@wetron/onnx

ONNX model parser for wetron. Reads .onnx files and returns a ModelGraph IR. Inline initializer bytes use ModelGraph.weights.kind === "available"; initializers with data_location = EXTERNAL use kind === "external" and must be fetched separately.

Install

pnpm add @wetron/onnx

Included automatically when you install @wetron/core or @wetron/react.

API

import { parseOnnx } from '@wetron/onnx';

const bytes = new Uint8Array(await file.arrayBuffer());
const graph = parseOnnx(bytes);

Throws ParseError from @wetron/common/ir on malformed input.

External data loading

For models where initializers use data_location = EXTERNAL, fetch the external files and build a WeightSource:

import { loadOnnxExternalWeightsFromUrl } from '@wetron/onnx';

const weights = await loadOnnxExternalWeightsFromUrl(modelBytes, 'https://.../model-dir');
// weights.get("init_name") -> Uint8Array | undefined

const graphWithWeights = { ...graph, weights: { kind: 'available' as const, source: weights } };

Each unique location filename is fetched once from ${baseUrl}/${location} and shared across initializers that slice it. If any initializer is external, parseOnnx does not expose a partial source for inline initializers. The loader returns an empty WeightSource when the model has no EXTERNAL initializers and throws ParseError on non-ok responses.

What gets parsed

• All nodes with op type, inputs, outputs, and attributes
• Initializer shapes and dtypes (raw weight data is not read)
• Intermediate tensor shapes from value_info
• Opset versions per domain

dtype mapping

| ONNX enum | dtype | | --------- | ---------- | | 1 | float32 | | 2 | uint8 | | 3 | int8 | | 6 | int32 | | 7 | int64 | | 10 | float16 | | 11 | float64 | | 16 | bfloat16 |