Worker API

The Worker API is used inside your torch() function. These tools handle state persistence across hot reloads and communication with the UI.

`spark.persist()`

Preserves an object across code changes during development. This is essential for keeping model weights and optimizer states alive while you edit your training logic.

const model = spark.persist('model_key', () => {
  return nn.Sequential(nn.Linear(10, 10));
});

`spark.expose()`

Defines the interface that the React UI can see and interact with.

// These will be available on the proxy returned by spark.use()
spark.expose({
  train,
  state,
  resetModel: () => spark.clearPersisted('model_key')
});

`spark.checkpoint()`

The most important function for cooperative multitasking. It yields execution back to Spark, allowing it to:

Pause the loop if the user clicked "Pause".
Inject new code if you just saved your file (Hot Reload).
Report progress back to the main thread.

async function train() {
  while (true) {
    await spark.checkpoint(); // Yield control;
    runOneIteration();
  }
}

Frequency Matters: Call spark.checkpoint() at least once per training batch. If you call it too rarely, your UI will feel sluggish or unresponsive to control commands.

Next Steps

Client API - How to consume these exposed bindings.
File I/O - Persisting your work permanently.