:telemetry events emitted by EMLX.

All span-style events use :telemetry.span/3 semantics, so attaching to *:start, *:stop, and *:exception is sufficient for histograms and error tracking.

Events

Evaluation boundaries

[:emlx, :eval, :start | :stop | :exception] — EMLX.eval/1. Spans the full round-trip through the resolved EMLX.CommandQueue worker, including the blocking wait for the worker to finish mlx::core::eval/1 (the real latency, not just NIF dispatch). The :stop event carries :duration (monotonic native units).

[:emlx, :to_binary, :start | :stop | :exception] — EMLX.Backend.to_binary/2 (the Nx.to_binary/1 path). Spans the sync-forcing blob copy. Metadata: :shape, :dtype, :byte_size (byte size of the binary actually returned to the caller).

Memory stats (poll-driven)

[:emlx, :memory, :stats] — discrete event, not a span. Call EMLX.Telemetry.memory_stats/0 to sample; measurements:

• :active_memory — bytes currently allocated and in use
• :peak_memory — highest active memory since the last EMLX.reset_peak_memory/0
• :cache_memory — bytes in the allocator cache (freed but not returned to the OS)

Wire this into a periodic task (e.g. Process.send_after/3 loop) to graph memory drift in a long-running serving.

Attaching a handler

:telemetry.attach(
  "emlx-eval-log",
  [:emlx, :eval, :stop],
  fn _event, measurements, _metadata, _config ->
    IO.inspect(measurements.duration)
  end,
  nil
)