Skip to main content

WebAssembly Deployment

Timber can emit WebAssembly Text Format (WAT) for browser and edge deployment.

Generating WASM Output​

from timber.codegen.wasm import WasmEmitter
from timber.frontends.auto_detect import parse_model

ir = parse_model("model.json")
emitter = WasmEmitter(ir)
files = emitter.emit()

# Write files
for filename, content in files.items():
    with open(filename, "w") as f:
        f.write(content)

This produces:

  • model.wat — WebAssembly Text Format with tree traversal logic
  • timber_model.js — JavaScript bindings with a simple predict() API

Browser Usage​

<!DOCTYPE html>
<html>
<head><title>Timber WASM Inference</title></head>
<body>
  <script src="timber_model.js"></script>
  <script>
    loadTimberModel("model.wat").then(model => {
      const features = [17.99, 10.38, 122.8, /* ... 30 features */];
      const prediction = model.predict(features);
      console.log("Prediction:", prediction);
    });
  </script>
</body>
</html>

Converting to Binary WASM​

The emitter produces WAT (text format). Convert to binary .wasm with:

# Install wabt (WebAssembly Binary Toolkit)
brew install wabt    # macOS
apt install wabt     # Ubuntu

# Convert
wat2wasm model.wat -o model.wasm

How It Works​

The WASM emitter generates:

  1. Linear memory layout — tree data (thresholds, feature indices, children, leaf values) packed into WASM linear memory
  2. $traverse_tree function — iterative tree traversal matching the C99 emitter's logic
  3. $timber_infer_single export — the main inference entry point
  4. Sigmoid activation — for binary classification, implemented via $exp_neg helper

Limitations​

  • Multi-class (softmax) is not yet supported in WASM
  • Binary encoding requires external wat2wasm tool
  • No SIMD instructions yet (future: wasm-simd proposal)