onnx 1.23.0 on Python PyPI

ONNX v1.23.0 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.

What's Changed

IR Version 14

IR Version is changed from 13->14

Breaking Changes and Deprecations

perf(reference): vectorize Conv implementation by @justinchuby with @Copilot in #8366

Spec and Operator

New Operator SwiGLU is added [SwiGLU]

Fix Attention is_causal causal-mask alignment + composed is_causal/attn_mask NaN robustness for external (static) KV cache (#8054) by @titaiwangms in #8068
Add Celu-28 with float16/bfloat16/double support by @cyyever in #8107
docs: clarify Transpose operator spec by @pctablet505 in #8197
Add SwiGLU as a function-op by @gramalingam in #8202
Mark Range operator as deterministic by @take-cheeze in #8232
docs: fix ScatterND updates shape formula by @uditjainstjis in #8212
Add local_window_size attribute to Attention opset 25 by @FuZoe in #8108
Fix TensorScatter circular mode wrapping non-sequence coordinates by @justinchuby in #8353
Attention (opset 23/24): fix bf16 type mismatch in causal-mask and padding-mask paths by @justinchuby with @Copilot in #7840
feat: add FLOAT6E2M3 and FLOAT6E3M2 quantization data types by @andife in #8349
feat(defs): support bfloat16 in Compress and related ops by @justinchuby with @Copilot in #8078
Add mode attribute (DCR/CRD) to SpaceToDepth in opset 28, add function bodies to SpaceToDepth and DepthToSpace by @justinchuby with @Copilot in #7848
Allow float input to Mod when fmod=0 by @justinchuby with @Copilot in #7521
Add Einsum-28 with bfloat16 support by @mcollinswisc in #8313
refactor(defs): centralize shared schema docs by @justinchuby in #8387
Improve STFT operator documentation by @justinchuby with @Copilot in #7905
Add BitShift-28 with signed integer type support by @mcollinswisc in #8270
feat(defs): support all IR14 types in Optional ops by @youzow3 in #8176
fix(defs): restrict ReduceLogSum operators to float types in opset 28 by @malakazlan in #8384
RFC Proposal for Grouped MatMul operator by @gramalingam in #8193

Reference Implementation

Fix ScatterND reference impl for min/max with multi-axis indices by @Ankita7033 in #8099
Register latest QuantizeLinear and DequantizeLinear reference implementations by @napronald in #8182
Fix GatherElements output shape for empty indices by @kiwigitops in #8298
fix(reference): handle scalar NonZero inputs by @kiwigitops in #8296
Fix ReduceMax on empty boolean tensors by @kiwigitops in #8312
fix(reference): handle zero-length ReverseSequence slices by @kiwigitops in #8304
Support higher-rank ScatterElements evaluation by @kiwigitops in #8299
fix(reference): reject out-of-range Concat axes by @arcz in #8327
Runtime checks of shape annotations in reference evaluator by @gramalingam in #8307
fix(reference): use accurate exponential functions by @justinchuby with @Copilot in #8355

Utilities and Tools

Fix heap buffer over-read in Conv/Pool shape inference for weight rank mismatch by @titaiwangms with @Copilot in #8037
Fix make_attribute_ref to set ref_attr_name by @devin-lai in #8009
Address shape inference issues in Conv by @yuslepukhin in #8051
Fix annotated float attrs parsed from int literals by @fengjikui in #8065
Fix locale-dependent float parsing in C++ parser by @yuslepukhin in #8112
test: fix two copy-paste bugs that left code paths untested by @cyyever in #8159
Add loop test case with static loop bound by @gramalingam in #8146
Update shape inference tests to use parser by @gramalingam in #8168
Validate Scan num_scan_inputs to prevent shape-inference underflow by @cyyever in #8114
fix(version_converter): harden Scan adapters and domain-index lookup against malformed models by @andife in #8084
improve checker which fails on a particular mismatch between type and shape by @xadupre in #8211
Preserve inputs with no shape in Scan 8 -> 9 by @mcollinswisc in #8216
Fix OneHot shape inference when depth is empty by @cbourjau in #8177
Add Resize 18 to 17 version converter adapter by @napronald in #8209
Update hasInput and hasOutput methods of InferenceContextImpl by @gramalingam in #8271
Fix shape inference overflow by @andife in #8277
fix(shape_inference): handle missing function inputs in subgraphs by @arcz in #8305
fix(shape_inference): prevent use-after-scope during function inference by @arcz in #8303
fix(shape_inference): reject zero scalar SplitToSequence split by @arcz in #8322
fix(shape_inference): bound inferred rank materialization by @arcz in #8321
fix(version_converter): prevent initializer-name use-after-free by @arcz in #8334
fix: guard ExtendSupportedTypes against missing output in version converter by @andife in #8274
fix(version_converter): validate Captured node outputs by @arcz in #8329
fix(shape_inference): validate TreeEnsemble leaf weight rank by @arcz in #8336
Update shape inference test to use check type option by @gramalingam in #8323
fix(version_converter): validate Reshape output in Softmax downgrade by @arcz in #8326
fix(shape_inference): prevent use-after-scope during node inference by @arcz in #8330
fix(inliner): handle Constant nodes without outputs by @arcz in #8325
fix(checker): reject non-canonical FLOAT6 encodings by @justinchuby in #8386
fix(shape_inference): reject Split nodes with no outputs by @arcz in #8328

Build, CI and Tests

Update renovate config for Pixi lock file update by @cbourjau in #8022
Remove redundant min-deps tests in release process by @cbourjau in #8034
Replace windows-latest with windows-2022 to fix CI by @cbourjau in #8074
Update screenshot with correct configuration by @cbourjau in #8073
Use a cache for the link checker by @cbourjau in #8098
Clean-up tests by @cbourjau in #8154
Fix pyodide/emscripten CMake configure after nanobind 2.13.0 bump by @andife in #8157
Fix Python test bugs by @cyyever in #8160
Migrate to pytest's asserts by @cbourjau in #8155
Bump protobuf/abseil-cpp versions used to build protobuf from source in CI by @andife in #8172
Build Pyodide wheel for Python 3.14 by @metalmancode in #8192
Add Pad shape inference coverage for value_ints by @nightcityblade in #8206
Remove submodules option from checkout step by @andife in #8226
fix(common): route tensor dim errors through throw_tensor_error by @rvandermeulen in #8220
ci: add Linux GCC/Clang build with exceptions disabled by @andife in #8225
Add TSAN and TypeSAN sanitizer support; fix ASAN CI env-var propagation by @cyyever in #7471
Remove node test artifacts by @cbourjau in #7959
Add tests for direction attribute of RNN ops by @mcollinswisc in #7935
Fix b_zero_point for int8 in QLinearMatMul 3D test case by @andife with @Copilot in #7991
Fix Unique backend test output shapes by @DarkBall123 in #8180
Precut Version Update for Release 1.23.0 by @sfatimar in #8368
ci: add a clang-cl Windows job to catch MSVC-only permissive extensions by @andife in #8280
Pin ml_dtypes==0.5.4 in numpy<2.0 CI test jobs by @andife in #8378

Documentation

Add ROADMAP.md: 12-month roadmap template for SIGs and Working Groups by @andife in #8030
Expand roadmap with SIG details for 2026-2027 by @andife in #8059
Update operator sig roadmap by @gramalingam in #8070
Add C-API definition to roadmap by @cbourjau in #8097
Bump version number to 1.23.0 by @andife in #8102
Add jax2onnx to converter documentation by @enpasos in #8103
Update ROADMAP.md by @andife in #8120
fix calender link description (working-groups.md) by @andife in #7663
Remove onnx.tools.net_drawer from repository by @andife in #8129
docs: Refresh CONTRIBUTING/README docs and move Model Zoo links to Hugging Face by @andife in #8123
docs: update release tracking docs for the 1.22.0 release by @andife in #8190
docs: recommend Conventional Commits for agent-authored PR titles by @andife in #8188
docs: add REUSE annotation for AGENTS.md symlink by @andife in #8224
Update RELEASE-MANAGEMENT.md by @andife in #8234
docs: correct model validation assurances by @andife in #8276
docs: fix cross-platform environment variable syntax by @andife in #8283
docs: require labels on agent-authored PRs by @andife in #8285
docs: correct Protobuf build version guidance by @andife in #8286
Fix warnings from docs-build and ensure we fail on warnings in the future by @cbourjau in #8294
docs: refresh community governance docs, add History section by @andife in #8297
docs: clarify release and reproducibility guidance by @andife in #8284
docs: clarify spurious unknown variable in 'Slice' operator by @BW-SK in #8335
test: fix typo flagged by the typos lint hook by @cyyever in #8339
docs(community): add GitHub Organization Seat Policy by @andife in #8223
fix(docs): restore operator examples by @kiwigitops in #8332
docs: add llms.txt for AI-agent discoverability by @andife in #8348
Update RELEASE-MANAGEMENT.md by @sfatimar in #8406
Version change to rc1 by @sfatimar in #8407
Version changed to 1.23.0 by @sfatimar in #8469

Full Changelog: v1.22.0...v1.23.0

onnx 1.23.0 v1.23.0 on Python PyPI