M10 (partial): quantized inference primitives

Binds MLX's affine int4/int8 group-wise quantization at the Native
and Elixir levels plus a direct-call helper for eager use. Scope
narrowed vs. PLAN.md's original M10 — the Axon-layer dispatch,
Bumblebee AWQ conformance, and Backend.dot/7 routing are deferred
to M10.5 because `Nx.Defn.Evaluator` has no public hook to inject a
custom op that isn't an `Nx.Backend` callback, and `Nx.dot/2`
collapses multi-tensor containers before reaching the backend. See
RELEASE.md and PLAN.md §M10.5 for the three candidate defn-
integration approaches.

- Native NIFs (c_src/ops/linalg.cpp): quantize/3 returning a
  3-tuple {w_q, scales, biases}, dequantize/5, quantized_matmul/7
  with explicit transpose (AWQ layouts need transpose=false).
- Emily.QuantizedWeight — Nx.Container-derived struct with keep:
  so scalar metadata (group_size, bits, transpose) survives
  container traversal. Validates rank, last-axis divisibility,
  dtype, bit count before dispatch.
- Emily.Quantization.quantized_matmul/2 — direct-call helper for
  eager/benchmark use over materialized tensors.
- Tests: +7 cases in test/emily/native_test.exs, 13 cases across
  test/emily/quantization/, and a 1k-iter memory soak in
  test/soak/quantized_memory_test.exs asserting packed-weight
  footprint invariants.

mix precommit: 238 tests, 0 failures, credo clean.

Author: ausimian

PLAN.md

@@ -277,7 +277,7 @@ and back. Lift to native MLX:
 - `gather` → `mlx::core::gather`
 
 Window reductions stay on `via_binary` in M9 — pool-based conv
-training is scoped to M10.
+training is scoped to M17.
 
 **Testing — Layers 4 (Grad) and 5 (Training):**
 
@@ -341,47 +341,116 @@ user-visible value, not difficulty. Headline rationale:
    without hand-holding.
 5. **1.0 release** (M22) ships the result.
 
-### M10 — Quantized inference
+### M10 — Quantized inference primitives (partial)
 
 Quantization is the single largest gap between Emily and "actually run
 Qwen3 on a 16 GB MacBook". MLX ships native int4/int8 affine
 quantization (`mx::quantize`, `mx::dequantize`,
-`mx::quantized_matmul`) and Bumblebee can already load quantized
-checkpoints — Emily just silently can't consume them today. Without
-this, the library's headline value proposition only works for users
-with enough RAM for fp16 weights.
-
-- **Native bindings**: `Native.quantize/3`, `Native.dequantize/4`,
-  `Native.quantized_matmul/6` over the MLX C++ functions. Pass group
-  size and bit-width as integers; default to MLX's `group_size=64,
-  bits=4` but expose overrides.
-- **Backend routing**: detect quantized operand structs at the Nx
-  layer (Bumblebee tags them in the parameter map) and dispatch the
-  matmul callback to `Native.quantized_matmul` rather than
-  `Native.matmul`. Non-matmul ops (norms, residual adds, embeddings)
-  stay on the existing fp16/bf16 fast paths — only the linear
-  projection is quantized.
-- **Memory accounting**: quantized inference is allocator-pattern
-  different from fp16 — packed weights load once and never
-  re-quantize. Add a soak case asserting peak memory matches the
-  expected packed-weight footprint within ~10%.
-
-**Testing**:
-- Native unit tests with hand-computed packed-weight expected values
-  at small group sizes so the bit-packing is checkable.
-- Backend property test: `quantize → dequantize → matmul` against
-  `Nx.BinaryBackend` matmul on the dequantized weights, asserting
-  agreement within the documented quantization-error bound.
-- Conformance: add `Qwen/Qwen3-0.6B-AWQ` (or the MLX-community
-  quantized variant) as `:qwen3_quant_full`. Greedy-decode the same
-  prompt as `qwen3_full` and assert the completion matches a
-  checked-in reference produced by MLX's own Python bindings on the
-  same quantized weights — *not* by the f16 model, because the whole
-  point is to catch quantization-specific drift.
-
-**Exit:** Qwen3-0.6B-AWQ greedy-decodes end-to-end on Emily; Backend
-property tests green; quantized soak harness asserts allocator
-invariants.
+`mx::quantized_matmul`); M10 binds it at the Native and Elixir levels
+and ships a direct-call helper for eager use. The Bumblebee-integrated
+conformance path is split out to M10.5 — see **Scope note** below.
+
+**Shipped**:
+
+- **Native bindings**: `Native.quantize/3`, `Native.dequantize/5`,
+  `Native.quantized_matmul/7` over the MLX C++ functions. `quantize`
+  returns a 3-tuple `{w_q, scales, biases}`. `quantized_matmul/7` takes
+  `transpose` as an explicit boolean rather than PLAN's original `/6`:
+  AWQ packed layouts need `transpose=false` while fresh-from-dense
+  weights use `transpose=true`, and MLX exposes it as a required
+  parameter.
+- **`Emily.QuantizedWeight`** (`lib/emily/quantized_weight.ex`) —
+  `Nx.Container`-derived struct with `{value, scales, biases,
+  group_size, bits, transpose}`. Scalar metadata survives container
+  traversal via `Nx.Container`'s `keep:` option. `from_dense/2`
+  validates rank, last-axis divisibility, dtype, and bit count.
+- **`Emily.Quantization.quantized_matmul/2`** — direct-call helper that
+  extracts refs from an input tensor and a `%QuantizedWeight{}` and
+  dispatches the fused kernel. Intended for eager/benchmark use and as
+  the substrate for M10.5's defn-integration path.
+- **Memory soak** (`test/soak/quantized_memory_test.exs`) — 1000-iter
+  quantized-matmul loop asserts active memory returns within 4 MB of
+  baseline after `Native.clear_cache/0`. Kept separate from the fp16
+  memory soak because quantized inference is allocator-pattern
+  different: packed weights load once and never re-quantize.
+- **Native unit tests + Backend property tests** — see
+  `test/emily/native_test.exs` (+7 cases) and
+  `test/emily/quantization/` (two new files). Round-trip `quantize →
+  dequantize` and `quantized_matmul` vs. `matmul(x, dequantize(…))`
+  oracles for both `transpose=true` and `transpose=false` layouts.
+
+**Scope note — why no Backend routing / Axon integration / conformance
+test in M10**:
+
+- **`Backend.dot/7` dispatch doesn't work.** PLAN'd approach was
+  "detect quantized operand structs at the Nx layer (Bumblebee tags
+  them in the parameter map) and dispatch the matmul callback to
+  `Native.quantized_matmul`". But `Nx.dot/2` calls
+  `Nx.LazyContainer.traverse/3` expecting a single `%T{}`; a
+  three-tensor `%QuantizedWeight{}` container raises before reaching
+  `Backend.dot/7`.
+- **Axon layer-op dispatch doesn't work either.** `Axon.layer` ops run
+  at `Nx.Defn.jit` trace time with `Nx.Defn.Expr` inputs;
+  `Nx.Defn.Evaluator` walks those expressions dispatching `Nx.Backend`
+  callbacks with materialized refs. There is no public hook to inject
+  a custom op like `Native.quantized_matmul` that isn't already a
+  `Nx.Backend` callback, and `deftransform` / `hook` / metadata all
+  run at trace time (no refs available).
+- **Bumblebee has no AWQ loader yet.** The exploration for M10
+  confirmed `deps/bumblebee` has zero quantization-loading code (no
+  AWQ, GPTQ, MLX-format paths). PLAN.md's "Bumblebee can already load
+  quantized checkpoints" is aspirational.
+
+All three of these are meaningful scope. M10 ships the substrate they
+all need; M10.5 picks the defn-integration strategy and ships the
+conformance test.
+
+**Exit**: Native NIFs green under unit + property tests; QuantizedWeight
+container property tests green; direct-call helper green under oracle
+comparison; quantized memory soak clean.
+
+### M10.5 — Bumblebee quantized inference integration
+
+Closes the gap M10 left open: getting `Native.quantized_matmul`
+reachable from `Nx.Defn.jit`-traced Axon forward passes so Bumblebee's
+AWQ-loading (when it lands) routes through the fused kernel.
+
+Approach choices (pick before starting):
+
+1. **Defn-native dequantize** — implement MLX's int4/int8 affine
+   dequantize using Nx bit primitives (right-shift + mask + multiply +
+   add). `Emily.Quantization.Layers.quantized_dense/3` becomes
+   `Nx.dot(x, dequantize_defn(qw))`. Correct and unblocks the full
+   Axon/Bumblebee path, but uses two kernels (dequantize + matmul)
+   instead of MLX's fused one — M11's fast-kernel work subsumes the
+   perf gap.
+2. **Emily.Compiler custom-op intercept** — fork `Nx.Defn.Evaluator`
+   under Emily to recognise a sentinel `Expr` node and route to
+   `Native.quantized_matmul`. Full fused-kernel story but large
+   surface; fragile against upstream Nx evolution.
+3. **Upstream Nx extension** — add a custom-backend-op hook to
+   `Nx.Defn.Compiler` / `Nx.Defn.Evaluator`. Cleanest long-term
+   solution; slowest to land because it needs upstream review/merge.
+
+Also in scope for M10.5:
+
+- **Test-only AWQ loader** (`test/support/awq_loader.ex`) — reads
+  `Qwen/Qwen3-0.6B-AWQ` safetensors, extracts `qweight`, `scales`,
+  `qzeros`, maps to MLX's `(w_q, scales, biases)` layout. The
+  trickiest bit is the AWQ zero-point → MLX bias conversion
+  (`biases = -scales * zero_points`) and the AWQ `[in, out/pack]` vs.
+  MLX `[out, in]` layout difference.
+- **`:qwen3_quant_full` conformance test** — greedy-decode
+  Qwen3-0.6B-AWQ on Emily, assert the completion matches a checked-in
+  reference produced by MLX's Python bindings on the same quantized
+  weights.
+- **Bumblebee upstream contribution (optional, follow-up to M10.5)** —
+  upstream the AWQ loader into `deps/bumblebee` so the test-only path
+  becomes unnecessary.
+
+**Exit**: Qwen3-0.6B-AWQ greedy-decodes end-to-end on Emily under
+`Nx.Defn.jit`; conformance test green; Axon-integrated quantization
+documented.
 
 ### M11 — `mlx::fast::*` fused kernels
 

RELEASE.md

@@ -2,6 +2,86 @@
 
 ## Added
 
+- M10 (partial) — Quantized inference primitives. Exposes MLX's affine
+  int4/int8 group-wise quantization at the Native and Elixir levels, plus
+  a direct-call helper for eager use. Enough to quantize a dense weight,
+  store it packed at rest, and dispatch a fused quantized matmul against
+  it in plain Elixir code.
+  - **Native `Native.quantize/3`, `Native.dequantize/5`,
+    `Native.quantized_matmul/7`** (`c_src/ops/linalg.cpp`) over
+    `mx::quantize`, `mx::dequantize`, `mx::quantized_matmul`. `quantize`
+    returns a 3-tuple `{w_q, scales, biases}` — first NIF in the tree
+    returning a tuple of resources. `quantized_matmul` takes the
+    `transpose` flag as an explicit arg rather than the PLAN-spec'd `/6`
+    signature: AWQ packed layouts need `transpose=false` while
+    fresh-from-dense weights use `transpose=true`, and MLX exposes the
+    flag as a required parameter on its C++ API.
+  - **`Emily.QuantizedWeight`** (`lib/emily/quantized_weight.ex`) —
+    `Nx.Container`-derived struct bundling `{value, scales, biases}`
+    with `{group_size, bits, transpose}` metadata (the latter flagged
+    `keep:` so the scalars survive container traversal — defn trace,
+    `backend_transfer`, parameter-map walks). `from_dense/2` validates
+    rank ≥ 2, last-axis divisibility, dtype ∈ {f16, bf16, f32}, and bits
+    ∈ {2,3,4,6,8} before dispatch. `to_dense/1` inverses via
+    `Native.dequantize`.
+  - **`Emily.Quantization.quantized_matmul/2`**
+    (`lib/emily/quantization.ex`) — direct-call helper over
+    materialized tensors. Extracts refs from the input `%T{}` and the
+    three tensors inside a `%QuantizedWeight{}`, dispatches to the
+    Native NIF, and rewraps the result. Input dtype must match
+    `qw.scales.type` (raises with a targeted error otherwise);
+    BinaryBackend inputs are transferred transparently. Used by the
+    soak and property tests today; the fused-kernel perf win is
+    available here but not yet from defn-traced forward passes (see
+    below).
+  - **`test/emily/native_test.exs`** (+7 cases) — hand-computed
+    quantize-shape checks (`group_size=64, bits=4` → 8 u32/row; bits=8
+    → 16 u32/row; smaller group_size → more scale/bias rows),
+    `dequantize` round-trip within int4 step tolerance, validation of
+    indivisible last axes, and `quantized_matmul` equivalence vs.
+    `matmul(x, dequantize(…))` for both `transpose=true` and
+    `transpose=false` layouts.
+  - **`test/emily/quantization/quantized_weight_test.exs`** (new) —
+    `from_dense/2` metadata/shape/dtype assertions, validation-raise
+    cases (indivisible axis, unsupported bits, unsupported dtype,
+    rank < 2), bits=8-tighter-than-bits=4 property, and an
+    `Nx.Container` traversal test confirming `keep:` preserves
+    `group_size`/`bits`/`transpose` across `backend_transfer`. Property
+    test: random group-shaped weights round-trip `from_dense |>
+    to_dense` within a 0.15 tolerance band.
+  - **`test/emily/quantization/quantized_matmul_test.exs`** (new) —
+    property test for `transpose=true` against a
+    `Nx.dot(x, Nx.transpose(to_dense(qw)))` oracle on BinaryBackend;
+    explicit `transpose=false` case asserting `Nx.dot(x, to_dense(qw))`
+    convention; dtype-mismatch raise test; BinaryBackend-input
+    auto-transfer test.
+  - **`test/soak/quantized_memory_test.exs`** (`@moduletag :soak`,
+    default suite) — 1000-iteration quantized-matmul soak asserting
+    active memory returns within 4 MB of baseline after `clear_cache/0`.
+    Separate from `memory_test.exs` because quantized inference is
+    allocator-pattern different from fp16: packed weights load once
+    (not re-quantized per call), and the per-iter activation/output
+    budget is smaller.
+  - **Scope narrowed from PLAN.md.** PLAN spec'd a `Backend.dot/7`
+    dispatch path that inspects the operand struct, plus an Axon
+    layer-replacement (`Emily.Quantization.Layers.quantized_dense` +
+    `Emily.Quantization.quantize/2`) and a `Qwen/Qwen3-0.6B-AWQ`
+    conformance test. Investigation uncovered that `Nx.dot/2` calls
+    `Nx.LazyContainer.traverse/3` expecting a single `%T{}` — a
+    three-tensor `%QuantizedWeight{}` container raises before reaching
+    `Backend.dot/7`. The alternative — an Axon layer op calling
+    `Native.quantized_matmul` during forward pass — fails for a
+    different reason: Axon layer ops run at `Nx.Defn.jit` trace time
+    with `Nx.Defn.Expr` inputs, and `Nx.Defn.Evaluator` has no public
+    hook to inject a custom op that isn't a `Nx.Backend` callback.
+    Closing the gap requires either (a) a defn-native dequantize built
+    from Nx bit primitives (loses the fused kernel), or (b) an
+    Emily-specific `Nx.Defn.Compiler` variant that recognizes a
+    sentinel `Expr` node and routes to `Native.quantized_matmul`. Both
+    are meaningful scope; tracked as M10.5. The Native + container +
+    direct-call helper surface shipped in M10 is the substrate that
+    either of those approaches will build on.
+
 - M9 — Gradient conformance and training primitives. Makes
   `Nx.Defn.grad` usable on Emily by lifting the three ops that grad
   lands on most heavily off the `via_binary` fallback, and adds the

c_src/ops/linalg.cpp

@@ -1,11 +1,13 @@
-// Linear algebra: matmul, tensordot, outer, inner.
+// Linear algebra: matmul, tensordot, outer, inner, and affine int4/int8
+// quantization primitives (quantize / dequantize / quantized_matmul).
 
 #include "../emily/tensor.hpp"
 
 #include <fine.hpp>
 #include <mlx/mlx.h>
 
 #include <cstdint>
+#include <tuple>
 #include <vector>
 
 namespace mx = mlx::core;
@@ -51,4 +53,71 @@ fine::ResourcePtr<Tensor> inner(
 }
 FINE_NIF(inner, 0);
 
+// Affine quantization along the last axis.
+//
+// Returns {w_q, scales, biases}:
+//   - w_q is packed uint32 with (last_dim * bits / 32) elements per row
+//   - scales and biases share shape (..., last_dim / group_size), dtype
+//     matching the input.
+//
+// MLX requires last_dim % group_size == 0; bits ∈ {2, 3, 4, 6, 8}.
+std::tuple<fine::ResourcePtr<Tensor>,
+           fine::ResourcePtr<Tensor>,
+           fine::ResourcePtr<Tensor>>
+quantize(
+    ErlNifEnv *,
+    fine::ResourcePtr<Tensor> w,
+    int64_t group_size,
+    int64_t bits) {
+  auto triple = mx::quantize(
+      w->array, static_cast<int>(group_size), static_cast<int>(bits));
+  return std::make_tuple(
+      wrap(std::move(std::get<0>(triple))),
+      wrap(std::move(std::get<1>(triple))),
+      wrap(std::move(std::get<2>(triple))));
+}
+FINE_NIF(quantize, 0);
+
+// Inverse of quantize. Reconstructs a dense tensor from packed w_q plus
+// per-group scales and biases.
+fine::ResourcePtr<Tensor> dequantize(
+    ErlNifEnv *,
+    fine::ResourcePtr<Tensor> w_q,
+    fine::ResourcePtr<Tensor> scales,
+    fine::ResourcePtr<Tensor> biases,
+    int64_t group_size,
+    int64_t bits) {
+  return wrap(mx::dequantize(
+      w_q->array,
+      scales->array,
+      biases->array,
+      static_cast<int>(group_size),
+      static_cast<int>(bits)));
+}
+FINE_NIF(dequantize, 0);
+
+// Matmul against a quantized weight. `transpose` is wired through
+// explicitly because AWQ-style packed checkpoints ship in a different
+// layout than freshly-quantized weights, and MLX's kernel selection
+// depends on the flag.
+fine::ResourcePtr<Tensor> quantized_matmul(
+    ErlNifEnv *,
+    fine::ResourcePtr<Tensor> x,
+    fine::ResourcePtr<Tensor> w_q,
+    fine::ResourcePtr<Tensor> scales,
+    fine::ResourcePtr<Tensor> biases,
+    bool transpose,
+    int64_t group_size,
+    int64_t bits) {
+  return wrap(mx::quantized_matmul(
+      x->array,
+      w_q->array,
+      scales->array,
+      biases->array,
+      transpose,
+      static_cast<int>(group_size),
+      static_cast<int>(bits)));
+}
+FINE_NIF(quantized_matmul, 0);
+
 } // namespace

lib/emily/native.ex

@@ -244,6 +244,36 @@ defmodule Emily.Native do
   @spec inner(tensor(), tensor()) :: tensor()
   def inner(_a, _b), do: nif()
 
+  # --- Quantization ------------------------------------------------
+
+  @spec quantize(tensor(), integer(), integer()) ::
+          {tensor(), tensor(), tensor()}
+  def quantize(_w, _group_size, _bits), do: nif()
+
+  @spec dequantize(tensor(), tensor(), tensor(), integer(), integer()) ::
+          tensor()
+  def dequantize(_w_q, _scales, _biases, _group_size, _bits), do: nif()
+
+  @spec quantized_matmul(
+          tensor(),
+          tensor(),
+          tensor(),
+          tensor(),
+          boolean(),
+          integer(),
+          integer()
+        ) :: tensor()
+  def quantized_matmul(
+        _x,
+        _w_q,
+        _scales,
+        _biases,
+        _transpose,
+        _group_size,
+        _bits
+      ),
+      do: nif()
+
   # --- Sort --------------------------------------------------------
 
   @spec sort(tensor(), integer()) :: tensor()