Tinkex.Regularizer.Executor (Tinkex v0.3.4)

View Source

Manages regularizer execution with process-based parallelism.

This module handles both synchronous and async regularizers, with optional parallel execution using Elixir Tasks.

Execution Modes

Sequential (:parallel => false)

Executes regularizers one at a time in the order specified. Useful for debugging and deterministic behavior.

Parallel (:parallel => true, default)

Executes regularizers concurrently using Task.async_stream/3. Provides better throughput for CPU-bound regularizers.

Async Regularizers

Regularizers with async: true in their spec should return a Task.t() instead of the result directly. The executor will Task.await/2 the result.

Telemetry

Emits the following events:

  • [:tinkex, :regularizer, :compute, :start] - Before each regularizer
  • [:tinkex, :regularizer, :compute, :stop] - After each regularizer
  • [:tinkex, :regularizer, :compute, :exception] - On failure

Summary

Functions

execute_all(regularizers, data, logprobs, opts)

Execute all regularizers and collect outputs.

execute_one(spec, data, logprobs, opts)

Execute a single regularizer and return output.

Functions

execute_all(regularizers, data, logprobs, opts)

@spec execute_all(
  [Tinkex.Types.RegularizerSpec.t()],
  [Tinkex.Types.Datum.t()],
  Nx.Tensor.t(),
  keyword()
) :: {:ok, [Tinkex.Types.RegularizerOutput.t()]} | {:error, term()}

Execute all regularizers and collect outputs.

Options

  • :parallel - Run in parallel (default: true)
  • :timeout - Execution timeout in ms (default: 30_000)
  • :track_grad_norms - Compute gradient norms (default: false)
  • :max_concurrency - Max parallel tasks (default: schedulers_online)

Returns

  • {:ok, list(RegularizerOutput.t())} on success
  • {:error, term()} on failure

Examples

{:ok, outputs} = Executor.execute_all(regularizers, data, logprobs,
  parallel: true,
  timeout: 60_000,
  track_grad_norms: true
)

execute_one(spec, data, logprobs, opts)

@spec execute_one(
  Tinkex.Types.RegularizerSpec.t(),
  [Tinkex.Types.Datum.t()],
  Nx.Tensor.t(),
  keyword()
) :: {:ok, Tinkex.Types.RegularizerOutput.t()} | {:error, term()}

Execute a single regularizer and return output.

Options

  • :timeout - Timeout for async regularizers (default: 30_000)
  • :track_grad_norms - Compute gradient norm (default: false)

Returns

  • {:ok, RegularizerOutput.t()} on success
  • {:error, {:regularizer_failed, name, exception}} on failure
  • {:error, {:regularizer_exit, name, reason}} on exit