End To End Examples

End To End Examples#

This page shows practical workflows for MLX -> Graph IR -> ONNX -> WebGPU.

Minimal script example#

require "json"
require "stringio"
require "mlx"

mx = MLX::Core
x = mx.array([[1.0, 2.0]], mx.float32)
y = mx.array([[0.5, 0.25]], mx.float32)

trace = ->(lhs, rhs) { MLX::Core.add(lhs, rhs) }
payload = MLX::ONNX.export_graph_ir(trace, x, y)
onnx_json = MLX::ONNX.graph_ir_to_onnx_json(payload, model_name: "minimal_add")

report = JSON.parse(
  MLX::ONNX::Native.ir_compatibility_report_json(payload)
)
abort("unsupported ops: #{report.fetch('unsupported_ops').inspect}") unless report.fetch("unsupported_nodes").zero?

MLX::ONNX.graph_ir_to_onnx("artifacts/model.onnx", payload, model_name: "minimal_add")
MLX::ONNX::WebGPUHarness.export_onnx_webgpu_harness(
  "artifacts/web_harness",
  payload,
  model_name: "minimal_add"
)

telemetry = MLX::ONNX::WebGPUHarness.smoke_test_onnx_webgpu_harness(
  "artifacts/web_harness",
  mock_ort: true
)
puts telemetry.fetch("format")

Repository task examples#

The repository already contains end-to-end export flows:

  • tasks/web_assets_task/export_gpt2_assets.rb

  • tasks/web_assets_task/export_nanogpt_assets.rb

  • tasks/web_assets_task/export_stable_diffusion_assets.rb

Run the web asset pipeline:

bundle exec rake web:assets

Browser demo integration#

Generated assets are consumed by:

  • web/demo/gpt2/

  • web/demo/nanogpt/

  • web/demo/stable_diffusion/

Use the local server task to validate demos after export:

bundle exec rake web:serve

Parity and harness checks#

Useful parity coverage examples:

  • test/onnx/integration/export_onnx_webgpu_harness_test.rb

  • test/onnx/integration/onnx_webgpu_harness_smoke_test.rb

  • test/onnx/integration/onnx_webgpu_harness_real_runtime_smoke_test.rb

Contents