Graph IR To ONNX

Graph IR To ONNX#

MLX Ruby supports native JSON ONNX export for debugging plus native binary ONNX model assembly.

Ownership boundary#

Use MLX::ONNX as the public API:

  • MLX::ONNX.graph_ir_to_onnx

  • MLX::ONNX.graph_ir_to_onnx_json

  • MLX::ONNX.export_onnx

  • MLX::ONNX.export_onnx_json

Implementation modules:

  • MLX::ONNX::Native owns Graph IR/ONNX export and conversion runtime logic.

Generate ONNX JSON from Graph IR payload#

Use graph_ir_to_onnx_json when you already have Graph IR payload/source.

onnx_json = MLX::ONNX.graph_ir_to_onnx_json(
  payload,
  opset: 18,
  model_name: "demo_graph"
)

File.binwrite(
  "artifacts/model_stub.json",
  JSON.pretty_generate(JSON.parse(onnx_json))
)

Generate ONNX JSON directly from trace#

Use export_onnx_json to capture Graph IR + lower to ONNX JSON in one call.

onnx_json = MLX::ONNX.export_onnx_json(
  trace,
  x,
  y,
  opset: 18,
  model_name: "demo_graph"
)

Export binary ONNX#

Use graph_ir_to_onnx when you already have Graph IR payload/source, or export_onnx for direct trace -> ONNX binary export.

MLX::ONNX.graph_ir_to_onnx("artifacts/model.onnx", payload)

MLX::ONNX.export_onnx(
  "artifacts/model.onnx",
  trace,
  x,
  y,
  opset: 18,
  model_name: "demo_graph"
)

Both methods require a path-like target and return the written path.

External data mode#

For large models, enable external initializer data:

MLX::ONNX.graph_ir_to_onnx(
  "artifacts/model.onnx",
  payload,
  external_data: true,
  external_data_size_threshold: 1024,
  external_data_file: "model.data"
)

This writes model.onnx plus model.data in the target directory.

External-data notes:

  • graph_ir_to_onnx and export_onnx require a path-like target (not IO-like).

  • external_data_size_threshold must be a non-negative integer.

  • If external_data_file is omitted, the default is <target_basename>.data.

Next step#

Continue with WebGPU Harness And Smoke.

Contents