[WebGPU EP] grouped conv audio underflow fix

[Prathik Rao (prathikr)]

Fix padded WebGPU grouped convolution coordinates

Issue

Gemma 4 audio processing could abort on macOS while Dawn compiled the WebGPU GroupedConv shader:

std::out_of_range: absl::container_internal::raw_hash_map<>::at

The generated shader computed padded convolution coordinates as u32. When an output position required a negative input coordinate, subtracting the padding underflowed to a large unsigned value. The subsequent x < 0u check could never succeed. This affected the padded grouped-convolution path used by the audio frontend and could trigger a Metal/Dawn shader-compilation failure.

Fix

GroupedConv now:

• Computes convolution coordinates as signed i32 values.
• Performs bounds checks before indexing tensor data.
• Converts coordinates to u32 only after they are known to be in range.

Regression test

Added a WebGPU test using a two-group, 1D-style convolution represented as a padded 2D convolution. The test exercises negative left-border coordinates and verifies the expected zero-padding behavior.

Validation:

Conv_WebGPU.GroupedConvWithPaddingUsesSignedCoordinates: PASSED

[github-actions]

You can commit the suggested changes from lintrunner.

[Copilot]

🟢 Approval recommended

The signed-coordinate fix is correctly bounded and covered by a focused regression test.

Pull request overview

Fixes WebGPU grouped convolution padding underflow during shader generation.

Changes:

• Uses signed WGSL coordinates for padded convolution bounds checks.
• Converts validated coordinates to unsigned indices.
• Adds a grouped-convolution padding regression test.

No actionable findings identified.

File summaries

File	Description
onnxruntime/core/providers/webgpu/nn/grouped_conv.cc	Corrects padded coordinate arithmetic and indexing.
onnxruntime/test/providers/webgpu/grouped_conv_padding_test.cc	Verifies grouped convolution zero-padding behavior.

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Balanced

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.