Integrate ONNX 1.22.0 (opset 27) — issue #28752 - #28754

Merged

titaiwangms merged 32 commits into

microsoft:mainfrom

titaiwangms:integrate-onnx-1.22.0rc1

Jun 16, 2026

Commits on Jun 15, 2026

Integrate ONNX 1.22.0rc1 version plumbing (T1)

titaiwangms
and
Copilot
committed
T4: bump onnx pin 1.21.0 -> 1.22.0rc1 across CI requirements.txt files

titaiwangms
and
Copilot
committed
Enable ONNX opset 27 in ORT C++: bump transpose-optimizer max opset to 27 and register Range at opset 27

titaiwangms
and
Copilot
committed
Add opset-27 ONNX backend test filters and regenerate webgl operator doc (ONNX 1.22.0rc1, issue #28752)

titaiwangms
and
Copilot
committed
T6: opset-27 (ONNX 1.22) optimizer + EP guard updates for Range

titaiwangms
and
Copilot
committed
T6: add Range opset-27 to GatherToSlice fusion + opset-27 test

titaiwangms
and
Copilot
committed
T7: regenerate docs/OperatorKernels.md for opset-27 Range (ONNX 1.22.0rc1)

titaiwangms
and
Copilot
committed
Narrow opset-27 ONNX backend test filters: drop over-broad CausalConvWithState/LinearAttention/Range-fp16-bf16 exclusions (ONNX 1.22.0rc1, issue #28752)

titaiwangms
and
Copilot
committed
review-polish A4: clarify opset-27 Range comment that fp16/bf16 models work today via ONNX function-body expansion (ONNX 1.22.0rc1, issue #28752)

titaiwangms
and
Copilot
committed
review-polish: make GPU/CUDA/DML-unverified status explicit in opset-27 filter comment (ONNX 1.22.0rc1, issue #28752)

titaiwangms
and
Copilot
committed
Add onnx-opset-bump-checklist skill doc

titaiwangms
and
Copilot
committed
Bump CUDA Range kernel to opset 27 (mirror CPU)

titaiwangms
and
Copilot
committed
onnx-opset-bump skill: encode 'bump ALL execution providers together' tradition

titaiwangms
and
Copilot
committed
Update FunctionTest cycle-rejection assertions for ONNX 1.22 checker message

titaiwangms
and
Copilot
committed
Fix opset-27 GatherToSliceFusion test on strict ALLOW_RELEASED_ONNX_OPSET_ONLY legs

titaiwangms
and
Copilot
committed
Filter unimplemented ONNX 1.22 FlexAttention + dtype-256 backend tests

titaiwangms
and
Copilot
committed
Pin released opset in symlink-data test to avoid ONNX 1.22 opset-27 default

titaiwangms
and
Copilot
committed
Bump ONNX pin 1.22.0rc1 -> 1.22.0rc2; drop local Xcode workaround

titaiwangms
and
Copilot
committed
docs(skill): document x-block-origin 404 failure mode + per-RC re-mirror gotcha in onnx-opset-bump-checklist

titaiwangms
and
Copilot
committed
docs(skill): add CI 404 failure signature + consolidated fix options for vcpkg asset-mirror gotcha

titaiwangms
and
Copilot
committed
Conform ConvTranspose output_shape to ONNX spec (drop non-spec rank+2 leniency)

titaiwangms
and
Copilot
committed
Fix second rc2 shape-inference regression in test_op_split.py

titaiwangms
and
Copilot
committed
Handle rc2 backend conformance: skip expanded softcap-neginf attention + atol override for fp16 causal_conv

titaiwangms
and
Copilot
committed
Fix WebGPU binary-elementwise broadcast shared-dim over-count underflow (#28969)

titaiwangms
and
Copilot
committed
Un-skip expanded softcap-neginf Attention backend tests (#28969 fixed)

titaiwangms
and
Copilot
committed
docs(skill): close 3 gaps in onnx-opset-bump-checklist (webgpu CI scope, filter-vs-override, latent bugs)

titaiwangms
and
Copilot
committed
docs(skill): add webgpu-local-testing skill for running WebGPU tests on Linux without a GPU

titaiwangms
and
Copilot
committed
Bump pinned ONNX 1.22.0rc2 -> 1.22.0 (final release)

titaiwangms
and
Copilot
committed
Load current-opset fusion tests with allow_released_opsets_only=false

titaiwangms
and
Copilot
committed
Document opset-27 under-development test handling and bump gotchas

titaiwangms
and
Copilot
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Integrate ONNX 1.22.0 (opset 27) — issue #28752 - #28754

Integrate ONNX 1.22.0 (opset 27) — issue #28752#28754
titaiwangms merged 32 commits into
microsoft:mainfrom
titaiwangms:integrate-onnx-1.22.0rc1

Commits on Jun 15, 2026

Integrate ONNX 1.22.0rc1 version plumbing (T1)

T4: bump onnx pin 1.21.0 -> 1.22.0rc1 across CI requirements.txt files

Enable ONNX opset 27 in ORT C++: bump transpose-optimizer max opset to 27 and register Range at opset 27

Add opset-27 ONNX backend test filters and regenerate webgl operator doc (ONNX 1.22.0rc1, issue #28752)

T6: opset-27 (ONNX 1.22) optimizer + EP guard updates for Range

T6: add Range opset-27 to GatherToSlice fusion + opset-27 test

T7: regenerate docs/OperatorKernels.md for opset-27 Range (ONNX 1.22.0rc1)

Narrow opset-27 ONNX backend test filters: drop over-broad CausalConvWithState/LinearAttention/Range-fp16-bf16 exclusions (ONNX 1.22.0rc1, issue #28752)

review-polish A4: clarify opset-27 Range comment that fp16/bf16 models work today via ONNX function-body expansion (ONNX 1.22.0rc1, issue #28752)

review-polish: make GPU/CUDA/DML-unverified status explicit in opset-27 filter comment (ONNX 1.22.0rc1, issue #28752)

Add onnx-opset-bump-checklist skill doc

Bump CUDA Range kernel to opset 27 (mirror CPU)

onnx-opset-bump skill: encode 'bump ALL execution providers together' tradition

Update FunctionTest cycle-rejection assertions for ONNX 1.22 checker message

Fix opset-27 GatherToSliceFusion test on strict ALLOW_RELEASED_ONNX_OPSET_ONLY legs

Filter unimplemented ONNX 1.22 FlexAttention + dtype-256 backend tests

Pin released opset in symlink-data test to avoid ONNX 1.22 opset-27 default

Bump ONNX pin 1.22.0rc1 -> 1.22.0rc2; drop local Xcode workaround

docs(skill): document x-block-origin 404 failure mode + per-RC re-mirror gotcha in onnx-opset-bump-checklist

docs(skill): add CI 404 failure signature + consolidated fix options for vcpkg asset-mirror gotcha

Conform ConvTranspose output_shape to ONNX spec (drop non-spec rank+2 leniency)

Fix second rc2 shape-inference regression in test_op_split.py

Handle rc2 backend conformance: skip expanded softcap-neginf attention + atol override for fp16 causal_conv

Fix WebGPU binary-elementwise broadcast shared-dim over-count underflow (#28969)

Un-skip expanded softcap-neginf Attention backend tests (#28969 fixed)

docs(skill): close 3 gaps in onnx-opset-bump-checklist (webgpu CI scope, filter-vs-override, latent bugs)

docs(skill): add webgpu-local-testing skill for running WebGPU tests on Linux without a GPU

Bump pinned ONNX 1.22.0rc2 -> 1.22.0 (final release)

Load current-opset fusion tests with allow_released_opsets_only=false

Document opset-27 under-development test handling and bump gotchas

Commits on Jun 16, 2026

Code polish: named opset-allow constant + broadcast util doc/rename (exec-11)

Drop dead leading opset 1 from Range fusion version lists (exec-12)