M3: Bumblebee DistilBERT end-to-end on Emily.Backend

Every Nx op on the transformer critical path now runs natively on MLX;
a full DistilBERT forward pass matches HuggingFace Transformers (PyTorch)
reference values within f32 tolerance across six architecture variants.

Emily.Backend changes
  - Native batched dot/7 via permute + 3-D reshape + Native.matmul,
    replacing the BinaryBackend bounce. Hits 12x per DistilBERT forward
    (2x per attention block x 6 layers).
  - Cast binary-op operands to out.type at the backend boundary. MLX's
    cross-type promotion for mixed integer widths (e.g. u64+s32) falls
    back to float32 and then rejects bitwise ops; Nx.Random.key hits
    this transitively via Axon's dropout defn.
  - slice starts can now be scalar tensors (dynamic slicing under defn);
    materialised to concrete ints on the fly.
  - bitcast implemented via mx::view (zero-copy reinterpret between
    equal-width dtypes). Removes the hard raise; unblocks Nx.Random.
  - argmax/argmin derive keep-axis from out.shape vs input rank so
    behaviour is stable across Nx's :keep_axis / :keep_axes drift.

Conformance suite
  - test/emily/conformance/distilbert_test.exs ports Bumblebee's own
    DistilBERT tests verbatim: :base, :for_masked_language_modeling,
    :for_sequence_classification, :for_token_classification,
    :for_question_answering, :for_multiple_choice, plus an
    Nx.Serving.batched_run smoke test exercising tokenizer -> model ->
    postprocess end to end.
  - @moduletag :conformance, excluded by default; run explicitly with
    mix test --only conformance. CI runs it as a separate step after
    mix precommit, with ~/Library/Caches/bumblebee cached across runs.
  - Batched-dot property tests added to backend_test.exs covering 1-
    and 2-axis batch cases plus scalar-output and multi-free-axis
    shapes.

Deps
  - bumblebee ~> 0.6 and tokenizers ~> 0.5 added as test-only deps.
  - nx pinned ~> 0.10 (down from 0.11) to match Bumblebee's constraint;
    Emily's own API is unaffected.

Out of scope for this milestone: native conv translation. PLAN.md lists
it under M3 but DistilBERT and M4's Qwen3 don't use it; the
BinaryBackend fallback remains until a CV model lands on Emily.

Author: ausimian

.github/workflows/ci.yml

@@ -50,6 +50,24 @@ jobs:
           key: build-${{ runner.os }}-${{ hashFiles('mix.lock', 'c_src/**', 'Makefile', '.tool-versions') }}
           restore-keys: build-${{ runner.os }}-
 
+      # The Bumblebee cache holds tiny-random HuggingFace fixtures
+      # downloaded by the conformance suite (~3 MB across 7 repos).
+      # Cached by conformance test content so new fixtures invalidate;
+      # restore-keys lets unrelated changes reuse the existing cache.
+      - name: Cache Bumblebee fixtures
+        uses: actions/cache@v4
+        with:
+          path: ~/Library/Caches/bumblebee
+          key: bumblebee-${{ runner.os }}-${{ hashFiles('test/emily/conformance/**') }}
+          restore-keys: bumblebee-${{ runner.os }}-
+
       - run: mix deps.get
 
       - run: mix precommit
+
+      # Conformance tests are excluded from the default suite because
+      # they require network access on a cold cache (see
+      # test/test_helper.exs). In CI we always want them green — a
+      # DistilBERT forward pass is the canonical integration signal
+      # that no Nx op on the transformer critical path has regressed.
+      - run: mix test --only conformance

RELEASE.md

@@ -110,6 +110,58 @@
   last-axis fast path stays on MLX; interior-axis usage is rare on
   our M3/M4 critical path (transformer inference doesn't need it).
 
+- M3 — DistilBERT end-to-end on Bumblebee. Every Nx op on the
+  transformer critical path now runs natively on MLX; the full
+  forward pass matches HuggingFace Transformers (PyTorch) reference
+  values within f32 tolerance.
+  - **Native batched `dot/7`** in `Emily.Backend`, replacing the
+    BinaryBackend bounce. Permutes operands to
+    `[batch… , free… , contract…]`/`[batch…, contract…, free…]`,
+    collapses to 3-D, dispatches to `Native.matmul` (which treats
+    leading dims as batch), reshapes to Nx's canonical
+    `batch ++ free_a ++ free_b` layout. Hits 12× per DistilBERT
+    forward pass (2× per attention layer × 6 layers). Falls back to
+    BinaryBackend for non-float dtypes — MLX matmul is float-only.
+  - **Binary op type promotion** fixed at the Backend boundary.
+    MLX's cross-type promotion for mixed integer widths (e.g.
+    `right_shift(u64, s32)`) falls to float32 and then rejects the
+    op. `Emily.Backend` now casts both operands to the Nx-computed
+    output type (for arithmetic/bitwise) or merged input type (for
+    compare/logical) before dispatching to MLX. Unblocks
+    `Nx.Random.key`, which is pulled in transitively even in
+    inference-only models via Axon's dropout defn.
+  - **Dynamic `slice` starts.** Nx passes scalar-tensor starts under
+    `defn` evaluation; `Emily.Backend.slice` now materialises them
+    to their concrete values on the fly.
+  - **`bitcast`** implemented via `mx::view` (zero-copy reinterpret
+    cast between equal-width dtypes). Required by `Nx.Random` to
+    move between f32 and u32 bit patterns.
+  - **`argmax`/`argmin` keep-axis robustness.** Derive the keep-axis
+    flag from `out.shape` vs input rank instead of trusting the raw
+    opts key (Nx's user-facing API uses `:keep_axis`, singular, while
+    some callers pass `:keep_axes`).
+  - **`test/emily/conformance/distilbert_test.exs`**
+    (`@moduletag :conformance`, excluded by default; run with
+    `mix test --only conformance`) — ports Bumblebee's own DistilBERT
+    tests verbatim. Six architecture variants (`:base`,
+    `:for_masked_language_modeling`,
+    `:for_sequence_classification`, `:for_token_classification`,
+    `:for_question_answering`, `:for_multiple_choice`) plus an
+    `Nx.Serving.batched_run` smoke test exercising the QA pipeline
+    end-to-end (tokenizer → model → postprocess).
+  - **CI runs the conformance suite** on every push/PR as a separate
+    step after `mix precommit`. `~/Library/Caches/bumblebee` is
+    cached across runs so the ~3 MB HF fixture download happens
+    once. Local `mix test` remains opt-in via `--only conformance`
+    so a fresh-clone/offline contributor isn't blocked by network.
+  - **Batched-dot property tests** added to
+    `test/emily/backend_test.exs` — 1- and 2-axis batch cases plus
+    edge shapes (scalar output, multi-free-axis both sides).
+  - **Test-only deps:** `bumblebee ~> 0.6`, `tokenizers ~> 0.5`
+    (both `only: :test`). Nx pinned to `~> 0.10` (down from 0.11)
+    to match Bumblebee's current constraint; emily's own API is
+    unaffected.
+
 ## Notes
 
 - Ops files use anonymous namespaces to prevent NIF function names
@@ -120,3 +172,6 @@
   `hadamard_transform`, quantized matmul, `linalg.*` decompositions
   (LU, QR, Cholesky, SVD). These will be added opportunistically when
   M2/M3 callers need them.
+- Deferred beyond M3: native `conv` translation (PLAN lists it under
+  M3, but DistilBERT and M4's Qwen3 don't use it; the BinaryBackend
+  fallback remains until a CV model lands on Emily).

c_src/ops/cast.cpp

@@ -23,4 +23,14 @@ fine::ResourcePtr<Tensor> astype(
 }
 FINE_NIF(astype, 0);
 
+// bitcast: reinterpret the bits as a different dtype of the same
+// element size. MLX exposes this as `mx::view`.
+fine::ResourcePtr<Tensor> bitcast(
+    ErlNifEnv *,
+    fine::ResourcePtr<Tensor> a,
+    std::tuple<fine::Atom, int64_t> dtype) {
+  return wrap(mx::view(a->array, to_mlx_dtype(dtype)));
+}
+FINE_NIF(bitcast, 0);
+
 } // namespace

lib/emily/backend.ex

@@ -12,7 +12,7 @@ defmodule Emily.Backend do
     * `{:f, 64}` is not supported — Metal cannot execute f64. Operations
       that would allocate an f64 tensor raise `ArgumentError` with a
       message pointing to f32.
-    * `bitcast`, `from_pointer`, `to_pointer`, `population_count`, and
+    * `from_pointer`, `to_pointer`, `population_count`, and
       `count_leading_zeros` raise `ArgumentError` — MLX has no primitive.
     * Window operations (`window_sum`, `window_scatter_max`, etc.) and
       advanced linalg (`lu`, `svd`, `qr`, `cholesky`, `eigh`, `solve`,
@@ -303,9 +303,13 @@ defmodule Emily.Backend do
     t |> ref() |> Native.astype(type) |> wrap(out)
   end
 
+  # bitcast: reinterpret the bits as the output dtype (same element
+  # size). MLX exposes this as `mx::view`. Used by Nx.Random to move
+  # between uint and float of matching width.
   @impl true
-  def bitcast(_out, _t),
-    do: raise(ArgumentError, "Emily.Backend does not implement bitcast (MLX has no primitive)")
+  def bitcast(%T{type: type} = out, t) do
+    t |> ref() |> Native.bitcast(type) |> wrap(out)
+  end
 
   # =================================================================
   # Unary ops
@@ -393,7 +397,15 @@ defmodule Emily.Backend do
   # Binary ops
   # =================================================================
 
-  @renamed_binary [
+  # Arithmetic + bitwise: cast both operands to `out.type` before
+  # handing to MLX. Two reasons:
+  #   1. MLX's cross-type promotion for mixed integer widths (e.g.,
+  #      u64 + s32) falls back to float32 — which then fails on
+  #      integer-only ops like right_shift. `Nx.Random.key` hits this.
+  #   2. `divide` has `out.type = float` even for integer operands
+  #      (`Nx.Type.to_floating/1`); casting to out.type first produces
+  #      the float division Nx promises.
+  @renamed_arith_binary [
     subtract: :subtract,
     multiply: :multiply,
     divide: :divide,
@@ -402,36 +414,62 @@ defmodule Emily.Backend do
     atan2: :arctan2,
     min: :minimum,
     max: :maximum,
+    bitwise_and: :bitwise_and,
+    bitwise_or: :bitwise_or,
+    bitwise_xor: :bitwise_xor,
+    left_shift: :left_shift,
+    right_shift: :right_shift
+  ]
+
+  for {nx_name, native_name} <- @renamed_arith_binary do
+    @impl true
+    def unquote(nx_name)(%T{type: type} = out, a, b) do
+      ra = Native.astype(ref(a), type)
+      rb = Native.astype(ref(b), type)
+      Native.unquote(native_name)(ra, rb) |> wrap(out)
+    end
+  end
+
+  # Compare + logical: out.type is `{:u, 8}` (pred), but MLX still
+  # needs the operands at a matched non-pred type to compare. Cast to
+  # `Nx.Type.merge(a, b)` so MLX sees a consistent arithmetic type.
+  @renamed_pred_binary [
     equal: :equal,
     not_equal: :not_equal,
     less: :less,
     less_equal: :less_equal,
     greater: :greater,
     greater_equal: :greater_equal,
     logical_and: :logical_and,
-    logical_or: :logical_or,
-    bitwise_and: :bitwise_and,
-    bitwise_or: :bitwise_or,
-    bitwise_xor: :bitwise_xor,
-    left_shift: :left_shift,
-    right_shift: :right_shift
+    logical_or: :logical_or
   ]
 
-  for {nx_name, native_name} <- @renamed_binary do
+  for {nx_name, native_name} <- @renamed_pred_binary do
     @impl true
     def unquote(nx_name)(out, a, b) do
-      Native.unquote(native_name)(ref(a), ref(b)) |> wrap(out)
+      target = Nx.Type.merge(a.type, b.type)
+      ra = Native.astype(ref(a), target)
+      rb = Native.astype(ref(b), target)
+      Native.unquote(native_name)(ra, rb) |> wrap(out)
     end
   end
 
   @impl true
-  def add(out, a, b), do: Native.add(ref(a), ref(b)) |> wrap(out)
+  def add(%T{type: type} = out, a, b) do
+    ra = Native.astype(ref(a), type)
+    rb = Native.astype(ref(b), type)
+    Native.add(ra, rb) |> wrap(out)
+  end
 
   # See moduledoc: we intentionally use MLX floor_divide for quotient,
   # matching its rounding (floor toward -inf) rather than Nx's
   # truncate-toward-zero. The two agree for non-negative operands.
   @impl true
-  def quotient(out, a, b), do: Native.floor_divide(ref(a), ref(b)) |> wrap(out)
+  def quotient(%T{type: type} = out, a, b) do
+    ra = Native.astype(ref(a), type)
+    rb = Native.astype(ref(b), type)
+    Native.floor_divide(ra, rb) |> wrap(out)
+  end
 
   # logical_xor: MLX has no direct op. Compose via `not_equal` on
   # booleanised inputs. With u8 inputs, `not_equal(0)` yields truthy
@@ -543,10 +581,17 @@ defmodule Emily.Backend do
 
   @impl true
   def slice(%T{} = out, t, starts, lengths, strides) do
+    # Nx passes starts as either integers or scalar tensors (dynamic
+    # slicing). MLX's slice takes integer bounds; under the evaluator
+    # we materialise scalar-tensor starts to their concrete value.
+    starts = Enum.map(starts, &slice_start/1)
     stops = Enum.zip_with(starts, lengths, fn s, l -> s + l end)
     Native.slice(ref(t), starts, stops, strides) |> wrap(out)
   end
 
+  defp slice_start(i) when is_integer(i), do: i
+  defp slice_start(%T{} = t), do: t |> Nx.backend_copy(Nx.BinaryBackend) |> Nx.to_number()
+
   # put_slice: MLX has no direct primitive; route via BinaryBackend.
   @impl true
   def put_slice(out, t, slice, starts),
@@ -609,10 +654,14 @@ defmodule Emily.Backend do
     end
   end
 
+  # Nx's argmax/argmin take `:keep_axis` (singular) on user-facing API
+  # but the backend callback exposes raw opts whose spelling has drifted
+  # across Nx versions. Derive `keep` from the shape invariant instead:
+  # if `out.shape` has the same rank as the input, the axis was kept.
   @impl true
   def argmax(%T{} = out, t, opts) do
     axis = opts[:axis] || 0
-    keep = opts[:keep_axes] || false
+    keep = tuple_size(out.shape) == tuple_size(t.shape)
 
     ref(t)
     |> Native.argmax(axis, keep)
@@ -623,7 +672,7 @@ defmodule Emily.Backend do
   @impl true
   def argmin(%T{} = out, t, opts) do
     axis = opts[:axis] || 0
-    keep = opts[:keep_axes] || false
+    keep = tuple_size(out.shape) == tuple_size(t.shape)
 
     ref(t)
     |> Native.argmin(axis, keep)
@@ -666,20 +715,78 @@ defmodule Emily.Backend do
   # Dot product
   # =================================================================
 
-  # Non-batched tensor contraction: tensordot. Batched case: we
-  # reshape-merge batch axes, matmul, reshape back. For now only
-  # handle the non-batched case fully; batched falls back to
-  # BinaryBackend.
+  # Non-batched: tensordot. Batched: permute to [batch, free, contract]
+  # on a and [batch, contract, free] on b, flatten to 3-D, hand to
+  # MLX matmul (which treats leading dims as batch), reshape back to
+  # Nx's canonical `batch ++ free_a ++ free_b` layout.
   @impl true
   def dot(%T{} = out, a, contract_a, [], b, contract_b, []) do
     Native.tensordot(ref(a), ref(b), contract_a, contract_b) |> wrap(out)
   end
 
-  # Batched dot is on the transformer-attention critical path. M3
-  # replaces this BinaryBackend bounce with a native matmul-with-
-  # transpose pattern before Bumblebee lands.
-  def dot(out, a, contract_a, batch_a, b, contract_b, batch_b),
-    do: via_binary(out, [a, b], &Nx.dot(&1, contract_a, batch_a, &2, contract_b, batch_b))
+  def dot(%T{type: type} = out, a, contract_a, batch_a, b, contract_b, batch_b) do
+    # MLX matmul is float-only; ints/preds fall through to BinaryBackend.
+    # In practice every transformer-attention call is float, so this is
+    # the hot path.
+    if float_like?(type) do
+      batched_matmul(out, a, contract_a, batch_a, b, contract_b, batch_b)
+    else
+      via_binary(out, [a, b], &Nx.dot(&1, contract_a, batch_a, &2, contract_b, batch_b))
+    end
+  end
+
+  defp float_like?({kind, _}) when kind in [:f, :bf, :c], do: true
+  defp float_like?(_), do: false
+
+  # Nx guarantees batch axes on both tensors are [0, 1, ..., k-1] in
+  # increasing order, so the permutation simplifies: batch dims stay
+  # at the front, free axes sort in positional order, contract axes
+  # in the Nx-given pairing order.
+  defp batched_matmul(
+         %T{shape: out_shape} = out,
+         %T{shape: as} = a,
+         contract_a,
+         batch_a,
+         %T{shape: bs} = b,
+         contract_b,
+         _batch_b
+       ) do
+    a_rank = tuple_size(as)
+    b_rank = tuple_size(bs)
+    k = length(batch_a)
+
+    contract_set_a = MapSet.new(contract_a)
+    contract_set_b = MapSet.new(contract_b)
+
+    free_a = for i <- k..(a_rank - 1)//1, not MapSet.member?(contract_set_a, i), do: i
+    free_b = for i <- k..(b_rank - 1)//1, not MapSet.member?(contract_set_b, i), do: i
+
+    b_prod = dim_product(batch_a, as)
+    m = dim_product(free_a, as)
+    n = dim_product(free_b, bs)
+    k_prod = dim_product(contract_a, as)
+
+    perm_a = batch_a ++ free_a ++ contract_a
+    perm_b = batch_a ++ contract_b ++ free_b
+
+    ra =
+      a
+      |> ref()
+      |> Native.transpose(perm_a)
+      |> Native.reshape([b_prod, m, k_prod])
+
+    rb =
+      b
+      |> ref()
+      |> Native.transpose(perm_b)
+      |> Native.reshape([b_prod, k_prod, n])
+
+    Native.matmul(ra, rb)
+    |> Native.reshape(shape_list(out_shape))
+    |> wrap(out)
+  end
+
+  defp dim_product(axes, shape), do: Enum.reduce(axes, 1, &(elem(shape, &1) * &2))
 
   # =================================================================
   # Sort / argsort / top_k / all_close / take / take_along_axis

lib/emily/native.ex

@@ -57,6 +57,9 @@ defmodule Emily.Native do
   @spec astype(tensor(), dtype()) :: tensor()
   def astype(_a, _dtype), do: nif()
 
+  @spec bitcast(tensor(), dtype()) :: tensor()
+  def bitcast(_a, _dtype), do: nif()
+
   # --- Unary -------------------------------------------------------
 
   unary_ops = [

mix.exs

@@ -52,6 +52,8 @@ defmodule Emily.MixProject do
       {:elixir_make, "~> 0.9"},
       {:fine, "~> 0.1"},
       {:nx, "~> 0.10"},
+      {:bumblebee, "~> 0.6", only: :test},
+      {:tokenizers, "~> 0.5", only: :test},
       {:stream_data, "~> 1.1", only: [:dev, :test]},
       {:credo, "~> 1.7", only: [:dev, :test], runtime: false},
       {:ex_doc, "~> 0.34", only: :docs, runtime: false}