haliax 1.4.dev396__tar.gz → 1.4.dev398__tar.gz

{haliax-1.4.dev396 → haliax-1.4.dev398}/.playbooks/wrap-non-named.md

@@ -47,3 +47,6 @@ Some functions need bespoke handling. For example `jnp.unique` returns several a
 
 ## Testing
 Add tests to ensure that named and unnamed calls produce the same results and that axis names are preserved or removed correctly.
+
+## Documentation
+Once your wrapper works, document it. Add `::: haliax.your_function` (replacing the name) to `docs/api.md` so users can find the new API.

{haliax-1.4.dev396 → haliax-1.4.dev398}/AGENTS.md

@@ -41,7 +41,8 @@ repository. Follow these notes when implementing new features or fixing bugs.
   `[Object 1][full.path.object1]`)
 * **Documentation**: When adding new features, ensure that the documentation is updated accordingly.
   This includes updating the Mkdocs files and any relevant docstrings. If you add a new module or
-  significant functionality, consider adding a dedicated section in the documentation.
+  significant functionality, consider adding a dedicated section in the documentation. When you
+  wrap a new JAX function, add a reference to it in `docs/api.md` so users can discover it.
 
 ## Testing
 

{haliax-1.4.dev396 → haliax-1.4.dev398}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: haliax
-Version: 1.4.dev396
+Version: 1.4.dev398
 Summary: Named Tensors for Legible Deep Learning in JAX
 Project-URL: Homepage, https://github.com/stanford-crfm/haliax
 Project-URL: Bug Tracker, https://github.com/stanford-crfm/haliax/issues/

{haliax-1.4.dev396 → haliax-1.4.dev398}/docs/api.md

@@ -256,6 +256,7 @@ These are all more or less directly from JAX's NumPy API.
 
 ### Other Operations
 
+::: haliax.bincount
 ::: haliax.clip
 ::: haliax.isclose
 ::: haliax.pad

haliax-1.4.dev398/src/haliax/__about__.py

@@ -0,0 +1 @@
+__version__ = "1.4.dev398"

{haliax-1.4.dev396 → haliax-1.4.dev398}/src/haliax/__init__.py

@@ -79,6 +79,7 @@ from .ops import (
     unique_counts,
     unique_inverse,
     unique_all,
+    bincount,
     where,
 )
 from .partitioning import auto_sharded, axis_mapping, fsdp, named_jit, shard, shard_with_axis_mapping
@@ -1053,6 +1054,7 @@ __all__ = [
     "unique_counts",
     "unique_inverse",
     "unique_all",
+    "bincount",
     "clip",
     "tril",
     "triu",

{haliax-1.4.dev396 → haliax-1.4.dev398}/src/haliax/ops.py

@@ -430,6 +430,33 @@ def unique_all(
     return values, indices, inverse, counts
 
 
+def bincount(
+    x: NamedArray,
+    Counts: Axis,
+    *,
+    weights: NamedArray | ArrayLike | None = None,
+    minlength: int = 0,
+) -> NamedArray:
+    """Named version of `jax.numpy.bincount`.
+
+    The output axis is specified by ``Counts``.
+    """
+
+    if x.ndim != 1:
+        raise ValueError("bincount only supports 1D arrays")
+
+    w_array = None
+    if weights is not None:
+        if isinstance(weights, NamedArray):
+            weights = haliax.broadcast_to(weights, x.axes)
+            w_array = weights.array
+        else:
+            w_array = jnp.asarray(weights)
+
+    result = jnp.bincount(x.array, weights=w_array, minlength=minlength, length=Counts.size)
+    return NamedArray(result, (Counts,))
+
+
 __all__ = [
     "trace",
     "where",
@@ -444,4 +471,5 @@ __all__ = [
     "unique_counts",
     "unique_inverse",
     "unique_all",
+    "bincount",
 ]

haliax-1.4.dev398/tests/test_moe_linear.py

@@ -0,0 +1,70 @@
+import jax
+import jax.random as jrandom
+from jax import numpy as jnp
+
+import haliax as hax
+from haliax.nn import MoELinear
+
+
+def _expected_moe_linear_output(moe: MoELinear, x: hax.NamedArray, group_sizes: hax.NamedArray):
+    dim_numbers = jax.lax.RaggedDotDimensionNumbers(
+        (
+            ((x.axis_indices(moe.In),), (moe.weight.axis_indices(moe.In),)),
+            ((), ()),
+        ),
+        x.axis_indices(hax.axis.without_axes(x.axes, moe.In)),
+        (moe.weight.axis_indices(moe.Experts),),
+    )
+    out_raw = jax.lax.ragged_dot_general(
+        lhs=x.array,
+        rhs=moe.weight.array,
+        group_sizes=group_sizes.array,
+        ragged_dot_dimension_numbers=dim_numbers,
+    )
+    out_axes = hax.replace_axis(x.axes, moe.In, moe.Out)
+    out = hax.named(out_raw, out_axes)
+    if moe.bias is not None:
+        out = out + moe.bias
+    return out
+
+
+def test_moe_linear_matches_ragged_dot_general():
+    B, In, Out, E = hax.make_axes(B=3, In=4, Out=5, E=2)
+    key = jrandom.PRNGKey(0)
+    moe = MoELinear.init(E, In, Out, key=key)
+
+    x = hax.random.normal(jrandom.PRNGKey(1), (B, In))
+    group_sizes = hax.named(jnp.array([2, 1], dtype=jnp.int32), (E,))
+
+    actual = moe(x, group_sizes)
+    expected = _expected_moe_linear_output(moe, x, group_sizes)
+
+    assert actual.axes == expected.axes
+    assert jnp.allclose(actual.array, expected.array, rtol=1e-5, atol=1e-5)
+
+
+def test_moe_linear_out_first_property():
+    E, In, Out = hax.make_axes(E=2, In=4, Out=3)
+    moe = MoELinear.init(E, In, Out, key=jrandom.PRNGKey(0), out_first=True)
+    assert moe.out_first
+    assert moe.weight.axes[:3] == (E, Out, In)
+
+    moe2 = MoELinear.init(E, In, Out, key=jrandom.PRNGKey(1), out_first=False)
+    assert not moe2.out_first
+    assert moe2.weight.axes[:3] == (E, In, Out)
+
+
+def test_moe_linear_gmm_matches_ragged_dot_general():
+    B, In, Out, E = hax.make_axes(B=3, In=4, Out=5, E=2)
+    moe = MoELinear.init(E, In, Out, key=jrandom.PRNGKey(0), use_gmm=True)
+
+    x = hax.random.normal(jrandom.PRNGKey(1), (B, In))
+    group_sizes = hax.named(jnp.array([2, 1], dtype=jnp.int32), (E,))
+
+    with jax.sharding.Mesh(jax.devices(), ("data",)):
+        actual = moe(x, group_sizes)
+
+    expected = _expected_moe_linear_output(moe, x, group_sizes)
+
+    assert actual.axes == expected.axes
+    assert jnp.allclose(actual.array, expected.array, rtol=1e-5, atol=1e-5)

{haliax-1.4.dev396 → haliax-1.4.dev398}/tests/test_ops.py

@@ -407,3 +407,19 @@ def test_unique_shortcuts():
     assert jnp.all(ia.array == ia_exp.array)
     assert jnp.all(ina.array == ina_exp.array)
     assert jnp.all(ca.array == ca_exp.array)
+
+
+def test_bincount():
+    X = Axis("X", 6)
+    x = hax.named([0, 1, 1, 2, 3, 1], (X,))
+    B = Axis("B", 5)
+
+    out = hax.bincount(x, B)
+    expected = jnp.bincount(x.array, length=B.size)
+    assert out.axes == (B,)
+    assert jnp.all(out.array == expected)
+
+    w = hax.arange((X,), dtype=jnp.float32)
+    out_w = hax.bincount(x, B, weights=w)
+    expected_w = jnp.bincount(x.array, weights=w.array, length=B.size)
+    assert jnp.allclose(out_w.array, expected_w)

haliax-1.4.dev396/src/haliax/__about__.py

@@ -1 +0,0 @@
-__version__ = "1.4.dev396"