onnx2webnn

ONNX → WebNN lowering crate extracted from webnn-graph. ONNX operators lower directly to rustnn MLGraphBuilder; full-graph validation runs via ORT CPU build() (onnx-runtime feature). There is no JSON IR and no on-disk graph export — success means builder.build() returns Ok(MLGraph).

Supported ONNX opset range: 1–26 (see MIN_SUPPORTED_OPSET / MAX_SUPPORTED_OPSET in src/onnx/convert.rs).

Build

cargo build
# or
cargo build --release

make build, make test, make fmt, and make check are defined in the repo Makefile.

Convert

cargo run -- convert --input model.onnx --optimize --override-dim batch_size=1

Dynamic ONNX inputs (unresolved symbolic dims kept as WebNN dynamic metadata):

cargo run -- convert --input model.onnx `
  --experimental-dynamic-inputs `
  --override-dim batch_size=1 `
  --override-dim sequence_length=1

If model.dims.json sits beside the ONNX file and no overrides were passed on the CLI, dimension bindings are loaded from that sidecar (freeDimensionOverrides or a flat JSON object).

Merged decoders (optimum's decoder_model_merged*.onnx) branch at runtime on use_cache_branch. WebNN has no runtime If, so pin the input and convert each branch separately:

cargo run -- convert --input decoder_model_merged.onnx --optimize `
  --pin-input use_cache_branch=false `
  --override-dim batch_size=1 --override-dim decoder_sequence_length=4 `
  --override-dim past_decoder_sequence_length=0 ...

Pinned inputs become constants, the chosen If branch is inlined, and inputs the branch never reads (e.g. the KV cache in the prefill branch) and zero-size dummy outputs are dropped.

Flag	Purpose
--input	Input .onnx path (required)
--optimize	Constant folding and shape propagation
--override-dim NAME=VALUE	Bind a symbolic dim (repeatable); unnamed zero dims are addressed as <input>_dim<axis>
--override-dims-file	JSON overrides (freeDimensionOverrides or flat object)
--pin-input NAME=VALUE	Freeze a graph input to true/false/an integer (repeatable)
--allow-missing-external-data	Zero-fill external tensors whose data file is absent (weight-stripped skeleton models)
--experimental-dynamic-inputs	Preserve unresolved symbolic dims as dynamic metadata
--debug	Verbose conversion logging (global)

On success the CLI prints ✓ ORT graph build succeeded for ….

Model sweep

tests/models/manifest.json lists the transformers.js exports the converter is expected to handle, with their dimension overrides and pinned inputs; tests/model_skeletons.rs converts each entry and builds it in ORT. No weights are downloaded: the test reads each export from the Hugging Face Hub with HTTP range requests, keeps the graph and small constants, and points every large initializer at a file that does not exist, which the converter zero-fills. A 1.4 GB export becomes a ~0.2 MB skeleton for ~10 MB of traffic. Skeletons are kept in target/model-skeletons (or O2W_SKELETON_CACHE), about 40 MB for the whole manifest, and CI caches that directory keyed on the manifest and scanner source.

O2W_MODELS selects the source: hub (the default when CI is set), dir=<path> for full local downloads laid out as <org>--<repo>/onnx/<file>.onnx, or strip=<path> to run local files through the skeleton scanner. Unset outside CI, the sweep is skipped. O2W_MODEL_FETCH_JOBS (default 8) sets how many skeletons are fetched at once, O2W_MODEL_TEST_JOBS (default 4) how many convert at once, and O2W_MODEL_TEST_SKIP_HEAVY skips the entries that need more than 10 GB of RAM.

$env:O2W_MODELS = "dir=..\transformers_js_experiments\models"; cargo test --release --test model_skeletons

Library API:

use onnx2webnn::{convert_onnx, ConvertOptions};

let graph = convert_onnx("model.onnx", ConvertOptions::default())?;

Layout

Path	Purpose
src/onnx/convert.rs	ONNX load, optional folding, lowering, ORT build()
src/onnx/builder.rs	OnnxBuilder — operand map and MLGraphBuilder bridge
src/onnx/builder_helpers.rs	Shared lowering helpers
src/onnx/shape_inference.rs	Static shape/type propagation
src/onnx/constant_folding.rs	Constant folding driver (with --optimize)
src/onnx/constant_folding/evaluators/	Per-op fold evaluators
src/onnx/ops/	ONNX op handlers (activation, conv, pool, reshape, …)
src/protos.rs	ONNX protobuf types
src/debug.rs	Debug logging toggle

Dependencies

• rustnn (../rustnn, onnx-runtime) — MLGraphBuilder, shape inference, ORT build() validation
• webnn-onnx-utils — ONNX protos, op names, data types

Related

• webnn-graph — DSL parser, validator, JS/HTML emit (source of the extracted lowering code)