WebGPU Harness And Smoke

WebGPU Harness And Smoke#

After ONNX export, package browser assets and run smoke verification.

Ownership boundary#

  • Public facade API: MLX::ONNX::WebGPUHarness.export_onnx_webgpu_harness and MLX::ONNX::WebGPUHarness.smoke_test_onnx_webgpu_harness.

  • Implementation module: MLX::ONNX::WebGPUHarness.

Package harness assets#

Use export_onnx_webgpu_harness to emit browser-ready files.

manifest = MLX::ONNX::WebGPUHarness.export_onnx_webgpu_harness(
  "artifacts/web_harness",
  payload,
  model_name: "demo_graph",
  execution_providers: %w[webgpu wasm],
  benchmark_warmup_runs: 1,
  benchmark_measure_runs: 5,
  external_data: false
)

puts manifest.fetch("format") # => "onnx_webgpu_harness_v1"
puts manifest.fetch("execution_providers").inspect # => ["webgpu", "wasm"]

Generated files include:

  • model.onnx

  • harness.manifest.json

  • inputs.example.json

  • index.html

  • harness.js

  • optional external data file (for example model.data)

Manifest fields include:

  • format (onnx_webgpu_harness_v1)

  • model (currently model.onnx)

  • execution_providers (webgpu/wasm)

  • benchmark (warmup_runs and measure_runs)

  • inputs (name/shape/dtype for each model input)

  • optional external_data file list when external-data export is enabled

Run smoke test#

Use smoke_test_onnx_webgpu_harness to validate harness wiring and runtime selection.

telemetry = MLX::ONNX::WebGPUHarness.smoke_test_onnx_webgpu_harness(
  "artifacts/web_harness",
  timeout_seconds: 30,
  mock_ort: false,
  local_ort: true,
  node_bin: "node"
)

puts telemetry.fetch("format") # => "onnx_webgpu_telemetry_v1"
puts telemetry.fetch("selected_provider", "unknown")
puts telemetry.fetch("requested_providers").inspect
puts telemetry.fetch("fallback_used")
puts telemetry.fetch("run_timings_ms").inspect

Notes#

  • Smoke tests require Node.js.

  • Real runtime checks require Playwright + onnxruntime-web in the harness environment.

  • bundle exec rake deps:web installs/checks onnx and the web runtime dependencies used by real-runtime smoke checks.

  • Harness export only accepts webgpu and wasm execution providers.

Contents