[WebGPU] Fix SkipSimplifiedLayerNormalization bias

[xenova]

Description

The op spec for com.microsoft.SkipSimplifiedLayerNormalization puts the optional 1-D bias at input slot 3 (the simplified variant has no beta), but onnxruntime/contrib_ops/webgpu/bert/skip_layer_norm.cc hardcodes beta = Input(3); bias = Input(4) regardless of the simplified template parameter — so the user-supplied bias is silently consumed as beta and the kernel reads slot 4 (out of range for this op) for the actual bias. The CPU EP handles the slot correctly; only WebGPU diverges (and produces NaN in larger graphs where slot 4's memory is non-zero).

Motivation and Context

Closes #28424

[Copilot]

Pull request overview

Fixes a WebGPU-specific input indexing bug in com.microsoft.SkipSimplifiedLayerNormalization where the optional 1-D bias was read from the wrong input slot (and could lead to incorrect results/NaNs), bringing WebGPU behavior in line with the operator schema and other EPs.

Changes:

• Correct WebGPU SkipSimplifiedLayerNormalization optional input handling so bias is read from slot 3 (and beta is absent for the simplified variant).
• Strengthen WebGPU shader program caching by including hasBeta/hasBias in the program cache hint to avoid cache collisions across optional-input combinations.
• Add a regression unit test that exercises SkipSimplifiedLayerNormalization with bias in FP16.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
onnxruntime/contrib_ops/webgpu/bert/skip_layer_norm.cc	Fixes simplified-vs-non-simplified beta/bias input slot selection and updates shader cache hint to account for optional inputs.
onnxruntime/test/contrib_ops/skiplayernorm_op_test.cc	Adds an FP16 regression test validating correct bias handling for SkipSimplifiedLayerNormalization.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[guschmue]

DirectML EP is not happy with the new unit test.
We don't maintain that EP anymore and don't want to touch it.
Maybe easiest way out is to disable the test for DirectML.

[xenova]

Any pointers on how to disable it only for DirectML? 👀

[guschmue]

yeap, copy this line:

onnxruntime/onnxruntime/test/contrib_ops/attention_op_test.cc

Line 75 in a49da87

bool enable_dml = (nullptr != DefaultDmlExecutionProvider().get()) && !disable_dml;

[xenova]

thanks 👍 936cafb should hopefully fix that.