NVIDIA/DALI v1.40.0 on GitHub

Key Features and Enhancements

This DALI release includes the following key features and enhancements:

Added operators: fn.zeros , fn.zeros_like , fn.ones , fn.ones_like , fn.full and fn.full_like (#5505).
Added support for H264, H265, and AV1 video formats to fn.plugin.video (#5504).
Added support for CUDA 12.5U1 (#5545).

Fixed Issues

Fixed following issues with S3 files reading:
- Fixed handling of file names with whitespaces in TFRecord reader (#5525).
- Fixed loading when no GPU is available (#5533).
- Fixed handling of TFrecord index file (#5515).

Improvements

Dependency update 07/2024 (#5556)
Move checkpoint to IterationData. Remove ExecIterData. (#5555)
Remove pruning from the Executor. (#5553)
Move most of Operator to OperatorBase. Unify and simplify operator interfaces. (#5548)
Move graph visiting utilities to a separate file. (#5549)
Move to CUDA 12.5U1 (#5545)
Extend the external source signature to include all arguments (#5541)
Update DALI_deps version (#5536)
Pin numpy to <1.24 in TensorFlow examples (#5534)
Use new graph in Pipeline (#5520)
Deps update 06/24 (#5514)
Revert reducing the number of epoch in SBSA training test case (#5531)
Add AV1 support (#5504)
Removes MXNet support from DALI (#5526)
Video decoder in plugin (#5477)
Checkpoint refactoring - recognize checkpoints by operator instance name. (#5503)
Keep separate per-pipeline operator counters. Error out when "stealing" subgraphs from other pipelines results in duplicate names. (#5506)
Graph lowering. (#5496)
Use "device" and "preserve" built-in arguments in OpGraph2. (#5516)
Add fn.zeros, fn.zeros_like, fn.ones, fn.ones_like, fn.full and fn.full_like (#5505)

Bug Fixes

Support spaces in S3 paths (#5525)
Fix device ID in s3_client_manager (#5533)
Add failure tests for stealing subgraphs. Minor fix in pipeline validation. (#5518)
TFRecord to support S3 index URIs (#5515)
Exclude docs line length adjustment PR from the blame history (#5509)
Fix keras compat mode for ResNet50 tensorflow example (#5530)

Breaking API changes

DALI 1.39 was the final release to support the MXNet integration.

Deprecated features

No features were deprecated in this release.

Known issues:

Starting with DALI 1.39, a performance regression was observed in hardware-accelerated image decoders for setups with high number of worker threads. The nvImageCodec hardware decoder pre-allocation uses higher mini-batch size, causing extra cuMemFree calls that may slowing down decoding in some iterations. The issue will be fixed in the upcoming release.
The following operators: experimental.readers.fits, experimental.decoders.video, and experimental.inputs.video do not currently support checkpointing.
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.
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.
The DALI TensorFlow plugin might not be compatible with TensorFlow versions 1.15.0 and later.
To use DALI with the TensorFlow version that does not have a prebuilt plugin binary shipped with DALI, make sure that the compiler that is used to build TensorFlow exists on the system during the plugin installation. (Depending on the particular version, you can use GCC 4.8.4, GCC 4.8.5, or GCC 5.4.)
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 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.

Binary builds

NOTE: DALI builds for CUDA 12 dynamically link the CUDA toolkit. To use DALI, install the latest CUDA toolkit.

CUDA 11.0 and CUDA 12.0 builds use CUDA toolkit enhanced compatibility. 
They are built with the latest CUDA 11.x/12.x toolkit respectively but they can run on the latest, 
stable CUDA 11.0/CUDA 12.0 capable drivers (450.80 or later and 525.60 or later respectively).
However, using the most recent driver may enable additional functionality. 
More details can be found in enhanced CUDA compatibility guide.

Install via pip for CUDA 12.0:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda120==1.40.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda120==1.40.0

or just:

pip install nvidia-dali-cuda120==1.40.0
pip install nvidia-dali-tf-plugin-cuda120==1.40.0

For CUDA 11:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda110==1.40.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda110==1.40.0

or just:

pip install nvidia-dali-cuda110==1.40.0
pip install nvidia-dali-tf-plugin-cuda110==1.40.0

Or use direct download links (CUDA 12.0):

Or use direct download links (CUDA 11.0):

FFmpeg source code:

This software uses code of FFmpeg licensed under the LGPLv2.1 and its source can be downloaded here

Libsndfile source code:

https://developer.download.nvidia.com/compute/redist/nvidia-dali/libsndfile-1.2.2.tar.gz

NVIDIA/DALI v1.40.0 DALI v1.40.0 on GitHub