Emily

Elixir bindings and Nx backend for Apple's MLX.

Overview

Emily runs Nx computations on Apple Silicon through MLX. Installing it as the default Nx backend is enough to get Bumblebee models executing on the Metal GPU with no further integration work — DistilBERT, Qwen3, ViT, and Whisper all run against pinned reference outputs in the conformance suite today.

The library is structured as four thin layers, each independently testable against its own oracle:

Emily.Compiler    (Nx.Defn.Compiler) — validates opts, pins the result backend

Emily.Backend     (Nx.Backend)       — op-by-op translation to Native

Emily.Native      (NIF shim)         — one function per MLX op, no policy

MLX C++                              — statically linked into libemily; mlx.metallib alongside

Installation

Add :emily to your mix.exs deps:

def deps do

  [

    {:emily, "~> 0.7"}

  ]

end

On first mix compile Emily downloads the precompiled NIF for your OS/arch/variant (libemily.{so,dylib} + mlx.metallib) from this repo's GitHub releases into $EMILY_CACHE (default $(getconf DARWIN_USER_CACHE_DIR)emily/ on macOS, ${XDG_CACHE_HOME:-~/.cache}/emily/ on Linux) and drops it into priv/. No cmake, Xcode, or C++ toolchain is required on the consumer side — nothing is compiled locally. See Building for details.

Run mix emily.doctor after compiling to verify the host platform, configured MLX variant, native artifacts, NIF loadability, and a tiny Emily backend smoke test.

Features

Prerequisites

No Xcode, Metal toolchain, cmake, or C++ compiler is required on the consumer side — Emily downloads a precompiled NIF from GitHub Releases on first mix compile.

Building

As a hex consumer

Add {:emily, "~> 0.7"} to mix.exs, then:

mix deps.get

mix compile

On a cold build Emily downloads the matching precompiled tarball (emily-nif-<version>-<variant>-<target>.tar.gz) from this repo's GitHub release for the pinned version, verifies its SHA256 against the .sha256 sidecar fetched alongside it (no checksums baked into mix.exs — the sidecar is the source of truth for the published asset), and extracts libemily.{so,dylib} + mlx.metallib into priv/. Subsequent builds reuse the cached tarball under $EMILY_CACHE (default $(getconf DARWIN_USER_CACHE_DIR)emily/ on macOS, ${XDG_CACHE_HOME:-~/.cache}/emily/ on Linux). The sidecar is re-fetched on every compile and the cached tarball is re-verified against it, so a republish with a new checksum invalidates a stale cache automatically.

Override the cache location with EMILY_CACHE=/some/path mix compile.

From source (contributors)

git clone https://github.com/ausimian/emily.git

cd emily

mix deps.get    # also clones ml-explore/mlx into deps/mlx_src at the pinned tag

mix compile

The in-repo checkout keeps c_src/ on disk, so mix compile takes the source-build path: scripts/build-mlx.sh cmake-builds libmlx.a + mlx.metallib out of deps/mlx_src/ into $EMILY_CACHE/mlx-<version>-<variant>/, then elixir_make links the NIF against it. Xcode + the Metal toolchain are required.

Force an MLX rebuild with mix compile.emily_mlx --force after editing scripts/build-mlx.sh or bumping @mlx_version.

MLX JIT (optional)

MLX can ship its Metal kernels either AOT-compiled into mlx.metallib, or as source strings that are JIT-compiled on first use. Emily defaults to the AOT variant. To switch, add to config/config.exs:

config :emily, variant: :jit

This changes which prebuilt tarball is downloaded; both variants can coexist under $EMILY_CACHE.

Artefact sizes on an M-series Mac (release optimisations):

Modelibemily.somlx.metallibpriv/ total
:aot (default)~20 MB~154 MB~175 MB
:jit~22 MB~3.5 MB~25 MB

With JIT on, kernels are compiled on first invocation, so there's a small per-kernel warm-up cost at runtime; subsequent calls are cached in-process. All of Emily's test suite passes under both variants.

The JIT prebuilt is built against the macOS 26.2+ SDK (MLX's NAX kernel sources transitively include <MetalPerformancePrimitives/MetalPerformancePrimitives.h>, which only ships in that SDK) and references half-float intrinsics such as __fmaxf16 that aren't present in older macOS libSystem. It therefore also requires macOS 26.2+ at runtime; older macOS hosts should stick with the default :aot variant, whose prebuilt is built against macOS 14 and runs anywhere.

Usage

Install Emily as the global Nx backend and use Nx normally:

Nx.global_default_backend({Emily.Backend, device: :gpu})



Nx.tensor([[1.0, 2.0], [3.0, 4.0]])

|> Nx.dot(Nx.tensor([[5.0], [6.0]]))

|> Nx.to_flat_list()

# => [17.0, 39.0]

Use Emily.Compiler for defn / Nx.Serving:

Nx.Defn.global_default_options(compiler: Emily.Compiler)

Bumblebee inference works with no further configuration once the backend is installed — see the conformance suites under test/emily/conformance/ for worked DistilBERT, Qwen3, ViT, and Whisper pipelines, and the Livebooks section of the HexDocs nav for runnable Livebooks.

The low-level tensor API (Emily.from_binary/3, to_binary/1, shape/1, dtype/1, eval/1) remains available for diagnostics and direct MLX round-trips, but most users should go through Nx.

Native compilation

Emily.Compiler has three modes. native is on by default; opt into a different mode per-call, globally via Nx.Defn.global_default_options/1, or app-wide with config :emily, native: false | true (a per-call native: option always overrides the app-env default):

# Default: native single-NIF replay. Lowers the traced Nx.Defn.Expr to

# a flat IR and replays the whole forward graph in one NIF call per

# invocation — ~5× decode throughput on Qwen3-0.6B, bit-identical to

# the evaluator. Anything the IR can't lower routes through

# Nx.Defn.Evaluator under the default `native_fallback: :eval` (with a

# `[:emily, :compiler, :fallback]` telemetry event).

Nx.Defn.jit(&forward/1, compiler: Emily.Compiler).(input)



# Op-by-op dispatch through Emily.Backend — opt out of native with

# `native: false`. Bit-identical to the Nx evaluator, with no

# whole-graph compile step, so no one-shot compile-time memory spike;

# use it as a numerics reference or on memory-constrained hosts.

Nx.Defn.jit(&forward/1, compiler: Emily.Compiler, native: false).(input)



# Native + mx::compile kernel fusion. Fuses elementwise runs the plain

# replay leaves as separate kernels (RMSNorm / softmax / SiLU gating /

# residual adds) — ~1.1× over the plain native lane on Qwen3-0.6B

# greedy decode (~5.4× the evaluator overall). For a `defn while`,

# the loop body is fused under mx::compile and cached per stream so

# it cache-hits across iterations.

Nx.Defn.jit(&forward/1, compiler: Emily.Compiler, native: true, fuse: true).(input)

Trade-off for :fuse: mx::compile reassociates f32, so logits drift by a few ULP and the output is not bit-identical to the evaluator. Greedy argmax is empirically robust to that drift (Qwen3-0.6B token ids matched the evaluator's exactly in our benchmarks), but sampling strategies will diverge from the evaluator even with a fixed seed.

Choosing a mode.native is the default and the right one for model inference: it's bit-identical to the evaluator, safe globally (un-lowerable ops route through the evaluator), and the single biggest win — eager → native is the largest jump in every benchmark tier. Opt out with native: false only when you need the op-by-op evaluator — a numerics reference, or a memory-constrained host where the one-shot compile peak is too large. fuse: true is not a universal add-on. It pays off only when the fused callable is reused, which in practice means autoregressive decode: the defn while body is mx::compiled once and cache-hits every step (the best lane on Qwen3-0.6B decode, 1.67× EXLA). On a one-shot forward pass — a single encoder/classifier/vision call — there's nothing to amortize the mx::compile build cost against, so fuse is neutral-to-slightly negative (cf. ViT-base and Whisper-tiny in §5–6 of the benchmark). Rule of thumb: fuse for decode loops, plain native for one-shot forwards.

Pass native_fallback: :raise to fail rather than silently degrade to the Evaluator — the conformance suites use this to prove a model lowers fully native. See the Emily.Compiler moduledoc for the full option list (:device, :hooks, :max_concurrency, etc.) and the trade-offs around :fuse in depth.

For autoregressive decode specifically, Emily.Generation is a model-agnostic loop driver that JIT-compiles a caller-supplied shape-stable per-token forward and runs the loop from Elixir with KV-cache threading, stop conditions, and per-token streaming.

Performance

Emily targets GPU-friendly model inference on Apple Silicon. The Emily-vs-EXLA benchmark compares Emily (MLX, Metal GPU) against EXLA — which on macOS arm64 ships no GPU client and runs on the CPU. So the practical choice most Elixir-on-Apple-Silicon users face is GPU-via-Emily vs CPU-via-EXLA, and the two have opposite cost structures: the GPU has a higher fixed per-op latency floor (~160–280 µs — a BEAM↔worker hop, a Metal command-buffer commit, and a GPU sync) but far more throughput at scale; the CPU has a low (~80–110 µs) floor but caps out sooner. The crossover is per-op tensor size, not model kind:

Workload (M4 Pro, f32)Best Emily lane vs EXLA-CPU
Large matmul (2048²)5.0× faster
ViT-base image classification2.2× faster
Qwen3-0.6B greedy decode1.67× faster (fuse)
DistilBERT QA (one encoder forward)~parity (1.06×)
Whisper-tiny transcription11× slower
Elementwise / matmul ≤ ~512 per dimup to ~2.3× slower

Rule of thumb: reach for Emily when the per-op tensors are large — hidden dim ≥ 768 and matmuls ≥ 1024 per dimension, which covers Qwen3, ViT, and most modern transformers at real sizes. Small models built from many tiny kernels (Whisper-tiny, hidden dim 384) or workloads dominated by small tensors can be slower than EXLA-CPU, because the GPU's per-kernel launch latency dominates and its parallelism goes unused. Notably, every model in the benchmark lowered fully native with zero fallbacks, so these gaps are kernel/dispatch efficiency, not coverage holes.

For decode, use the native compiler rather than the eager backend: eager Qwen3 decode is 3.5× slower than EXLA where native is 1.67× faster — a 5.8× swing that is purely the per-op dispatch floor, paid once per tiny op across thousands of decode steps.

Re-run the benchmark with elixir bench/emily_vs_exla.exs; the full write-up is in bench/emily_vs_exla_report.md and the raw numbers in bench/emily_vs_exla_results.md.

Livebooks

End-to-end Livebooks under livebooks/. Each one declares its own Mix.install/2 block and pins Emily.Backend as the default Nx backend, so they're self-contained — open in Livebook and run.

Requires Apple Silicon + macOS 14+ (MLX / Metal) — these notebooks won't run on Linux or Intel hosts, including the hosted Livebooks a "Run in Livebook" badge would open. To get a copy, open its page in the Livebooks section of these docs and use the View Source link (top-right) to view or download the .livemd from GitHub, then run it in Livebook on a Mac.

Concurrency model

MLX dispatches GPU work through Metal command queues. Emily owns one worker thread per command queue; each worker is a dedicated OS thread that runs the MLX ops on behalf of BEAM processes. NIFs return immediately after enqueueing their work on a worker: the worker runs the op, then posts {ref, {:ok, result}} back to the caller via enif_send, and the caller's public wrapper awaits that message with a plain receive. No BEAM scheduler (regular or dirty) blocks on MLX work — callers see the same synchronous semantics as before, but the scheduler is free to run other processes while the GPU is busy.

Because the MLX stream is pinned to its worker thread, MLX's per-thread CommandEncoder state stays consistent regardless of how the BEAM migrates Elixir processes between schedulers.

By default, every op uses the default worker owned by the Emily.MlxStream.Default GenServer under the application supervisor. That single queue serialises all GPU work across the VM — correct and simple, but a bottleneck under concurrent inference.

Stream-per-worker, shared weights

The recommended pattern for concurrent inference: load the model once, create a pool of streams at boot, and route each request to a worker that owns one of those streams. Weights live in one MLX buffer that every stream reads; the only per-stream cost is the Metal command buffer.

# 1. Load weights once, at application start.

{:ok, model}     = Bumblebee.load_model({:hf, "Qwen/Qwen3-0.6B"})

{:ok, tokenizer} = Bumblebee.load_tokenizer({:hf, "Qwen/Qwen3-0.6B"})

{:ok, config}    = Bumblebee.load_generation_config({:hf, "Qwen/Qwen3-0.6B"})



serving =

  Bumblebee.Text.generation(model, tokenizer, config,

    defn_options: [compiler: Emily.Compiler]

  )



# 2. Start N workers; each owns one Emily.Stream for its lifetime.

defmodule MyApp.StreamWorker do

  use GenServer



  def start_link({id, serving}),

    do: GenServer.start_link(__MODULE__, serving, name: via(id))



  def run(id, input), do: GenServer.call(via(id), {:run, input}, :infinity)



  @impl true

  def init(serving) do

    {:ok, %{serving: serving, stream: Emily.Stream.new(:gpu)}}

  end



  @impl true

  def handle_call({:run, input}, _from, %{stream: s, serving: sv} = state) do

    result = Emily.Stream.with_stream(s, fn -> Nx.Serving.run(sv, input) end)

    {:reply, result, state}

  end



  defp via(id), do: {:via, Registry, {MyApp.StreamRegistry, id}}

end



# 3. Dispatch each request to any free worker (round-robin, poolboy,

#    a Registry lookup, etc.). Calls to different workers run

#    concurrently on distinct Metal command queues.

MyApp.StreamWorker.run(pick_worker(), "The quick brown fox…")

Create streams once at worker init, not per-request — Emily.Stream.new/1 spawns an OS thread.

Stream lifecycle.Emily.Stream has no explicit release API; cleanup piggybacks on BEAM GC of the NIF resource held in the struct. In the pattern above, the stream lives as long as its owning worker process: when the worker terminates (crash, supervisor shutdown, or GenServer.stop/1), the process heap is reclaimed, the resource's refcount drops to zero, and the NIF destructor joins the dedicated OS thread. A supervised restart therefore drops the old stream and allocates a fresh one in the child's init/1. To drop a stream deliberately, terminate the process that owns it.

Pooled servings — K weight copies, one default queue

For small models where duplicating weights is cheap, start K Nx.Serving instances behind poolboy / Registry / etc. Each instance holds its own copy of the weights. No Emily.Stream is involved, so every instance dispatches onto the same default Metal command queue — requests run sequentially at the GPU even though multiple servings exist at the BEAM level. The pool buys parallelism for CPU-side serving work (pre/post-processing, batching) but not for GPU-side compute. Memory scales linearly with K.

Combine the two if you need both: K servings for CPU parallelism, each with its own Emily.Stream for GPU parallelism.

Using Emily with Nx.Serving

Nx.Serving itself is stream-agnostic — it calls into Emily.Compiler which dispatches to whatever MLX stream is installed in the calling process. That gives three viable configurations:

ConfigurationWeights in GPU memoryGPU queuesWhen to use
Single serving, default stream1×1 (shared)Default. Simplest; fine for single-user / batched workloads.
Single serving + pool of Emily.Streams1×N (per ws)Concurrent inference on a shared model. Large models.
K servings (pooled), default streamK×1 (shared)Small models where CPU serving work dominates GPU compute.

In every case Nx.Serving.run/2 / Nx.Serving.batched_run/2 is the caller-facing API; the only difference is whether the calling process wraps the call in Emily.Stream.with_stream/2 and whether you run one serving or many.

See Emily.Stream for the API and the qwen3_quantized livebook under Livebooks for a worked multi-stream example.

Observability

Emily emits :telemetry events at the evaluation boundary ([:emily, :eval, *], [:emily, :to_binary, *]) and at every Nx.BinaryBackend fallback ([:emily, :fallback, *]). Attach a handler to graph hotspots or detect silent performance cliffs — see Emily.Telemetry for the full event catalogue.

When a backend callback has no native MLX path, Emily transparently falls back to Nx.BinaryBackend. The fallback is ~100× slower; configure per-fallback behaviour with :fallback:

# config/dev.exs — one-shot Logger.warning per {op, shapes} pair

config :emily, fallback: :warn



# config/test.exs (CI) — fail loud if a hot path goes via BinaryBackend

config :emily, fallback: :raise

The default is :silent, so library consumers and CI logs stay quiet unless they opt in. The [:emily, :fallback, *] telemetry events fire regardless in :silent/:warn mode; :raise raises on entry and skips the span.

The legacy config :emily, :warn_on_fallback, true boolean is still honoured when :fallback is unset (true → :warn). Prefer :fallback in new code.

Memory

MLX buffers live outside the BEAM heap, so long-running serving and training processes should observe MLX allocator state directly. Emily.Memory.stats/0 returns active, peak, and cached bytes and emits the same [:emily, :memory, :stats] telemetry event:

stats = Emily.Memory.stats()

# %{active: ..., peak: ..., cache: ...}

Use Emily.Memory.reset_peak/0 before a benchmark or soak window, and Emily.Memory.clear_cache/0 when you want MLX to release reusable cached buffers. clear_cache/0 does not free live tensors or binaries returned by Nx.to_binary/1; those buffers are released only after the owning BEAM references are garbage collected.

Debug assertions

Two compile-time flags re-enable runtime checks that MLX (and every other GPU backend) skips by default. Both are off by default with zero runtime cost when off — the guarded branches are dead-code eliminated by the Elixir compiler.

# config/dev.exs

config :emily,

  debug_bounds_check: true,

  debug_detect_nan_inf: true

Each check is a per-op MLX reduction plus a scalar readback — a worker sync that breaks lazy-graph fusion. Leave off in release builds. See the Emily moduledoc for the full opt-in snippet.

Testing

mix test                           # fast suite (unit + property)

mix test --only conformance        # + Bumblebee tiny-random suites

mix test --only qwen3_full         # full Qwen3-0.6B checkpoint (~1.5 GB)

mix test --only qwen3_quant_full   # quantized Qwen3-0.6B end-to-end

mix test --only vit_full           # full ViT-base (~330 MB)

mix test --only whisper_full       # full whisper-tiny (~150 MB)

mix test --only training_full      # MNIST convergence canary

mix test --only soak               # memory + concurrency soak harnesses

Each layer has its own oracle: hand-computed expected values at the Native layer, Nx.BinaryBackend on the same inputs at the Backend layer, Emily.Backend in non-defn mode at the Compiler layer, and HuggingFace Transformers reference slices end-to-end. A bug can only be introduced in the layer where its test fails.

Documentation

Acknowledgements

Emily stands on the shoulders of two projects without which it would not exist:

Thanks also to Cocoa Xu, whose MLX build scripts and prebuilt-binary tooling underpin much of the macOS NIF supply chain in the Elixir ML ecosystem.

License

MIT