Release Notes#

This document describes the key features, software enhancements and improvements, and known issues for DALI 1.44.0. For previously released DALI documentation, see DALI Archives.

Overview#

DALI offers both performance and flexibility of accelerating different data pipelines (graphs that can have multiple outputs and inputs), as a single library, that can be easily integrated into different deep learning training and inference applications.

Using DALI#

Note

DALI builds for NVIDIA® CUDA® 12 dynamically link the CUDA toolkit. To use DALI, install the latest CUDA toolkit.

To upgrade to DALI 1.44.0 from a previous version of DALI, follow the installation and usage information in the DALI User Guide.

Note

The internal DALI C++ API used for operator’s implementation, and the C++ API that enables using DALI as a library from native code, is not yet officially supported. Hence these APIs may change in the next release without advance notice.

Key Features and Enhancements#

This DALI release includes the following key features and enhancements:

  • The dynamic executor (exec_dynamic) is no longer experimental. It supports GPU to CPU transfers and reduces memory consumption. (#5704)

  • Added support for zero-copy outputs transfer with dynamic executor. (#5684, #5673)

  • Eliminated the outputs copy in PyTorch plugin. (#5699)

  • Added dynamic executor support to TF plugin. (#5686)

  • Optimized pipeline’s output contiguity handling. (#5677)

Fixed Issues#

The following issues were fixed in this release:

  • Restricted nvImageCodec version in DALI wheel dependencies list, as the most recent nvImageCodec (0.4.0) is incompatible. (#5709)

  • Fixed custom stream handling on non-default device in fn.external_source (#5690)

  • Fixed problem with using DALI with Python3.12 with no distutils/setuptools installed.

  • Fixed incorrect stream usage in fn.experimental.inputs.video (#5682)

  • Fixed possible hang in video decoder when rewinding near the last keyframe (#5676, #5669)

  • Fixed dont_use_mmap option handling in fn.readers.webdataset (#5683)

  • Fixed redundant usage of pinned memory in the CPU fn.readers.numpy reader (#5678)

  • Fixed dynamic executor’s handling of operators that produce no outputs (#5674)

Breaking Changes#

  • There are no breaking changes in this DALI release.

Deprecated Features#

  • No issues were deprecated in this release.

Known Issues#

This DALI release includes the following known issues:

  • The most recent nvImageCodec (0.4.0) is currently incompatible with DALI. Python wheel for DALI 1.44 pins the dependency to 0.3.0, but older releases do not specify the required version explicitly. Users of previous DALI releases may need to manually install older nvImageCodec in order to use fn.experimental.decoders.image.* or, for DALI 1.39 and 1.40, fn.decoders.image.*. Install the compatible version by running the pip install nvidia-nvimgcodec-cu12~=0.3.0 command.

  • The following operators do not currently support checkpointing: experimental.readers.fits, experimental.decoders.video, experimental.inputs.video, and experimental.decoders.image_random_crop.

  • The video loader operator requires that the key frames occur, at a minimum, every 10 to 15 frames of the video stream.

    If the key frames occur at a frequency that is less than 10-15 frames, the returned frames might be out of sync.

  • The experimental VideoReaderDecoder does not support open GOP.

    It will not report an error and might produce invalid frames. VideoReader uses a heuristic approach to detect open GOP and should work in most common cases.

  • In experimental debug and eager modes, the GPU external source is not properly synchronized with DALI internal streams.

    As a workaround, you can manually synchronize the device before returning the data from the callback.

  • Due to some known issues with meltdown/spectra mitigations and DALI, DALI shows the best performance when running in Docker with escalated privileges, for example:

    • privileged=yes in Extra Settings for AWS data points

    • --privileged or --security-opt seccomp=unconfined for bare Docker

On this page