Skip to content

webgpu-op-support.md

Latest commit

 

History

History
34 lines (27 loc) · 1.46 KB

webgpu-op-support.md

File metadata and controls

34 lines (27 loc) · 1.46 KB
Raw

Operator Support

The current source of truth for WebGPU operator support is the runtime registry under backends/webgpu/runtime/ops. Operator tests are under backends/webgpu/test/ops, and the generated native test catalog is defined in backends/webgpu/test/op_tests/cases.py.

Compatibility Considerations

  • WebGPUPartitioner is a thin wrapper over VulkanPartitioner; it preserves Vulkan partitioning behavior and does not add WebGPU capability policy.
  • Support can depend on tensor rank, shape, layout, and data type in addition to the operator name.
  • FP16 shader paths require a WebGPU adapter exposing shader-f16.
  • Dynamic shapes allocate buffers at the exported maximum shape. Operators must also implement runtime resize propagation.
  • Quantized and LLM-specific operators generally require the corresponding Vulkan custom-op export pattern.

Use the native operator-test framework to generate fixtures for specific operators. For example, generate the verified add cases with:

python -m executorch.backends.webgpu.test.op_tests.generate_op_tests \
    --output /tmp/webgpu_op_tests --ops add

The full build-and-run flow is available through:

bash backends/webgpu/test/test_build_webgpu.sh