Running ONNX Runtime (Silero VAD) in Cloudflare Workers — complete solution with 0.7ms inference

[CosteGieF]

Running onnxruntime-web + Silero VAD in workerd

After a lot of digging, I got onnxruntime-web v1.24.1 + Silero VAD v5 running natively in a Cloudflare Worker. Avg inference time: 0.7ms. Sharing the full solution here since there was no prior art on this.

---

Why it fails out of the box

Three things block ort in workerd:

1. WebAssembly.compile() is blocked — workerd does not allow dynamic WASM compilation at runtime
2. import.meta.url is empty — ort uses it to resolve the .wasm file path and throws
3. Dynamic import(variable) is rejected — workerd rejects non-string dynamic imports at module analysis time

---

The solution — 3 patches at build time

Patch 1 — Pre-compile WASM via wrangler.toml rules

[[rules]]
type = "CompiledWasm"
globs = ["**/*.wasm"]

[[rules]]
type = "Data"
globs = ["**/*.onnx"]

This gives you a WebAssembly.Module (already compiled at deploy time). Inject it via instantiateWasm before ort initializes:

ortConfig.instantiateWasm = (imports, callback) => {
  const instance = new WebAssembly.Instance(wasmModule, imports);
  callback(instance);
  return instance.exports;
};

Patch 2 — Silence the URL resolution error

ort.env.wasm.wasmPaths = { wasm: "dummy.wasm" };

Patch 3 — Kill dynamic import(variable) in the bundled ort source

Replace all await import(nonStringVariable) patterns with Promise.reject(new Error('dynamic import not supported')) using a regex patch in your build script.

---

Benchmark results

	Time
Session init (one-time / isolate warm-up)	~424ms
First inference	~36ms
Subsequent inferences	~0.7ms avg

---

Resources

• Standalone example repo: https://github.com/CosteGieF/ort-cloudflare-workers — minimal reproduction with full build script
• Gist (annotated build script): https://gist.github.com/CosteGieF/1f98738b3eab2592593ecb2018a98211
• ONNX Runtime discussion: Running onnxruntime-web in Cloudflare Workers via pre-compiled WASM + instantiateWasm callback microsoft/onnxruntime#27400

Hope this saves someone the debugging time!