Skip to content

Fill CUDA opset gap for Elu, Selu, ThresholdedRelu to opset 22 - #27753

Draft
Tianlei Wu (tianleiwu) with Copilot wants to merge 1 commit into
mainfrom
copilot/update-elus-selus-thresholdedrelus
Draft

Tianlei Wu (tianleiwu) with Copilot wants to merge 1 commit into
mainfrom
copilot/update-elus-selus-thresholdedrelus

Conversation

Copilot AI commented Mar 19, 2026

Copy link
Copy Markdown
Contributor

Description

Extends CUDA kernel registrations for Elu (6→22), Selu (6→22), and ThresholdedRelu (10→22) to cover the latest ONNX opset, following the established HardSigmoid/HardSwish pattern.

  • activations.cc: Cap existing non-versioned kernels to versioned 6/10–21, add non-versioned opset 22 registrations with BFloat16 support
  • cuda_execution_provider.cc: Update forward declarations and BuildKernelCreateInfo entries from ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAMEONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME for old ranges; add opset 22 entries for float, double, MLFloat16, BFloat16

Motivation and Context

CUDA EP had opset registration gaps for these three activation operators — models exported at opset ≥7 (Elu/Selu) or ≥11 (ThresholdedRelu) could not find a matching CUDA kernel. Tracked in #27729.

Co-authored-by: tianleiwu <30328909+tianleiwu@users.noreply.github.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants