ONNX v1.18.0 is now available with exciting new features! We would like to thank everyone who contributed to this release!
Please visit onnx.ai to learn more about ONNX and associated projects.

Key Updates

ai.onnx Opset 23

Attention, Cast, CastLike, Constant, ConstantOfShape, DequantizeLinear, Flatten, Identity, If, Loop, Pad, QuantizeLinear, RMSNormalization, Reshape, RotaryEmbedding, Scan, Shape, Size, Squeeze, Transpose, Unsqueeze

IR Version 11

Added FLOAT4E2M1 and multi-device configuration support
Relaxed naming requirements (#6652)

Python support

Support Python 3.13
Experimental support for Python 3.13t (Windows, Mac)
Removed support for Python 3.8
Windows Arm64

Build

Minimum protobuf version is upgraded to v25.1
A new option ONNX_BUILD_CUSTOM_PROTOBUF is added for CMAKE (#6495)

What's Changed

Breaking Changes and Deprecations

Remove/raise exception when external file exists during onnx.save by @tonypottera24 in #6497
Remove python 3.8 workflows (python 3.8 is eol) by @cyyever in #6434
Deprecate all type casting functions by @justinchuby in #6639
GroupNormalization-18 is now deprecated and replaced by GroupNormalization-23 due to an incorrect definition (#6358)

Spec and Operator

Add FLOAT4E2M1 data type by @yuanyao-nv in #6318
Add FLOAT4E2M1 support to relevant operators by @yuanyao-nv in #6283
Fix typo in the GroupNorm description by @yuanyao-nv in #6358
clarify clip with min>max by @AlexandreEichenberger in #6395
Update IR spec to clarify optional input/output availability by @justinchuby in #6435
Fix GlobalLpPool input types by @justinchuby in #6503
Fix reference implementation for TopK by @neNasko1 in #6593
Clarify that FLOAT4E2M1 can be in int32_data by @justinchuby in #6640
Relax naming requirements in IR spec by @justinchuby in #6652
Add Rotary Embedding op to ONNX opset 23 by @shubhambhokare1 in #6461
Format rotary embedding documentation by @justinchuby in #6717
Add RMSNormalization to ONNX opset 23 by @shubhambhokare1 in #6443
Add Attention Op to ONNX Opset 23 by @shubhambhokare1 in #6501
Add multi-device execution support in ONNX by @kevinch-nv in #6641

Reference Implementation

Fix NonMaxSuppression default values in ReferenceEvaluator by @xadupre in #6354
Update op_pool_common.py, add missing "| None" by @andife in #6421
Allow 1D vector for w_scale in QLinearConv by @mcollinswisc in #6460
Fix pooling pads issues by @titaiwangms in #6650

Utilities and Tools

Fix shape inference for Squeeze-1,11 with dynamic input shape by @Yosshi999 in #6314
Extend printer and parser to support invalid identifiers by @gramalingam in #6346
Support node labels in parser and printer by @gramalingam in #6349
Fix parser to handle empty optional parameters in edge cases by @gramalingam in #6427
Optimize DFS by @tonypottera24 in #6440
Separate the infer_shapes option from the check_model option by @tonypottera24 in #6441
Support models larger than 2GB by @tonypottera24 in #6438
Add missing CHECK_PARSER_STATUS by @cyyever in #6448
Fix MAXIMUM_PROTOBUF size (2GiB, not 2GB) by @xenova in #6556
Fix numpy_helper to_array by @justinchuby in #6638
Add version converter softmax 13 -> 12 by @seungwoo-ji-03 in #6608

Build, CI and Tests

Update Protobuf to latest version (Win) by @liqunfu in #6362
Use CMake python module by @cyyever in #6381
Revert Python finding logic and some CMake fixes for Windows by @cyyever in #6382
Enable more compiler warnings by @cyyever in #6383
Remove pypi publishing inside "release_" by @andife in #6483
Fix g++-13 build errors by @cyyever in #6509
Move to using protogen's .pyi file generator by @justinchuby in #6462
Python 3.13: builds successful for all os by @andife in #6370
Simplify CMake argument handling in setup.py by @cyyever in #6545
Remove struct and leverage ml_dtypes in helper tests by @justinchuby in #6631
Add cmake dependency only when system cmake is not available (backend version) by @mgorny in #6643
Remove onnxruntime tests by @justinchuby in #6709
python313t builds in main_freethreading.yml by @andife in #6706
Improve CMake summary by @cyyever in #6704
Add protobuf local build option by @cyyever in #6495
Harmonize protobuf versions by upgrading to a minimum protobuf version of 25.1 and fix CI error by @cyyever in #6725
Include backend.py in source distribution by @mgorny in #6755

Documentation

Move INSTALL instruction to separate file by @andife in #6560
Fix Pad operator example in Python API docs (correct attribute usage) by @kolasaniv1996 in #6702

Other Changes

Remove unused variables by @cyyever in #6303
Fix main url checks by @roborags in #6312
Bumped main VERSION_NUMBER to 1.18.0 by @roborags in #6315
Bump ai.onnx opset to 23 by @roborags in #6316
Combine different release pipelines by the use of reusable workflows by @andife in #6277
Bump actions/upload-artifact from 3 to 4 by @andife in #6319
Fix missing secrets for publishing of onnxweekly by @andife in #6321
Update main.yml (upgrade github actions download/upload artifact) by @andife in #6320
BF: fix condition for publishing to testpypi (Update create_release.yml) by @andife in #6338
Fix OOB in data propagation of math ops when input is broadcasted to zero by @Yosshi999 in #6323
The latest protobuf pkg 5.28.0 is failing on Windows. use the one pre… by @liqunfu in #6342
Set up codecov test analysis by @justinchuby in #6345
Fix model extraction utility by @gramalingam in #6344
Add missing matrix.target-architecture ? (Update release_mac.yml) by @andife in #6350
Some performance fixes in C++ code, use of emplace instead of insert by @cyyever in #6304
Fix some clang-tidy warnings by @cyyever in #6353
BugFix: make sure that unique files names are used before uploading the wheels by @andife in #6355
Simplify create_release.yml by @andife in #6357
Fix more static analyser warnings by @cyyever in #6360
Fix more static analyzer warnings by @cyyever in #6363
Not all release runs have to be successul for publishing weekly to testpypi by @andife in #6372
CI: Restoring the possibility to initiate a run via labels by @andife in #6371
Fix wrong publishing target (onnx instead of onnx-weekly) by @andife in #6375
Trigger release pipeline when label "run release..." was added to PR by @andife in #6376
Fix clang-tidy warnings by @cyyever in #6378
Avoid returning const vectors by @cyyever in #6391
Fix more clang-tidy warnings by @cyyever in #6392
Use std::unordered_{map,set} by @cyyever in #6390
Set hidden visibility on onnx target by @cyyever in #6393
Fix more static analyser warnings by @cyyever in #6396
Add clang-tidy checks and fix warnings by @cyyever in #6398
Use -Wall -Wextra on APPLE platforms by @cyyever in #6399
Support official Python launcher on Windows by @petewarden in #6407
Add misc checks and turn some functions and variables into static by @cyyever in #6401
Enable more clang-tidy checks by @cyyever in #6429
Update DFS function to non-recursive version by @tonypottera24 in #6424
Refactor parser code by @cyyever in #6397
Create a workflow to suggest lintrunner fixes by @justinchuby in #6377
Fix ODR violations by @cyyever in #6439
Fix invalid use of const_cast by @cyyever in #6451
Use enum class for StatusCategory and StatusCode by @cyyever in #6450
Fix static analyzer warnings by @cyyever in #6449
Fix excpetion handling by @cyyever in #6453
Fix python linter warnings in setup.py by @cyyever in #6459
Move suggest fixes to a separate CI workflow by @justinchuby in #6455
Optimize _collect_new_io_core in extract_model by @tonypottera24 in #6466
Enhance CenterCropPad documentation by @jantonguirao in #6464
Simplify protobuf linking by @cyyever in #6444
Print correct Python version in CMake summary by @cyyever in #6467
Fix ODR violations by @cyyever in #6468
Migrate to ubuntu-24.04 runner by @andife in #6373
Add type annotations and fix some lint warnings in Python code by @cyyever in #6470
Optimize dfs in extract_model by dict by @tonypottera24 in #6472
More robust utf-8 string conversions on Windows by @cyyever in #6454
Optimize collect_referred_local_functions in extract_model by @tonypottera24 in #6478
Move output map to a function in extract_model by @tonypottera24 in #6485
Add type annotations and fix some lint warnings in Python code by @cyyever in #6473
Update ruff to 0.7.0 by @cyyever in #6482
Add more type annotations and fix warnings in Python code by @cyyever in #6474
Remove __STDC_FORMAT_MACROS by @cyyever in #6487
Fix clang -Weverything warnings by @cyyever in #6493
change onnx-weekly destination from test.pypi to pypi by @andife in #6491
Bump pybind11 to 2.13.6 by @cyyever in #6486
Bump protobuf to 5.28.3 by @cyyever in #6330
Fix protobuf version unmatch by @cyyever in #6498
Upload the srcdist to pypi as a separate step by @andife in #6488
Another attempt to fix protobuf version problem by @cyyever in #6508
Bugfix fix file name in create_release.yml by @andife in #6510
Fix pipeline, limit to onnxruntime <1.20 for docs build by @andife in #6524
Add LANGUAGE to CMake project by @cyyever in #6527
Fix reference implementation for ConcatFromSequence by @xadupre in #6369
Add publishing of devbuild to testpypi by @andife in #6535
Update page deployment CI to test building pages in PRs by @justinchuby in #6530
Use localy build protobuf in doc pages by @cyyever in #6539
Migration to FindPython3.cmake by @ilya-lavrenov in #6526
Fix upload endpoint (pypi) by @andife in #6541
Update calender link by @andife in #6528
Remove usage of Mac12-Runner by @andife in #6542
Add python 3.13 for release_* workflows by @andife in #6543
Fix getsize bug in extract_model by @tonypottera24 in #6496
Improving the environment name by @andife in #6544
Fix extra-semi warnings on public headers by @cyyever in #6538
Fix of the devbuild publish-workflow by @andife in #6546
Removing unit test involving onnxruntime by @andife in #6549
Publish to test.pypi only for testing (with workflow dispatch) by @andife in #6548
Use rank in data propagation by @gramalingam in #6557
Fix handling of tensor rank in concat data propagation by @gramalingam in #6570
Update gather doc by @gramalingam in #6585
Fix: skip empty input output names in Extractor by @wine99 in #6587
Upgrade to Python 3.9 by @cyyever in #6595
add grammatical corrections to documentation by @rudrOwO in #6599
Fix wrong comparison between numpy dtypes to None by @xadupre in #6588
Fix reference implementation for CumSum by @neNasko1 in #6597
Fix reference implementation for Unique by @neNasko1 in #6596
Bump Protobuf to 5.29.2 by @cyyever in #6600
Use ruff as the code formatter by @justinchuby in #6504
Install missing pytest; Update release_test_weekly_sdist.yml by @andife in #6617
Enable more readability checks by @cyyever in #6618
Allow scale type in QuantizeLinear and DequantizeLinear to be different from the high-precision type by @yuanyao-nv in #6531
Improve attribute checking in CategoryMapper by @neNasko1 in #6619
fix typo: Update README.md by @andife in #6620
Expose cxx standard by @ilya-lavrenov in #6625
Remove vscode settings by @justinchuby in #6628
Improve INSTALL.md by @cyyever in #6629
Enable isort rules in ruff formatter by @justinchuby in #6632
Fix the bfloat16 test in helper for min numpy version by @justinchuby in #6636
Fix link in readme by @andife in #6644
Fix literal operator in Clang17+ by @fs-eire in #6633
Unify names of github environments by @andife in #6651
Improve Names in CI pipeline by @andife in #6657
Improve name in CI-Pipeline by @andife in #6659
Remove unnecessary secrets in pipeline by @andife in #6661
Hardening of checkout by @andife in #6662
Use static dbg_count variables by @cyyever in #6494
Use DeprecationWarning by @justinchuby in #6667
Improve release_linux_*.yml by @andife in #6671
Check if wheels should be published before using a job with a deployment environment by @andife in #6688
Bump absl cpp to 20240722.1 by @cyyever in #6672
remove twine from requirements-dev-file by @andife in #6692
Correct protobuf install instructions on Windows by @natke in #6695
Update INSTALL.md to show build onnx on Windows by @natke in #6696
Check the input of ConstantOfShape to be non-negative by @sgjzfzzf in #6691
TypoFix in float8.md by @mehmetcanfarsak in #6699
Update recommended Protobuf version and other improvements in INSTALL.md by @cyyever in #6701
Use Python3_add_library to simplify build code by @cyyever in #6705
Hardcoding the docker image version, adding the information which docker images are used by @andife in #6663
Fix a tensor type inaccuracy in the docs by @nenad1002 in #6707
Add dtype tests for various operators by @bili2002 in #6687
Improve names, add debug variables by @andife in #6714
Fix condition at download step for publishing manuel preview build by @andife in #6715
Fix mismatched parenthesis in ci by @andife in #6716
Fix condition in CI for sdist download by @andife in #6719
Update Jupyter notebook examples by @emmanuel-ferdman in #6729
check_for_publish_preview_build_to_testpypi_weekly in CI should fail for PullRequests by @andife in #6727
Modify axis description for RMSNormalization op by @shubhambhokare1 in #6732
Set version 4.25.1 as min protobuf python version by @andife in #6722
update auto_update_doc and create sdist_test.yml by @andife in #6723
Update preview_source_dist_test.yml by @andife in #6724
Publish to testpypi only for workflow_dispatch event by @andife in #6750
Update dependabot.yml by @andife in #6753
Add copyright header to pyproject.toml by @andife in #6752
define top-level permission by @andife in #6747
add windows 3.13t build to main_freethreading.yml (Updates of external dependencies now enable this) by @andife in #6749
Simplify checkout submodules by @andife in #6762
Integrate mac freethreading whl to release process by @andife in #6761
Clarify minimum Protobuf version in INSTALL.md by @cyyever in #6760
Improve condition in create_release.yml by @andife in #6765
Remove separate onnx-directory in release_win.yml by @andife in #6766
Update RotaryEmbedding 4D input shape dimension order by @shubhambhokare1 in #6767
chore(config): set persist credentials to false by @andife in #6774
Use CMAKE_COMPILER_IS_GNUCXX for g++ by @cyyever in #6770
fix: names of jobs in create_release.yml by @andife in #6769
Improve Python code typing by @cyyever in #6763
Enable flake8-type-checking by @cyyever in #6775
Merge requirements-release_freethreading into requirements-release by @cyyever in #6776
Bump ruff to 0.9.10 by @cyyever in #6780
Adapt requirements files as 3.9 is the minimal supported python version by @andife in #6777
feat: freethreading python wheels for windows by @andife in #6764
refactor: remove unnecessary x86 check in win freethreading by @andife in #6785
Update sdist_test.yml by @andife in #6786
fix: Use correct requirements file for release_mac_freethreading.yml by @andife in #6768
Update auto_update_doc.yml by @andife in #6792
fix: dont' push wheels with the same filename to pypi twice by @andife in #6790
Remove dependabot configs for dockerfiles by @andife in #6789
fix non existing requirements-file by @andife in #6796
fix wrong use of always in release_linux_x86_64.yml by @andife in #6797
Remove ignore_type suppression by @cyyever in #6788
Update gtest to version > v1.16.0 (commit of 8March25) by @andife in #6783
hard code docker image docker://quay.io/pypa/manylinux2014_x86_64:2025.02.28-1 in release pipeline by @andife in #6799
windows-latest -> windows-2022 by @andife in #6800
[2/N]Remove ignore_type suppression by @cyyever in #6798
Remove duplicates in test by @andife in #6791
Remove wrong usage of "always" in aarch64 release pipeline... by @andife in #6806
fix condition in pipeline by @andife in #6810
Bump minimum numpy to 1.22 by @andife in #6812
Improve condition in rotary embedding to control 3D/4D inputs/outputs by @shubhambhokare1 in #6815
Fix MSVC warnings by @cyyever in #6820
Fix pylint warnings by @cyyever in #6778
Update version to 1.18.0rc1 by @ramkrishna2910 in #6866
1.18.0 cherry pick by @ramkrishna2910 in #6867
Fix missing inputs in create_release.yml by @andife in #6868
update versions by @ramkrishna2910 in #6880
Add wheel build for windows arm64 by @andife in #6912
Export symbols used by PyTorch by @cyyever in #6917
Pick CMake changes by @cyyever in #6922
Update VERSION_NUMBER to 1.18.0rc2 by @andife in #6921
update version to 1.18.0 by @ramkrishna2910 in #6954

Contributors

Andreas Fehlner, Atanas Dimitrov, Bilyana Indzheva, Emmanuel Ferdman, G. Ramalingam, Ilya Lavrenov, Joaquin Anton, Joshua Lochner, Justin Chu, Kevin Chen, Liu Jinjie, Mehmet Can Farsak, Michał Górny, Nat Kershaw (MSFT), Nenad Banfic, Pete Warden, Po-Chu Hsu, Ramakrishnan Sivakumar, Rudro, Shubham Bhokare, Ti-Tai Wang, Xavier Dupré, Yuan Yao, Yulong Wang, Zijun Yu, cyyever, dependabot[bot], kolasaniv1996, mcollinswisc, seungwoo-ji

New Contributors

@Yosshi999 made their first contribution in #6314
@petewarden made their first contribution in #6407
@tonypottera24 made their first contribution in #6424
@mcollinswisc made their first contribution in #6460
@xenova made their first contribution in #6556
@wine99 made their first contribution in #6587
@rudrOwO made their first contribution in #6599
@neNasko1 made their first contribution in #6597
@mgorny made their first contribution in #6643
@titaiwangms made their first contribution in #6650
@sgjzfzzf made their first contribution in #6691
@mehmetcanfarsak made their first contribution in #6699
@kolasaniv1996 made their first contribution in #6702
@nenad1002 made their first contribution in #6707
@bili2002 made their first contribution in #6687
@seungwoo-ji-03 made their first contribution in #6608
@emmanuel-ferdman made their first contribution in #6729

Full Changelog: v1.17.0...v1.18.0

onnx 1.18.0 v1.18.0 on Python PyPI