Register SimplifiedLayerNormalization to MS domain

[justinchuby]

Register SimplifiedLayerNormalization to MS domain. Previously it was incorrectly registered to the ONNX domain, causing onnx models to be invalid.

Fix #7573

[justinchuby]

There are errors

cuda_contrib_kernels.obj : error LNK2001: unresolved external symbol "struct onnxruntime::KernelCreateInfo __cdecl onnxruntime::contrib::cuda::BuildKernelCreateInfo<class onnxruntime::contrib::cuda::kCudaExecutionProvider_SimplifiedLayerNormalization_kMSDomain_ver1_float_float_float>(void)" (??$BuildKernelCreateInfo@VkCudaExecutionProvider_SimplifiedLayerNormalization_kMSDomain_ver1_float_float_float@cuda@contrib@onnxruntime@@@cuda@contrib@onnxruntime@@YA?AUKernelCreateInfo@2@XZ) [E:\_work\_temp\build\RelWithDebInfo\onnxruntime_providers_cuda.vcxproj]

Is it the way I registered the ops?

[github-actions]

You can commit the suggested changes from lintrunner.

[fs-eire]

Is there any following up for this PR?

[justinchuby]

I still want to merge but there’s some build errors I don’t understand. Any pointers you can share?

[fs-eire]

maybe merge to latest main and check latest error message (if any)? the current code is a little bit out of date

[MaanavD]

@justinchuby (cleanup) - can we close this issue / unassign me pls :)

[justinchuby]

@MaanavD can you unsubscribe from the pr for now? The reviewer list doesn't load for me so I am not able to make changes.

[justinchuby]

@copilot please update from main and fix breaking tests.

[Copilot]

@copilot please update from main and fix breaking tests.

Updated from main and resolved conflicts in onnxruntime/contrib_ops/webgpu/webgpu_contrib_kernels.cc and onnxruntime/contrib_ops/rocm/rocm_contrib_kernels.cc (ROCm contrib ops were removed entirely on main, so that file is deleted). Verified the CUDA kernel declarations/registrations for SimplifiedLayerNormalization in both kMSDomain and the backward-compat kOnnxDomain are still consistent (15 declarations match 15 registrations).

[Copilot]

Pull request overview

This PR corrects the operator domain for SimplifiedLayerNormalization so it is registered in the MS contrib domain (kMSDomain) instead of only (incorrectly) being exposed as an ONNX-domain op, while keeping ONNX-domain registrations for backward compatibility with existing (invalid) models.

Changes:

• Register SimplifiedLayerNormalization kernels for MS domain in CUDA, JS, and WebGPU contrib registries (and keep ONNX-domain registrations for compatibility).
• Add MS-domain kernel definitions for SimplifiedLayerNormalization in the JS and WebGPU implementations.
• Add CUDA MS-domain kernel registry entries for all supported type combinations.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
onnxruntime/contrib_ops/webgpu/webgpu_contrib_kernels.cc	Adds MS-domain SimplifiedLayerNormalization to the WebGPU contrib kernel table; retains ONNX-domain entry for compatibility.
onnxruntime/contrib_ops/webgpu/bert/layer_norm.cc	Adds MS-domain kernel definition for SimplifiedLayerNormalization on WebGPU.
onnxruntime/contrib_ops/js/layer_norm.cc	Adds MS-domain kernel definition for SimplifiedLayerNormalization on JS EP.
onnxruntime/contrib_ops/js/js_contrib_kernels.cc	Registers MS-domain SimplifiedLayerNormalization in the JS contrib kernel registry.
onnxruntime/contrib_ops/cuda/layer_norm.cc	Adds MS-domain typed kernel registrations for SimplifiedLayerNormalization in CUDA.
onnxruntime/contrib_ops/cuda/cuda_contrib_kernels.cc	Registers MS-domain typed kernels for SimplifiedLayerNormalization in CUDA contrib kernel registry.
onnxruntime/contrib_ops/cpu/layer_norm.cc	Adds MS-domain typed kernel registrations for SimplifiedLayerNormalization in CPU contrib implementation (but registry wiring still needed).

Suppressed comments (1)

onnxruntime/contrib_ops/cpu/layer_norm.cc:29

• This adds an MS-domain kernel registration for SimplifiedLayerNormalization, but the CPU contrib kernel registry (onnxruntime/contrib_ops/cpu/cpu_contrib_kernels.cc) still only registers the kOnnxDomain versions. As-is, MS-domain SimplifiedLayerNormalization will fail to resolve on CPU (fallback) despite having a kernel implementation here.

Please add the corresponding class declarations and BuildKernelCreateInfo entries for kMSDomain, while keeping the kOnnxDomain entries for backward compatibility.

  ONNX_OPERATOR_TYPED_KERNEL_EX(SimplifiedLayerNormalization, kMSDomain, 1, T, kCpuExecutionProvider,       \
                                KernelDefBuilder()                                                          \
                                    .TypeConstraint("T", DataTypeImpl::GetTensorType<T>())                  \
                                    .TypeConstraint("U", DataTypeImpl::GetTensorType<T>())                  \
                                    .TypeConstraint("V", DataTypeImpl::GetTensorType<T>()),                 \

[justinchuby]

@copilot also npm run build:doc

[tianleiwu]

@justinchuby, you can use ort-ci skill to fix the doc gen CI.