 |\n| 4 | 64 | 640 | 11,773.5 | 21.73 | 21.92 | 22.06 | 22.68 |\n| 8 | 64 | 640 | 22,604.7 | 22.63 | 22.92 | 23.08 | 23.56 |\n| 16 | 64 | 640 | 41,481.6 | 24.67 | 24.83 | 24.99 | 25.73 |\n| 32 | 64 | 640 | 58,556.9 | 34.95 | 35.13 | 35.24 | 35.85 |\n\n**FP16, TorchScript**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 64 | 640 | 5,199.9 | 12.31 | 12.59 | 12.65 | 12.98 |\n| 2 | 64 | 640 | 9,802.5 | 13.06 | 13.30 | 13.42 | 13.82 |\n| 4 | 64 | 640 | 19,609.4 | 13.05 | 13.17 | 13.24 | 13.94 |\n| 8 | 64 | 640 | 37,598.7 | 13.61 | 13.71 | 13.77 | 14.62 |\n| 16 | 64 | 640 | 57,840.2 | 17.69 | 17.73 | 17.76 | 18.36 |\n| 32 | 64 | 640 | 66,955.9 | 30.57 | 30.78 | 30.86 | 30.96 |\n\n**FP32, pure Python**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 64 | 640 | 2,940.0 | 21.79 | 22.23 | 22.42 | 25.52 |\n| 2 | 64 | 640 | 5,652.0 | 22.66 | 23.00 | 23.20 | 26.86 |\n| 4 | 64 | 640 | 10,526.0 | 24.30 | 24.62 | 24.72 | 25.03 |\n| 8 | 64 | 640 | 15,767.2 | 32.45 | 32.67 | 32.78 | 33.32 |\n| 16 | 64 | 640 | 20,303.2 | 50.39 | 50.82 | 50.89 | 51.07 |\n| 32 | 64 | 640 | 21,707.1 | 94.26 | 94.76 | 94.94 | 95.26 |\n\n**FP32, TorchScript**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 64 | 640 | 4,974.1 | 12.88 | 13.25 | 13.37 | 13.69 |\n| 2 | 64 | 640 | 9,625.3 | 13.30 | 13.58 | 13.72 | 14.15 |\n| 4 | 64 | 640 | 15,069.9 | 16.98 | 17.27 | 17.35 | 17.54 |\n| 8 | 64 | 640 | 18,269.8 | 28.00 | 28.23 | 28.28 | 28.37 |\n| 16 | 64 | 640 | 20,884.5 | 48.99 | 49.46 | 49.50 | 49.63 |\n| 32 | 64 | 640 | 22,289.2 | 91.80 | 92.25 | 92.56 | 92.67 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and setup \nthe container, and then proceed to the \n[Inference performance benchmark](#inference-performance-benchmark) section for \ninstruction on how to launch the benchmark.\n\n###### Large model\n\nOur results were obtained by running the\n`pytorch/scripts/inference_benchmark.sh` inferencing benchmarking script in the\npytorch-20.06-py3 NGC container on NVIDIA DGX-1 with 1x V100 16GB GPU.\n\nThe command to launch the inference performance benchmark is provided in the\n[Inference performance benchmark](#inference-performance-benchmark) section.\n\n**FP16, pure Python**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 128 | 1,600 | 5,119.6 | 25.00 | 25.47 | 25.66 | 26.12 |\n| 2 | 128 | 1,600 | 8,676.1 | 29.49 | 29.81 | 29.94 | 30.88 |\n| 4 | 128 | 1,600 | 12,960.9 | 39.47 | 39.84 | 39.91 | 40.69 |\n| 8 | 128 | 1,600 | 14,870.6 | 68.81 | 69.28 | 69.42 | 69.76 |\n| 16 | 128 | 1,600 | 15,528.5 | 131.78 | 132.74 | 132.86 | 133.07 |\n| 32 | 128 | 1,600 | 15,649.4 | 261.54 | 262.45 | 262.99 | 271.10 |\n\n**FP16, TorchScript**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 128 | 1,600 | 8,718.2 | 14.68 | 15.01 | 15.07 | 15.50 |\n| 2 | 128 | 1,600 | 12,157.8 | 21.04 | 21.29 | 21.31 | 21.38 |\n| 4 | 128 | 1,600 | 14,534.8 | 35.20 | 35.48 | 35.53 | 35.93 |\n| 8 | 128 | 1,600 | 15,863.8 | 64.50 | 64.90 | 65.15 | 65.31 |\n| 16 | 128 | 1,600 | 16,674.0 | 122.73 | 123.34 | 123.66 | 123.92 |\n| 32 | 128 | 1,600 | 17,154.1 | 238.60 | 239.48 | 239.73 | 247.48 |\n\n**FP32, pure Python**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 128 | 1,600 | 3,009.8 | 42.52 | 43.01 | 43.09 | 43.53 |\n| 2 | 128 | 1,600 | 3,838.4 | 66.64 | 67.24 | 67.45 | 67.83 |\n| 4 | 128 | 1,600 | 4,265.3 | 119.94 | 120.87 | 121.00 | 121.39 |\n| 8 | 128 | 1,600 | 4,646.5 | 220.19 | 221.30 | 221.50 | 221.68 |\n| 16 | 128 | 1,600 | 4,805.4 | 426.39 | 426.25 | 426.47 | 427.25 |\n| 32 | 128 | 1,600 | 4,787.4 | 855.09 | 854.95 | 855.46 | 912.05 |\n\n**FP32, TorchScript**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 128 | 1,600 | 3,319.0 | 38.56 | 38.91 | 39.01 | 39.19 |\n| 2 | 128 | 1,600 | 3,925.2 | 65.16 | 65.74 | 65.89 | 66.12 |\n| 4 | 128 | 1,600 | 4,344.1 | 117.76 | 118.46 | 118.55 | 118.69 |\n| 8 | 128 | 1,600 | 4,716.2 | 216.94 | 217.99 | 218.27 | 218.69 |\n| 16 | 128 | 1,600 | 4,922.1 | 415.72 | 417.16 | 417.32 | 417.59 |\n| 32 | 128 | 1,600 | 4,965.2 | 824.98 | 821.79 | 831.71 | 952.47 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and setup \nthe container, and then proceed to the \n[Inference performance benchmark](#inference-performance-benchmark) section for \ninstruction on how to launch the benchmark.\n\n##### Inference performance: NVIDIA T4\n\n###### Base model\n\nOur results were obtained by running the\n`pytorch/scripts/inference_benchmark.sh` inferencing benchmarking script in the\npytorch-20.06-py3 NGC container on NVIDIA T4.\n\nThe command to launch the inference performance benchmark is provided in the\n[Inference performance benchmark](#inference-performance-benchmark) section.\n\n**FP16, pure Python**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 64 | 640 | 3,775.3 | 16.97 | 17.51 | 17.84 | 18.18 |\n| 2 | 64 | 640 | 6,417.4 | 19.96 | 20.49 | 20.56 | 21.52 |\n| 4 | 64 | 640 | 9,988.6 | 25.64 | 26.07 | 26.14 | 27.32 |\n| 8 | 64 | 640 | 11,878.9 | 43.07 | 43.42 | 43.46 | 44.24 |\n| 16 | 64 | 640 | 13,630.0 | 75.07 | 75.26 | 75.32 | 76.07 |\n| 32 | 64 | 640 | 14,511.2 | 141.01 | 141.38 | 141.41 | 142.16 |\n\n**FP16, TorchScript**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 64 | 640 | 6,132.5 | 10.47 | 10.93 | 11.31 | 11.45 |\n| 2 | 64 | 640 | 8,319.4 | 15.39 | 15.89 | 15.92 | 16.10 |\n| 4 | 64 | 640 | 11,259.1 | 22.74 | 23.16 | 23.23 | 23.30 |\n| 8 | 64 | 640 | 13,120.3 | 38.99 | 39.35 | 39.37 | 39.42 |\n| 16 | 64 | 640 | 15,120.0 | 67.67 | 67.90 | 67.94 | 68.06 |\n| 32 | 64 | 640 | 16,158.1 | 126.65 | 126.97 | 127.03 | 127.18 |\n\n**FP32, pure Python**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 64 | 640 | 2,323.1 | 27.59 | 29.39 | 29.56 | 29.86 |\n| 2 | 64 | 640 | 3,094.8 | 41.39 | 42.49 | 42.78 | 43.47 |\n| 4 | 64 | 640 | 3,889.8 | 65.82 | 66.60 | 66.71 | 67.57 |\n| 8 | 64 | 640 | 4,270.1 | 119.80 | 120.61 | 120.68 | 120.89 |\n| 16 | 64 | 640 | 4,765.7 | 214.68 | 215.87 | 216.01 | 216.14 |\n| 32 | 64 | 640 | 4,985.2 | 410.43 | 413.58 | 413.67 | 413.92 |\n\n**FP32, TorchScript**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 64 | 640 | 2,486.3 | 25.78 | 27.52 | 27.66 | 27.92 |\n| 2 | 64 | 640 | 3,260.7 | 39.28 | 40.32 | 40.49 | 40.84 |\n| 4 | 64 | 640 | 4,033.3 | 63.48 | 64.28 | 64.35 | 64.56 |\n| 8 | 64 | 640 | 4,411.4 | 115.96 | 116.74 | 116.85 | 116.89 |\n| 16 | 64 | 640 | 4,924.9 | 207.74 | 208.91 | 209.04 | 209.21 |\n| 32 | 64 | 640 | 5,163.1 | 396.29 | 399.42 | 399.50 | 399.70 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and setup \nthe container, and then proceed to the \n[Inference performance benchmark](#inference-performance-benchmark) section for \ninstruction on how to launch the benchmark.\n\n###### Large model\n\nOur results were obtained by running the\n`pytorch/scripts/inference_benchmark.sh` inferencing benchmarking script in the\npytorch-20.06-py3 NGC container on NVIDIA T4.\n\nThe command to launch the inference performance benchmark is provided in the\n[Inference performance benchmark](#inference-performance-benchmark) section.\n\n**FP16, pure Python**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 128 | 1,600 | 2,978.0 | 42.99 | 43.40 | 43.44 | 44.40 |\n| 2 | 128 | 1,600 | 3,161.4 | 80.98 | 81.38 | 81.45 | 81.75 |\n| 4 | 128 | 1,600 | 3,459.3 | 147.89 | 148.11 | 148.14 | 148.49 |\n| 8 | 128 | 1,600 | 3,657.8 | 279.74 | 279.82 | 279.86 | 280.48 |\n| 16 | 128 | 1,600 | 3,762.9 | 543.92 | 543.48 | 543.55 | 544.43 |\n| 32 | 128 | 1,600 | 3,794.4 | 1079.15 | 1076.23 | 1076.37 | 1158.93 |\n\n**FP16, TorchScript**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 128 | 1,600 | 3,066.4 | 41.74 | 42.08 | 42.12 | 42.19 |\n| 2 | 128 | 1,600 | 3,399.2 | 75.31 | 75.54 | 75.57 | 75.64 |\n| 4 | 128 | 1,600 | 3,721.5 | 137.47 | 137.65 | 137.70 | 137.82 |\n| 8 | 128 | 1,600 | 3,932.9 | 260.19 | 260.23 | 260.29 | 260.50 |\n| 16 | 128 | 1,600 | 4,057.9 | 504.43 | 503.97 | 504.01 | 504.14 |\n| 32 | 128 | 1,600 | 4,117.8 | 994.54 | 991.40 | 991.46 | 1079.17 |\n\n**FP32, pure Python**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 128 | 1,600 | 786.9 | 162.7 | 163.2 | 163.3 | 163.9 |\n| 2 | 128 | 1,600 | 889.6 | 287.8 | 288.1 | 288.2 | 288.4 |\n| 4 | 128 | 1,600 | 992.1 | 515.6 | 516.0 | 516.0 | 516.5 |\n| 8 | 128 | 1,600 | 1,047.0 | 977.2 | 977.6 | 977.6 | 977.8 |\n| 16 | 128 | 1,600 | 1,069.3 | 1913.5 | 1914.7 | 1914.7 | 1915.0 |\n| 32 | 128 | 1,600 | 1,069.5 | 3826.3 | 3823.7 | 3823.8 | 3915.8 |\n\n**FP32, TorchScript**\n\n|**Batch size**|**Sequence length**|**Memory length**|**Throughput Avg (tok/s)**|**Latency Avg (ms)**|**Latency 90% (ms)**|**Latency 95% (ms)**|**Latency 99% (ms)**|\n|-------------:|------------------:|----------------:|-------------------------:|-------------------:|-------------------:|-------------------:|-------------------:|\n| 1 | 128 | 1,600 | 792.5 | 161.5 | 161.9 | 162.0 | 162.2 |\n| 2 | 128 | 1,600 | 904.7 | 283.0 | 283.3 | 283.3 | 283.4 |\n| 4 | 128 | 1,600 | 1,009.0 | 507.0 | 507.3 | 507.4 | 507.5 |\n| 8 | 128 | 1,600 | 1,065.0 | 960.7 | 961.1 | 961.1 | 961.2 |\n| 16 | 128 | 1,600 | 1,088.6 | 1879.7 | 1880.9 | 1881.0 | 1881.1 |\n| 32 | 128 | 1,600 | 1,102.0 | 3713.7 | 3710.0 | 3718.1 | 3819.0 |\n\nTo achieve these same results, follow the steps in the \n[Quick Start Guide](#quick-start-guide) to download the dataset and setup \nthe container, and then proceed to the \n[Inference performance benchmark](#inference-performance-benchmark) section for \ninstruction on how to launch the benchmark.","hasPerformance",[],{"_85":182,"_112":234,"_113":-5,"_121":34,"_120":34,"_122":233,"_217":90},{"_85":28,"_112":28,"_113":-5,"_121":34,"_120":34,"_122":233,"_217":90},[],"Latest (20.06.0)","nvidia/dle/transformerxl_pyt","key","nvidia/dle/resources/transformerxl_pyt","resourceId","transformerxl_pyt","shortDescription","Transformer-XL is a transformer-based language model with a segment-level recurrence and a novel relative positional encoding.","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/transformerxl-pyt)\n\n\nThis repository provides an implementation of the Transformer-XL model in\n[PyTorch](https://pytorch.org) from the paper [Transformer-XL: Attentive\nLanguage Models Beyond a Fixed-Length\nContext](https://arxiv.org/abs/1901.02860). Transformer-XL is a\ntransformer-based language model with a segment-level recurrence and a novel\nrelative positional encoding. Enhancements introduced in Transformer-XL help\ncapture better long-term dependencies by attending to tokens from multiple\nprevious segments.\n\nOur implementation is based on the\n[codebase](https://github.com/kimiyoung/transformer-xl) published by the\nauthors of the Transformer-XL paper.\nOur implementation uses a modified model architecture. Our\nmodifications were made to achieve better hardware utilization and to take\nadvantage of Tensor Cores. Similar modifications were also proposed in an\nimplementation available from\n[github.com/cybertronai/transformer-xl](https://github.com/cybertronai/transformer-xl).\nRefer to the [Model architecture](#model-architecture) section for more\ndetails.\n\nThis model is trained with mixed precision using Tensor Cores on NVIDIA Volta\nand the NVIDIA Ampere GPU architectures and evaluated on Volta, Turing and the\nNVIDIA Ampere GPU architectures.\nTherefore, researchers can get results up to 2.5x faster than training without\nTensor Cores, while experiencing the benefits of mixed precision training. This\nmodel is tested against each NGC monthly container release to ensure consistent\naccuracy and performance over time.\n\n### Model architecture\n\nThe Transformer-XL \"base\" model for WikiText-103 dataset available in this\nrepository was modified to use the following hyperparameter values:\n\n|**Hyperparameter**|**Description**|**Original setting for the base model**|**Our modification for the base model**|\n|------------------|---------------|--------------------------------------:|--------------------------------------:|\n| `d_model` | hidden size | 410 | 512 |\n| `n_head` | number of attention heads | 10 | 8 |\n| `d_head` | size of each attention head | 41 | 64 |\n| `d_inner` | hidden size in fully-connected layers | 2100 | 2048 |\n| `tgt_len` | number of tokens to predict during training | 150 | 192 |\n| `mem_len` | number of tokens cached from previous iterations during training | 150 | 192 |\n\nChanges described above were made to align certain hyperparameters with powers\nof two, with this modification, the model is able to achieve better hardware\nutilization, and therefore higher training throughput.\n\nThe Transformer-XL \"large\" model for WikiText-103 dataset available in this\nrepository uses the original hyperparameters from the [reference\nimplementation](https://github.com/kimiyoung/transformer-xl).\n\nThe following table lists the hyperparameters for the large and the base\nTransformer-XL models for WikiText-103 dataset available in this repository.\n\n| **Hyperparameter** | **Description** | **Base model** | **Large model** |\n| ------------------ | ---------------------------------------------------------------- | -------------: | ---------------: |\n| `n_layer` | number of layers | 16 | 18 |\n| `d_model` | hidden size | 512 | 1024 |\n| `n_head` | number of attention heads | 8 | 16 |\n| `d_head` | size of each attention head | 64 | 64 |\n| `d_inner` | inner hidden size in fully-connected layers | 2048 | 4096 |\n| `dropout` | dropout | 0.1 | 0.2 |\n| `dropatt` | dropout after softmax in the attention | 0.0 | 0.2 |\n| `lr` | base learning rate | 0.01 | 0.01 |\n| `eta_min` | minimum learning rate (for cosine decay) | 0.001 | 0.0001 |\n| `max_step` | number of training steps | 40,000 | 100,000 |\n| `warmup_step` | number of learning rate warmup steps | 1,000 | 16,000 |\n| `batch_size` | training batch size | 256 | 128 |\n| `tgt_len` | number of tokens to predict during training | 192 | 384 |\n| `mem_len` | number of tokens cached from previous iterations during training | 192 | 384 |\n\nThe Transformer-XL model addresses the limitations of vanilla transformer-based\nlanguage models, which are only able to use relatively short context, bounded\nby the segment length. The Transformer-XL introduces a recurrence mechanism,\nwhich is able to use a cached hidden state from previous segments. During\ntraining, the context consists of a concatenation of current segment's hidden\nstate and cached states from previous iterations. Gradients are backpropagated\nonly through the current segment, although the model is able to take advantage\nof the extra information stored in the cache and therefore is able to model\nlong-term dependencies.\n\nAn illustration of the recurrence mechanism taken from the [Transformer-XL\npaper](https://arxiv.org/abs/1901.02860) is shown below.\n![model](https://raw.githubusercontent.com/NVIDIA/DeepLearningExamples/master/PyTorch/LanguageModeling/Transformer-XL//pytorch/img/model.png)\n\n\n### Default configuration\n\nThe following features were implemented in this model:\n\n* general\n * single-node or multi-node, data-parallel multi-GPU training\n * training and inference with mixed precision using Tensor Cores\n * mixed precision training implemented using \n [Apex AMP](https://nvidia.github.io/apex/amp.html), with `O2` optimization\n level and with a dynamic loss scaling\n* model\n * 16-layer base Transformer-XL model with hidden size 512, 8 attention heads,\n each head with hidden size 64\n * 18-layer large Transformer-XL model with hidden size 1024, 16 attention\n heads, each head with hidden size 64\n * the model trained on\n [WikiText-103](https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/)\n dataset, using word-level vocabulary and\n adaptive softmax\n * embedding weights are tied with weights in the classifier\n* training\n * training with [LAMB](https://arxiv.org/abs/1904.00962) optimizer, the\n implementation of the optimizer uses\n [TorchScript](https://pytorch.org/docs/stable/jit.html), which enables\n the fusion of elementwise operations and accelerates the training\n * support for training with a gradient accumulation\n * base model:\n * linear learning rate warmup for 1,000 iterations, followed by the cosine\n learning rate schedule, the initial learning rate is set to 0.01, and the final\n learning rate is set to 0.001\n * training for 40,000 steps, using a batch size of 256\n * large model:\n * single node:\n * linear learning rate warmup for 16,000 iterations, followed by the cosine\n learning rate schedule, the initial learning rate is set to 0.01, and the final\n learning rate is set to 0.0001\n * training for 100,000 steps, using a batch size of 128\n * multi node:\n * linear learning rate warmup for 16,000 iterations, followed by the cosine\n learning rate schedule, the initial learning rate is set to 0.02, and the final\n learning rate is set to 0.0002\n * training for 25,000 steps, using a batch size of 512\n* inference\n * support for multi-gpu inference\n * support for [TorchScript](https://pytorch.org/docs/stable/jit.html) and\n pure Python inference\n * each token is using the same size of the context from previous time steps.\n * base model:\n * target length is set to 64, length of memory is set to 640\n * positional embeddings are clamped after 400 time steps\n * large model:\n * target length is set to 128, length of memory is set to 1,600\n * positional embeddings are clamped after 1,000 time steps\n\n### Feature support matrix\n\nThe following features are supported by this model:\n\n| **Feature** | **Transformer-XL** |\n|:------------|-------------------:|\n|[Apex AMP](https://nvidia.github.io/apex/amp.html) | Yes |\n|[PyTorch DistributedDataParallel](https://pytorch.org/docs/stable/nn.html#torch.nn.parallel.DistributedDataParallel) | Yes |\n|[LAMB](https://arxiv.org/abs/1904.00962v3) | Yes |\n| Inference with [TorchScript](https://pytorch.org/docs/stable/jit.html) | Yes |\n| Multi-node training | Yes |\n\n#### Features\n\n[Apex AMP](https://nvidia.github.io/apex/amp.html) - a tool that enables Tensor\nCore-accelerated training. Refer to the [Enabling mixed\nprecision](#enabling-mixed-precision) section for more details.\n\n[PyTorch\nDistributedDataParallel](https://pytorch.org/docs/stable/nn.html#torch.nn.parallel.DistributedDataParallel) - a module\nwrapper that enables easy multiprocess distributed data-parallel\ntraining.\n\n[LAMB](https://arxiv.org/abs/1904.00962v3) - stands for Layerwise Adaptive\nMoments Based optimizer, is a large batch optimization technique that helps\naccelerate training of deep neural networks using large minibatches.\n\n[TorchScript](https://pytorch.org/docs/stable/jit.html) - is a way to create\nserializable and optimizable models from PyTorch code. Any TorchScript program\ncan be saved from a Python process and loaded in a process where there is no\nPython dependency.\n\n### Mixed precision training\n\nMixed precision is the combined use of different numerical precisions in a\ncomputational method.\n[Mixed precision](https://arxiv.org/abs/1710.03740) training offers significant\ncomputational speedup by performing operations in half-precision format while\nstoring minimal information in single-precision to retain as much information\nas possible in critical parts of the network. Since the introduction of [Tensor\nCores](https://developer.nvidia.com/tensor-cores) in Volta, and following with \nboth the Turing and Ampere architectures, significant training speedups are \nexperienced by switching to\nmixed precision -- up to 3x overall speedup on the most arithmetically intense\nmodel architectures. Using mixed precision training previously required two\nsteps:\n\n1. Porting the model to use the FP16 data type where appropriate.\n2. Manually adding loss scaling to preserve small gradient values.\n\nThe ability to train deep learning networks with lower precision was introduced\nin the Pascal architecture and first supported in [CUDA\n8](https://devblogs.nvidia.com/parallelforall/tag/fp16/) in the NVIDIA Deep\nLearning SDK.\n\nFor information about:\n* How to train using mixed precision, see the [Mixed Precision\n Training](https://arxiv.org/abs/1710.03740) paper and [Training With Mixed\n Precision](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html)\n documentation.\n* Techniques used for mixed precision training, see the [Mixed-Precision\n Training of Deep Neural\n Networks](https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/)\n blog.\n* APEX tools for mixed precision training, see the [NVIDIA Apex: Tools for Easy\n Mixed-Precision Training in\n PyTorch](https://devblogs.nvidia.com/apex-pytorch-easy-mixed-precision-training/)\n .\n\n#### Enabling mixed precision\nThe `pytorch/train.py` training script launches mixed precision training\nwith Tensor Cores if the flag `--fp16` is set.\n\nMixed precision is enabled in PyTorch by using the Automatic Mixed Precision\n(AMP), library from [APEX](https://github.com/NVIDIA/apex) that casts variables\nto half-precision upon retrieval, while storing variables in single-precision\nformat. Furthermore, to preserve small gradient magnitudes in backpropagation,\na [loss\nscaling](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html#lossscaling)\nstep must be included when applying gradients. In PyTorch, loss scaling can be\neasily applied by using `scale_loss()` method provided by AMP. The scaling\nvalue to be used can be\n[dynamic](https://nvidia.github.io/apex/amp.html#apex.amp.initialize) or fixed.\n\nFor an in-depth walk through on AMP, check out sample usage\n[here](https://nvidia.github.io/apex/amp.html#).\n[APEX](https://github.com/NVIDIA/apex) is a PyTorch extension that contains\nutility libraries, such as AMP, which require minimal network code changes to\nleverage Tensor Cores performance.\n\nThe following steps were needed to enable mixed precision training in\nTransformer-XL:\n\n1. Import AMP from APEX:\n\n```\nfrom apex import amp\n```\n\n2. Initialize AMP and wrap the model and the optimizer before starting the\n training:\n\n```\nmodel, optimizer = amp.initialize(\n model,\n optimizer,\n opt_level='O2',\n )\n```\n\n3. Apply `scale_loss` context manager:\n\n```\nwith amp.scale_loss(loss, optimizer) as scaled_loss:\n scaled_loss.backward()\n```\n\n4. Apply gradient clipping on single precision master weights:\n\n```\ntorch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.clip)\n```\n\n#### Enabling TF32\nTensorFloat-32 (TF32) is the new math mode in [NVIDIA\nA100](https://www.nvidia.com/en-us/data-center/a100/) GPUs for handling the\nmatrix math also called tensor operations. TF32 running on Tensor Cores in A100\nGPUs can provide up to 10x speedups compared to single-precision floating-point\nmath (FP32) on Volta GPUs. \n\nTF32 Tensor Cores can speed up networks using FP32, typically with no loss of\naccuracy. It is more robust than FP16 for models which require high dynamic\nrange for weights or activations.\n\nFor more information, refer to the [TensorFloat-32 in the A100 GPU Accelerates\nAI Training, HPC up to\n20x](https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/)\nblog post.\n\nTF32 is supported in the NVIDIA Ampere GPU architecture and is enabled by\ndefault.","displayName","Transformer-XL for PyTorch",["D",1680632080303],"publisher","NVIDIA Deep Learning Examples","latestVersion","collectionCounts","imageUrl","/api/image/nvidia/dle/resources/transformerxl_pyt","rawImageUrl","https://assets.nvidiagrid.net/ngc/logos/Deep-Learning-Examples.png","canGuestDownload","latestVersionSizeInBytes","licenseTerms","isMultinodeEnabled","isMultiArchitecture","bias","explainability","privacy","safetyAndSecurity","featureLabels",[],"displayLabels",[274],"policyLabels",[],"systemLabels",[],"products",[],"hasNimMetadata",{"_243":275,"_276":277,"_278":279,"_280":87},"Natural Language Understanding","value","uscs_natural_language_understanding","filter",{"_281":282,"_276":277},"category","field","application",{"_151":1266,"_153":1267},"ngcApps",[1219],"csaApiBaseUrl","https://www.nvidia.com/ai-agents/csa-agent/dcsm","datadogRum",{"_1200":1201,"_1202":1203,"_1204":1205,"_1206":1207,"_1208":294,"_109":322,"_1209":1210,"_1211":59,"_1212":90,"_1213":90,"_1214":1215,"_1216":90,"_1217":1218},"amplitude",{"_1197":1198,"_834":-7,"_1199":34},"disableJs","environment","prod","featureFlagsBlob","JTdCJTIydGVzdEZsYWclMjIlM0FmYWxzZSUyQyUyMnNob3dTZWFyY2hMYWJlbFNldEZpbHRlcnMlMjIlM0FmYWxzZSUyQyUyMmVuYWJsZUNzYVdpZGdldCUyMiUzQWZhbHNlJTJDJTIyZW5hYmxlUHJvZHVjdEdyb3VwaW5nQnlTdWJzY3JpcHRpb24lMjIlM0FmYWxzZSUyQyUyMmNvbnRhaW5lclNjYW5uaW5nUmVkZXNpZ25NMiUyMiUzQWZhbHNlJTdE","featureFlagsGuarded","isDevLikeEnv","isPlaywright","isMobileDevice","categorizedOrgs",{"_1191":1192,"_1193":1194,"_1195":1196},"currentOrgContext","currentContextRoles",[],"enterpriseOrgs",[],"manageableExternalUserGroups",[],"locale","en","ngcApiOrigin","ngcHelmOrigin","origin","https://catalog.ngc.nvidia.com","usesMocks","theme","year",2026,"now",1785955791111,"1.11.0","defaultSearchSuggestions",{"_1074":402,"_1075":1076,"_1077":1078,"_1079":1080},"announcements",[],"consentState",{"_359":1064,"_1065":90,"_1066":34,"_1067":34,"_1068":1069},"consentConfig",{"_341":342,"_343":344,"_345":346,"_347":348,"_349":350,"_351":352,"_353":354,"_355":356,"_357":358,"_359":360},"showContactAdmin",1785955791113,[],[],[336,337,338,339,340],{"_48":63,"_50":74,"_52":146,"_54":-5,"_56":-5,"_58":59},{"_48":63,"_50":70,"_52":146,"_54":-5,"_56":-5,"_58":59},{"_48":63,"_50":64,"_52":146,"_54":-5,"_56":-5,"_58":59},{"_48":60,"_50":61,"_52":146,"_54":-5,"_56":-5,"_58":59},{"_48":49,"_50":51,"_52":53,"_54":55,"_56":62,"_58":59},"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",{"_1056":1057,"_1058":1059,"_1060":1061,"_1062":1063},"bannerAcceptLabel","Accept All","bannerRejectLabel","Reject Optional","bannerSettingsLabel","Manage Settings","preferenceSaveLabel","Save Choices","preferenceRejectLabel","Decline All","alwaysActiveText","Always Active","groups",[361,362,363,364,365],{"_366":746,"_112":747,"_211":748,"_370":90,"_371":16,"_373":749,"_375":750},{"_366":649,"_112":650,"_211":651,"_370":34,"_371":57,"_373":652,"_375":653},{"_366":557,"_112":558,"_211":559,"_370":34,"_371":62,"_373":560,"_375":561},{"_366":377,"_112":378,"_211":379,"_370":34,"_371":380,"_373":381,"_375":382},{"_366":367,"_112":368,"_211":369,"_370":34,"_371":372,"_373":374,"_375":376},"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,[499,500,501,502,503,504,505,506,507,508,509,510,511,512,513,514,515,516,517,518,519,520],[383,384,385,386,387,388,389,390,391,392,393,394,395,396,397,398,399],{"_400":494,"_243":494,"_211":402,"_373":495},{"_400":490,"_243":490,"_211":402,"_373":491},{"_400":485,"_243":485,"_211":402,"_373":486},{"_400":481,"_243":481,"_211":402,"_373":482},{"_400":476,"_243":476,"_211":402,"_373":477},{"_400":471,"_243":471,"_211":402,"_373":472},{"_400":463,"_243":463,"_211":402,"_373":464},{"_400":455,"_243":455,"_211":402,"_373":456},{"_400":450,"_243":450,"_211":402,"_373":451},{"_400":445,"_243":445,"_211":402,"_373":446},{"_400":440,"_243":440,"_211":402,"_373":441},{"_400":435,"_243":435,"_211":402,"_373":436},{"_400":430,"_243":430,"_211":402,"_373":431},{"_400":423,"_243":423,"_211":402,"_373":424},{"_400":418,"_243":418,"_211":402,"_373":419},{"_400":411,"_243":411,"_211":402,"_373":412},{"_400":401,"_243":401,"_211":402,"_373":403},"hostName","amazon-adsystem.com","",[404,405],{"_112":410,"_407":401,"_408":34,"_211":409},{"_112":406,"_407":401,"_408":34,"_211":409},"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",[413,414],{"_112":417,"_407":411,"_408":34,"_211":416},{"_112":415,"_407":411,"_408":34,"_211":416},"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",[420],{"_112":421,"_407":418,"_408":34,"_211":422},"MUID","This domain is owned by Mircosoft - it is the site for the search engine Bing.","assets.adobedtm.com",[425,426],{"_112":429,"_407":423,"_408":34,"_211":428},{"_112":427,"_407":423,"_408":34,"_211":428},"kndctr_F207D74D549850760A4C98C6_AdobeOrg_*","If you know anything about this host and its uses please add your contributions here. ","kndctr_F207D74D549850760A4C98C6_AdobeOrg_consent","tiktok.com",[432],{"_112":433,"_407":430,"_408":34,"_211":434},"_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",[437],{"_112":438,"_407":435,"_408":34,"_211":439},"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",[442],{"_112":443,"_407":440,"_408":34,"_211":444},"sp","Testing","google.com",[447],{"_112":448,"_407":445,"_408":34,"_211":449},"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",[452],{"_112":453,"_407":450,"_408":34,"_211":454},"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",[457,458],{"_112":461,"_407":455,"_408":34,"_211":462},{"_112":459,"_407":455,"_408":34,"_211":460},"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",[465,466],{"_112":469,"_407":463,"_408":34,"_211":470},{"_112":467,"_407":463,"_408":34,"_211":468},"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",[473],{"_112":474,"_407":471,"_408":34,"_211":475},"demdex","This cookie helps Adobe Audience Manger perform basic functions such as visitor identification, ID synchronization, segmentation, modeling, reporting, etc.","im-apps.net",[478],{"_112":479,"_407":476,"_408":34,"_211":480},"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",[483],{"_112":484,"_407":481,"_408":34,"_211":444},"R","6sc.co",[487],{"_112":488,"_407":485,"_408":34,"_211":489},"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",[492],{"_112":493,"_407":490,"_408":34,"_211":444},"_hly_gvid","t.contentsquare.net",[496],{"_112":497,"_407":494,"_408":90,"_211":498},"_cs_*","This is a session cookie. This is a pattern type cookie where a random 13-digit number is appended to the prefix _cs_.",{"_112":555,"_407":522,"_408":34,"_211":556},{"_112":553,"_407":522,"_408":34,"_211":554},{"_112":551,"_407":522,"_408":34,"_211":552},{"_112":433,"_407":522,"_408":34,"_211":550},{"_112":527,"_407":522,"_408":90,"_211":529},{"_112":549,"_407":522,"_408":34,"_211":402},{"_112":548,"_407":525,"_408":34,"_211":402},{"_112":546,"_407":522,"_408":34,"_211":547},{"_112":544,"_407":522,"_408":34,"_211":545},{"_112":543,"_407":522,"_408":34,"_211":402},{"_112":541,"_407":522,"_408":34,"_211":542},{"_112":527,"_407":540,"_408":90,"_211":529},{"_112":538,"_407":525,"_408":34,"_211":539},{"_112":538,"_407":528,"_408":34,"_211":539},{"_112":497,"_407":522,"_408":34,"_211":498},{"_112":536,"_407":522,"_408":34,"_211":537},{"_112":443,"_407":534,"_408":34,"_211":535},{"_112":532,"_407":522,"_408":34,"_211":533},{"_112":530,"_407":522,"_408":34,"_211":531},{"_112":527,"_407":528,"_408":90,"_211":529},{"_112":524,"_407":525,"_408":34,"_211":526},{"_112":521,"_407":522,"_408":90,"_211":523},"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.",[591,592,593,594,595,596,597,598,599,600,601,602,603,604,605,606,607,608,609,610,611,612,613,614,615],[562,563,564,565,566],{"_400":476,"_243":476,"_211":402,"_373":586},{"_400":450,"_243":450,"_211":402,"_373":581},{"_400":576,"_243":576,"_211":402,"_373":577},{"_400":571,"_243":571,"_211":402,"_373":572},{"_400":567,"_243":567,"_211":402,"_373":568},"nvidia.partners",[569],{"_112":570,"_407":567,"_408":34,"_211":444},"edge","app.hushly.com",[573],{"_112":574,"_407":571,"_408":34,"_211":575},"*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",[578],{"_112":579,"_407":576,"_408":34,"_211":580},"_octo","This domain is owned by Github. The main business activity is: Source code repository and collaboration platform",[582,583],{"_112":585,"_407":450,"_408":34,"_211":454},{"_112":584,"_407":450,"_408":34,"_211":454},"lidc","li_gc",[587,588],{"_112":590,"_407":476,"_408":34,"_211":480},{"_112":589,"_407":476,"_408":34,"_211":480},"imid_*","imid",{"_112":616,"_407":525,"_408":34,"_211":618},{"_112":647,"_407":522,"_408":34,"_211":648},{"_112":616,"_407":646,"_408":34,"_211":618},{"_112":644,"_407":522,"_408":90,"_211":645},{"_112":642,"_407":522,"_408":34,"_211":643},{"_112":641,"_407":624,"_408":90,"_211":402},{"_112":616,"_407":620,"_408":34,"_211":618},{"_112":639,"_407":522,"_408":34,"_211":640},{"_112":616,"_407":528,"_408":34,"_211":618},{"_112":627,"_407":525,"_408":34,"_211":402},{"_112":616,"_407":638,"_408":34,"_211":618},{"_112":636,"_407":522,"_408":34,"_211":637},{"_112":633,"_407":634,"_408":34,"_211":635},{"_112":631,"_407":632,"_408":34,"_211":402},{"_112":616,"_407":628,"_408":34,"_211":618},{"_112":630,"_407":522,"_408":34,"_211":402},{"_112":616,"_407":629,"_408":34,"_211":618},{"_112":627,"_407":628,"_408":34,"_211":402},{"_112":626,"_407":624,"_408":90,"_211":402},{"_112":616,"_407":625,"_408":34,"_211":618},{"_112":616,"_407":624,"_408":34,"_211":618},{"_112":616,"_407":623,"_408":34,"_211":618},{"_112":622,"_407":522,"_408":90,"_211":618},{"_112":619,"_407":620,"_408":90,"_211":621},{"_112":616,"_407":617,"_408":34,"_211":618},"_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.",[674,675,676,677,678,679,680,681,682,683,684,685,686,687,688,689,690,691,692,693,694,695,696,697,698,699,700,701,702],[654,655,656,657],{"_400":669,"_243":669,"_211":402,"_373":670},{"_400":664,"_243":664,"_211":402,"_373":665},{"_400":463,"_243":463,"_211":402,"_373":661},{"_400":576,"_243":576,"_211":402,"_373":658},[659],{"_112":660,"_407":576,"_408":90,"_211":580},"_gh_sess",[662],{"_112":663,"_407":463,"_408":90,"_211":470},"YSC","bizible.com",[666],{"_112":667,"_407":664,"_408":34,"_211":668},"_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",[671],{"_112":672,"_407":669,"_408":34,"_211":673},"tuid","This domain is owned by Ookla which provides internet speed testing services through its Speedtest application.",{"_112":744,"_407":522,"_408":34,"_211":745},{"_112":729,"_407":743,"_408":90,"_211":730},{"_112":712,"_407":525,"_408":34,"_211":637},{"_112":712,"_407":625,"_408":34,"_211":637},{"_112":741,"_407":522,"_408":34,"_211":742},{"_112":729,"_407":528,"_408":90,"_211":730},{"_112":739,"_407":522,"_408":34,"_211":740},{"_112":729,"_407":525,"_408":90,"_211":730},{"_112":737,"_407":624,"_408":34,"_211":738},{"_112":729,"_407":646,"_408":90,"_211":730},{"_112":709,"_407":525,"_408":34,"_211":711},{"_112":735,"_407":522,"_408":90,"_211":736},{"_112":733,"_407":522,"_408":34,"_211":734},{"_112":731,"_407":522,"_408":90,"_211":732},{"_112":729,"_407":620,"_408":90,"_211":730},{"_112":727,"_407":522,"_408":34,"_211":728},{"_112":725,"_407":522,"_408":34,"_211":726},{"_112":723,"_407":522,"_408":90,"_211":724},{"_112":721,"_407":629,"_408":34,"_211":722},{"_112":709,"_407":620,"_408":34,"_211":711},{"_112":719,"_407":522,"_408":34,"_211":720},{"_112":717,"_407":522,"_408":90,"_211":718},{"_112":715,"_407":522,"_408":34,"_211":716},{"_112":713,"_407":522,"_408":34,"_211":714},{"_112":712,"_407":522,"_408":34,"_211":637},{"_112":709,"_407":710,"_408":34,"_211":711},{"_112":707,"_407":522,"_408":34,"_211":708},{"_112":705,"_407":522,"_408":90,"_211":706},{"_112":703,"_407":522,"_408":34,"_211":704},"_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.",[891,892,893,894,895,896,897,898,899,900,901,902,903,904,905,906,907,908,909,910,911,912,913,914,915,916,917,918,919,920,921,922,923,924,925,926,927,928,929,930,931,932,933,934,935,936,937,938,939,940,941,942,943,944,945,946,947,948,949,950,951,952,953,954,955,956,957,958,959,960,961,962,963],[751,752,753,754,755,756,757,758,759,760,761,762,763,764,765,766,767,768,769,770,771,772,773,774,775],{"_400":888,"_243":888,"_211":402,"_373":889},{"_400":882,"_243":882,"_211":402,"_373":883},{"_400":879,"_243":879,"_211":402,"_373":880},{"_400":867,"_243":867,"_211":402,"_373":868},{"_400":864,"_243":864,"_211":402,"_373":865},{"_400":861,"_243":861,"_211":402,"_373":862},{"_400":856,"_243":856,"_211":402,"_373":857},{"_400":463,"_243":463,"_211":402,"_373":853},{"_400":850,"_243":850,"_211":402,"_373":851},{"_400":455,"_243":455,"_211":402,"_373":848},{"_400":450,"_243":450,"_211":402,"_373":846},{"_400":841,"_243":841,"_211":402,"_373":842},{"_400":836,"_243":836,"_211":402,"_373":837},{"_400":831,"_243":831,"_211":402,"_373":832},{"_400":828,"_243":828,"_211":402,"_373":829},{"_400":576,"_243":576,"_211":402,"_373":825},{"_400":435,"_243":435,"_211":402,"_373":822},{"_400":571,"_243":571,"_211":402,"_373":815},{"_400":567,"_243":567,"_211":402,"_373":809},{"_400":803,"_243":803,"_211":402,"_373":804},{"_400":798,"_243":798,"_211":402,"_373":799},{"_400":793,"_243":793,"_211":402,"_373":794},{"_400":789,"_243":789,"_211":402,"_373":790},{"_400":784,"_243":784,"_211":402,"_373":785},{"_400":776,"_243":776,"_211":402,"_373":777},"polyfill.io",[778,779],{"_112":782,"_407":776,"_408":34,"_211":783},{"_112":780,"_407":776,"_408":34,"_211":781},"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",[786],{"_112":787,"_407":784,"_408":34,"_211":788},"_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",[791],{"_112":792,"_407":789,"_408":34,"_211":444},"__cflb","instagram.com",[795],{"_112":796,"_407":793,"_408":34,"_211":797},"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",[800],{"_112":801,"_407":798,"_408":34,"_211":802},"__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",[805,806],{"_112":808,"_407":803,"_408":34,"_211":444},{"_112":807,"_407":803,"_408":34,"_211":444},"AWSALBTG","AWSALBTGCORS",[810,811],{"_112":813,"_407":567,"_408":34,"_211":814},{"_112":812,"_407":567,"_408":34,"_211":444},"ak_bmsc","bm_mi","This domain is likely associated with NVIDIA's partner program, which could involve marketing, sales, and support services for partners.",[816,817,818],{"_112":821,"_407":571,"_408":34,"_211":444},{"_112":820,"_407":571,"_408":90,"_211":444},{"_112":819,"_407":571,"_408":34,"_211":444},"AWSALBCORS","JSESSIONID","AWSALB",[823],{"_112":801,"_407":435,"_408":34,"_211":824},"This is a CloudFoundry cookie",[826],{"_112":827,"_407":576,"_408":34,"_211":580},"logged_in","hcaptcha.com",[830],{"_112":801,"_407":828,"_408":34,"_211":824},"tracking.intentsify.io",[833],{"_112":834,"_407":831,"_408":34,"_211":835},"userId","Used to track usage of the embedded video player.","static.reo.dev",[838],{"_112":839,"_407":836,"_408":34,"_211":840},"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",[843],{"_112":844,"_407":841,"_408":34,"_211":845},"dd_cookie_test_","dd_cookie_test",[847],{"_112":801,"_407":450,"_408":34,"_211":824},[849],{"_112":801,"_407":455,"_408":34,"_211":824},"supabase.co",[852],{"_112":801,"_407":850,"_408":34,"_211":824},[854],{"_112":855,"_407":463,"_408":34,"_211":470},"VISITOR_PRIVACY_METADATA","dosa-prd.azurewebsites.net",[858],{"_112":859,"_407":856,"_408":34,"_211":860},"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",[863],{"_112":801,"_407":861,"_408":34,"_211":824},"cdn.readme.io",[866],{"_112":801,"_407":864,"_408":34,"_211":824},"queue-it.net",[869,870,871,872,873],{"_112":878,"_407":867,"_408":34,"_211":402},{"_112":877,"_407":867,"_408":34,"_211":402},{"_112":876,"_407":867,"_408":34,"_211":402},{"_112":875,"_407":867,"_408":34,"_211":402},{"_112":874,"_407":867,"_408":34,"_211":402},"Queue-it-{queueId}","Queue-it-{dmonvidia }________{eventId}","Queue-it-token","queueitsoftblock_{queueId}","Queue-it-{nvidia}________{eventId}","x.com",[881],{"_112":801,"_407":879,"_408":34,"_211":802},"api.hcaptcha.com",[884,885],{"_112":887,"_407":882,"_408":34,"_211":444},{"_112":792,"_407":882,"_408":34,"_211":886},"This domain is associated with hCaptcha, a service that provides CAPTCHA and reCAPTCHA alternatives for user verification and bot protection.","hmt_id","vimeo.com",[890],{"_112":801,"_407":888,"_408":34,"_211":824},{"_112":1024,"_407":522,"_408":34,"_211":1055},{"_112":1053,"_407":617,"_408":90,"_211":1054},{"_112":1046,"_407":1015,"_408":90,"_211":1047},{"_112":1052,"_407":710,"_408":90,"_211":402},{"_112":1050,"_407":522,"_408":34,"_211":1051},{"_112":808,"_407":1036,"_408":34,"_211":975},{"_112":983,"_407":1045,"_408":90,"_211":985},{"_112":807,"_407":974,"_408":34,"_211":975},{"_112":986,"_407":1045,"_408":34,"_211":987},{"_112":1048,"_407":522,"_408":90,"_211":1049},{"_112":1046,"_407":1022,"_408":90,"_211":1047},{"_112":980,"_407":1045,"_408":34,"_211":982},{"_112":1043,"_407":522,"_408":34,"_211":1044},{"_112":807,"_407":966,"_408":34,"_211":975},{"_112":971,"_407":1042,"_408":34,"_211":973},{"_112":819,"_407":972,"_408":34,"_211":988},{"_112":821,"_407":972,"_408":34,"_211":1039},{"_112":986,"_407":981,"_408":34,"_211":987},{"_112":1040,"_407":620,"_408":34,"_211":1041},{"_112":821,"_407":646,"_408":34,"_211":1039},{"_112":1037,"_407":522,"_408":34,"_211":1038},{"_112":813,"_407":646,"_408":34,"_211":1032},{"_112":1009,"_407":625,"_408":90,"_211":402},{"_112":807,"_407":1036,"_408":34,"_211":975},{"_112":1034,"_407":974,"_408":90,"_211":1035},{"_112":998,"_407":522,"_408":34,"_211":989},{"_112":971,"_407":522,"_408":34,"_211":973},{"_112":1033,"_407":534,"_408":34,"_211":402},{"_112":813,"_407":522,"_408":34,"_211":1032},{"_112":821,"_407":966,"_408":34,"_211":1031},{"_112":1029,"_407":629,"_408":90,"_211":1030},{"_112":1027,"_407":522,"_408":90,"_211":1028},{"_112":1026,"_407":522,"_408":34,"_211":402},{"_112":1024,"_407":972,"_408":34,"_211":1025},{"_112":1023,"_407":534,"_408":34,"_211":402},{"_112":1014,"_407":1022,"_408":90,"_211":1016},{"_112":1020,"_407":625,"_408":34,"_211":1021},{"_112":1018,"_407":710,"_408":34,"_211":1019},{"_112":1017,"_407":528,"_408":34,"_211":402},{"_112":1014,"_407":1015,"_408":90,"_211":1016},{"_112":1012,"_407":981,"_408":90,"_211":1013},{"_112":1010,"_407":522,"_408":34,"_211":1011},{"_112":1009,"_407":525,"_408":90,"_211":402},{"_112":980,"_407":984,"_408":34,"_211":982},{"_112":812,"_407":999,"_408":34,"_211":989},{"_112":1007,"_407":525,"_408":34,"_211":1008},{"_112":1005,"_407":522,"_408":34,"_211":1006},{"_112":1003,"_407":522,"_408":34,"_211":1004},{"_112":1002,"_407":522,"_408":34,"_211":402},{"_112":844,"_407":525,"_408":34,"_211":1001},{"_112":998,"_407":646,"_408":34,"_211":989},{"_112":801,"_407":617,"_408":34,"_211":1000},{"_112":998,"_407":999,"_408":34,"_211":989},{"_112":997,"_407":534,"_408":34,"_211":402},{"_112":996,"_407":522,"_408":34,"_211":969},{"_112":994,"_407":981,"_408":34,"_211":995},{"_112":807,"_407":992,"_408":34,"_211":993},{"_112":991,"_407":534,"_408":34,"_211":402},{"_112":990,"_407":522,"_408":34,"_211":402},{"_112":983,"_407":981,"_408":90,"_211":985},{"_112":812,"_407":646,"_408":34,"_211":989},{"_112":819,"_407":646,"_408":34,"_211":988},{"_112":986,"_407":984,"_408":34,"_211":987},{"_112":983,"_407":984,"_408":90,"_211":985},{"_112":980,"_407":981,"_408":34,"_211":982},{"_112":978,"_407":974,"_408":90,"_211":979},{"_112":976,"_407":632,"_408":34,"_211":977},{"_112":808,"_407":974,"_408":34,"_211":975},{"_112":971,"_407":972,"_408":34,"_211":973},{"_112":801,"_407":624,"_408":34,"_211":970},{"_112":968,"_407":522,"_408":34,"_211":969},{"_112":819,"_407":966,"_408":34,"_211":967},{"_112":964,"_407":522,"_408":34,"_211":965},"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.

",{"_746":16,"_649":59,"_557":59,"_377":59,"_367":59},"hasConsented","gpcApplied","showBanner","geolocation",{"_1070":1071,"_1072":1073},"country","US","state","OR","query","categories",[1093,1094,1095,1096,1097],"optionUrls",[1081,1082,1083,1084,1085,1086,1087,1088,1089,1090,1091,1092],"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/-",{"_20":1171,"_1099":1172,"_1101":59},{"_20":1136,"_1099":1137,"_1101":57},{"_20":1119,"_1099":1120,"_1101":1121},{"_20":21,"_1099":1112,"_1101":1113},{"_20":1098,"_1099":1100,"_1101":19},"helm-charts","items",[1102,1103],"startIndex",{"_20":1098,"_243":1109,"_205":206,"_207":-7,"_238":1110,"_246":1106,"_271":1111,"_1108":1091},{"_20":1098,"_243":1104,"_205":206,"_207":-7,"_238":1105,"_246":1106,"_271":1107,"_1108":1092},"NVIDIA NIM Operator","k8s-nim-operator","NVIDIA",[],"url","GPU Operator","gpu-operator",[],[1114],9,{"_20":21,"_243":1115,"_205":206,"_207":1116,"_238":1117,"_246":1106,"_271":1118,"_1108":1090},"Riva Skills Quick Start","riva","riva_quickstart",[],"models",[1122,1123,1124],6,{"_20":1119,"_243":1132,"_205":206,"_207":1133,"_238":1134,"_246":1106,"_271":1135,"_1108":1087},{"_20":1119,"_243":1129,"_205":206,"_207":1126,"_238":1130,"_246":1106,"_271":1131,"_1108":1088},{"_20":1119,"_243":1125,"_205":206,"_207":1126,"_238":1127,"_246":1106,"_271":1128,"_1108":1089},"TrafficCamNet","tao","trafficcamnet",[],"PeopleNet","peoplenet",[],"StyleGAN3 pretrained models","research","stylegan3",[],"collections",[1138,1139,1140],{"_20":1136,"_243":1168,"_205":206,"_207":1142,"_238":1169,"_246":1106,"_271":1170,"_1108":1084},{"_20":1136,"_243":1165,"_205":206,"_207":-7,"_238":1166,"_246":1106,"_271":1167,"_1108":1085},{"_20":1136,"_243":1141,"_205":206,"_207":1142,"_238":1143,"_246":1106,"_271":1144,"_1108":1086},"Omniverse Kit App Streaming","omniverse","kit-appstreaming-collection",[1145,1146,1147],{"_112":1159,"_243":1160,"_240":1161,"_250":1162,"_1152":1163},{"_112":1142,"_243":1155,"_240":1156,"_250":1151,"_1152":1157},{"_112":1148,"_243":1149,"_240":1150,"_250":1151,"_1152":1153},"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",[1154],"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.",[1154,1158],"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",[1154,1158,1164],"NGC_CONSOLE","DeepStream SDK","deepstream_sdk",[],"Omniverse Kit (FB)","kit",[1145,1146,1147],"containers",[1173,1174,1175],{"_20":1171,"_243":1187,"_205":1180,"_207":1188,"_238":1189,"_246":1106,"_271":1190,"_1108":1081},{"_20":1171,"_243":1179,"_205":1180,"_207":206,"_238":1181,"_246":1106,"_271":1182,"_1108":1082},{"_20":1171,"_243":1176,"_205":206,"_207":-7,"_238":1177,"_246":1106,"_271":1178,"_1108":1083},"PyTorch","pytorch",[],"Llama-3.1-Nemotron-70B-Instruct","nim","llama-3.1-nemotron-70b-instruct",[1183,1145],{"_112":1184,"_243":1149,"_240":1150,"_250":1185,"_1152":1186},"nim-dev","https://assets.ngc.nvidia.com/products/catalog/images/product-nim-dev.jpg",[1154],"DeepSeek-R1","deepseek-ai","deepseek-r1",[1183,1145],"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",{"_1220":1221,"_1222":315,"_1223":34,"_1224":1225,"_1099":1226,"_1227":1228,"_1229":1221,"_1230":34,"_1231":34,"_1232":34,"_1233":1234},"appName","ngc-catalog","appOrigin","hideOnPrivateOrg","icon","ObjectsModules2",[1235,1236,1237,1238,1239,1240,1241],"menuIcon","editor-swatches","productName","requiresAllScopes","requiresLogin","requiresOrgLevelScopesOnly","title","Catalog",{"_1233":1263,"_50":1264,"_1227":1265},{"_1256":1257,"_1233":1258,"_50":1259,"_1229":1159},{"_1233":1254,"_50":1255,"_1227":1228},{"_1233":1251,"_50":1252,"_1227":1253},{"_1233":1248,"_50":1249,"_1227":1250},{"_1233":1245,"_50":1246,"_1227":1247},{"_1233":1242,"_50":1243,"_1227":1244},"Resources","/resources","misc-library","Models","/models","misc-neural-network","Helm Charts","/helm-charts","misc-ship-wheel","Containers","/containers","3d-cube","Collections","/collections","access",[1260],"Enterprise Catalog","/enterprise",[1261,1262],"NVAIE","READ","Explore Catalog","/","common-home",{"_156":157,"_158":159,"_160":161,"_162":163,"_164":165,"_166":167},{"_155":34}]