Releases · NVIDIA/DALI

Release list

DALI v2.2.0 Latest

Latest

stiepan released this 29 Jun 12:39

v2.2.0

5a6c01c

Key Features and Enhancements

This DALI release includes the following key features and enhancements:

Improved DALI dynamic:
- Added transparent pipelining (#6382, #6357, #6363, #6323, #6301, #6298)
- Added support for checkpointing (#6368, #6340)
- Added static metadata inference (#6317, #6280, #6311)
- Added support for numpy dtypes (#6350)
- Improved dynamic mode control features (#6312)
- Improved random number generators (#6334, #6332)
- Improved performance (#6329)
Improved Torchvision API (#6383, #6383, #6313, #6364, #6297, #6372)
- Added operators RandomResizedCrop, Crop, RandomApply, RandomGrayscale (#6369, #6353, #6342, #6335)
Improved agent skills (#6376, #6358, #6375, #6368, #6374, #6271)
Made nvImageCodec 0.8 a default image decoding backend. (#6366, #6349, #6344, #6343, #6337, #6293)
Added support for CUDA 13.3 (#6371)
Added support for nvCOMP to 5.2.0 (#6292)
Improved DLPack support (#6261, #6307)
Extended arithmetic operators coverage: bit shifts, bit not and modulo (#6294)
Added support for shuffle_after_epoch parameter in WebDataset, TFRecord, and MXNet readers (#6288)

Fixed Issues

Removed unnecessary samples' meta-data copies in FileReader (#6379)
- Thanks @fallintoplace for your contribution!
Fixed parsing of malformed .npy headers numpy reader (#6380)
Added missing validation for COCO annotation bbox size (#6360)
Improved validation/error reporting for batch permutation in ndd (#6359)
Fixed validation for all operands residing on the same device in ndd (#6347)
Fixed handling of scalar-like values with new NumPy. (#6331)
Fixed memory leak when accessing Python attributes of Tensor/TensorList (data_ptr, stream, __(cuda_)array_interface__["data"]) (#6322)
Fixed per-sample tensor_resize argument handling (#6318)
Ensured consistent reporting of empty layouts in ndd (#6304)
Fixed handling of NVML_ERROR_NOT_SUPPORTED on partial-NVML platforms (#6296)

Improvements

Update VERSION to 2.2.0
Port libsound fixes (#6384)
Accept constant local variables as arguments with compile=True (#6382)
[Torchvision API] Remove Torchvision's dependency - InterpolationMode (#6383)
Improved NDD skills (#6376)
Torchvsion API documentation page update (#6313)
Document that the dali-dynamic-mode skill exists (#6358)
[Torchvision API] resized_crop and RandomResizedCrop (#6369)
Add more Python-related leak suppressions. (#6370)
Update video decoder & reader supported format list (#6377)
[Torchvision API] Input metadata (#6364)
Make skills/ the source of truth for skills (#6375)
Move to CUDA 13.3 (#6371)
Add checkpointing examples and skill. (#6368)
Add NVSkills CI workflow (#6374)
[Torchvision API] crop (#6353)
Reenable 1ch torchvision grayscale test (#6372)
Include reader in pipelines in transparent pipelining (#6357)
Hide WorkerThread implementation details (#6362)
Restore default hw_decoder_load to 0.65 in imgcodec decoder (#6366)
Document dynamic operator wrapper arguments (#6363)
Deps update/2026 05 14 (#6351)
Improve DLPack support for external tensor consumption (#6261)
Bump DALI_DEPS_VERSION (#6356)
Drop opencv_imgcodecs from libdali (#6349)
Convert numpy types to DALIDataType (#6350)
Update DALI dependencies to include security hardening in the build script. (#6346)
Update QA package versions for Python 3.14 (#6330)
Improve call site resolution (#6323)
Rename DALI dynamic mode skill and move to .agents directory (#6328)
Use upstream FFmpeg in Conda builds (#6345)
Torchvision API RandomApply implementation (#6342)
imgcodec: work around nvImageCodec ROI/orientation contract bug (#6344)
test: disable HW JPEG decoder in free-threading multithreading test (#6343)
Dynamic Mode checkpointing (#6340)
Source nvImageCodec, OpenCV, and libjpeg-turbo from conda-forge (#6315)
Exclude CodeQL cyclic import diagnostics (#6341)
Torchvision API RandomGrayscale operator (#6335)
Keep nvimgcodec wheel path in LD_LIBRARY_PATH for CPU-only test (#6338)
Surface dlerror() per attempted path in nvimgcodec wrap (#6337)
Coverity static analysis fixes - May 2026 (#6333)
Expose Philox4x32_10 to Python and use it in dynamic mode RNG. Fix cloning semantics. (#6334)
Add starter .greptile/ config for DALI reviews (#6326)
Move Philox RNG to dali_core. (#6332)
Cumulative performance fixes for Dynamic Mode (#6329)
Make Tegra DALI wheel declare nvimgcodec dependency (#6321)
Add transparent pipelining in dynamic mode (#6301)
Add static type inference to arithmetic ops. (#6317)
Coverity static analysis fixes (2026-04-21) (#6305)
Make nvImageCodec the default decoder and remove legacy (#6306)
Add dynamic mode control features docs and example (#6312)
OpSchema & factory refactoring part 1 (#6311)
Move to nose2 only (#6146)
Update nvCOMP to 5.2.0 (#6292)
Upgrade dlpack protocol. (#6307)
Adapt DALI to nvImageCodec 0.8.0 (#6293)
Update torch and torchvision package versions for CUDA 13.x (#6297)
ci: exclude py/unsafe-cyclic-import from CodeQL reporting (#6302)
Add shuffle_after_epoch to WebDataset, TFRecord, and MXNet readers (#6288)
Rework call site identification (#6298)
Move to CUDA 13.2 update 1 (#6299)
Add missing operators: bit shifts, bit not and modulo (#6294)
Fix formatting issue in keyword argument pluralization (#6300)
OpSchema/OpSpec metadata inference (#6280)
Add skill and evals for dynamic mode usage (#6271)

Bug Fixes

Bind FileReader samples by reference (#6379)
Missing _invoke in RandomApply, RandomGrayscale and Pad (#6385)
Fix escaped string parsing in numpy headers (#6380)
[Torchvision API] Exception for isolated objective operator execution (#6373)
Validate COCO annotation bbox size (#6360)
Require batch size for dynamic batch permutation (#6359)
Fix ndd mixing device ids (#6347)
Fix test_random_state_arg and add it to CI (#6339)
Adjust handling of scalar-like values with new NumPy. (#6331)
Fix and improve DeviceGuard. (#6327)
Fix leaked PyLong refs in pointer bindings (#6322)
Fix per-sample tensor_resize argument handling (#6318)
Fix a race condition when querying op metadata (#6309)
Fix handling of empty layouts in ndd.Invocation (#6304)
Disable default metadata policy for Python-based operators. (#6308)
Make NDD RN50 test runnable more than once. (#6303)
Handle NVML_ERROR_NOT_SUPPORTED gracefully on partial-NVML platforms (#6296)

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 -...

Contributors

fallintoplace

Assets 2

DALI v2.1.1

stiepan released this 15 Jun 18:55

v2.1.1

d575bee

Key Features and Enhancements

There are no new features in this release

Fixed Issues

Updated CPU-based media support (#6352)

Bug Fixes

Bump up DALI version to 2.1.1
Fix DALI build with Clang 21 (#6348)
Update CPU-based media support (#6352)
Fix Outdated GitHub Links and Hardcoded Artifacts (#6365)
Fix generated docs see also links (#6355)
Fix docs version selector paths for DALI 2.x (#6354)

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==2.1.1
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda130==2.1.1

or just:

pip install nvidia-dali-cuda130==2.1.1
pip install nvidia-dali-tf-plugin-cuda130==2.1.1

For CUDA 12:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda120==2.1.1
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda120==2.1.1

or just:

pip install nvidia-dali-cuda120==2.1.1
pip install nvidia-dali-tf-plugin-cuda120==2.1.1

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

Assets 2

DALI v2.1.0

stiepan released this 28 Apr 18:16

v2.1.0

fd3f55f

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):

Assets 2

DALI v2.0.0

stiepan released this 03 Mar 16:32

v2.0.0

a807a5a

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 o...

Contributors

tonyreina

Assets 2

DALI v1.53.0

stiepan released this 07 Jan 15:31

v1.53.0

55113c6

Key Features and Enhancements

This DALI release includes the following key features and enhancements:

Improved Dynamic Mode (ndd)
- Added ndd.imread function for reading and decoding images (#6092)
- Support dali random operators in dynamic mode (#6110, #6101, #6100, #6096, #6093, #6088)
- Added automatic argument type conversion (#6095)
- Improved documentation, annotations, and stubs (#6085, #6089, #6080, #6078)
Added fn.bbox_rotate operator (#5979)
- Thank you @5had3z for your contribution!
Added support for nvImageCodec 0.7.0 (#6105)
Improved streams re-usage to decrease spurious depenedencies between kernels (#6072)
Migrated to C++20 (#5962)

Fixed Issues

Fixed broadcasting in constants and fn.full. (#6104)
Fixed large batch handling in GPU random ops. (#6082)
Fixed device tracking in dynamic mode. (#6090)
Fixed synchronization in GPU Tensor and TensorList copy. (#6071)

Improvements

Support random ops in dynamic mode stub files (#6110)
Update VERSION to 1.53.0
Update nvcomp to libnvcomp 5.1.0.21 (#6111)
Update numba-cuda to 0.22.0, fix TL0_python-self-test-operators_1 test (#6106)
Update nvImageCodec to 0.7.0 (#6105)
Add Python RNG API for DALI dynamic mode random operators (#6101)
Fix in TL1_cutom_src_pattern_build test (#6103)
Fix numba-cuda installation for Python 3.9 (#6097)
Random operator rework (#6100)
Dependency update 2025.11 (#6094)
Add deprecation warning for Python 3.9 (#6098)
Create stub files for the dynamic mode (#6089)
Add ndd.imread function for reading and decoding images (#6092)
Remove the name "dynamic executor" from the docs and error messages (#6084)
Poisson distribution using cuRAND and STL (#6096)
Generate Dynamic Mode documentation (#6085)
Add automatic argument type conversion to Dynamic API. (#6095)
Random distributions for host and device. (#6093)
Fix clang CUDA runtime wrapper patching for multiple versions (#6091)
Add hidden _random_state argument for Dynamic Mode random operators (#6087)
Add Philox32x4_10 generator for CPU. (#6088)
Fill gaps in dynamic mode documentation - Tensor, Batch, EvalContext, Device and Readers. (#6080)
Remove texture-based video processing in NvDecoder (#6076)
Coveriity check 2025.11. (#6081)
Dynamic mode operator docstrings (#6078)
Add --no-build-isolation flag to DALI plugins (#6077)
Document executor flags StreamPolicy and OperatorConcurrency (#5988)
Add a busy list to CUDAStreamPool. Don't return busy streams from the pool. (#6072)
Make semaphore implementation selectable. Default is POSIX. (#6074)
fn.bbox_rotate (#5979)
Switch to C++20 (#5962)
Fix warnings reported by flake8 (#6059)

Bug Fixes

Fix TL3_EfficientDet_convergence & TL3_YOLO_convergence tests (#6118)
Add proper broadcasting logic to constant_value operator family. (#6104)
Fix Python 3.9 compatibility in imread type annotations (#6099)
Make _signatures.py compatible with Python 3.9 (#6102)
Add EvalContext tracking to Invocation to prevent silent device switching (#6090)
Fix large batch handling in GPU random ops. (#6082)
Always use proper CUDA stream for GPU Tensor(List) copy. Don't use stream 0. (#6071)
Fix usage of std::optional in bbox rotate. Fix conversion to numpy in bbox_rotate tests. (#6073)

Breaking API changes

There are no breaking changes in this DALI release.

Deprecated features

DALI 1.53 is the last release to support Python 3.9.

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==1.53.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda130==1.53.0

or just:

pip install nvidia-dali-cuda130==1.53.0
pip install nvidia-dali-tf-plugin-cuda130==1.53.0

For CUDA 12:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda120==1.53.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda120==1.53.0

or just:

pip install nvidia-dali-cuda120==1.53.0
pip install nvidia-dali-tf-plugin-cuda120==1.53.0

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

Contributors

5had3z

Assets 2

DALI v1.52.0

stiepan released this 27 Oct 14:10

v1.52.0

85806ed

Key Features and Enhancements

This DALI release includes the following key features and enhancements:

Introduced experimental Dynamic Mode: imperative execution model with lazy evaluation for easier integration into Python workflows. (#6066, #6064, #6060, #6056, #6042, #6039, #6037, #6036, #5954)
- Dynamic mode: add augmentation gallery (#6057)
- DALI Dynamic docs main page (#6052)
Added pipeline ZOO - snippets and examples for common image and video processing use cases. (#5922)
Added support for CUDA 13U2 (#6063)
Added fn.decoders.numpy (#5953) and CPU fn.paste operators (#5968).
Thank you @5had3z for your contributions.
Exposed knobs for pipeline dynamic executor:
- Exposed executor's stream_policy and concurrency options (#5983)
- Environment variable to control executor threads. (#5949)

Fixed Issues

Fixed stream ordering in Tensor::Copy and Tensor(List)GPU.as_cpu (#6070)
Fixed conversion of pinned tensors to DLPack. (#6061)
Fixed DLPack stride check if stride pointer is NULL
Fixed handling of videos without keyframes and reuse of old indices (#6058)
Fixed resize_crop_mirror video output shape (#5957)

Improvements

Update to FFmpeg 8.0
Dynamic mode: add augmentation gallery (#6057)
Add dynamic API for math functions + tests. (#6066)
Rename DALI2 to dynamic (#6064)
Move to CUDA 13.0 U2 (#6063)
Dynamic mode: operator base classes and operator call generator (#6060)
Update VERSION to 1.52.0
Update deps 25/10 (#6053)
Dynamic Mode: Tensor and Batch Types (#6056)
Remove CMake from acknowledgements. (#6020)
DALI Dynamic docs main page (#6052)
Reduce minimum throughput for experimental decoder in TL1_decoder_perf (#6050)
Fix TL0_video_plugin to run with sanitizer (#6040)
Imperative mode: Invocation (#6042)
Update LD_PRELOAD in sanitizer configuration, exclude more numba tests (#6041)
Imperative mode: EvalContext, EvalMode, Type and Device (#6039)
Update the test environment to Ubuntu 24.04 (#6033)
Update curl 3.15 -> 3.16 (#6038)
Add TensorList broadcasting constructor. (#6037)
Backend changes for imperative mode (#6036)
Add nvcc/nvjitlink version compatibility check to numba CUDA test (#6035)
Unify minimum required CMake version. (#6022)
Fix installation of Horovod in TL1_tensorflow-dali_test (#6024)
Remove confusing warning on host decoder fallback (#6029)
Add stream argument to TensorGPU DLPack constructor. (#6015)
Cumulative dependency update for September 2025. (#6017)
Silence false warnings in sanitized build (#6018)
Lower the 5% threshold in image decoder perf test to 15% to account for off iterations (#6021)
Bump CMake to 3.25.2 (#6019)
Move to CUDA 13.0 U1 (#6016)
Move to the gcc-toolset-14 (#6014)
Update test packages (#6010)
Correct support matrix entry for Orin (#6008)
Silence a false positive warning triggered by GCC 12.2.1 (#6002)
Fix CVE-2024-13978 and CVE-2025-8534 in libtiff (#6007)
Bump up OpenCV version to 4.12 in conda (#6005)
Move to the latest nvJPEG2k (#6000)
Enable more aggressive binary compression (#6001)
Use subprocess.run in get_tf_compiler_version to avoid CalledProcessError on grep (#5991)
Add functions that change the type of the tensor or tensor list to a different type of the same size. (#5995)
Update OpenCV version in tests (#5987)
Improve performance of experimental.resize (#5662)
Expose executor policy flags (#5983)
Pin CMake to max 4.0.3 in jupter_conda tests. (#5985)
Add driver version check to the usage of numba_cuda (#5982)
Fix nvComp installation in tests (#5984)
Update DALI_DEPS_VERSION to use patched libtiff (#5981)
Improve creating image batches in CV-CUDA ops (#5966)
Dependency update 07-2025 (#5978)
Make the numba operator compatible with the numba-cuda package (#5975)
Adjust TF plugin build dependencies (#5976)
fn.paste CPU impl (#5968)
Make sure that protobuf always uses own absl version instead of system one (#5974)
Thread pool with semaphore and spinlock (#5970)
Extend GetInputDevice in OpSchema python bindings. (#5972)
Remove data preparation instructions from the video superres use case (#5965)
Added fn.decoders.numpy (#5953)
Pipeline zoo - initial commit (#5922)
Expose Stream, Operator and Workspace in Python (#5954)
Fix nvcc not working with sanitizer (#5959)
Make the number of dynamic executor threads configurable via environment variables. (#5949)

Bug Fixes

Fix stream ordering in Tensor::Copy and Tensor(List)GPU.as_cpu
Fix conversion of pinned tensors to DLPack. (#6061)
Fix DLPack tests to use HWC layout instead of NHWC (#6062)
Fix handling of videos without keyframes and reuse of old indices (#6058)
Refactor layout handling in Python backend + add layout dimensionality checks in Tensor and TensorList python bindings (#6054)
Fix standalone op output streams. (#6055)
Remove EvalContext destructor. (#6043)
Fix static analysis issues (#6032)
Install newer CMake in TL0_jupyter (#6034)
Disable PYBIND11_FINDPYTHON in CMakeLists.txt (#6031)
Remove a custom patch for PyCuda, add numba_cuda version constrain (#6023)
Bugfix: Skip DLPack stride check if stride pointer is NULL
Improve error handling in ThreadPool (#6011)
Fix test_backend_impl launch command (#6003)
Remove unnecessary default values from optional arguments. (#5992)
Add missing backslash in test scripts. (#5986)
Fixes outdated DALI mannylinux tag (#5980)
resize_crop_mirror - invalid video output shape fix (#5957)

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==1.52.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda130==1.52.0

or just:

pip install nvidia-dali-cuda130==1.52.0
pip install nvidia-dali-tf-plugin-cuda130==1.52.0

For CUDA 12:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda120==1.52.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda120==1.52.0

or just:

pip install nvidia-dali-cuda120==1.52.0
pip install nvidia-dali-tf-plugin-cuda120==1.52.0

Or use direct download links (CUDA 13.0):

Contributors

5had3z

Assets 2

DALI v1.51.2

stiepan released this 13 Aug 10:26

v1.51.0

a103d4a

Key Features and Enhancements

This DALI release includes the following key features and enhancements:

Added support for CUDA 13 and CUDA 12.9U1. (#5946)
Added support for nvImageCodec 0.6.0.
Improved CPU multithreading efficiency. (#5960, #5963, #5961)
- Reduced lock contention on ARM CPUs.
- Reduced number of mutex locks in ThreadPool.
- Optimized spinlock hot path.
Made the new (dynamic) executor a default. (#5936, #5944)
Improved memory management in nvImageCodec based decoders (#5948, #5945)

Improvements

Optimize spinlock hot path. (#5961)
Improve ThreadPool efficiency (#5963)
Reduce the number of mutex locks in ThreadPool. (#5960)
Fix model weight path for TL1_superres_pytorch test (#5955)
Update VERSION to 1.51.0
Dependencies update 06.2025 (#5951)
Fix problem of installing numpy 2 in some tests (#5952)
Move to CUDA 12.9U1 (#5946)
New executor performance fixes. (#5944)
Update TensorFlow and Numba versions in tests (#5942)
Use std::move instead of copy where applicable (#5940)
BLD: Silence warning from setuptools about packages config (#5939)
Make dynamic executor the default choice. (#5936)
Update DALI_DEPS_VERSION (#5934)
Add a guard for out-of-range memory write in sw_scale (#5931)
Removes duplicated document version selector (#5933)
Update submodule dependencies (#5927)
Enable tfrecord2idx script to convert a tfrecord from Object Storage into the index file, which is also stored in Object Storage (#5918)
Disable conda tests when sanitizers are enabled (#5923)
Enable conda build with AWS SDK (#5917)
Add POST_BUILD to custom commands and include stdexcept in wrap files (#5903)
Enable nvJPEG2k in conda build (#5920)

Bug Fixes

Don't make the image decoder output forcibly non-contiguous. (#5948)
nvImageCodec decoder - allocate whole batch (#5945)
Set prefetch_queue_depth=1 parameter in test_crop_window_warning test pipeline (#5938)
BLD: Set CMake Policy 175 to NEW (#5937)
Suppress Warning when No Boxes in Sample (#5932)
Fix ResNet50 test for DALI proxy (#5925)

Breaking API changes

DALI 1.50 was the last release to support CUDA 11.
Support for architectures of compute capability lower than 75 was dropped in CUDA 13 builds.

Deprecated features

No features were deprecated in this release.

Known issues:

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.

CUDA 12.0 and CUDA 13.0 builds use CUDA toolkit enhanced compatibility. 
They are built with the latest CUDA 12.x/13.x toolkit respectively but they can run on any, 
stable drivers from the respective CUDA family (525 and 580 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 13.0:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda130==1.51.2
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda130==1.51.2

or just:

pip install nvidia-dali-cuda130==1.51.2
pip install nvidia-dali-tf-plugin-cuda130==1.51.2

For CUDA 12:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda120==1.51.2
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda120==1.51.2

or just:

pip install nvidia-dali-cuda120==1.51.2
pip install nvidia-dali-tf-plugin-cuda120==1.51.2

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

Assets 2

DALI v1.50.0

stiepan released this 27 May 17:43

v1.50.0

d5c7b54

Key Features and Enhancements

This DALI release includes the following key features and enhancements:

Added support for CUDA 12.9 (#5908)
Added option to disable SSL verification for S3 bucket (#5907)
Thank you @dimabasow for your contribution.
Added support for loading nvComp from a Python wheel (#5894, #5889, #5909)
Improved error messages in video loader with file name in the message (#5910)

Fixed Issues

Fixed handling of multiple frames per packet in video decoder (#5911)
Fixed sparse tensor handling in TF plugin (#5916, #5887)
Fixed serialization of default seeds in operators (#5919)
Fixed handling of empty inputs in GPU reductions (#5914)
Fixed handling of stdin descriptor in CUFileDriverScope (#5902)

Improvements

Make Python 3.10 a default version for the build.sh (#5913)
Make library bundling errors easier to find in the log. (#5915)
Move to CUDA 12.9 (#5908)
Improve error messages in video loader with file name in the message (#5910)
Add an ability to disable SSL verification for S3 bucket (#5907)
Migrate DALI TF plugin to C API 2.0 (#5904)
BLD: Use CMake nvimgcodec module if available to get headers (#5906)
C API changes required for TF plugin. (#5898)
Remove redundant imports from the augmentation_gallery (#5900)
Move to externally provided nvComp (#5894)
Remove Python 3.8 support due to EOL (#5896)
Extend EfficientNet readme (#5895)
Fix memory consumption by PyTorch in dlpack zero-copy perf test. (#5891)
Add handling for NVMLError_NotSupported in get_device_memory_info (#5890)
Enable nvComp for SBSA platform (#5889)
experimental video reader to drop frames with negative display timestamps (#5885)

Bug Fixes

Coverity check 04.2025 (#5912)
frames decoder fixes: avoid overflow, handle multiple frames per packet (#5911)
Fix sparse tensor construction in TF plugin. (#5916)
Do not serialize default seed (#5919)
Fix gpu empty reductions (#5914)
Make sure that nvComp is bundled also when WITH_DYNAMIC_CUDA_TOOLKIT is off (#5909)
Improve conda build recipe (#5905)
Fix stdin handling in CUFileDriverScope (#5902)
Remove squeezing from C API - it probably never worked anyway. (#5893)
Fix invalid stack read in legacy C API (#5892)
Use Polygon(..., closed=true) instead of Polygon(..., true) (#5842)
Fix handling of scalars in TF sparse tensors. (#5887)

Breaking API changes

DALI 1.49 was the last release to support Python 3.8

Deprecated features

Support for CUDA 11 will end in the upcoming releases.

Known issues:

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==1.50.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda120==1.50.0

or just:

pip install nvidia-dali-cuda120==1.50.0
pip install nvidia-dali-tf-plugin-cuda120==1.50.0

For CUDA 11:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda110==1.50.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda110==1.50.0

or just:

pip install nvidia-dali-cuda110==1.50.0
pip install nvidia-dali-tf-plugin-cuda110==1.50.0

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

Contributors

dimabasow

Assets 2

DALI v1.49.0

stiepan released this 29 Apr 14:58

v1.49.0

ccb1902

Key Features and Enhancements

This DALI release includes the following key features and enhancements:

Improved new (experimental) C API (#5879, #5872, #5866, #5857, #5835, #5868)
Added support for CUDA 12.8U1 (#5850)
Added CPU support to dali.fn.experimental.debayer (#5832)
Thank you @5had3z for your contribution!
Added support for nvImageCodec 0.5.0 (#5854)

Fixed Issues

Fixed race-condition in experimental image decoder (#5856)

Improvements

Update VERSION to 1.49.0
C API 2.0 Checkpointing + unblock dali.h (#5879)
Temporarily disable failing test (#5882)
Experimental Video Reader Refactoring and API Improvements (#5839)
Move to LLVM 20.1.2 (#5870)
C API 2.0: External source info (#5872)
Add _zmq.cpython to the address sanitizer suppression list (#5873)
Set minimum CMake policy version for Horovod build (#5871)
Pipeline refactoring (#5866)
Add multi-configuration performance benchmarking (#5858)
Sort out Python 3.8 support (#5867)
Moves to manylinux_2_28 (#5608)
Adjust test compatibility with numpy 2.x (#5862)
Bump up the minimum version of CMake required by ffts (#5864)
Remove unnecessary global declarations and add noqa comments (#5865)
Add fallback for missing source info in check_batch (#5861)
Bump nvimagecodec requirement to 0.5.0 (#5854)
Skip C API2 test using Mixed ImageDecoder when it's not registered. (#5857)
C API 2.0 Pipeline & Pipeline Outputs (#5835)
Update six package version constraint (#5855)
Add info about GIT sha to the documentation (#5853)
Bump up the Black version to 25.x (#5849)
Bump OpenCV version in conda to 4.11 (#5851)
Improve sanitizer configuration and suppress false positives (#5795)
Add Debayer CPU based on OpenCV (#5832)
FramesDecoder boundary handling, video utils (#5844)
Move to CUDA 12.8 U1 (#5850)
Added warp perpective tests to other test suites. (#5847)
Add operator statefulness info to OpSchema (#5848)
Bump up support TF version to 2.18 (#5840)

Bug Fixes

Fixed race-condition in experimental image decoder (#5856)

Breaking API changes

There are no breaking changes in this DALI release.

Deprecated features

DALI 1.49 is the last release to support Python 3.8

Known issues:

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==1.49.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda120==1.49.0

or just:

pip install nvidia-dali-cuda120==1.49.0
pip install nvidia-dali-tf-plugin-cuda120==1.49.0

For CUDA 11:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda110==1.49.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda110==1.49.0

or just:

pip install nvidia-dali-cuda110==1.49.0
pip install nvidia-dali-tf-plugin-cuda110==1.49.0

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

Contributors

5had3z

Assets 2

DALI v1.48.0

stiepan released this 25 Mar 17:57

v1.48.0

211abfa

Key Features and Enhancements

This DALI release includes the following key features and enhancements:

Improved fn.experimental.decoders.video video decoder: (#5830, #5814)
- Improved seeking and reset behavior
- Added support for frame padding with configurable modes
- Added frame selection options
- Added build_index option to control the generation of a frame index
Added CPU support to dali.fn.experimental.warp_perspective (#5829, #5815)
- Thank you @5had3z for your contribution!
Introduced new (experimental) C API (#5796, #5797, #5798, #5799)

Fixed Issues

Introduced AvUniquePtr to avoid memory leaks in frames decoder (#5834)
Removed an unnecessary host sync in operators taking pinned inputs. (#5822)
Fixed host-side access to pinned CPU buffers produced with non-host order (#5820)
Fixed handling of empty batches in GPU arithmetic operators. (#5818)

Improvements

Fix data paths in TL3 short tests (#5845)
Revert change of batch size in SSD LT3 to 64 due to convergence problem (#5846)
Update VERSION to 1.48.0
Fix coverity issues 25/03 (#5843)
Bump up FFmpeg to 7.1.1 (#5838)
Reorganize video decoder sources (#5836)
Dependency update 2025-03 (#5833)
C API 2.0 Tensor and TensorList (#5799)
Update documentation of audio decoder operator (supported formats) (#5803)
Removes RN50 benchmark tests, move to DALI_EXTRA for RN50 DL FW iter tests (#5824)
Improve video decoder seeking and reset behavior (#5830)
Warp Perspective CPU Impl (#5829)
Remove ScratchpadAllocator and ScratchpadEstimator (#5810)
Code modernization and refactoring in Pipeline, OpSpec and InputOperator (#5826)
fn.experimental.decoders.video improvements (#5814)
C API 2.0 helpers (#5798)
C API 2.0 initialization and error handling (#5797)
Limit the max. tensor list size in TensorTest (#5823)
Relax DisplacementTest.Sphere constraints from 0.005 to 0.006 (#5821)
Restrict dm-tree version for Python 3.8 and 3.9 (#5819)
Add C API header and C language build test. (#5796)
Expose DLPack support in the docs (#5817)

Bug Fixes

Fix usage of unique_ptr for arrays in data_objects_test (#5837)
Introduce AvUniquePtr to avoid memory leaks in frames decoder (#5834)
tensor_shape.h warning fix (#5831)
Enhanced Video Codec Support and Error Handling (#5825)
Fix documentation for warp_perspective, requires 3x3 shape, not flattened 1D (#5815)
Remove an unwanted potential host sync in operators taking pinned inputs. (#5822)
Fix host-side access to pinned CPU buffers produced with non-host order (#5820)
Fix handling of empty batches in GPU arithmetic operators. (#5818)

Breaking API changes

There are no breaking changes in this DALI release.

Deprecated features

No features were deprecated in this release.

Known issues:

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==1.48.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda120==1.48.0

or just:

pip install nvidia-dali-cuda120==1.48.0
pip install nvidia-dali-tf-plugin-cuda120==1.48.0

For CUDA 11:
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-cuda110==1.48.0
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/ nvidia-dali-tf-plugin-cuda110==1.48.0

or just:

pip install nvidia-dali-cuda110==1.48.0
pip install nvidia-dali-tf-plugin-cuda110==1.48.0

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

Contributors

5had3z

Assets 2

Uh oh!

Releases: NVIDIA/DALI

Release list

DALI v2.2.0

Key Features and Enhancements

Fixed Issues

Improvements

Bug Fixes

Breaking API changes

Deprecated features

Known issues:

Binary builds

Contributors

Uh oh!

DALI v2.1.1

Key Features and Enhancements

Fixed Issues

Bug Fixes

Breaking API changes

Deprecated features

Known issues:

Binary builds

Uh oh!

DALI v2.1.0

Key Features and Enhancements

Fixed Issues

Improvements

Bug Fixes

Breaking API changes

Deprecated features

Known issues:

Binary builds

Uh oh!

DALI v2.0.0

Key Features and Enhancements

Fixed Issues

Improvements

Bug Fixes

Breaking API changes

Deprecated features

Known issues:

Binary builds

Contributors

Uh oh!

DALI v1.53.0

Key Features and Enhancements

Fixed Issues

Improvements

Bug Fixes

Breaking API changes

Deprecated features

Known issues:

Binary builds

Contributors

Uh oh!

DALI v1.52.0

Key Features and Enhancements

Fixed Issues

Improvements

Bug Fixes

Breaking API changes

Deprecated features

Known issues:

Binary builds

Contributors

Uh oh!

DALI v1.51.2

Key Features and Enhancements

Improvements

Bug Fixes

Breaking API changes

Deprecated features

Known issues:

Binary builds

Uh oh!

DALI v1.50.0

Key Features and Enhancements

Fixed Issues

Improvements

Bug Fixes