Repository navigation
Add opset 28 SpaceToDepth support - #32362
Open
Justin Chu (justinchuby) wants to merge 16 commits into
Open
Justin Chu (justinchuby) wants to merge 16 commits into
Justin Chu (justinchuby) wants to merge 16 commits into
Conversation
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Base automatically changed from
justinchuby-onnx-1-23-opset-28-base
to
main
September 30, 2026 15:34
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Contributor
There was a problem hiding this comment.
Copilot review overview
🟡 Changes recommended
The ONNX pin is inconsistent with the final wheel, other opset-28 schema changes remain unsupported, and generated documentation and provider coverage are incomplete.
Review effort: Balanced
Findings: 1
Open (4)
What changed in this PR
Adds ONNX opset-28 SpaceToDepth/DepthToSpace support and updates the ONNX dependency.
Changes:
- Implements DCR/CRD behavior across CPU, CUDA, DML, and internal NHWC paths.
- Splits opset-28 kernel registrations and raises provider opset ceilings.
- Adds validation, function-parity tests, and updates ONNX package pins.
| File | Description |
|---|---|
.github/workflows/linux_ci.yml |
Updates ONNX version documentation. |
cmake/deps.txt |
Pins ONNX 1.23 source. |
cmake/onnxruntime_unittests.cmake |
Parses commit-pin version metadata. |
cmake/vcpkg-ports/onnx/portfile.cmake |
Updates vcpkg ONNX source and hash. |
cmake/vcpkg-ports/onnx/vcpkg.json |
Bumps vcpkg ONNX version. |
onnxruntime/core/graph/contrib_ops/internal_nhwc_onnx_schemas.cc |
Registers opset-28 NHWC schemas. |
onnxruntime/core/optimizer/transpose_optimization/optimizer_api.h |
Enables opset-28 optimization. |
onnxruntime/core/providers/coreml/builders/impl/base_op_builder.h |
Raises CoreML opset ceiling. |
onnxruntime/core/providers/cpu/cpu_execution_provider.cc |
Registers opset-28 CPU kernels. |
onnxruntime/core/providers/cpu/tensor/space_depth_ops.cc |
Implements CPU CRD SpaceToDepth. |
onnxruntime/core/providers/cpu/tensor/space_depth_ops.h |
Reads and stores SpaceToDepth mode. |
onnxruntime/core/providers/cuda/cuda_execution_provider.cc |
Registers opset-28 CUDA kernels. |
onnxruntime/core/providers/cuda/cuda_nhwc_kernels.cc |
Registers opset-28 NHWC CUDA kernels. |
onnxruntime/core/providers/cuda/tensor/space_depth_ops.cc |
Implements CUDA CRD permutations. |
onnxruntime/core/providers/cuda/tensor/space_depth_ops.h |
Stores CUDA SpaceToDepth mode. |
onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorSpaceToDepth.cpp |
Adds ordered DML SpaceToDepth. |
onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorRegistration.cpp |
Registers opset-28 DML operators. |
onnxruntime/core/providers/dml/OperatorAuthorHelper/OperatorVersions.h |
Defines DML opset-28 versions. |
onnxruntime/core/providers/js/js_execution_provider.cc |
Registers split JS kernels. |
onnxruntime/core/providers/js/operators/depth_to_space.cc |
Splits JS DepthToSpace registration. |
onnxruntime/core/providers/nnapi/nnapi_builtin/builders/impl/base_op_builder.h |
Raises NNAPI opset ceiling. |
onnxruntime/core/providers/qnn/builder/opbuilder/simple_op_builder.cc |
Rejects unsupported QNN CRD mode. |
onnxruntime/core/providers/vsinpu/builders/impl/base_op_builder.h |
Raises VSINPU opset ceiling. |
onnxruntime/core/providers/webgpu/tensor/depth_to_space.cc |
Splits WebGPU registrations. |
onnxruntime/core/providers/webgpu/webgpu_execution_provider.cc |
Registers opset-28 WebGPU kernels. |
onnxruntime/core/providers/webnn/builders/impl/base_op_builder.h |
Raises WebNN opset ceiling. |
onnxruntime/test/contrib_ops/function_ops_test.cc |
Adds function/native parity tests. |
onnxruntime/test/providers/cpu/tensor/space_depth_ops_test.cc |
Adds opset-28 mode and validation tests. |
onnxruntime/test/python/onnx_node_test_equivalence_test.py |
Reads explicit ONNX version metadata. |
onnxruntime/test/python/requirements.txt |
Pins ONNX 1.23. |
requirements-materialize-onnx-node-tests.txt |
Updates materialization dependency. |
tools/ci_build/github/linux/docker/inference/aarch64/python/cpu/scripts/requirements.txt |
Updates AArch64 ONNX pin. |
tools/ci_build/github/linux/docker/scripts/lort/requirements.txt |
Updates LORT ONNX pin. |
tools/ci_build/github/linux/docker/scripts/manylinux/requirements.txt |
Updates manylinux ONNX pin. |
tools/ci_build/github/linux/docker/scripts/requirements.txt |
Updates Docker ONNX pin. |
tools/ci_build/github/linux/python/requirements.txt |
Updates Linux ONNX pin. |
tools/ci_build/github/windows/python/requirements.txt |
Updates Windows ONNX pin. |
tools/python/materialize_onnx_node_tests.py |
Enforces exact fixture arity. |
💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Regenerate the WebGPU operator table and let the new opset 28 coverage bypass released-opset validation on a per-test basis. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Correct CRD CPU layout and test vectors, permit internal NHWC schemas through the latest ONNX opset, and propagate unreleased-opset policy through function model construction. Keep minimal-build model loading on its supported buffer path and constrain CPU error assertions to the CPU provider. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Run SpaceToDepth mode parity only on its CPU kernel and apply the unreleased-opset test policy while constructing QNN test models. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Contributor
|
/azp run Linux Android Emulator QNN CI Pipeline |
|
Azure Pipelines will not run the associated pipelines, because the pull request was updated after the run command was issued. Review the pull request again and issue a new run command. |
Contributor
|
/azp run Linux Android Emulator QNN CI Pipeline |
|
Azure Pipelines will not run the associated pipelines, because the pull request was updated after the run command was issued. Review the pull request again and issue a new run command. |
Mark Harfouche (hmaarrfk)
added a commit
to hmaarrfk/onnxruntime-feedstock
that referenced
this pull request
Oct 6, 2026
<details><summary>Claude's draft</summary>
libonnx 1.23.1 on conda-forge reports opset 28 as the last released
ai.onnx opset, so models built at the default opset are now opset 28.
onnxruntime 1.30.0 with the 0013 backport only has opset-28 kernels for
Mod and BitShift; the other re-versioned operators have open-ended
registrations, which KernelRegistry::VerifyVersion does not match
against a newer schema version. On the last CI run that failed 24 unit
tests on linux-64, win-64 and osx-arm64 CPU builds (more on macOS) with
Failed to find kernel for Cast(28)
Add patch 0014. For the 15 operators whose opset-28 schema differs
from the previous one only in its type constraints, give the newest CPU
registration an explicit end version of 28: Cast, Celu, Compress,
DepthToSpace, DequantizeLinear, QuantizeLinear, Einsum, OneHot,
Optional, OptionalHasElement, OptionalGetElement, ReduceLogSum,
ReduceLogSumExp, ReverseSequence and Unique. No kernel code changes.
Not covered, and still rejected at session creation rather than
computed wrongly: the types opset 28 adds (bfloat16, float6), and
SpaceToDepth, whose new mode attribute the kernel does not implement.
CUDA registrations are untouched, so an opset-28 node of these
operators runs on the CPU kernel in GPU builds.
This is a conda-forge patch, not an upstream one. Upstream's per-operator
work is in the open PRs microsoft/onnxruntime#32362, #32372, #32378,
#32381 and #32393.
Built locally: linux-64 CPU, Python 3.13, against libonnx 1.23.1.
ctest reports 100% tests passed (11 of 11) and both package test
phases pass.
Resume this Claude session:
```
cd /home/mark/git/feedstock/onnxruntime-feedstock
claude --resume c16b7120-814d-4a38-9b1f-2afd9f75d054
```
</details>
Claude-Session: https://claude.ai/code/session_01XZa8BwzZpVJtREiCoVo1k6
Mark Harfouche (hmaarrfk)
added a commit
to hmaarrfk/onnxruntime-feedstock
that referenced
this pull request
Oct 11, 2026
<details><summary>Claude's draft</summary>
libonnx 1.23.1 on conda-forge reports opset 28 as the last released
ai.onnx opset, so models built at the default opset are now opset 28.
onnxruntime 1.30.0 with the 0013 backport only has opset-28 kernels for
Mod and BitShift; the other re-versioned operators have open-ended
registrations, which KernelRegistry::VerifyVersion does not match
against a newer schema version. On the last CI run that failed 24 unit
tests on linux-64, win-64 and osx-arm64 CPU builds (more on macOS) with
Failed to find kernel for Cast(28)
Add patch 0014. For the 15 operators whose opset-28 schema differs
from the previous one only in its type constraints, give the newest CPU
registration an explicit end version of 28: Cast, Celu, Compress,
DepthToSpace, DequantizeLinear, QuantizeLinear, Einsum, OneHot,
Optional, OptionalHasElement, OptionalGetElement, ReduceLogSum,
ReduceLogSumExp, ReverseSequence and Unique. No kernel code changes.
Not covered, and still rejected at session creation rather than
computed wrongly: the types opset 28 adds (bfloat16, float6), and
SpaceToDepth, whose new mode attribute the kernel does not implement.
CUDA registrations are untouched, so an opset-28 node of these
operators runs on the CPU kernel in GPU builds.
This is a conda-forge patch, not an upstream one. Upstream's per-operator
work is in the open PRs microsoft/onnxruntime#32362, #32372, #32378,
#32381 and #32393.
Built locally: linux-64 CPU, Python 3.13, against libonnx 1.23.1.
ctest reports 100% tests passed (11 of 11) and both package test
phases pass.
Resume this Claude session:
```
cd /home/mark/git/feedstock/onnxruntime-feedstock
claude --resume c16b7120-814d-4a38-9b1f-2afd9f75d054
```
</details>
Claude-Session: https://claude.ai/code/session_01XZa8BwzZpVJtREiCoVo1k6
Ti-Tai Wang (titaiwangms)
pushed a commit
that referenced
this pull request
Oct 12, 2026
### Description <!-- Describe your changes. --> Reintegrate the ONNX 1.23 base plumbing originally introduced by #32359 and reverted by #33073, now using ONNX **1.23.2**. - Coordinate source/submodule, FetchContent, vcpkg, and Python dependency pins and raise the relevant opset ceilings to 28. - Restore Float6 public element-type mappings and conversion coverage, preserving subsequent independent type-support changes. - Update existing-type Cast registrations, CPU/CUDA Mod boundaries, signed BitShift support, provider compatibility, backend filters/tolerances, and operator documentation. - Materialize the removed upstream node-test corpus with exact-version and minimum-case guards, and cover backend serialization and released-opset policy. - Keep matching FetchContent/vcpkg compatibility patches; remove the released-opset correction now fixed upstream in 1.23.2. - Address review feedback: OpenVINO rejects BitShift-28; CPU/CUDA legacy SpaceToDepth kernels stop at 27; DirectML accepts only safe DCR; unverified graph providers defer SpaceToDepth-28. Full builds use ONNX function expansion for CRD, with literal-output and routing regressions. Avoid creating unsupported SpaceToDepth-28 nodes in the post-expansion Slice/Concat fusion. - Use shared ABI-safe tensor-element conversion in the native ONNX node runner, fixing INT2/UINT2/FLOAT8E8M0 corpus decoding after the main merge. ### Motivation and Context <!-- - Why is this change required? What problem does it solve? - If it fixes an open issue, please link to the issue here. --> This is a redo of #32359 after its rollback in #33073. ONNX 1.23.2 includes the external tensor read/re-save fix from onnx/onnx#8472 for onnx/onnx#8471 and the released-opset metadata fix from onnx/onnx#8544 for onnx/onnx#8529. The redundant local `LastReleaseVersionMap` patch has been removed. Remaining FetchContent/vcpkg patches are byte-identical and apply against the pinned release. **Scope:** This PR establishes dependency/runtime base integration. Full per-operator opset-28 support remains separate follow-up work; see #32362, #32369, #32372, #32378, #32381, and #32393. These references describe staged scope, not a claim that every follow-up has landed or every opset-28 operator is supported. Runtime function expansion is a full-build fallback, not minimal/ORT-format runtime coverage. ### Local validation Completed on Windows with retained CPU-only Ninja/MSVC `RelWithDebInfo` builds and strict released-opset policy: - Built `onnxruntime_provider_test`, `onnxruntime_test_all`, `onnx_test_runner`, and Python bindings against ONNX 1.23.2; retained build remains available for incremental work. - Source-built and installed the exact patched ONNX 1.23.2 Python wheel. - Separate `--minimal_build extended` shared runtime configured, compiled, and linked successfully with ML/contrib ops and unit tests disabled. This exercises `ONNX_MINIMAL_BUILD=ON`; it is not a minimal-runtime function-op execution test. - Mod/BitShift/SpaceToDepth/Cast provider tests: **247 passed**, no skips. SpaceToDepth checks default DCR, explicit DCR, and CRD with distinct literal expected outputs, expansion, and legacy opset boundaries. - Slice/Concat fusion and type-info/tensor-proto framework tests: **27 passed**, no skips. - Python backend tests: **22 passed**, plus **2 subtests passed**. - Full Python backend series: **4,066 tests**, **2,685 skipped**, zero failures. Skipped/unsupported coverage is not a pass. - Materialized **1,884 node cases**. Native CPU runner with `-m 1500` collected **1,760 cases**: **1,590 succeeded**, **170 not implemented**, **zero failed**, exit 0. These results do not establish complete opset-28 support. - Generated CPU operator documentation matches the rebuilt kernel registry; verified archive hashes, patch identity/application, submodule SHA, and changed-file lint. **Remaining validation/infrastructure limits:** GPU and other EP execution relies on CI; optional-provider routing checks are compiled into the relevant builds but were unavailable locally. Full unfiltered C++ suites and the ancillary broad `ALL` target were not rerun. The oracle-dependent corpus equivalence test skips because upstream no longer ships the corpus. The new ONNX 1.23.2 tarball currently returns **404 from the Microsoft vcpkg asset mirror**; it still needs Terrapin/infra seeding before read-only `x-block-origin` vcpkg legs can download it. FetchContent and the successful minimal build are mirror-independent. New-head CI is pending. --------- Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: gramalingam <10075881+gramalingam@users.noreply.github.com>
Keep opset-28 provider support compatible with the reintegrated ONNX base, including DirectML CRD fallback and updated generated kernel docs. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Contributor
|
/azp run Linux Android Emulator QNN CI Pipeline |
|
Azure Pipelines: No pipelines were found matching this branch/path. |
PR #33134 added SpaceToDepthOpset28ModesAndProviderRouting with an expectation that opset 28 nodes expand via ONNX function expansion (EXPECT_FALSE(has_space_to_depth)), because native opset 28 kernels were intentionally deferred to this PR (#32362). Now that native opset 28 SpaceToDepth kernels are implemented for CPU and CUDA, update the test to assert that the node remains SpaceToDepth (EXPECT_TRUE), routes to CPU for CPUExecutionProvider, routes to CUDA for CUDAExecutionProvider, and falls back from DML to CPU for CRD mode while DCR routes to DML. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Check provider assignment for DirectML and OpenVINO fused nodes instead of requiring the original SpaceToDepth operator. Preserve native CPU/CUDA and opset-28 CPU fallback checks, excluding inserted memory-copy nodes from the computation count. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
The CUDA provider does not support opset-28 SpaceToDepth natively, so ONNX function expansion can leave Transpose/Reshape nodes assigned to CUDA. Keep checking the provider assignment and output, but require the native SpaceToDepth node only for CPU and explicit CPU fallbacks. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Distinguish the NCHW and NHWC CUDA provider instances, which report the same provider type. Require one native SpaceToDepth node for both, plus exactly two layout transposes for NHWC. Preserve provider assignment checks for every computation and layout node and retain the existing graph-EP fallback checks. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
SpaceToDepthsupport for bothDCRandCRDmodes on CPU and CUDA, including internal NHWC kernels.SpaceToDepthandDepthToSpaceregistrations at opset 28 across CPU, CUDA, DML, JS, and WebGPU; add DML orderedSpaceToDepthsupport and make QNN decline unsupported CRD nodes for fallback.Validation
git diff --checkbuild.py --config Debug --build --build_dir build\\windows-spacedepth --cmake_generator "Visual Studio 17 2022" --target onnxruntime_provider_testis blocked by the base branch's unrelated FP6TensorProtodatatype-map static assertion intensorprotoutils.h, before producing a test binary.