:telemetry events emitted by Emily.

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

[:emily, :eval, :start | :stop | :exception] — Emily.eval/1. The :stop event carries :duration (monotonic native units).

[:emily, :to_binary, :start | :stop | :exception] — fires for both Emily.to_binary/1 and the Nx.to_binary/1 path on Emily.Backend. Metadata: :byte_size, :shape, :dtype.

Fallback entry

[:emily, :fallback, :start | :stop | :exception] — fires whenever an op routes through Nx.BinaryBackend because the MLX path is not wired. Metadata: :op, :input_shapes, :input_dtypes.

Per-fallback behaviour is configurable via :fallback:

config :emily, fallback: :silent | :warn | :raise

• :silent (default) — span events still fire, no log, no raise. Library consumers and CI logs stay quiet.
• :warn — one-shot Logger.warning per {op, input_shapes} pair. A Bumblebee user sees "indexed_put on shape [...] fell back to Nx.BinaryBackend" once per shape, not every forward pass. Typically set in config/dev.exs while chasing performance regressions.
• :raise — raises a RuntimeError carrying the op, input shapes, and input dtypes. Use in CI to fail builds when a hot path unexpectedly routes through Nx.BinaryBackend.

In :raise mode the :start/:stop span events do not fire because the raise happens on entry; :silent and :warn preserve the full span.

The legacy :warn_on_fallback boolean is still honoured when :fallback is unset (true → :warn, false → :silent). Prefer :fallback in new code; if both are set, :fallback wins.

Memory stats (poll-driven)

[:emily, :memory, :stats] — discrete event, not a span. Call Emily.Memory.stats/0 to sample; measurements:

• :active — bytes currently held by MLX
• :peak — high-water mark since last Emily.Memory.reset_peak/0
• :cache — bytes cached for reuse

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(
  "emily-fallback-log",
  [:emily, :fallback, :stop],
  fn _event, measurements, metadata, _config ->
    IO.inspect({metadata.op, measurements.duration})
  end,
  nil
)