WACS.WASI.NN.OnnxRuntime 0.3.2

WACS.WASI.NN.OnnxRuntime

ONNX Runtime backend for WACS.WASI.NN. Implements IBackend for graph-encoding.onnx directly against Microsoft.ML.OnnxRuntime — no ML.NET wrapper, just ORT.

This is the default wasi-nn backend for the WACS CLI. wacs run --wasip2 --wasi-nn auto-loads it; embedders who don't want the ~50 MB of ORT native binaries can use one of the other backends instead.

Install

dotnet add package WACS.WASI.NN.OnnxRuntime

CLI

# Bundled with WACS.Cli — works out of the box.
wacs run my.component.wasm --wasip2 --wasi-nn -d ./models::/models

The Gemma 3 270M ONNX SLM is the canonical end-to-end test target: docs/COMPONENT_CHAINING.md walks through it.

Embedder

Interpreter / one-line:

using Wacs.Core.Runtime;
using Wacs.WASI.NN;
using Wacs.WASI.NN.OnnxRuntime;
using Wacs.WASI.NN.Types;

var runtime = new WasmRuntime();
runtime.UseWasiNN(b => b.AddBackend(GraphEncoding.ONNX, new OnnxBackend()));

Transpiler-direct-link / DI:

services
 .AddWasiPreview2()
 .AddWasiNN(b => b.AddBackend(GraphEncoding.ONNX, new OnnxBackend()))
 .AddWasiPreview2NNBundle();

(WasiPreview2RuntimeScope auto-wires OnnxBackend when this assembly is on the load path — no explicit AddBackend needed.)

What it provides

• OnnxBackend : IBackend — implements LoadGraph(builders, target) for byte-loaded ONNX models. Suitable for the SLM / inference workflow where the guest reads model bytes and passes them through wasi:nn/graph.load
• OnnxBackendOptions / OnnxExecutionProvider — typed config for execution-provider selection (CoreML / CUDA / DirectML / ROCm, with auto-detect + CPU fallback)
• WasiNNOnnxBindable : IBindable — parameterless adapter for --bind. Auto-pulled by the CLI's --wasi-nn shorthand
• [assembly: WasiHostPackage] — picked up by runtime.AutoDiscoverHostPackages()

Hardware acceleration

Default is CPU — hardware acceleration is opt-in via the WACS_WASINN_ONNX_EP env var or OnnxBackendOptions.ExecutionProvider. CPU default avoids the silent op-coverage issues seen with the CoreML / DirectML EPs against generative-LLM ops (e.g., GroupQueryAttention in Gemma 3): partition- and-fallback inside ORT can produce numerically wrong results without raising an error, which manifests as "the LLM doesn't respond" against the Gemma 3 270M SLM workflow. Pin the EP per-model after you've verified your model works with it.

EP-append failure silently falls back to CPU regardless — out-of-box behavior favors "inference still works" over "EP misconfiguration is loud".

Enable via environment:

# Platform-best pick (CoreML on macOS, DirectML on Windows, CUDA on Linux)
WACS_WASINN_ONNX_EP=auto wacs run my.wasm --wasip2 --wasi-nn

# Force a specific provider
WACS_WASINN_ONNX_EP=coreml wacs run my.wasm --wasip2 --wasi-nn
WACS_WASINN_ONNX_EP=cuda WACS_WASINN_ONNX_CUDA_DEVICE=1 wacs run my.wasm --wasip2 --wasi-nn
WACS_WASINN_ONNX_EP=dml wacs run my.wasm --wasip2 --wasi-nn
WACS_WASINN_ONNX_EP=rocm wacs run my.wasm --wasip2 --wasi-nn

# Explicitly stay on CPU (default — no env var also gets you CPU)
WACS_WASINN_ONNX_EP=cpu wacs run my.wasm --wasip2 --wasi-nn

Override via typed config (library embedders):

using Wacs.WASI.NN.OnnxRuntime;
using Microsoft.ML.OnnxRuntime;

var backend = new OnnxBackend(new OnnxBackendOptions
{
 ExecutionProvider = OnnxExecutionProvider.CoreML,
 CoreMLFlags = CoreMLFlags.COREML_FLAG_USE_CPU_AND_GPU
 | CoreMLFlags.COREML_FLAG_ONLY_ENABLE_DEVICE_WITH_ANE,
 FallbackToCpu = true,
});

Full escape hatch (custom SessionOptions factory):

var backend = new OnnxBackend(() =>
{
 var opts = new SessionOptions();
 opts.AppendExecutionProvider_CUDA(deviceId: 0);
 opts.GraphOptimizationLevel = GraphOptimizationLevel.ORT_ENABLE_ALL;
 return opts; // factory wins over OnnxBackendOptions
});

OnnxBackendOptions.FallbackToCpu = false propagates EP-append failures as ErrorCode.RuntimeError at graph.load time — useful for environments where silent CPU fallback would mask a misconfiguration.

Backend choice

Documentation

• docs/WASI_NN_USAGE.md — unified usage guide (CLI flags, env vars, programmatic embedding, worked examples)
• docs/COMPONENT_CHAINING.md
• Wacs.WASI/Wacs.WASI.NN/README.md — backend matrix + package layout

License

Apache-2.0