webgpu support for qwen3.5

[guschmue]

webgpu support for qwen3.5, adding LinearAttention and CausalConvWithState ops based on this proposal:
from onnx/onnx#7767

The model can be created with model builder from https://github.com/microsoft/onnxruntime-genai/blob/main/src/python/py/models/builder.py.

For example for the text only flavor:

python builder.py -m Qwen/Qwen3.5-0.8B  -o Qwen3.5-0.8B -e webgpu -p int4 --extra_options int4_accuracy_level=4 exclude_embeds=False

[Copilot]

Pull request overview

Adds WebGPU execution-provider support needed to run Qwen3.5 models by introducing two new contrib ops and wiring them into kernel registration.

Changes:

• Fixes boolean Expand vec4/bool-pack indexing in the WebGPU Expand shader generation.
• Adds WebGPU contrib kernels for LinearAttention and CausalConvWithState (C++ + WGSL templates) and registers them.
• Implements fused recurrent linear attention and causal depthwise conv-with-state shaders for Qwen3.5-style decode/prefill.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
onnxruntime/core/providers/webgpu/tensor/expand.cc	Corrects boolx4 input offset calculation for vec4-path expand.
onnxruntime/contrib_ops/webgpu/webgpu_contrib_kernels.cc	Registers new WebGPU contrib kernels for LinearAttention and CausalConvWithState.
onnxruntime/contrib_ops/webgpu/bert/linear_attention.h	Declares LinearAttention kernel/program and update-rule parsing.
onnxruntime/contrib_ops/webgpu/bert/linear_attention.cc	Implements LinearAttention kernel setup, validation, dispatch, and shader templating.
onnxruntime/contrib_ops/webgpu/bert/linear_attention.wgsl.template	Adds fused linear attention WGSL implementation (incl. delta/gated-delta paths).
onnxruntime/contrib_ops/webgpu/bert/causal_conv_with_state.h	Declares CausalConvWithState kernel/program and activation parsing.
onnxruntime/contrib_ops/webgpu/bert/causal_conv_with_state.cc	Implements CausalConvWithState kernel setup, validation, and dispatch.
onnxruntime/contrib_ops/webgpu/bert/causal_conv_with_state.wgsl.template	Adds WGSL for depthwise causal convolution with optional state/bias/SiLU.

---

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