Integrate ONNX 1.23.0 base plumbing

[Justin Chu (justinchuby)]

Summary

• pin ONNX Runtime to ONNX c78026fa617f273fe5c685c74f9de3e6e96504ed (1.23.0, opset 28 development)
• raise base and CoreML, NNAPI, VSINPU, and WebNN opset ceilings to 28
• add the mechanically required Float6 public element-type mappings and exhaustive conversion coverage
• materialize the removed ONNX node-test corpus at build time, with version and minimum-case tripwires in the materializer, C++ runner, and Python harness

Scope

This stack base intentionally does not add per-operator opset-28 registrations or kernel implementations. Follow-up stacks own coordinated version splits across every registering EP. The materialized CPU runner keeps the existing Python ReduceMax empty-bool exclusion in parity.

Validation

• verified target archive hashes and onnx.patch / binskim.patch byte identity and application
• source-built exact onnx==1.23.0 wheel
• materialized 1,884 node cases; direct tensor/map/sequence/optional fixture serialization smoke test passed
• --minimal_build extended build passed
• regular full build passed
• onnx_test_node_materialized CTest passed: 1,762 CPU cases, 0 failures, with ALLOW_RELEASED_ONNX_OPSET_ONLY=0

External dependency

The vcpkg artifact mirror for the new SHA512 currently returns HTTP 404 and requires Terrapin/infra seeding. No privileged upload was attempted.

[Copilot]

🟡 Changes recommended

Opset-28 dispatch can mis-handle valid models, and the iOS build and formatting paths contain concrete regressions.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Updates ONNX Runtime’s dependency and infrastructure for ONNX 1.23.0 and development opset 28.

Changes:

• Pins ONNX sources and CI dependencies to the 1.23 release line.
• Adds Float6 public type mappings and opset-28 ceilings.
• Materializes and validates the removed ONNX node-test corpus.

File summaries

File	Description
.github/workflows/linux_ci.yml	Updates materialization dependency documentation.
cmake/deps.txt	Pins the ONNX source archive.
cmake/external/onnx	Advances the ONNX submodule.
cmake/onnxruntime_unittests.cmake	Updates materialization version parsing and dependencies.
cmake/patches/onnx/onnx.patch	Rebases ONNX build and iOS patches.
cmake/vcpkg-ports/onnx/binskim.patch	Mirrors the rebased ONNX patch.
cmake/vcpkg-ports/onnx/fix-dependency-protobuf.patch	Rebases protobuf patch context.
cmake/vcpkg-ports/onnx/portfile.cmake	Updates vcpkg source reference and hash.
cmake/vcpkg-ports/onnx/vcpkg.json	Sets ONNX vcpkg version metadata.
include/onnxruntime/core/graph/schema_registry.h	Matches the updated ONNX registry signature.
include/onnxruntime/core/session/onnxruntime_c_api.h	Exposes Float6 element types.
js/web/docs/webgl-operators.md	Regenerates opset documentation.
onnxruntime/core/framework/onnxruntime_map_type_info.h	Extends element-type mapping validation.
onnxruntime/core/framework/tensor_type_and_shape.cc	Maps Float6 types to public enums.
onnxruntime/core/framework/tensorprotoutils.h	Adds Float6 logical element sizes.
onnxruntime/core/optimizer/embed_layer_norm_fusion.cc	Recognizes updated opset versions.
onnxruntime/core/optimizer/transpose_optimization/optimizer_api.h	Enables opset-28 optimization.
onnxruntime/core/providers/coreml/builders/impl/base_op_builder.h	Raises the CoreML ceiling.
onnxruntime/core/providers/nnapi/nnapi_builtin/builders/impl/base_op_builder.h	Raises the NNAPI ceiling.
onnxruntime/core/providers/vsinpu/builders/impl/base_op_builder.h	Raises the VSINPU ceiling.
onnxruntime/core/providers/webnn/builders/impl/base_op_builder.h	Raises the WebNN ceiling.
onnxruntime/test/framework/tensorutils_test.cc	Covers Float6 element sizes.
onnxruntime/test/framework/type_info_test.cc	Covers Float6 type conversion.
onnxruntime/test/onnx/TestCase.cc	Mirrors the ReduceMax exclusion.
onnxruntime/test/optimizer/layout_transformation_potentially_added_ops_test.cc	Temporarily exempts opset-28 Q/DQ.
onnxruntime/test/providers/cpu/controlflow/scan_test.cc	Updates expected ONNX diagnostics.
onnxruntime/test/python/onnx_node_test_equivalence_test.py	Reads commit-pin version metadata.
onnxruntime/test/python/requirements.txt	Pins the ONNX RC wheel.
onnxruntime/test/testdata/onnx_backend_test_series_filters.jsonc	Defers unsupported opset-28 cases.
onnxruntime/test/unittest_util/op_tester.cc	Adds a released ONNX import for custom-domain tests.
requirements-materialize-onnx-node-tests.txt	Updates the materializer ONNX pin.
tools/ci_build/github/linux/docker/inference/aarch64/python/cpu/scripts/requirements.txt	Updates the AArch64 ONNX pin.
tools/ci_build/github/linux/docker/scripts/lort/requirements.txt	Updates the LORT ONNX pin.
tools/ci_build/github/linux/docker/scripts/manylinux/requirements.txt	Updates the manylinux ONNX pin.
tools/ci_build/github/linux/docker/scripts/requirements.txt	Updates the Linux image ONNX pin.
tools/ci_build/github/linux/python/requirements.txt	Updates the Linux CI ONNX pin.
tools/ci_build/github/windows/python/requirements.txt	Updates the Windows CI ONNX pin.
tools/python/materialize_onnx_node_tests.py	Enforces exact dataset arity.

Review details

Suppressed comments (1)

onnxruntime/test/testdata/onnx_backend_test_series_filters.jsonc:118

• These skips hide a runtime correctness regression: the existing CPU/CUDA Mod registration starting at opset 13 is open-ended, so it binds opset-28 nodes, but its floating implementation rejects fmod=0, which opset 28 now defines as floor-mod. Version-cap the legacy registrations at 27 or include the opset-28 implementation before enabling this schema; otherwise valid models throw during execution.

        // TODO(#32369): Remove these exact skips when the opset-28 Mod kernels land.
        "^test_mod_broadcast(_(cpu|cuda))?$",
        "^test_mod_float16_mixed_sign_fmod_0(_(cpu|cuda))?$",
        "^test_mod_float32_mixed_sign_fmod_0(_(cpu|cuda))?$",
        "^test_mod_float64_mixed_sign_fmod_0(_(cpu|cuda))?$",

• Files reviewed: 38/38 changed files
• Comments generated: 6
• Review effort level: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🔵 Needs a closer look

The source/wheel mismatch, incomplete Float6 conversion, and missing Cast-28 matcher updates can cause test drift, incorrect API type reporting, and optimizer regressions.

Review details

Suppressed comments (3)

Previously missed (3) — in code that hasn't changed since the last review.

cmake/deps.txt:39

• The source and wheel pins are not actually in lockstep. The pinned c78026f source includes the newer LRN node-test generator ([1,5,5,5] input), while the 1.23.0rc1 release branch still generates [5,5,5,5]; the release-base-only checks accept both as 1.23.0. As a result, CI materializes and runs a corpus from different ONNX code than ORT compiles, and the post-#7959 tree has no disk oracle to catch that drift. Pin a wheel built from the exact source revision (or pin the source to rc1) and enforce revision/exact-version parity.
  onnxruntime/core/framework/tensor_type_and_shape.cc:195
• These cases update only one Proto-to-C-API conversion path. utils::CApiElementTypeFromProtoType still falls through to UNDEFINED for both Float6 values (tensorprotoutils.cc:2023-2059), and that helper feeds custom-op shape inference at custom_ops.cc:86. Thus a custom shape-inference callback sees UNDEFINED instead of either newly public enum. Add both cases there (and in its standalone test-runner counterpart) with coverage.
  onnxruntime/core/optimizer/embed_layer_norm_fusion.cc:250
• Cast now has schema version 28, but this is the only fusion matcher updated. Other Cast-sensitive optimizers still stop at 25—for example layer_norm_fusion.cc:148, fast_gelu_fusion.cc:178, isinf_reducesum_fusion.cc:48, and pad_fusion.cc:135—so equivalent opset-28 graphs silently lose those fusions. Audit all Cast version lists and extend the compatible matchers/tests together.

• Files reviewed: 45/45 changed files
• Comments generated: 0 new
• Review effort level: Balanced

[Ti-Tai Wang (titaiwangms)]

I have uploaded official 1.23. We should use official 1.23 now.