NVIDIA/DALI v2.1.0 on GitHub

Key Features and Enhancements

This DALI release includes the following key features and enhancements:

Added torchvision API:
- Enabled composing DALI pipelines with torchvision-like API (#6281, #6278, #6276, #6275, #6272, #6266, #6229)
- Added utilities for torch tensor/PIL image conversions (#6287)
Improved DALI dynamic:
- Added new thread pool for better threads utilization and sharing across process (#4635, #6219, #6245, #6224, #6254)
- Improved error-reporting (#6210, #6260)
- Improved deletion order handling to avoid unnecessary syncs (#6277)
- Improved readers API (#6252)
Improved video readers:
- Added uniform_sample option to VideoReaderDecoder (#6258)
- Added enable_frame_num='sequence' mode to video readers. (#6237)
Improved free-threaded Python support:
- Support free-threaded Python in DALI python_function (#6289)
- Replaced dm-tree with optree dependency (#6225)
Added support for CUDA 13.2 (#6249)
Added support for instantiating operators and building pipelines in C API (#6253)
Updated JAX integration to support JAX 0.9. (#6238, #6286, #6259, #6256, #6247)

Fixed Issues

Documented workaround for CUDA graph capture clash between JAX and DALI (#6286)
Fixed out-of-bounds access and key handling in Caffe/Caffe2 reader (#6211)
Fixed range clamping in subscript operator. (#6242)
Fixed too strict contiguity check when importing tensors via DLPack (#6285)
Fixed inflate operator max output estimation (#6283)

Improvements

Update DALI_DEPS_VERSION (patch libtiff) (#6295)
Torchvision API to tensor/PIL image conversion operators (#6282)
Allow passing tensor arguments in reader constructors (#6252)
Update third-party dependencies (2026-04-09) (#6287)
Declare free-threaded Python support on the python_function plugin (#6289)
Torchvision API documentation (#6281)
Update VERSION to 2.1.0
Move dynamic API class constructor docs to class-level docstrings (#6273)
Torchvision normalize (#6278)
Pipeline building in C API (#6253)
Torchvision padding (#6276)
Torchvision gaussian blur (#6275)
Torchvision API - ColorJitter and Grayscale operators (#6272)
Add uniform_sample option to VideoReaderDecoder (#6258)
Torchvision API - center crop operator (#6266)
Add quiet argument to RandomBBoxCrop to suppress crop failure warning (#6270)
Fix Coverity detected defects (#6257)
Improve deadsnakes PPA key handling in aarch64-linux Dockerfile (#6268)
Use NewThreadPool in dynamic mode. Use only one default instance of ThreadPool per device. (#6254)
Change a way deadsnakes ppa is accessed (#6263)
Torchvision API infrastructure (#6229)
New ThreadPool + thread pool facade (#6224)
Make result of AtScopeExit non-discardable. (#6248)
Move to CUDA 13.2 (#6249)
Add an ability to skip in-test timestamps (#6250)
Update third-party dependencies (2026.03) (#6243)
Add enable_frame_num='sequence' mode to video readers. (#6237)
Update JAX plugin to JAX 0.9. (#6238)
Add non-cooperative jobs to new ThreadPool (#6245)
Add shuffle_after_epoch_seed argument to file-based readers. (#6236)
Add numpy missing dependency to TL1_custom_src_pattern_build (#6240)
Set TensorList deletion order in set_order when possible (#6235)
Improve NDD operator filtering. (#6239)
Add numpy as an explicit conda dependency for dali_python_bindings (#6232)
Remove experimental C++ API documentation page. (#6230)
Rework NVTX annotations in dynamic mode (#6227)
Replace dm-tree with optree (#6225)
Add dump_artifacts flag to avoid dumping artifacts for expected test failures (#6223)
Improve nvcomp header detection for dynamic nvcomp builds (#6226)
Raise exceptions when an EvalContext is active in multiple threads (#6221)
Cleanup after instance cache rework. (#6209)
Remove start_immediately parameter from AddWork. (#6219)
Remove python tests with forced new executor. (#6222)
New thread pool (#4635)
Add exception propagation for deferred and async execution (#6210)

Bug Fixes

Compile the function ahead of time in the JAX example (#6286)
Add torchvision module to exclusion list in conda jupyter notebook (#6291)
Fix glob_to_regex for Python 3.14 (#6290)
DLPack import: Relax stride check in unit dimensions. (#6285)
Fix inflate operator: Reset max output volume. Use size in bytes, not elements. (#6283)
Fix stream handling in cvcuda resize. (#6284)
Defer DLTensor deletion when CUDA graph capture is active. (#6259)
Fix call stack depth handling for error tracebacks in dynamic mode (#6262)
Remove misleading legacy CMN warning for video layouts (#6269)
SequenceOperator: Do not keep thread pool and output order from the 1st iteration. (#6264)
Fix missing NVML_ENABLED guards around nvml.h includes (#6255)
Fix DALIGenericPeekableIterator missing pmap_compatible parameter. (#6256)
Fix UB in JpegCompressionDistortion: use data() for past-the-end pointer (#6251)
Fix compatibility with flax-basic_example.ipynb after JAX update (#6247)
Fix documentation switcher generation for 2.+ releases (#6246)
Fix range clamping in subscript operator. (#6242)
Fix TL1_python-nvjpeg_test test (#6233)
Fix EvalContext reentrancy (#6220)
Dynamic vs Pipeline mode equivalence tests, part 2 (#6214)
Fix builds with non-dynamic nvCOMP. (#6231)
Add missing include directory to nvCOMP stubgen command. (#6228)
Add ArgValue broadcasting when using ShapeFromSize callback. (#6218)
Fix out-of-bounds access and key handling in Caffe/Caffe2 reader (#6211)

Breaking API changes

There are no breaking changes in this DALI release.

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==2.1.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda130==2.1.0

or just:

pip install nvidia-dali-cuda130==2.1.0
pip install nvidia-dali-tf-plugin-cuda130==2.1.0

For CUDA 12:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda120==2.1.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda120==2.1.0

or just:

pip install nvidia-dali-cuda120==2.1.0
pip install nvidia-dali-tf-plugin-cuda120==2.1.0

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 v2.1.0 DALI v2.1.0 on GitHub