What's New in Version 4.3

In January 2023, the HugeCTR team plans to deprecate semantic versioning, such as `v4.3`.
Afterward, the library will use calendar versioning only, such as `v23.01`.

• Support for BERT and Variants: This release includes support for BERT in HugeCTR. The documentation includes updates to the MultiHeadAttention layer and adds documentation for the SequenceMask layer. For more information, refer to the samples/bst directory of the repository in GitHub.

• HPS Plugin for TensorFlow integration with TensorFlow-TensorRT (TF-TRT): This release includes plugin support for integration with TensorFlow-TensorRT. For sample code, refer to the Deploy SavedModel using HPS with Triton TensorFlow Backend notebook.

• Deep & Cross Network Layer version 2 Support: This release includes support for Deep & Cross Network version 2. For conceptual information, refer to https://arxiv.org/abs/2008.13535. The documentation for the MultiCross Layer is updated.

• Enhancements to Hierarchical Parameter Server:

  • RedisClusterBackend now supports TLS/SSL communication. For sample code, refer to the Hierarchical Parameter Server Demo notebook. The notebook is updated with step-by-step instructions to show you how to setup HPS to use Redis with (and without) encryption. The Volatile Database Parameters documentation for HPS is updated with the enable_tls, tls_ca_certificate, tls_client_certificate, tls_client_key, and tls_server_name_identification parameters.
  • MultiProcessHashMapBackend includes a bug fix that prevented configuring the shared memory size when using JSON file-based configuration.
  • On-device input keys are supported now so that an extra host-to-device copy is removed to improve performance.
  • A dependency on the XX-Hash library is removed. The library is no longer used by HugeCTR.
  • Added the static table support to the embedding cache. The static table is suitable when the embedding table can be placed entirely in GPU memory. In this case, the static table is more than three times faster than the embedding cache lookup. The static table does not support embedding updates.

• Support for New Optimizers:

  • Added support for SGD, Momentum SGD, Nesterov Momentum, AdaGrad, RMS-Prop, Adam and FTRL optimizers for dynamic embedding table (DET). For sample code, refer to the test_embedding_table_optimizer.cpp file in the test/utest/embedding_collection/ directory of the repository on GitHub.
  • Added support for the FTRL optimizer for dense networks.

• Data Reading from S3 for Offline Inference: In addition to reading during training, HugeCTR now supports reading data from remote file systems such as HDFS and S3 during offline inference by using the DataSourceParams API. The HugeCTR Training and Inference with Remote File System Example is updated to demonstrate the new functionality.

• Documentation Enhancements:

  • The set up instructions for running the example notebooks are revised for clarity.
  • The example notebooks are also updated to show using a data preprocessing script that simplifies the user experience.
  • Documentation for the MLP Layer is new.
  • Several 2022 talks and blogs are added to the HugeCTR Talks and Blogs page.

• Issues Fixed:

  • The original CUDA device with NUMA bind before a call to some HugeCTR APIs is recovered correctly now. This issue sometimes lead to a problem when you mixed calls to HugeCTR and other CUDA enabled libraries.
  • Fixed the occasional CUDA kernel launch failure of embedding when installed HugeCTR with macro DEBUG.
  • Fixed an SOK build error that was related to TensorFlow v2.1.0 and higher. The issue was that the C++ API and C++ standard were updated to use C++17.
  • Fixed a CUDA 12 related compilation error.

• Known Issues:

  • HugeCTR can lead to a runtime error if client code calls the RMM rmm::mr::set_current_device_resource() method or rmm::mr::set_current_device_resource() method. The error is due to the Parquet data reader in HugeCTR also calling rmm::mr::set_current_device_resource(). As a result, the device becomes visible to other libraries in the same process. Refer to GitHub issue #356 for more information. As a workaround, you can set environment variable HCTR_RMM_SETTABLE to 0 to prevent HugeCTR from setting a custom RMM device resource, if you know that rmm::mr::set_current_device_resource() is called by client code other than HugeCTR. But be cautious because the setting can reduce the performance of Parquet reading.

  • HugeCTR uses NCCL to share data between ranks and NCCL can require shared system memory for IPC and pinned (page-locked) system memory resources. If you use NCCL inside a container, increase these resources by specifying the following arguments when you start the container:

      -shm-size=1g -ulimit memlock=-1
    

    See also the NCCL known issue and the GitHub issue #243.

  • KafkaProducers startup succeeds even if the target Kafka broker is unresponsive. To avoid data loss in conjunction with streaming-model updates from Kafka, you have to make sure that a sufficient number of Kafka brokers are running, operating properly, and are reachable from the node where you run HugeCTR.

  • The number of data files in the file list should be greater than or equal to the number of data reader workers. Otherwise, different workers are mapped to the same file and data loading does not progress as expected.

  • Joint loss training with a regularizer is not supported.

  • Dumping Adam optimizer states to AWS S3 is not supported.

What's New in Version 4.2

In January 2023, the HugeCTR team plans to deprecate semantic versioning, such as `v4.2`.
Afterward, the library will use calendar versioning only, such as `v23.01`.

• Change to HPS with Redis or Kafka: This release includes a change to Hierarchical Parameter Server and affects deployments that use RedisClusterBackend or model parameter streaming with Kafka. A third-party library that was used for HPS partition selection algorithm is replaced to improve performance. The new algorithm can produce different partition assignments for volatile databases. As a result, volatile database backends that retain data between application startup, such as the RedisClusterBackend, must be reinitialized. Model streaming with Kafka is equally affected. To avoid issues with updates, reset all respective queue offsets to the end_offset before you reinitialize the RedisClusterBackend.

• Enhancements to the Sparse Operation Kit in DeepRec: This release includes updates to the Sparse Operation Kit to improve the performance of the embedding variable lookup operation in DeepRec. The API for the lookup_sparse() function is changed to remove the hotness argument. The lookup_sparse() function is enhanced to calculate the number of non-zero elements dynamically. For more information, refer to the sparse_operation_kit directory of the DeepRec repository in GitHub.

• Enhancements to 3G Embedding: This release includes the following enhancements to 3G embedding:

  • The API is changed. The EmbeddingPlanner class is replaced with the EmbeddingCollectionConfig class. For examples of the API, see the tests in the test/embedding_collection_test directory of the repository in GitHub.
  • The API is enhanced to support dumping and loading weights during the training process. The methods are Model.embedding_dump(path: str, table_names: list[str]) and Model.embedding_load(path: str, list[str]). The path argument is a directory in file system that you can dump weights to or load weights from. The table_names argument is a list of embedding table names as strings.

• New Volatile Database Type for HPS: This release adds a db_type value of multi_process_hash_map to the Hierarchical Parameter Server. This database type supports sharing embeddings across process boundaries by using shared memory and the /dev/shm device file. Multiple processes running HPS can read and write to the same hash map. For an example, refer to the Hierarchcal Parameter Server Demo notebook.

• Enhancements to the HPS Redis Backend: In this release, the Hierarchical Parameter Server can open multiple connections in parallel to each Redis node. This enhancement enables HPS to take advantage of overlapped processing optimizations in the I/O module of Redis servers. In addition, HPS can now take advantage of Redis hash tags to co-locate embedding values and metadata. This enhancement can reduce the number of accesses to Redis nodes and the number of per-node round trip communications that are needed to complete transactions. As a result, the enhancement increases the insertion performance.

• MLPLayer is New: This release adds an MLP layer with the hugectr.Layer_t.MLP class. This layer is very flexible and makes it easier to use a group of fused fully-connected layers and enable the related optimizations. For each fused fully-connected layer in MLPLayer, the output dimension, bias, and activation function are all adjustable. MLPLayer supports FP32, FP16 and TF32 data types. For an example, refer to the dgx_a100_mlp.py in the samples/dlrm directory of the GitHub repository to learn how to use the layer.

• Sparse Operation Kit installable from PyPi: Version 1.1.4 of the Sparse Operation Kit is installable from PyPi in the merlin-sok package.

• Multi-task Model Support added to the ONNX Model Converter: This release adds support for multi-task models to the ONNX converter. This release also includes an enhancement to the preprocess_census.py script in samples/mmoe directory of the GitHub repository.

• Issues Fixed:

  • Using the HPS Plugin for TensorFlow with MirroredStrategy and running the Hierarchical Parameter Server Demo notebook triggered an issue with ReplicaContext and caused a crash. The issue is fixed and resolves GitHub issue #362.
  • The 4_nvt_process.py sample in the samples/din/utils directory of the GitHub repository is updated to use the latest NVTabular API. This update resolves GitHub issue #364.
  • An illegal memory access related to 3G embedding and the dgx_a100_ib_nvlink.py sample in the samples/dlrm directory of the GitHub repository is fixed.
  • An error in HPS with the lookup_fromdlpack() method is fixed. The error was related to calculating the number of keys and vectors from the corresponding DLPack tensors.
  • An error in the HugeCTR backend for Triton Inference Server is fixed. A crash was triggered when the initial size of the embedding cache is smaller than the allowed minimum size.
  • An error related to using a ReLU layer with an odd input size in mixed precision mode could trigger a crash. The issue is fixed.
  • An error related to using an asynchronous reader with the AsyncParam class and specifying an io_alignment value that is smaller than the block device sector size is fixed. Now, if the specified io_alignment value is smaller than the block device sector size, io_alignment is automatically set to the block device sector size.
  • Unreported memory leaks in the GRU layer and collectives are fixed.
  • Several broken documentation links related to HPS are fixed.

• Known Issues:

  • HugeCTR uses NCCL to share data between ranks and NCCL can require shared system memory for IPC and pinned (page-locked) system memory resources. If you use NCCL inside a container, increase these resources by specifying the following arguments when you start the container:

      -shm-size=1g -ulimit memlock=-1
    

    See also the NCCL known issue and the GitHub issue.

  • KafkaProducers startup succeeds even if the target Kafka broker is unresponsive. To avoid data loss in conjunction with streaming-model updates from Kafka, you have to make sure that a sufficient number of Kafka brokers are running, operating properly, and are reachable from the node where you run HugeCTR.

  • The number of data files in the file list should be greater than or equal to the number of data reader workers. Otherwise, different workers are mapped to the same file and data loading does not progress as expected.

  • Joint loss training with a regularizer is not supported.

  • Dumping Adam optimizer states to AWS S3 is not supported.

What's New in Version 4.1.1

• Simplified Interface for 3G Embedding Table Placement Strategy: 3G embedding now provides an easier way for you to configure an embedding table placement strategy. Instead of using JSON, you can configure the embedding table placement strategy by using function arguments. You only need to provide the shard_matrix, table_group_strategy, and table_placement_strategy arguments. With these arguments, 3G embedding can group different tables together and place them according to the shard_matrix argument. For an example, refer to dlrm_train.py file in the test/embedding_collection_test directory of the repository on GitHub. For comparison, refer to the same file from the v4.0 branch of the repository.

• New MMoE and Shared-Bottom Samples: This release includes a new shared-bottom model, an example program, preprocessing scripts, and updates to documentation. For more information, refer to the README.md, mmoe_parquet.py, and other files in the samples/mmoe directory of the repository on GitHub. This release also includes a fix to the calculation and reporting of AUC for multi-task models, such as MMoE.

• Support for AWS S3 File System: The Parquet DataReader can now read datasets from the Amazon Web Services S3 file system. You can also load and dump models from and to S3 during training. The documentation for the DataSourceParams class is updated. To view sample code, refer to the HugeCTR Training with Remote File System Example class is updated.

• Simplication for File System Usage: You no longer ’t need to pass DataSourceParams for model loading and dumping. The FileSystem class automatically infers the correct file system type, local, HDFS, or S3, based on the path URI that you specified when you built the model. For example, the path hdfs://localhost:9000/ is inferred as an HDFS file system and the path https://mybucket.s3.us-east-1.amazonaws.com/ is inferred as an S3 file system.

• Support for Loading Models from Remote File Systems to HPS: This release enables you to load models from HDFS and S3 remote file systems to HPS during inference. To use the new feature, specify an HDFS for S3 path URI in InferenceParams.

• Support for Exporting Intermediate Tensor Values into a Numpy Array: This release adds function check_out_tensor to Model and InferenceModel. You can use this function to check out the intermediate tensor values using the Python interface. This function is especially helpful for debugging. For more information, refer to Model.check_out_tensor and InferenceModel.check_out_tensor.

• On-Device Input Keys for HPS Lookup: The HPS lookup supports input embedding keys that are on GPU memory during inference. This enhancement removes a host-to-device copy by using the DLPack lookup_fromdlpack() interface. By using the interface, the input DLPack capsule of embedding key can be a GPU tensor.

• Documentation Enhancements:

  • The graphic for the Hierarchical Parameter Server library that shows relationship to other software packages is enhanced.
  • The sample notebook for Deploy SavedModel using HPS with Triton TensorFlow Backend is added to the documentation.
  • Style updates to the Hierarchical Parameter Server API documentation.

• Issues Fixed:

  • The InteractionLayer class is fixed so that it works correctly with num_feas > 30.
  • The cuBLASLt configuration is corrected by increasing the workspace size and adding the epilogue mask.
  • The NVTabular based preprocessing script for our samples that demonstrate feature crossing is fixed.
  • The async data reader is fixed. Previously, it would hang and cause a corruption issue due to an improper I/O block size and I/O alignment problem. The AsyncParam class is changed to implement the fix. The io_block_size argument is replaced by the max_nr_request argument and the actual I/O block size that the async reader uses is computed accordingly. For more information, refer to the AsyncParam class documentation.

• Known Issues:

  • HugeCTR uses NCCL to share data between ranks and NCCL can require shared system memory for IPC and pinned (page-locked) system memory resources. If you use NCCL inside a container, increase these resources by specifying the following arguments when you start the container:

      -shm-size=1g -ulimit memlock=-1
    

    See also the NCCL known issue and the GitHub issue.

  • KafkaProducers startup succeeds even if the target Kafka broker is unresponsive. To avoid data loss in conjunction with streaming-model updates from Kafka, you have to make sure that a sufficient number of Kafka brokers are running, operating properly, and are reachable from the node where you run HugeCTR.

  • The number of data files in the file list should be greater than or equal to the number of data reader workers. Otherwise, different workers are mapped to the same file and data loading does not progress as expected.

  • Joint loss training with a regularizer is not supported.

  • Dumping Adam optimizer states to AWS S3 is not supported.

What's New in Version 4.1

• Simplified Interface for 3G Embedding Table Placement Strategy: 3G embedding now provides an easier way for you to configure an embedding table placement strategy. Instead of using JSON, you can configure the embedding table placement strategy by using function arguments. You only need to provide the shard_matrix, table_group_strategy, and table_placement_strategy arguments. With these arguments, 3G embedding can group different tables together and place them according to the shard_matrix argument. For an example, refer to dlrm_train.py file in the test/embedding_collection_test directory of the repository on GitHub. For comparison, refer to the same file from the v4.0 branch of the repository.

• New MMoE and Shared-Bottom Samples: This release includes a new shared-bottom model, an example program, preprocessing scripts, and updates to documentation. For more information, refer to the README.md, mmoe_parquet.py, and other files in the samples/mmoe directory of the repository on GitHub. This release also includes a fix to the calculation and reporting of AUC for multi-task models, such as MMoE.

• Support for AWS S3 File System: The Parquet DataReader can now read datasets from the Amazon Web Services S3 file system. You can also load and dump models from and to S3 during training. The documentation for the DataSourceParams class is updated. To view sample code, refer to the HugeCTR Training with Remote File System Example class is updated.

• Simplication for File System Usage: You no longer ’t need to pass DataSourceParams for model loading and dumping. The FileSystem class automatically infers the correct file system type, local, HDFS, or S3, based on the path URI that you specified when you built the model. For example, the path hdfs://localhost:9000/ is inferred as an HDFS file system and the path https://mybucket.s3.us-east-1.amazonaws.com/ is inferred as an S3 file system.

• Support for Loading Models from Remote File Systems to HPS: This release enables you to load models from HDFS and S3 remote file systems to HPS during inference. To use the new feature, specify an HDFS for S3 path URI in InferenceParams.

• Support for Exporting Intermediate Tensor Values into a Numpy Array: This release adds function check_out_tensor to Model and InferenceModel. You can use this function to check out the intermediate tensor values using the Python interface. This function is especially helpful for debugging. For more information, refer to Model.check_out_tensor and InferenceModel.check_out_tensor.

• On-Device Input Keys for HPS Lookup: The HPS lookup supports input embedding keys that are on GPU memory during inference. This enhancement removes a host-to-device copy by using the DLPack lookup_fromdlpack() interface. By using the interface, the input DLPack capsule of embedding key can be a GPU tensor.

• Documentation Enhancements:

  • The graphic for the Hierarchical Parameter Server library that shows relationship to other software packages is enhanced.
  • The sample notebook for Deploy SavedModel using HPS with Triton TensorFlow Backend is added to the documentation.
  • Style updates to the Hierarchical Parameter Server API documentation.

• Issues Fixed:

  • The InteractionLayer class is fixed so that it works correctly with num_feas > 30.
  • The cuBLASLt configuration is corrected by increasing the workspace size and adding the epilogue mask.
  • The NVTabular based preprocessing script for our samples that demonstrate feature crossing is fixed.
  • The async data reader is fixed. Previously, it would hang and cause a corruption issue due to an improper I/O block size and I/O alignment problem. The AsyncParam class is changed to implement the fix. The io_block_size argument is replaced by the max_nr_request argument and the actual I/O block size that the async reader uses is computed accordingly. For more information, refer to the AsyncParam class documentation.

• Known Issues:

  • HugeCTR uses NCCL to share data between ranks and NCCL can require shared system memory for IPC and pinned (page-locked) system memory resources. If you use NCCL inside a container, increase these resources by specifying the following arguments when you start the container:

      -shm-size=1g -ulimit memlock=-1
    

    See also the NCCL known issue and the GitHub issue.

  • KafkaProducers startup succeeds even if the target Kafka broker is unresponsive. To avoid data loss in conjunction with streaming-model updates from Kafka, you have to make sure that a sufficient number of Kafka brokers are running, operating properly, and are reachable from the node where you run HugeCTR.

  • The number of data files in the file list should be greater than or equal to the number of data reader workers. Otherwise, different workers are mapped to the same file and data loading does not progress as expected.

  • Joint loss training with a regularizer is not supported.

  • Dumping Adam optimizer states to AWS S3 is not supported.

What's New in Version 4.0

• 3G Embedding Stablization: Since the introduction of the next generation of HugeCTR embedding in v3.7, several updates and enhancements were made, including code refactoring to improve usability. The enhancements for this release are as follows:

  • Optimized the performance for sparse lookup in terms of inter-warp load imbalance. Sparse Operation Kit (SOK) takes advantage of the enhancement to improve performance.
  • This release includes a fix for determining the maximum embedding vector size in the GlobalEmbeddingData and LocalEmbeddingData classes.
  • Version 1.1.4 of Sparse Operation Kit can be installed with Pip and includes the enhancements mentioned in the preceding bullets.

• Embedding Cache Initialization with Configurable Ratio: In previous releases, the default value for the cache_refresh_percentage_per_iteration parameter of the InferenceParams was 0.1.

  In this release, default value is 0.0 and the parameter provides an additional purpose. If you set the parameter to a value greater than 0.0 and also set use_gpu_embedding_cache to True for a model, when Hierarchical Parameter Server (HPS) starts, HPS initializes the embedding cache for the model on the GPU by loading a subset of the embedding vectors from the sparse files for the model. When embedding cache initialization is used, HPS creates log records when it starts at the INFO level. The logging records are similar to EC initialization for model: "<model-name>", num_tables: <int> and EC initialization on device: <int>. This enhancement reduces the duration of the warm up phase.

• Lazy Initialization of HPS Plugin for TensorFlow: In this release, when you deploy a SavedModel of TensorFlow with Triton Inference Server, HPS is implicitly initialized when the loaded model is executed for the first time. In previous releases, you needed to run hps.Init(ps_config_file, global_batch_size) explicitly. For more information, see the API documentation for hierarchical_parameter_server.Init.

• Enhancements to the HDFS Backend:

  • The HDFS Backend is now called IO::HadoopFileSystem.
  • This release includes fixes for memory leaks.
  • This release includes refactoring to generalize the interface for HDFS and S3 as remote filesystems.
  • For more information, see hadoop_filesystem.hpp in the include/io directory of the repository on GitHub.

• Dependency Clarification for Protobuf and Hadoop: Hadoop and Protobuf are true third_party modules now. Developers can now avoid unnecessary and frequent cloning and deletion.

• Finer granularity control for overlap behavior: We deperacated the old overlapped_pipeline knob and introduces four new knobs train_intra_iteration_overlap/train_inter_iteration_overlap/eval_intra_iteration_overlap/eval_inter_iteration_overlap to help user better control the overlap behavior. For more information, see the API documentation for Solver.CreateSolver

• Documentation Improvements:

  • Removed two deprecated tutorials triton_tf_deploy and dump_to_tf.
  • Previously, the graphics in the Performance page did not appear. This issue is fixed in this release.
  • Previously, the API documentation for the HPS Plugin for TensorFlow did not show the class information. This issue is fixed in this release.

• Issues Fixed:

  • Fixed a build error that was triggered in debug mode. The error was caused by the newly introduced 3G embedding unit tests.
  • When using the Parquet DataReader, if a parquet dataset file specified in metadata.json does not exist, HugeCTR no longer crashes. The new behavior is to skip the missing file and display a warning message. This change relates to GitHub issue 321.

• Known Issues:

  • HugeCTR uses NCCL to share data between ranks and NCCL can require shared system memory for IPC and pinned (page-locked) system memory resources. If you use NCCL inside a container, increase these resources by specifying the following arguments when you start the container:

      -shm-size=1g -ulimit memlock=-1
    

    See also the NCCL known issue and the GitHub issue.

  • KafkaProducers startup succeeds even if the target Kafka broker is unresponsive. To avoid data loss in conjunction with streaming-model updates from Kafka, you have to make sure that a sufficient number of Kafka brokers are running, operating properly, and are reachable from the node where you run HugeCTR.

  • The number of data files in the file list should be greater than or equal to the number of data reader workers. Otherwise, different workers are mapped to the same file and data loading does not progress as expected.

  • Joint loss training with a regularizer is not supported.

• fix compatibility issue of cudf 22.06
• some document refactors

What's New in Version 3.9

• Updates to 3G Embedding:

  • Sparse Operation Kit (SOK) is updated to use the HugeCTR 3G embedding as a developer preview feature. For more information, refer to the Python programs in the sparse_operation_kit/experiment/benchmark/dlrm directory of the repository on GitHub.
  • Dynamic embedding table mode is added. The mode is based on the cuCollection with some functionality enhancement. A dynamic embedding table grows its size when the table is full so that you no longer need to configure the memory usage information for embedding. For more information, refer to the embedding_storage/dynamic_embedding_storage directory of the repository on GitHub.

• Enhancements to the HPS Plugin for TensorFlow: This release includes improvements to the interoperability of SOK and HPS. The plugin now supports the sparse lookup layer. The documentation for the HPS plugin is enhanced as follows:

  • An introduction to the plugin is new.
  • New notebooks demonstrate how to use the HPS plugin are added.
  • API documentation for the plugin is new.

• Enhancements to the HPS Backend for Triton Inference Server This release adds support for integrating the HPS Backend and the TensorFlow Backend through the ensemble mode with Triton Inference Server. The enhancement enables deploying a TensorFlow model with large embedding tables with Triton by leveraging HPS. For more information, refer to the sample programs in the hps-triton-ensemble directory of the HugeCTR Backend repository in GitHub.

• New Multi-Node Tutorial: The multi-node training tutorial is new. The additions show how to use HugeCTR to train a model with multiple nodes and is based on our most recent Docker container. The tutorial should be useful to users who do not have a job-scheduler-installed cluster such as Slurm Workload Manager. The update addresses a issue that was first reported in GitHub issue 305.

• Support Offline Inference for MMoE: This release includes MMoE offline inference where both per-class AUC and average AUC are provided. When the number of class AUCs is greater than one, the output includes a line like the following example:

  [HCTR][08:52:59.254][INFO][RK0][main]: Evaluation, AUC: {0.482141, 0.440781}, macro-averaging AUC: 0.46146124601364136
  

• Enhancements to the API for the HPS Database Backend This release includes several enhancements to the API for the DatabaseBackend class. For more information, see database_backend.hpp and the header files for other database backends in the HugeCTR/include/hps directory of the repository. The enhancments are as follows:

  • You can now specify a maximum time budget, in nanoseconds, for queries so that you can build an application that must operate within strict latency limits. Fetch queries return execution control to the caller if the time budget is exhausted. The unprocessed entries are indicated to the caller through a callback function.
  • The dump and load_dump methods are new. These methods support saving and loading embedding tables from disk. The methods support a custom binary format and the RocksDB SST table file format. These methods enable you to import and export embedding table data between your custom tools and HugeCTR.
  • The find_tables method is new. The method enables you to discover all table data that is currently stored for a model in a DatabaseBackend instance. A new overloaded method for evict is added that can process the results from find_tables to quickly and simply drop all the stored information that is related to a model.

• Documentation Enhancements

  • The documentation for the max_all_to_all_bandwidth parameter of the HybridEmbeddingParam class is clarified to indicate that the bandwidth unit is per-GPU. Previously, the unit was not specified.

• Issues Fixed:

  • Hybrid embedding with IB_NVLINK as the communication_type of the HybridEmbeddingParam is fixed in this release.
  • Training performance is affected by a GPU routine that checks if an input key can be out of the embedding table. If you can guarantee that the input keys can work with the specified workspace_size_per_gpu_in_mb, we have a workaround to disable the routine by setting the environment variable HUGECTR_DISABLE_OVERFLOW_CHECK=1. The workaround restores the training performance.
  • Engineering discovered and fixed a correctness issue with the Softmax layer.
  • Engineering removed an inline profiler that was rarely used or updated. This change relates to GitHub issue 340.

• Known Issues:

  • HugeCTR uses NCCL to share data between ranks and NCCL can require shared system memory for IPC and pinned (page-locked) system memory resources. If you use NCCL inside a container, increase these resources by specifying the following arguments when you start the container:

      -shm-size=1g -ulimit memlock=-1
    

    See also the NCCL known issue and the GitHub issue.

  • KafkaProducers startup succeeds even if the target Kafka broker is unresponsive. To avoid data loss in conjunction with streaming-model updates from Kafka, you have to make sure that a sufficient number of Kafka brokers are running, operating properly, and are reachable from the node where you run HugeCTR.

  • The number of data files in the file list should be greater than or equal to the number of data reader workers. Otherwise, different workers are mapped to the same file and data loading does not progress as expected.

  • Joint loss training with a regularizer is not supported.

What's New in Version 3.8

• Sample Notebook to Demonstrate 3G Embedding: This release includes a sample notebook that introduces the Python API of the embedding collection and the key concepts for using 3G embedding. You can view HugeCTR Embedding Collection from the documentation or access the embedding_collection.ipynb file from the notebooks directory of the repository.

• DLPack Python API for Hierarchical Parameter Server Lookup: This release introduces support for embedding lookup from the Hierarchical Parameter Server (HPS) using the DLPack Python API. The new method is lookup_fromdlpack(). For sample usage, see the Lookup the Embedding Vector from DLPack heading in the "Hierarchical Parameter Server Demo" notebook.

• Read Parquet Datasets from HDFS with the Python API: This release enhances the DataReaderParams class with a data_source_params argument. You can use the argument to specify the data source configuration such as the host name of the Hadoop NameNode and the NameNode port number to read from HDFS.

• Logging Performance Improvements: This release includes a performance enhancement that reduces the performance impact of logging.

• Enhancements to Layer Classes:

  • The FullyConnected layer now supports 3D inputs
  • The MatrixMultiply layer now supports 4D inputs.

• Documentation Enhancements:

  • An automatically generated table of contents is added to the top of most pages in the web documentation. The goal is to provide a better experience for navigating long pages such as the HugeCTR Layer Classes and Methods page.
  • URLs to the Criteo 1TB click logs dataset are updated. For an example, see the HugeCTR Wide and Deep Model with Criteo notebook.

• Issues Fixed:

  • The data generator for the Parquet file type is fixed and produces consistent file names between the _metadata.json file and the actual dataset files. Previously, running the data generator to create synthetic data resulted in a core dump. This issue was first reported in the GitHub issue 321.
  • Fixed the memory crash in running a large model on multiple GPUs that occurred during AUC warm up.
  • Fixed the issue of keyset generation in the ETC notebook. Refer to the GitHub issue 332 for more details.
  • Fixed the inference build error that occurred when building with debug mode.
  • Fixed the issue that multi-node training prints duplicate messages.

• Known Issues:

  • Hybrid embedding with IB_NVLINK as the communication_type of the HybridEmbeddingParam class does not work currently. We are working on fixing it. The other communication types have no issues.

  • HugeCTR uses NCCL to share data between ranks and NCCL can require shared system memory for IPC and pinned (page-locked) system memory resources. If you use NCCL inside a container, increase these resources by specifying the following arguments when you start the container:

      -shm-size=1g -ulimit memlock=-1
    

    See also the NCCL known issue and the GitHub issue.

  • KafkaProducers startup succeeds even if the target Kafka broker is unresponsive. To avoid data loss in conjunction with streaming-model updates from Kafka, you have to make sure that a sufficient number of Kafka brokers are running, operating properly, and are reachable from the node where you run HugeCTR.

  • The number of data files in the file list should be greater than or equal to the number of data reader workers. Otherwise, different workers are mapped to the same file and data loading does not progress as expected.

  • Joint loss training with a regularizer is not supported.

What's New in Version 3.7

• 3G Embedding Developer Preview: The 3.7 version introduces next-generation of embedding as a developer preview feature. We call it 3G embedding because it is the new update to the HugeCTR embedding interface and implementation since the unified embedding in v3.1 version, which was the second one. Compared with the previous embedding, there are three main changes in the embedding collection.

  • First, it allows users to fuse embedding tables with different embedding vector sizes. The previous embedding can only fuse embedding tables with the same embedding vector size. The enhancement boosts both flexibility and performance.
  • Second, it extends the functionality of embedding by supporting the concat combiner and supporting different slot lookup on the same embedding table.
  • Finally, the embedding collection is powerful enough to support arbitrary embedding table placement which includes data parallel and model parallel. By providing a plan JSON file, you can configure the table placement strategy as you specify. See the dlrm_train.py file in the embedding_collection_test directory of the repository for a more detailed usage example.

• HPS Performance Improvements:

  • Kafka: Model parameters are now stored in Kafka in a bandwidth-saving multiplexed data format. This data format vastly increases throughput. In our lab, we measured transfer speeds up to 1.1 Gbps for each Kafka broker.
  • HashMap backend: Parallel and single-threaded hashmap implementations have been replaced by a new unified implementation. This new implementation uses a new memory-pool based allocation method that vastly increases upsert performance without diminishing recall performance. Compared with the previous implementation, you can expect a 4x speed improvement for large-batch insertion operations.
  • Suppressed and simplified log: Most log messages related to HPS have the log level changed to TRACE rather than INFO or DEBUG to reduce logging verbosity.

• Offline Inference Usability Enhancements:

  • The thread pool size is configurable in the Python interface, which is useful for studying the embedding cache performance in scenarios of asynchronous update. Previously it was set as the minimum value of 16 and std::thread::hardware_concurrency(). For more information, please refer to Hierarchical Parameter Server Configuration.

• DataGenerator Performance Improvements: You can specify the num_threads parameter to parallelize a Norm dataset generation.

• Evaluation Metric Improvements:

  • Average loss performance improvement in multi-node environments.
  • AUC performance optimization and safer memory management.
  • Addition of NDCG and SMAPE.

• Embedding Training Cache Parquet Demo: Created a keyset extractor script to generate keyset files for Parquet datasets. Provided users with an end-to-end demo of how to train a Parquet dataset using the embedding cache mode. See the Embedding Training Cache Example notebook.

• Documentation Enhancements: The documentation details for HugeCTR Hierarchical Parameter Server Database Backend are updated for consistency and clarity.

• Issues Fixed:

  • If slot_size_array is specified, workspace_size_per_gpu_in_mb is no longer required.
  • If you build and install HugeCTR from scratch, you can specify the CMAKE_INSTALL_PREFIX CMake variable to identify the installation directory for HugeCTR.
  • Fixed SOK hang issue when calling sok.Init() with a large number of GPUs. See the GitHub issue 261 and 302 for more details.

• Known Issues:

  • HugeCTR uses NCCL to share data between ranks and NCCL can require shared system memory for IPC and pinned (page-locked) system memory resources. If you use NCCL inside a container, increase these resources by specifying the following arguments when you start the container:

      -shm-size=1g -ulimit memlock=-1
    

    See also the NCCL known issue and the GitHub issue.

  • KafkaProducers startup succeeds even if the target Kafka broker is unresponsive. To avoid data loss in conjunction with streaming-model updates from Kafka, you have to make sure that a sufficient number of Kafka brokers are running, operating properly, and are reachable from the node where you run HugeCTR.

  • The number of data files in the file list should be greater than or equal to the number of data reader workers. Otherwise, different workers are mapped to the same file and data loading does not progress as expected.

  • Joint loss training with a regularizer is not supported.

  • The Criteo 1 TB click logs dataset that is used with many HugeCTR sample programs and notebooks is currently unavailable. Until the dataset becomes downloadable again, you can run those samples based on our synthetic dataset generator. For more information, see the Getting Started section of the repository README file.

  • Data generator of parquet type produces inconsistent file names between _metadata.json and actual dataset files, which will result in core dump fault when using the synthetic dataset.

What's New in Version 3.6

• Concat 3D Layer: In previous releases, the Concat layer could handle two-dimensional (2D) input tensors only. Now, the input can be three-dimensional (3D) and you can concatenate the inputs along axis 1 or 2. For more information, see the API documentation for the Concat Layer.

• Dense Column List Support in Parquet DataReader: In previous releases, HugeCTR assumes each dense feature has a single value and it must be the scalar data type float32. Now, you can mix float32 or list[float32] for dense columns. This enhancement means that each dense feature can have more than one value. For more information, see the API documentation for the Parquet dataset format.

• Support for HDFS is Re-enabled in Merlin Containers: Support for HDFS in Merlin containers is an optional dependency now. For more information, see HDFS Support.

• Evaluation Metric Enhancements: In previous releases, HugeCTR computes AUC for binary classification only. Now, HugeCTR supports AUC for multi-label classification. The implementation is inspired by sklearn.metrics.roc_auc_score and performs the unweighted macro-averaging strategy that is the default for scikit-learn. You can specify a value for the label_dim parameter of the input layer to enable multi-label classification and HugeCTR will compute the multi-label AUC.

• Log Output Format Change: The default log format now includes milliseconds.

• Documentation Enhancements:

  • These release notes are included in the documentation and are available at https://nvidia-merlin.github.io/HugeCTR/v3.6/release_notes.html.
  • The Configuration section of the Hierarchical Parameter Server information is updated with more information about the parameters in the configuration file.
  • The example notebooks that demonstrate how to work with multi-modal data are reorganized in the navigation. The notebooks are now available under the heading Multi-Modal Example Notebooks. This change is intended to make it easier to find the notebooks.
  • The documentation in the sparse_operation_kit directory of the repository on GitHub is updated with several clarifications about SOK.

• Issues Fixed:

  • The dlrm_kaggle_fp32.py file in the samples/dlrm/ directory of the repository is updated to show the correct number of samples. The num_samples value is now set to 36672493. This fixes GitHub issue 301.
  • Hierarchical Parameter Server (HPS) would produce a runtime error when the GPU cache was turned off. This issue is now fixed.

• Known Issues:

  • HugeCTR uses NCCL to share data between ranks and NCCL can require shared system memory for IPC and pinned (page-locked) system memory resources. If you use NCCL inside a container, increase these resources by specifying the following arguments when you start the container:

      -shm-size=1g -ulimit memlock=-1
    

    See also the NCCL known issue and the GitHub issue.

  • KafkaProducers startup succeeds even if the target Kafka broker is unresponsive. To avoid data loss in conjunction with streaming-model updates from Kafka, you have to make sure that a sufficient number of Kafka brokers are running, operating properly, and are reachable from the node where you run HugeCTR.

  • The number of data files in the file list should be greater than or equal to the number of data reader workers. Otherwise, different workers are mapped to the same file and data loading does not progress as expected.

  • Joint loss training with a regularizer is not supported.

  • The Criteo 1 TB click logs dataset that is used with many HugeCTR sample programs and notebooks is currently unavailable. Until the dataset becomes downloadable again, you can run those samples based on our synthetic dataset generator. For more information, see the Getting Started section of the repository README file.

What's New in Version 3.5

• HPS interface encapsulation and exporting as library: We encapsulate the Hierarchical Parameter Server(HPS) interfaces and deliver it as a standalone library. Besides, we prodvide HPS Python APIs and demonstrate the usage with a notebook. For more information, please refer to Hierarchical Parameter Server and HPS Demo.

• Hierarchical Parameter Server Triton Backend: The HPS Backend is a framework for embedding vectors looking up on large-scale embedding tables that was designed to effectively use GPU memory to accelerate the looking up by decoupling the embedding tables and embedding cache from the end-to-end inference pipeline of the deep recommendation model. For more information, please refer to Hierarchical Parameter Server.

• SOK pip release: SOK pip releases on https://pypi.org/project/merlin-sok/. Now users can install SOK via pip install merlin-sok.

• Joint loss and multi-tasks training support:: We support joint loss in training so that users can train with multiple labels and tasks with different weights. MMoE sample is added to show the usage here.

• HugeCTR documentation on web page: Now users can visit our web documentation.

• ONNX converter enhancement:: We enable converting MultiCrossEntropyLoss and CrossEntropyLoss layers to ONNX to support multi-label inference. For more information, please refer to HugeCTR to ONNX Converter.

• HDFS python API enhancement:

  • Simplified DataSourceParams so that users do not need to provide all the paths before they are really necessary. Now users only have to pass DataSourceParams once when creating a solver.
  • Later paths will be automatically regarded as local paths or HDFS paths depending on the DataSourceParams setting. See notebook for usage.

• HPS performance optimization: We use better method to determine partition number in database backends in HPS.

• Bug fixing:

  • HugeCTR input layer now can take dense_dim greater than 1000.

What's New in Version 3.4.1

• Support mixed precision inference for dataset with multiple labels: We enable FP16 for the Softmax layer and support mixed precision for multi-label inference. For more information, please refer to Inference API.

• Support multi-GPU offline inference with Python API: We support multi-GPU offline inference with the Python interface, which can leverage Hierarchical Parameter Server and enable concurrent execution on multiple devices. For more information, please refer to Inference API and Multi-GPU Offline Inference Notebook.

• Introduction to metadata.json: We add the introduction to _metadata.json for Parquet datasets. For more information, please refer to Parquet.

• Documents and tool for workspace size per GPU estimation: we add a tool named embedding_workspace_calculator to help calculate workspace_size_per_gpu_in_mb required by hugectr.SparseEmbedding. For more information, please refer to embedding_workspace_calculator/README.md and QA 24.

• Improved Debugging Capability: The old logging system, which was flagged as deprecated for some time has been removed. All remaining log messages and outputs have been revised and migrated to the new logging system (base/debug/logging.hpp/cpp). During this revision, we also adjusted log levels for log messages throughout the entire codebase to improve visibility of relevant information.

• Support HDFS Parameter Server in Training:

  • Decoupled HDFS in Merlin containers to make the HDFS support more flexible. Users can now compile HDFS related functionalities optionally.
  • Now supports loading and dumping models and optimizer states from HDFS.
  • Added a notebook to show how to use HugeCTR with HDFS.

• Support Multi-hot Inference on Hugectr Backend: We support categorical input in multi-hot format for HugeCTR Backend inference.

• Multi-label inference with mixed precision: Mixed precision training is enabled for softmax layer.

• Python Script and documentation demonstrating how to analyze model files: In this release, we provide a script to retreive vocabulary information from model file. Please find more details on the readme

• Bug Fixing:

  • Mirror strategy bug in SOK (see in https://github.com/NVIDIA-Merlin/HugeCTR/issues/291)
  • Can't import sparse operation kit in nvcr.io/nvidia/merlin/merlin-tensorflow-training:22.03 (see in https://github.com/NVIDIA-Merlin/HugeCTR/issues/296)
  • HPS: Fixed access violation that can occur during initialization when not configuring a volatile DB.

Known Issues

• HugeCTR uses NCCL to share data between ranks, and NCCL may require shared system memory for IPC and pinned (page-locked) system memory resources. When using NCCL inside a container, it is recommended that you increase these resources by issuing: -shm-size=1g -ulimit memlock=-1 See also NCCL's known issue. And the GitHub issue.

• KafkaProducers startup will succeed, even if the target Kafka broker is unresponsive. In order to avoid data-loss in conjunction with streaming model updates from Kafka, you have to make sure that a sufficient number of Kafka brokers is up, working properly and reachable from the node where you run HugeCTR.

• The number of data files in the file list should be no less than the number of data reader workers. Otherwise, different workers will be mapped to the same file and data loading does not progress as expected.

What's New in Version 3.4

• Supporting HugeCTR Development with Merlin Unified Container: From Merlin v22.02 we encourage you to develop HugeCTR under Merlin Unified Container (release container) according to the instructions in Contributor Guide to keep consistent.

• Hierarchical Parameter Server (HPS) Enhancements:

  • Missing key insertion feature: Via a simple flag, it is now possible to configure HugeCTR such that missed embedding-table entries during lookup are automatically inserted into volatile database layers such as the Redis and Hashmap backends.
  • Asynchronous timestamp refresh: In the last release we introduced the passing-of-time-aware eviction policies. These are policies that are applied to shrink database partitions through dropping keys if they grow beyond certain limits. However, the time-information utilized by these eviction policies represented the update time. Hence, an embedding was evicted based on the time passed since its last update. If you operate HugeCTR in inference mode, the embedding table is typically immutable. With the above-described missing key insertion feature we now support actively tuning the contents of volatile database layers to the data distribution during lookup. To allow time-based eviction to take place, it is now possible to enable timestamp refreshing for frequently used embeddings. Once enabled, refreshing is handled asynchronously using background threads. Hence, it won’t block your inference jobs. For most applications, the associated performance impact from enabling this feature is barely noticeable.
  • Support HDFS(Hadoop Distributed File System) Parameter Server in Training:
    • A new Python API DataSourceParams used to specify the file system and paths to data and model files.
    • Support loading data from HDFS to the local file system for HugeCTR training.
    • Support dumping trained model and optimizer states into HDFS.
  • Online seamless update of the parameters of the dense part of the model: HugeCTR Backend has supported online model version updating by the Load API of Triton (including the seamless update of the dense part and corresponding embedding inference cache for the same model), and the Load API is still fully compatible with online deployment of new models.

• Sparse Operation Kit Enhancements:

  • Mixed Precision Training: Mixed precision training can be enabled via TF’s pattern to enhance the training performance and lessen memory usage.
  • DLRM Benchmark: DLRM is a standard benchmark for recommendation model training. A note book is added to address the performance of SOK on this benchmark in this release.
  • Support Uint32_t / int64_t key dtype in SOK: Int64 or uint32 can be used as the key data type for SOK’s embedding. By default, it is int64.
  • Add TensorFlow initializers support: Tensorflow native initializer can be used in SOK now. e.g. sok.All2AllDenseEmbedding(embedding_initializer=tf.keras.initializers.RandomUniform())

• User Experience Enhancements

  • We have revised several notebooks and readme files to clarify instructions and make HugeCTR more accessible in general.
  • Thanks to GitHub user @MuYu-zhi , who brought to our attention that having configured too few shared memory can impact the proper operation of HugeCTR. We extended the SOK docker setup instructions to address how such issues can be resolved using the --shm-size setting of docker.
  • Although HugeCTR is designed for scalability, having a beefy machine is not necessary for smaller workloads and testing. We added information about the required specs for notebook testing environments in README.

• Inference for Multi-tasking： We support HugeCTR inference for multiple tasks. When the label dimension is the number of binary classification tasks and MultiCrossEntropyLoss is employed during training, the shape of inference results will be (batch_size*num_batches, label_dim). For more information, please refer to Inference API.

• Fix the Embedding Cache Issue for Super Small Embedding Tables

What's New in Version 3.3.1

• Hierarchical Parameter Server Enhancements:
  • Online deployment of new models and recycling of old models: In this release, HugeCTR Backend is fully compatible with the model control protocol of Triton. Adding the configuration of a new model to the HPS configuration file. The HugeCTR Backend has supported online deployment of new models by the Load API of Triton. The old models can also be recycled online by the Unload API.
  • Simplified database backend: Multi-nodes, single-node and all other kinds of volatile database backends can now be configured using the same configuration object.
  • Multi-threaded optimization of Redis code: ~2.3x speedup up over HugeCTR v3.3.
  • Fix to some issues: Build HPS test environment and implement unit test of each component; Access violation issue of online Kafka updates; Parquet data reader incorrectly parses the index of categorical features in the case of multiple embedded tables; HPS Redis Backend overflow handling not invoked upon single insertions.
• New group of fused fully connected layers: We support adding a group of fused fully connected layers when constructing the model graph. A concise Python interface is provided for users to adjust the number of layers, as well as to specify the output dimensions in each layer, which makes it easy to leverage the highly-optimized fused fully connected layer in HugeCTR. For more information, please refer to GroupDenseLayer
• Fix to some issues:
  • Warnning is added for the case users forget to import mpi before launching multi-process job
  • Removing massive log when runing with embedding training cache
  • Removing lagacy conda related informations from documents

What's New in Version 3.3

• Hierarchical Parameter Server:

  • Support Incremental Models Updating From Online Training: HPS now supports iterative model updating via Kafka message queues. It is now possible to connect HugeCTR with Apache Kafka deployments to update the model in-place in real-time. This feature is supported in both phases, training and inference. Please refer to the Demo Notebok.
  • Support Embedding keys Eviction Mechanism: In-memory databases such as Redis or CPU memory backed storage are used now as the feature memory management. Hence, when performing iterative updating, they will automatically evict infrequently used embeddings as training progresses.
  • Support Embedding Cache Asynchronous Refresh Mechanism: We have supported the asynchronous refreshing of incremental embedding keys into the embedding cache. Refresh operation will be triggered when completing the model version iteration or incremental parameters output from online training. The Distributed Database and Persistent Database will be updated by the distributed event streaming platform(Kafka). And then the GPU embedding cache will refresh the values of the existing embedding keys and replace them with the latest incremental embedding vectors. Please refer to the HPS README.
  • Other Improvements: Backend implementations for databases are now fully configurable. JSON interface parser can cope better with inaccurate parameterization. Less and if (hopefully) more meaningful jabber! Based on your requests, we revised the log levels for throughout the entire database backend API of the parameter server. Selected configuration options are now printed wholesomely and uniformly to the log. Errors provide more verbose information on the matter at hand. Improved performance of Redis cluster backend. Improved performance of CPU memory database backend.

• SOK TF 1.15 Support: In this version, SOK can be used along with TensorFlow 1.15. See README. Dedicated CUDA stream is used for SOK’s Ops, and kernel interleaving might be eliminated. Users can now install SOK via pip install SparseOperationKit, which no longer requires root access to compile SOK and no need to copy python scripts. There was a hanging issue in tf.distribute.MirroredStrategy when TensorFlow version greater than 2.4. In this version, this issue in TensorFlow 2.5+ is fixed.

• MLPerf v1.1 integration：

  • Hybrid-embedding indices pre-computing：The indices needed for hybrid embedding are pre-computed ahead of time and are overlapped with previous iterations.
  • Cached evaluation indices:：The hybrid-embedding indices for eval are cached when applicable, hence eliminating the re-computing of the indices at every eval iteration.
  • MLP weight/data gradients calculation overlap:：The weight gradients of MLP are calculated asynchronously with respect to the data gradients, enabling overlap between these two computations.
  • Better compute-communication overlap:：Better overlap between compute and communication has been enabled to improve training throughput.
  • Fused weight conversion:：The FP32-to-FP16 conversion of the weights are now fused into the SGD optimizer, saving trips to memory.
  • GraphScheduler:：GrapScheduler was added to control the timing of cudaGraph launching. With GraphScheduler, the gap between adjacent cudaGraphs is eliminated.

• Multi-node training support on the cluster without RDMA：We support multi-node training without RDMA now. You can specify allreduce algorithm as AllReduceAlgo.NCCL and it can support non-RDMA hardware. For more information, please refer to all_reduce_algo in CreateSolver API.

• SOK support device setting with tf.config：tf.config.set_visible_device can be used to set the visible GPUs for each process. Meanwhile, CUDA_VISIBLE_DEVICES can also be used to achieve the same purpose. When tf.distribute.Strategy is used, device argument must not be set.

• User defined name is supported in model dumping: We support specifying the model name with the training API CreateSolver, which will be dumped to the JSON configuration file with the API Model.graph_to_json. This feature will facilitate the Triton deployment of saved HugeCTR models, and help to distinguish between models when Kafka sends parameters from the training side to the inference side.

• Fine-grained control of the embedding layers: We support the fine-grained control of the embedding layers. Users can freeze or unfreeze the weights of a specific embedding layer with the APIs Model.freeze_embedding and Model.unfreeze_embedding. Besides, the weights of multiple embedding layers can be loaded independently, which enables the use case of loading pre-trained embeddings for a particular layer. For more information, please refer to Model API and Section 3.4 of HugeCTR Criteo Notebook.

What's New in Version 3.2.1

• Performance optimization on GPU embedding cache: We have optimized the performance of the GPU embedding cache stand-alone module. Now the performance has been significantly improved under small to medium batch sizes. For large batch sizes, the performance remains unchanged. This feature does not introduce any mandatory changes to the interface of the GPU embedding cache, so any existing code that uses this module does not need to change. For more information, please refer to the document of the GPU embedding cache under the gpu_cache folder.

• Host memory cache for HugeCTR embedding training cache: We have introduced the host memory cache (HMEM-Cache) based PS for the incremental training, which is a component of the Embedding Training Cache (MOS) and responsible for handling the case when the embedding table is too large to even fit into the host memory. We have provided the SSD-based PS for this scenario in former releases, but the SSD-based PS will be deprecated from the v3.3 release due to its unsatisfactory performance. Please check the Host Memory Cache in MOS for a detailed introduction. Compared with the former SSD-based PS, the loading and dumping bandwidth of the HMEM-Cache based PS can be substantially improved if it is properly configured, which contributes to the incremental training of models with huge embedding tables when using the MOS feature in HugeCTR. To ease the utilization of the MOS feature, we have also simplified the python interface of MOS. Specifically, we drop the use_host_memory_ps entry by providing the ps_types entry for choosing the HMEM-based PS or the HMEM-Cache based PS; A unified entry sparse_models is introduced, and you don’t need to use different entries to tell whether a pre-trained embedding table exists or not. For a detailed explanation of the python interface, please check the HugeCTR Python Interface.

• Debugging Capability Improvement：We have introduced a set of new debugging capability features, which include the multi-level logging, more informative throw and check. We also provide a set of kernel debugging functions. Based on these features, we are actively working on making the information and error messages from HugeCTR cleaner, so that our users are well informed about what is happening with their training & inference code at a desired level. Stay tuned! For more detailed information, check out comments in header files located at HugeCTR/include/base/debug.

• Embedding cache asynchronous insertion mechanism：We have supported the asynchronous insertion of missing embedding keys into the embedding cache. This feature can be activated automatically through user-defined hit rate threshold in configuration file.When the real hit rate of the embedding cache is higher than the user-defined threshold, the embedding cache will insert the missing key asynchronously, and vice versa, it will still be inserted in a synchronous way to ensure high accuracy of inference requests. Through the asynchronous insertion method, compared with the previous synchronous method, the real hit rate of the embedding cache can be further improved after the embedding cache reaches the user-defined threshold.

• Performance optimization of Parameter Server：We have added support for multiple database interfaces to our parameter server. In particular, we added an “in memory” database, that utilizes the local CPU memory for storing and recalling embeddings and uses multi-threading to accelerate look-up and storage. Further, we revised support for “distributed” storage of embeddings in a Redis cluster. This way, you can use the combined CPU-accessible memory of your cluster for storing embeddings. The new implementation is up over two orders of magnitude faster than the previous. Further, we performance-optimized support for the “persistent” storage and retrieval of embeddings via RocksDB through the structured use of column families. Creating a hierarchical storage (i.e. using Redis as distributed cache, and RocksDB as fallback), is supported as well. These advantages are free to end-users, as there is no need to adjust the PS configuration. We plan to further integrate the hierarchical parameter server with other features, such as the GPU backed embedding caches in upcoming releases. Stay tuned!

• Graph Analysis to internalize the Slice layer: The branch topology is inherently supported by the HugeCTR model graph, which requires users to explicitly insert a Slice layer with Python APIs to enable it. In order to simplify the usage, the Slice layer for the branch topology can be abstracted away in the Python interface. The graph analysis will be conducted to resolve the tensor dependency and the Slice layer will be internally inserted if the same tensor is consumed more than once to form the branch topology. The previous usage of explicitly adding the Slice layer is still supported, while using this new feature to internalize it is strongly recommended. Please refer to Getting Started to see how to construct a model graph with branches without the Slice layer. You can refer to Slice Layer for more details.

What's New in Version 3.2

• New HugeCTR to ONNX Converter: We’re introducing a new HugeCTR to ONNX converter in the form of a Python package. All graph configuration files are required and model weights must be formatted as inputs. You can specify where you want to save the converted ONNX model. You can also convert sparse embedding models. For more information, refer to HugeCTR to ONNX Converter and HugeCTR2ONNX Demo Notebook.

• New Hierarchical Storage Mechanicsm on the Parameter Server (POC): We’ve implemented a hierarchical storage mechanism between local SSDs and CPU memory. As a result, embedding tables no longer have to be stored in the local CPU memory. The distributed Redis cluster is being implemented as a CPU cache to store larger embedding tables and interact with the GPU embedding cache directly. The local RocksDB serves as a query engine to back up the complete embedding table on the local SSDs and assist the Redis cluster with looking up missing embedding keys. Please find more information here

• Parquet Format Support Within the Data Generator: The HugeCTR data generator now supports the parquet format, which can be configured easily using the Python API. For more information, refer to Data Generator API.

• Python Interface Support for the Data Generator: The data generator has been enabled within the HugeCTR Python interface. The parameters associated with the data generator have been encapsulated into the DataGeneratorParams struct, which is required to initialize the DataGenerator instance. You can use the data generator's Python APIs to easily generate the Norm, Parquet, or Raw dataset formats with the desired distribution of sparse keys. For more information, refer to Data Generator API and Data Generator Samples.

• Improvements to the Formula of the Power Law Simulator within the Data Generator: We've modified the formula of the power law simulator within the data generator so that a positive alpha value is always produced, which will be needed for most use cases. The alpha values for Long, Medium, and Short within the power law distribution are 0.9, 1.1, and 1.3 respectively. For more information, refer to Data Generator API.

• Support for Arbitrary Input and Output Tensors in the Concat and Slice Layers: The Concat and Slice layers now support any number of input and output tensors. Previously, these layers were limited to a maximum of four tensors.

• New Continuous Training Notebook: We’ve added a new notebook to demonstrate how to perform continuous training using the model oversubscription (also referred to as Embedding Training Cache) feature. For more information, refer to HugeCTR Continuous Training.

• New HugeCTR Contributor Guide: We've added a new HugeCTR Contributor Guide that explains how to contribute to HugeCTR, which may involve reporting and fixing a bug, introducing a new feature, or implementing a new or pending feature.

• Enhancements to Sparse Operation Kits (SOK): SOK now supports TensorFlow 2.5 and 2.6. We also added support for identity hashing, dynamic input, and Horovod within SOK. Lastly, we added a new SOK docs set to help you get started with SOK.

• Supporting Arbitrary Number of Inputs in Concat Layer and Slice Layer: The Concat and Slice layers now support any number of input and output tensors, respectively. Previously, these layers would be limited to a maximum of 4 tensors.

• Fix power law in Data Generator (Generalize the power law simulator in Data Generator): We’ve modified the formula of the power law simulator to make for the positive alpha value, which is more general in different use cases. Besides, the alpha values for Long, Medium and Short of power law distribution are 0.9, 1.1 and 1.3 respectively. For more information, see Data Generator API.

What's New in Version 3.1

• MLPerf v1.0 Integration: We've integrated MLPerf optimizations for DLRM training and enabled them as configurable options in Python interface. Specifically, we have incorporated AsyncRaw data reader, HybridEmbedding, FusedReluBiasFullyConnectedLayer, overlapped pipeline, holistic CUDA Graph and so on. The performance of 14-node DGX-A100 DLRM training with Python APIs is comparable to CLI usage. For more information, see HugeCTR Python Interface and DLRM Sample.

• Enhancements to the Python Interface: We’ve enhanced the Python interface for HugeCTR so that you no longer have to manually create a JSON configuration file. Our Python APIs can now be used to create the computation graph. They can also be used to dump the model graph as a JSON object and save the model weights as binary files so that continuous training and inference can take place. We've added an Inference API that takes Norm or Parquet datasets as input to facilitate the inference process. For more information, see HugeCTR Python Interface and HugeCTR Criteo Notebook.

• New Interface for Unified Embedding: We’re introducing a new interface to simplify the use of embeddings and datareaders. To help you specify the number of keys in each slot, we added nnz_per_slot and is_fixed_length. You can now directly configure how much memory usage you need by specifying workspace_size_per_gpu_in_mb instead of max_vocabulary_size_per_gpu. For convenience, mean/sum is used in combinators instead of 0 and 1. In cases where you don't know which embedding type you should use, you can specify use_hash_table and let HugeCTR automatically select the embedding type based on your configuration. For more information, see HugeCTR Python Interface.

• Multi-Node Support for Embedding Training Cache (MOS): We’ve enabled multi-node support for the embedding training cache. You can now train a model with a terabyte-size embedding table using one node or multiple nodes even if the entire embedding table can't fit into the GPU memory. We're also introducing the host memory (HMEM) based parameter server (PS) along with its SSD-based counterpart. If the sparse model can fit into the host memory of each training node, the optimized HMEM-based PS can provide better model loading and dumping performance with a more effective bandwidth. For more information, see HugeCTR Python Interface.

• Enhancements to the Multi-Nodes TensorFlow Plugin: The Multi-Nodes TensorFlow Plugin now supports multi-node synchronized training via tf.distribute.MultiWorkerMirroredStrategy. With minimal code changes, you can now easily scale your single GPU training to multi-node multi GPU training. The Multi-Nodes TensorFlow Plugin also supports multi-node synchronized training via Horovod. The inputs for embedding plugins are now data parallel, so the datareader no longer needs to preprocess data for different GPUs based on concrete embedding algorithms.

• NCF Model Support: We've added support for the NCF model, as well as the GMF and NeuMF variant models. With this enhancement, we're introducing a new element-wise multiplication layer and HitRate evaluation metric. Sample code was added that demonstrates how to preprocess user-item interaction data and train a NCF model with it. New examples have also been added that demonstrate how to train NCF models using MovieLens datasets.

• DIN and DIEN Model Support: All of our layers support the DIN model. The following layers support the DIEN model: FusedReshapeConcat, FusedReshapeConcatGeneral, Gather, GRU, PReLUDice, ReduceMean, Scale, Softmax, and Sub. We also added sample code to demonstrate how to use the Amazon dataset to train the DIN model.

• Multi-Hot Support for Parquet Datasets: We've added multi-hot support for parquet datasets, so you can now train models with a paraquet dataset that contains both one hot and multi-hot slots.

• Mixed Precision (FP16) Support in More Layers: The MultiCross layer now supports mixed precision (FP16). All layers now support FP16.

• Mixed Precision (FP16) Support in Inference: We've added FP16 support for the inference pipeline. Therefore, dense layers can now adopt FP16 during inference.

• Optimizer State Enhancements for Continuous Training: You can now store optimizer states that are updated during continuous training as files, such as the Adam optimizer's first moment (m) and second moment (v). By default, the optimizer states are initialized with zeros, but you can specify a set of optimizer state files to recover their previous values. For more information about dense_opt_states_file and sparse_opt_states_file, see Python Interface.

• New Library File for GPU Embedding Cache Data: We’ve moved the header/source code of the GPU embedding cache data structure into a stand-alone folder. It has been compiled into a stand-alone library file. Similar to HugeCTR, your application programs can now be directly linked from this new library file for future use. For more information, see our GPU Embedding Cache ReadMe.

• Embedding Plugin Enhancements: We’ve moved all the embedding plugin files into a stand-alone folder. The embedding plugin can be used as a stand-alone python module, and works with TensorFlow to accelerate the embedding training process.

• Adagrad Support: Adagrad can now be used to optimize your embedding and network. To use it, change the optimizer type in the Optimizer layer and set the corresponding parameters.

Release Notes

Bigger model and large scale training are always the main requirements in recommendation system. In v3.1, we provide a set of new optimizations for good scalability as below, and now they are available in this beta version.

• Distributed Hybrid embedding - Model/data parallel split of embeddings based on statistical access frequency to minimize embedding exchange traffic.
• Optimized communication collectives - Hierarchical multi-node all-to-all for NVLINK aggregation and oneshot algorithm for All-reduce.
• Optimized data reader - Async I/O based data reader to maximize I/O utilization, minimize interference with collectives and eval caching.
• MLP fusions - Fused GEMM + Relu + Bias fprop and GEMM + dRelu + bgrad bprop.
• Compute-communication overlap - Generalized embedding and bottom MLP overlap.
• Holistic CUDA graph - Full iteration graph capture to reduce launch latencies and jitter.

What's New in Version 3.0.1

• DLRM Inference Benchmark: We've added two detailed Jupyter notebooks to illustrate how to train and deploy a DLRM model with HugeCTR whilst benchmarking its performance. The inference notebook demonstrates how to create Triton and HugeCTR backend configs, prepare the inference data, and deploy a trained model by another notebook on Triton Inference Server. It also shows the way of benchmarking its performance (throughput and latency), based on Triton Performance Analyzer. For more details, check out our HugeCTR inference repository.
• FP16 Speicific Optimization in More Dense Layers: We've optimized DotProduct, ELU, and Sigmoid layers based on __half2 vectorized loads and stores, so that they better utilize device memory bandwidth. Now most layers have been optimized in such a way except MultiCross, FmOrder2, ReduceSum, and Multiply layers.
• More Finely Tunable Synthetic Data Generator: Our new data generator can generate uniformly distributed datasets in addition to power law based datasets. Instead of specifying vocabulary_size in total and max_nnz, you can specify such information per categorical feature. See our user guide to learn its changed usage.
• Decreased Memory Demands of Trained Model Exportation: To prevent the out of memory error from happening in saving a trained model including a very large embedding table, the actual amount of memory allocated by the related functions was effectively reduced.
• CUDA Graph Compatible Dropout Layer: HugeCTR Dropout Layer uses cuDNN by default, so that it can be used together with CUDA Graph. In the previous version, if Dropout was used, CUDA Graph was implicitly turned off.

Release Notes

What’s New in Version 3.0

• Inference Support: To streamline the recommender system workflow, we’ve implemented a custom HugeCTR backend on the NVIDIA Triton Inference Server. The HugeCTR backend leverages the embedding cache and parameter server to efficiently manage embeddings of different sizes and models in a hierarchical manner. For additional information, see our inference repository.

• New High-Level API: You can now also construct and train your models using the Python interface with our new high-level API. See our preview example code to grasp how it works.

• FP16 Support in More Layers: All the layers except MultiCross support mixed precision mode. We’ve also optimized some of the FP16 layer implementations based on vectorized loads and stores.

• Enhanced TensorFlow Embedding Plugin: Our embedding plugin now supports LocalizedSlotSparseEmbeddingHash mode. With this enhancement, the DNN model no longer needs to be split into two parts since it now connects with the embedding op through MirroredStrategy within the embedding layer.

• Extended Model Oversubscription: We’ve extended the model oversubscription feature to support LocalizedSlotSparseEmbeddingHash and LocalizedSlotSparseEmbeddingHashOneHot.

• Epoch-Based Training Enhancement: The num_epochs option in the Solver clause can now be used with the Raw dataset format.

• Deprecation of the eval_batches Parameter: The eval_batches parameter has been deprecated and replaced with the max_eval_batches and max_eval_samples parameters. In epoch mode, these parameters control the maximum number of evaluations. An error message will appear when attempting to use the eval_batches parameter.

• MultiplyLayer Renamed: To clarify what the MultiplyLayer does, it was renamed to WeightMultiplyLayer.

• Optimized Initialization Time: HugeCTR’s initialization time, which includes the GEMM algorithm search and parameter initialization, was significantly reduced.

• Sample Enhancements: Our samples now rely upon the Criteo 1TB Click Logs dataset instead of the Kaggle Display Advertising Challenge dataset. Our preprocessing scripts (Perl, Pandas, and NVTabular) have also been unified and simplified.

• Configurable DataReader Worker: You can now specify the number of data reader workers, which run in parallel, with the num_workers parameter. Its default value is 12. However, if you are using the Parquet data reader, you can't configure the num_workers parameter since it always corresponds to the number of active GPUs.

Known Issues

• Since the automatic plan file generator isn't able to handle systems that contain one GPU, you must manually create a JSON plan file with the following parameters and rename it using the name listed in the HugeCTR configuration file: {"type": "all2all", "num_gpus": 1, "main_gpu": 0, "num_steps": 1, "num_chunks": 1, "plan": [[0, 0]], and "chunks": [1]}.

• If using a system that contains two GPUs with two NVLink connections, the auto plan file generator will print the following warning message: RuntimeWarning: divide by zero encountered in true_divide. This is an erroneous warning message and should be ignored.

• The current plan file generator doesn't support a system where the NVSwitch or a full peer-to-peer connection between all nodes is unavailable.

• Users need to set an export CUDA_DEVICE_ORDER=PCI_BUS_ID environment variable to ensure that the CUDA runtime and driver have a consistent GPU numbering.

• LocalizedSlotSparseEmbeddingOneHot only supports a single-node machine where all the GPUs are fully connected such as NVSwitch.

• HugeCTR version 3.0 crashes when running the DLRM sample on DGX2 due to a CUDA Graph issue. To run the sample on DGX2, disable the CUDA Graph by setting the cuda_graph parameter to false even if it degrades the performance a bit. This issue doesn't exist when using the DGX A100.

• The HugeCTR embedding TensorFlow plugin only works with single-node machines.

• The HugeCTR embedding TensorFlow plugin assumes that the input keys are in int64 and its output is in float.

• If the number of samples in a dataset is not divisible by the batch size when in epoch mode and using the num_epochs instead of max_iter, a few remaining samples are truncated. If the training dataset is large enough, its impact can be negligible. If you want to minimize the wasted batches, try adjusting the number of data reader workers. For example, using a file list source, set the num_workers parameter to an advisor based on the number of data files in the file list.

Release Notes

What's New in Version 2.3

We’ve implemented the following enhancements to improve usability and performance:

• Python Interface: To enhance the interoperability with NVTabular and other Python-based libraries, we're introducing a new Python interface for HugeCTR. If you are already using HugeCTR with JSON, the transition to Python will be seamless for you as you'll only have to locate the hugectr.so file and set the PYTHONPATH environment variable. You can still configure your model in your JSON config file, but the training options such as batch_size must be specified through hugectr.solver_parser_helper() in Python. For additional information regarding how to use the HugeCTR Python API and comprehend its API signature, see our Jupyter Notebook tutorial.

• HugeCTR Embedding with Tensorflow: To help users easily integrate HugeCTR’s optimized embedding into their Tensorflow workflow, we now offer the HugeCTR embedding layer as a Tensorflow plugin. To better understand how to intall, use, and verify it, see our Jupyter notebook tutorial. It also demonstrates how you can create a new Keras layer EmbeddingLayer based on the hugectr_tf_ops.py helper code that we provide.

• Model Oversubscription: To enable a model with large embedding tables that exceeds the single GPU's memory limit, we added a new model prefetching feature, giving you the ability to load a subset of an embedding table into the GPU in a coarse grained, on-demand manner during the training stage. To use this feature, you need to split your dataset into multiple sub-datasets while extracting the unique key sets from them. This feature can only currently be used with a Norm dataset format and its corresponding file list. This feature will eventually support all embedding types and dataset formats. We revised our criteo2hugectr tool to support the key set extraction for the Criteo dataset. For additional information, see our Python Jupyter Notebook to learn how to use this feature with the Criteo dataset. Please note that The Criteo dataset is a common use case, but model prefetching is not limited to only this dataset.

• Enhanced AUC Implementation: To enhance the performance of our AUC computation on multi-node environments, we redesigned our AUC implementation to improve how the computational load gets distributed across nodes.

• Epoch-Based Training: In addition to max_iter, a HugeCTR user can set num_epochs in the Solver clause of their JSON config file. This mode can only currently be used with Norm dataset formats and their corresponding file lists. All dataset formats will be supported in the future.

• Multi-Node Training Tutorial: To better support multi-node training use cases, we added a new a step-by-step tutorial.

• Power Law Distribution Support with Data Generator: Because of the increased need for generating a random dataset whose categorical features follows the power-law distribution, we revised our data generation tool to support this use case. For additional information, refer to the --long-tail description here.

• Multi-GPU Preprocessing Script for Criteo Samples: Multiple GPUs can now be used when preparing the dataset for our samples. For additional information, see how preprocess_nvt.py is used to preprocess the Criteo dataset for DCN, DeepFM, and W&D samples.

Known Issues

• Since the automatic plan file generator is not able to handle systems that contain one GPU, a user must manually create a JSON plan file with the following parameters and rename using the name listed in the HugeCTR configuration file: {"type": "all2all", "num_gpus": 1, "main_gpu": 0, "num_steps": 1, "num_chunks": 1, "plan": [[0, 0]], "chunks": [1]}.
• If using a system that contains two GPUs with two NVLink connections, the auto plan file generator will print the following warning message: RuntimeWarning: divide by zero encountered in true_divide. This is an erroneous warning message and should be ignored.
• The current plan file generator doesn't support a system where the NVSwitch or a full peer-to-peer connection between all nodes is unavailable.
• Users need to set an export CUDA_DEVICE_ORDER=PCI_BUS_ID environment variable to ensure that the CUDA runtime and driver have a consistent GPU numbering.
• LocalizedSlotSparseEmbeddingOneHot only supports a single-node machine where all the GPUs are fully connected such as NVSwitch.
• HugeCTR version 2.2.1 crashes when running our DLRM sample on DGX2 due to a CUDA Graph issue. To run the sample on DGX2, disable the use of CUDA Graph with "cuda_graph": false even if it degrades the performance a bit. We are working on fixing this issue. This issue doesn't exist when using the DGX A100.
• The model prefetching feature is only available in Python. Currently, a user can only use this feature with the DistributedSlotSparseEmbeddingHash embedding and the Norm dataset format on single GPUs. This feature will eventually support all embedding types and dataset formats.
• The HugeCTR embedding TensorFlow plugin only works with single-node machines.
• The HugeCTR embedding TensorFlow plugin assumes that the input keys are in int64 and its output is in float.
• When using our embedding plugin, please note that the fprop_v3 function, which is available in tools/embedding_plugin/python/hugectr_tf_ops.py, only works with DistributedSlotSparseEmbeddingHash.

What’s New in Version 2.2.1

In HugeCTR version 2.2.1, we enriched the user-convenience features together with the refactoring efforts and bug fixes.

• Dataset in Parquet Format Support : HugeCTR data reader was extended to support Parquet format. The preprocessed dataset and its metadata can be generated with nvTabular

• GPU-Powered Preprocessing Script : The preprocessing script used for HugeCTR samples such as DCN, DeepFM and W&D was rewritten in nvTabular. This GPU accelerated script doesn’t use too much host memory anymore.

• Preprocessing Tool for DLRM Sample : To make it easier for the user to run our DLRM sample, a preprocessing tool written in CUDA C++ was added.

• Use of RAPIDS MLPrims : Some existing layers were rewritten to utilize the highly optimized machine learning primitives (MLPrims) of RAPIDs.

• Reorganization of Submodules : All the submodules were moved to third_party directory.

• Revived Pascal Support : The support for Pascal Architecture, e.g., P100 was added back. However, with a Pascal graphic card, InteractionLayer doesn’t support FP16.

• Compile Time Reduction : By modularizing the embedding related code into several files, HugeCTR compile time was improved.

• Refactoring of Tensor and GeneralBuffer : In HugeCTR, Tensor and GeneralBuffer are used for memory management and access control. In the version 2.2.1, they were refactored to clarify their responsibilities and support different memory kinds, .e.g, Host, Device and Unified. Check their interface changes if you are using them to add a new layer.

New Features in Version 2.2

HugeCTR version 2.2 adds a lot of features to enhance its usability and performance. HugeCTR is not only a high-performance refereence design for framework designers but also a self contained training framework.

• Algorithm Search : HugeCTR runs an exhaustive algorithm search for each fully connected layer to find the best performant one.
• AUC : An user can choose to use AUC as an evaluation metric in addition to AverageLoss. It is also possible to stop training when AUC reaches a speicifed threshold.
• Batch Shuffle in Training Data Set : Training data batch shuffling is supported.
• Different Batch Sizes for Training and Evaluation : An user can speicify the different batch sizes for training and evalation. It can be useful to tune overal performance.
• Full FP16 pipeline : In order to data and compute throughputs, We added the full FP16 pipeline.
• Fused Fully Connected Layer : In FP16 mode, you can choose to use a specilized fully connected layer fused with ReLU activation function.
• Evaluation Data Caching on Device : For GPUs with large memory capacity like A100, a user can choose to cache data batches for small evaluation data sets.
• Interaction Layer : We added Intearction layer used for popular models such as DLRM.
• Optimized Data Reader for Raw Data Format : RAW data format is supported to simplify the one hot data reading and achieve better performance.
• Deep Learning Recommendation Model (DLRM) : We eanbled and optimized the training of DLRM. Please find more details in samples/dlrm.
• Learning Rate Scheduling : Different learning rate scheduling is supported.
• Weight Initialization Methods : For each trainable layer, a use can choose which method ,e.g., XavierUnifrom, Zero, etc is used for its weight initialization.
• Ampere Support : We tested and optimized HugeCTR for Ampere Architecture.

HugeCTR is a high-efficiency GPU framework designed for Click-Through-Rate (CTR) estimation training. In version 2.2 beta release we introduce several important feature updates:

• Algorithm Search : Support algorithm selection in fully connected layers for better performance.
• AUC : Support AUC calculation for accuracy evaluation.
• Batch shuffle and last batch in eval : Support batch shuffle and the last batch during evaluation won’t be dropped.
• Different batch size in training and evaluation : Support this for best performance in evaluation.
• Full mixed precision training pipeline: Support full mixed precision training [1].
• Fused fully connected layer : Fused bias adding and relu activation into a single layer.
• Caching evaluation data on device : For the GPUs with large memory like A100, we can use caching data for small evaluation data sets.
• Interaction layer
• Optimized data reader for raw format
• Learning rate scheduling

[1] P. Micikevicius, S. Narang, J. Alben, G. Diamos, E. Elsen, D. Garcia, B. Ginsburg, M. Houston, O. Kuchaiev, G. Venkatesh, H. Wu Mixed Precision Training https://arxiv.org/abs/1710.03740

HugeCTR is a high-efficiency GPU framework designed for Click-Through-Rate (CTR) estimation training, and the new released NVIDIA A100 GPU has excellent acceleration on various scales for AI, data analysis and high performance computing (HPC), and meet extremely severe computing challenges. To demonstrate HugeCTR’s performance on A100 GPU, this version is developed to leverage new features of the latest GPU.

New Features:

• Supporting three important networks: Wide and Deep Learning (WDL), Deep Cross Network (DCN) and DeepFM.
• A new embedding implementation LocalizedSlotSparseEmbedding which reduces the memory transactions across GPUs and nodes resiliently to the number of GPUs.
• Supporting multiple Embeddings in one network.
• Supporting dense feature input.
• Supporting new layers like: Dropout / Split / Reshape / Multiply / FmOrder2 / MultCross / Add
• Check bits in data reader to enable data check and error skip.