[Reduction] Extend noop_with_empty_axes behavior to non-CPU EPs #26541
Description
Describe the issue
Follow-up to #26288 and PR #26436.
The spec-aligned noop_with_empty_axes behavior for Reduce operators is currently implemented only for the CPU Execution Provider.
Other EPs (e.g., CUDA, TensorRT, OpenVINO) still rely on the previous reduction logic and should be updated to match the ONNX specification.
This issue extends the spec-aligned behavior introduced in PR #26436 to all non-CPU providers for consistency across Execution Providers.
To reproduce
(After PR is merged, or, on branch fix/reduction-noop-empty-axes )
Run any Reduce operator (e.g. ReduceSumSquare) with axes=[] and noop_with_empty_axes=1 on a non-CPU EP
(e.g., CUDA, TensorRT, OpenVINO).
The output will differ from the CPU EP, which already follows the ONNX spec.
Urgency
No response
Platform
Windows
OS Version
Windows 11 Pro 25H2 (Build 26200.7019)
ONNX Runtime Installation
Built from Source
ONNX Runtime Version or Commit ID
ONNX Runtime API
Python
Architecture
X64
Execution Provider
Default CPU
Execution Provider Library Version
No response