NVIDIA/DALI v2.0.0 on GitHub

Key Features and Enhancements

This DALI release includes the following key features and enhancements:

Improved DALI dynamic mode:
- Added asynchronous and deferred execution (#6210, #6204, #6124, #6216, #6152)
- Improved multithreading, supporting no-gil Python 3.13t and Python 3.14. (#6200, #6174, #6136, #5884, #6201, #6202, #6164, #6142)
- Added TorchData integration (#6198)
- Improved usability and interoperability with other libraries (#6131, #6182, #6188, #6172, #6179, ##6143)
- Improved execution device specification and handling (#6194, #6165)
- Improved examples and documentation (#6140, #6189, #6170)
Added contrast-limited adaptive histogram equalization (CLAHE) operator (#6069)
- Thank you @tonyreina for your contribution!
Added support for CUDA 13.1U1 (#6163)
Improved slice, full, zeros, ones operators (#6159, #6109, #6169)

Fixed Issues

Added DALI_MAX_IMAGE_SIZE env var to limit decoded image size in CPU and GPU decoders. (#6208)
Fiedx out-of-bounds reads in image format detection. (#6207)
Fixed audio decoder handling of files over 2GB. (#6199)
Fixed random crop operators conforming to new random state passing. (#6190)
Fixed displacement filter occasionaly returning corrupted data due to missing synchronization. (#6168)
Replaced pickle with JSON in DALI checkpoints format. (#6154)
Fixed slicing with negative stride. (#6161)
Fixed memory leak (#6153) in fixed-size poll allocator. (#6158)

Improvements

Add a function that purges operator instance cache for an EvalContext. (#6216)
Add TorchData integration in dynamic mode and create examples (#6198)
Add exception propagation for deferred and async execution (#6210)
Update VERSION to 2.0.0
Add ndd.Stream.synchronize method and implement EvalMode.sync_full (#6204)
ndd vs fn tests part 1: utils and automated tests (#6191)
Add multithreading guide for dynamic mode (#6200)
Limit thread count to 32 in ndd multithreading tests. (#6201)
Fix the conda tests in free threaded env (#6202)
Improved device handling. Remove mixed device. Make DALI work without GPU (#6194)
Replace deprecated pkg_resources.require with packaging/importlib-based alternative (#6196)
Add first class batch to tensor conversion with optional padding (#6182)
Make DALI Dynamic and Pipeline APIs two separate sections (#6189)
Documentation for ndd.DType (#6170)
Add multithreaded tests for dynamic mode (#6164)
Exclude ndd readers from operator docs (#6173)
Update DALI_DEPS: libsound, openssl (#6185)
Broadcast lists of scalars into any shape in ArgValue. (#6188)
Add per-thread stream. Rework stream semantics. Add a real Python stream class. (#6174)
Hide deprecated operators from documentation (#6180)
Fix jupyter tests (#6184)
Move to CUDA 13.1U1 (#6163)
Improve the interoperability of dynamic mode with PyTorch (#6172)
Remove debug mode references from documentation (#6175)
Create examples showing ndd usage (#6140)
Add str and repr generic formatting utilites (#6167)
Add layout handling to full, zeros, ones operator family (#6159)
Make EvalMode.eager the default (#6152)
Default num_threads and stream for dynamic API (#6165)
Dependency update 2026-02 (#6155)
Unexperimentalize operators (#6134)
Adjust performance threshold for dynamic mode in TL1_decoder_perf (#6160)
Update PyTorch Lightning example notebook (#6145)
Fix O_DIRECT expected to read number of bytes numpy reader (#6148)
Add pkg_resources compatibility fallback using importlib.metadata (#6144)
Relax numpy version constraints (#6137)
Move inflate from experimental to decoders, fix doc hiding for ndd, bump deprecation cut-off for ndd to 2.0 (#6141)
Support asynchronous execution in dynamic mode. (#6124)
Fix conda free-threaded Python build (#6142)
Add experimental Python 3.14 support and remove Python 3.9 (#6136)
Add dynamic mode RN50 pipeline to hw decoder bench (#6115)
Add --no-build-isolation flag to cocoapi pip install (#6132)
Improve interoperability of ndd tensors with third party libraries (#6131)
Fix cuFFT linking to respect BUILD_FFTS option (#6135)
Enable cross-device copy with cudaMemcpyPeerAsync. (#6130)
Add support for Python 3.13t (#5884)
Upgrade GitHub Actions for Node 24 compatibility (#6133)
Add PyTorch DataLoader Evaluator plugin (#6112)
Hide ops API (#6123)
Add the information of deprecation version origin (#6127)
Change the defaults for build options in docker/build_helper.sh (#6129)
Allow non-copying TensorList construction from a list of tesnors. (#6128)
Move all internal dnn API class/object public members to private (#6120)
Support more border modes in Slice (#6109)
Contrast-limited adaptive histogram equalization (CLAHE) to DALI image operators (#6069)
Add USE_PREBUILD_PYBIND11 option to use system pybind11 (#6117)
Drop Python 3.9 support (#6119)
Move to cuda 13.1 (#6116)
Remove old eager mode. (#6113)

Bug Fixes

Allocate CPU outputs in host order. Reset workspace order to host whe… (#6217)
Fix workspace stream handling in CPU imgcodec decoders. (#6215)
Add missing pillow installation in TL0_self_test_Ampere test (#6213)
Add DALI_MAX_IMAGE_SIZE env var to limit decoded image size in CPU and GPU decoders (#6208)
Accept more types in BBoxRotate input_shape argument. (#6212)
Fix out-of-bounds reads in image format detection (#6207)
Rework instance cache. (#6206)
Use notify_all instead of notify for EvalMode.async (#6205)
Fix dynamic mode pyi files (#6187)
Add sharding support to dynamic mode Reader (#6197)
Fix audio decoder to support files over 2GB (#6199)
Improve type hints in dynamic mode (#6183)
Safely calling Operator._init_spec in invocation.py (#6193)
Rework random crop operators (#6190)
Fix batch creation from unevaluated tensors (#6178)
Forbid passing axes to expand_dims as an input. (#6181)
Fix stream handling in tensor join when called from Dynamic mode. (#6171)
Fix batch construction from a tensor and layout. Add ability to change batch layout in batch and as_batch. (#6179)
Add handling of default layouts in standalone operator calls. (#6176)
Prevent deadlocks with asynchronous execution (#6177)
Set the device of ndd tensor slices (#6169)
Add missing __syncthreads in displacement filter. (#6168)
Use JSON in pipeline checkpointing (#6154)
Limit the maximum number of items stored in a fixed-size poll allocator. (#6158)
Fix slicing with negative stride. (#6161)
Enable arithmetic operations between device tensors/batches and scalars (#6143)
Make TFRecord work with dynamic mode (#6151)
Fix enum handling in Dynamic Mode + require NumPy (#6150)
Make internal hidden operators names private by adding underscore (#6125)
Dynamic API: pass cross-device copies through host memory. (#6121)

Breaking API changes

DALI 1.53 was last release supporting Python 3.9 (#6119)
Experimental eager and debug modes were removed in favour of dynamic mode. (#6175, #6113)
Multiple fn operators were moved from fn.experimental namespace (#6134, #6141)

Deprecated features

No features were deprecated in this release.

Known issues:

In some cases, the pass-through parallel external source outputs may be corrupted when used with pipelined dynamic executor. The issue occurs when all four conditions are met: 1. the pipeline uses dynamic executor exec_dynamic=True (default), 2. the external_source runs in parallel mode (parallel=True), 3. the ES output is directly returned from the pipeline, 4. the ES output is a single contiguous chunk of memory (either batch=True or batch_size=1). Currently, as a workaround, user can specify exec_dynamic=False when instantiating pipeline or add an extra fn.copy to prevent directly returning ES outputs from the pipeline.
A problem with insufficient static TLS allocation size has been observed on Ubuntu 22.04 for aarch64 that can result in process crash when loading dynamic libraries. Updating glibc to 2.39 or newer, or specifying higher static TLS size with GLIBC_TUNABLES=glibc.rtld.optional_static_tls=10000 should resolve the issue.
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 dynamically link the CUDA toolkit. To use DALI, please install the latest (12.x or 13.x) CUDA toolkit.

DALI builds use CUDA toolkit enhanced compatibility: 
DALI is built with the latest CUDA 12.x/13.x toolkit but can be run on any stable drivers from the respective CUDA family (525 and 580).
Using the most recent driver may enable additional functionality. 
More details can be found in enhanced CUDA compatibility guide.

Install via pip for CUDA 13.0:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda130==2.0.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda130==2.0.0

or just:

pip install nvidia-dali-cuda130==2.0.0
pip install nvidia-dali-tf-plugin-cuda130==2.0.0

For CUDA 12:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda120==2.0.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda120==2.0.0

or just:

pip install nvidia-dali-cuda120==2.0.0
pip install nvidia-dali-tf-plugin-cuda120==2.0.0

Or use direct download links (CUDA 13.0):

Or use direct download links (CUDA 12.0):

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 v2.0.0 DALI v2.0.0 on GitHub