Effects And Handlers

Chelis tracks effects explicitly. Pure tensor code stays pure, and the places that touch randomness, host input and output, or a specific device are surfaced in the type and discharged by a handler. Effect inference runs after type checking: a function's effect set is the union of the effects of the operations in its body.

The effects

• Random comes from dropout and uniform_like, and from the Std.Init initializers that call them. It is discharged by with seed(...).
• IO is inferred from host operations such as print and the file builtins. It is permitted at the top level rather than requiring a handler.
• Resource("device") marks a region that runs on a named device. It is introduced by with device(...) and validated against the build target.

Annotating effects

A signature or a def carries its effect set as a ! { ... } suffix. The annotation is optional; the checker infers the set and verifies any annotation you supply.

sig predict: tensor[n, f32] -> tensor[n, f32] ! { Random }

In Deep the effect set is eff metadata on the function type:

(t-fn {eff: (effects {} random)}
  (t-tensor {} (d-name {} n) (t-prim {} f32))
  (t-tensor {} (d-name {} n) (t-prim {} f32)))

Handlers

A handler is a with block. with seed(...) takes an integer literal and makes the randomness inside it deterministic; an unhandled Random effect at the top level is a check error with repair guidance. with device(...) takes a string literal naming the device.

with seed(42) {
  dropout(x, 0.5)
}

Handlers nest. A region can sit on a device and seed its randomness at once:

with device("gpu:0") {
  with seed(42) {
    dropout(x, 0.5)
  }
}

The handled names are seed and device. For the effect-checking details see spec/04-type-system.md and the effect-checking crates and tests.