Skip to content

Parallel ​

Source: core/Parallel.luau

Parallel Luau support (Actors) for RoNetV4.1. Single source of truth for training/eval compute that can run EITHER in the calling thread ("single") OR on a pool of Actor workers ("multi"). The mode is a user-facing toggle: Parallel.setMode("single") -- default; everything in-process Parallel.setMode("multi") -- dispatch jobs to N Actor workers Parallel.setMode("auto") -- multi on a Roblox runtime, else single Grounded in the official Parallel Luau docs / API reference: * Each Actor owns a separate Luau VM. ModuleScripts required inside an Actor are re-required per VM, so every worker gets a private copy of the RoNet graph and module-level state NEVER races across VMs. * require() is illegal in a desynchronized (parallel) phase, so the worker script requires RoNet at its top level (serial phase), keeps the refs, then registers BindToMessageParallel. All math runs inside the parallel handler using already-loaded refs. * Instance mutation is illegal in parallel, so JOBS ARE PURE COMPUTE: they only build Lua tables and run Tensor/Serialize/math.* calls. The disposable MLP a job builds is fully local to the worker VM. * Instances cannot cross VM boundaries, so payloads travel as plain tables via Actor:SendMessage (copied by value), models travel as weight strings (Serialize.dump/load), and results come back through a pool-owned BindableEvent whose instance the worker resolves through the DataModel. * SINGLE/MULTI PARITY IS BY CONSTRUCTION: the same runJob() does the math. multi just slices the batch across workers and re-aggregates. Ops (identical code path in both modes): forward : outputs for a batch of rows loss : scalar loss over the batch (row-count-weighted mean across shards) backward : scalar loss + per-param gradients; the pool writes SUMMED grads into the master model's params so the caller's optimizer:step() works afterwards scoreGenomes : per-genome loss (GA fitness) - the headline op for running an evolutionary population eval in parallel "multi" requires a real Roblox runtime (game + Instance + Actor + task). In the Luau CLI (or anywhere those globals are absent) every mode falls back to "single", so this module always compiles and runs.

Methods ​

Class Parallel ​

available() -> boolean ​

True when running on a real Roblox runtime (game, Instance, Actor and task are all present), i.e. when "multi" mode can actually spawn Actor workers. Always false in the Luau CLI.

setMode(m: string, workers: number?) ​

Set execution mode. Unknown modes error; "multi" without a runtime downgrades to "single" with a warning.

getMode() -> string ​

The configured mode string ("single", "multi" or "auto"), exactly as last set. Use Parallel.effective() for the mode that will actually run.

setWorkers(n: number) ​

Set the worker count (>= 1) used by newPool/attach when no explicit worker count is given.

getWorkers() -> number ​

The configured worker count (>= 1).

setParent(p: any?) ​

Optional parent instance the pool folder is created under.

effective() -> string ​

Effective mode right now ("auto" resolves against the runtime).

weightsString(params: { any }) -> string ​

Portable weights string for a param list (Serialize.dump). Models (and GA genomes) travel across the Actor VM boundary as such strings.

modelFromWeights(layerSizes: { number }, act: string, weightsStr: string) -> any ​

Rebuild a fresh MLP from a portable weights string (used on workers and by the genome op). act must match the model the string came from.

genomeToWeightsString(genome: any) -> string ​

GA genome {layerSizes, weights, biases, act?} -> a Serialize.dump string (genomes therefore cross the VM boundary as ONE string each).

runJob(job: any) -> any ​

THE shared job implementation. Payload fields: op: "forward" | "loss" | "backward" | "genome" layerSizes / act / weightsStr : the single model everyone shares rows : this shard's rows ({ {number} }) targets : {number} or { {number} } slice (loss/backward/genome) lossName : "mse" | "bce" | "crossentropy" (loss/backward/genome) genomes : { string } weight-strings (genome only) Returns a plain-Lua table (numbers/tables only): safe to copy back.

newPool(opts: any?) -> any ​

Build a pool. opts: { workers?, mode?, parent?, modulePath? } modulePath: DataModel path of the requireable RoNet root ModuleScript (default "ReplicatedStorage.RoNet").

Class pool ​

shutdown() ​

Tear the pool down: disconnects the result listener and destroys the worker folder (and its Actor instances).

dispatch(jobs: { any }) -> { any } ​

Run a list of jobs on the pool and block until every result is back. Jobs run in shard order across workers; results are returned in the same order as the jobs argument. Only one dispatch may be in flight at a time. In "single" mode this just runs the jobs in-process.

Class Parallel ​

attach(model: any, opts: any?) -> any ​

One-shot API: attach the op facade to a live model using a default pool. Call handle:forward/loss/backward/scoreGenomes, then :shutdown().