jax-pyrefly-stubs 1.4.0.dev3__tar.gz → 1.5.0.dev1__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (19) hide show
  1. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/PKG-INFO +8 -3
  2. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/README.md +4 -1
  3. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/jax-stubs/__init__.pyi +65 -4
  4. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/jax-stubs/_array.pyi +190 -4
  5. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/jax-stubs/_shapes.pyi +83 -12
  6. jax_pyrefly_stubs-1.5.0.dev1/jax-stubs/_src/basearray.pyi +10 -0
  7. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/jax-stubs/lax/__init__.pyi +16 -0
  8. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/jax-stubs/numpy/__init__.pyi +125 -45
  9. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/jax-stubs/random.pyi +3 -2
  10. jax_pyrefly_stubs-1.5.0.dev1/jax-stubs/sharding.pyi +36 -0
  11. jax_pyrefly_stubs-1.5.0.dev1/jax-stubs/typing.pyi +14 -0
  12. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/pyproject.toml +4 -2
  13. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/LICENSE +0 -0
  14. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/jax-stubs/lax/linalg.pyi +0 -0
  15. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/jax-stubs/nn/__init__.pyi +0 -0
  16. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/jax-stubs/nn/initializers.pyi +0 -0
  17. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/jax-stubs/numpy/fft.pyi +0 -0
  18. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/jax-stubs/numpy/linalg.pyi +0 -0
  19. {jax_pyrefly_stubs-1.4.0.dev3 → jax_pyrefly_stubs-1.5.0.dev1}/jax-stubs/py.typed +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.5
2
2
  Name: jax-pyrefly-stubs
3
- Version: 1.4.0.dev3
3
+ Version: 1.5.0.dev1
4
4
  Summary: JAX type stubs with array-shape tracking, maintained by Pyrefly
5
5
  Project-URL: homepage, https://pyrefly.org
6
6
  Project-URL: documentation, https://pyrefly.org/en/docs/
@@ -33,7 +33,9 @@ Classifier: Programming Language :: Python
33
33
  Classifier: Programming Language :: Python :: 3
34
34
  Classifier: Typing :: Stubs Only
35
35
  Requires-Python: >=3.12
36
- Requires-Dist: pyrefly-shape-extensions==1.4.0.dev3
36
+ Requires-Dist: jax
37
+ Requires-Dist: pyrefly-numpy-stubs==1.5.0.dev1
38
+ Requires-Dist: pyrefly-shape-extensions==1.5.0.dev1
37
39
  Description-Content-Type: text/markdown
38
40
 
39
41
  # Pyrefly JAX shape stubs
@@ -41,7 +43,10 @@ Description-Content-Type: text/markdown
41
43
  This package is a PEP 561 stub-only distribution. It installs the `jax-stubs`
42
44
  stub package so Pyrefly can discover shape-aware stubs for the runtime `jax`
43
45
  package without replacing or shadowing JAX itself. It is versioned in lockstep
44
- with Pyrefly and depends on the matching `pyrefly-shape-extensions` package.
46
+ with Pyrefly and depends on the matching `pyrefly-shape-extensions` and
47
+ `pyrefly-numpy-stubs` packages. Installing it also installs `jax`, which provides
48
+ upstream modules used by these partial stubs. The `jax` dependency is unpinned,
49
+ so projects can choose their own JAX version.
45
50
 
46
51
  Shape-typed fixture stubs for a subset of JAX. This is a starting point for
47
52
  working with the JAX core team rather than a complete model: it covers array
@@ -3,7 +3,10 @@
3
3
  This package is a PEP 561 stub-only distribution. It installs the `jax-stubs`
4
4
  stub package so Pyrefly can discover shape-aware stubs for the runtime `jax`
5
5
  package without replacing or shadowing JAX itself. It is versioned in lockstep
6
- with Pyrefly and depends on the matching `pyrefly-shape-extensions` package.
6
+ with Pyrefly and depends on the matching `pyrefly-shape-extensions` and
7
+ `pyrefly-numpy-stubs` packages. Installing it also installs `jax`, which provides
8
+ upstream modules used by these partial stubs. The `jax` dependency is unpinned,
9
+ so projects can choose their own JAX version.
7
10
 
8
11
  Shape-typed fixture stubs for a subset of JAX. This is a starting point for
9
12
  working with the JAX core team rather than a complete model: it covers array
@@ -3,14 +3,76 @@
3
3
  # This source code is licensed under the MIT license found in the
4
4
  # LICENSE file in the root directory of this source tree.
5
5
 
6
+ from collections.abc import Callable, Iterable, Sequence
6
7
  from types import SimpleNamespace
7
- from typing import Any, Callable
8
+ from typing import Any, overload
8
9
 
9
10
  # Override jax.Array with shape-aware version.
10
11
  from jax._array import Array as Array
12
+ from jax.stages import Lowered as _Lowered, Traced as _Traced
13
+
14
+ class _Wrapped[**P, R]:
15
+ def __call__(self, *args: P.args, **kwargs: P.kwargs) -> R: ...
16
+ def lower(self, *args: P.args, **kwargs: P.kwargs) -> _Lowered: ...
17
+ def trace(self, *args: P.args, **kwargs: P.kwargs) -> _Traced: ...
18
+ def eval_shape(self, *args: P.args, **kwargs: P.kwargs) -> Any: ...
19
+ def clear_cache(self) -> None: ...
20
+
21
+ class _JitTransform:
22
+ def __call__[**P, R](self, fun: Callable[P, R], /) -> _Wrapped[P, R]: ...
11
23
 
12
24
  # Override jit and vmap because jax annotations fail with shaped arrays.
13
- def jit[F: Callable[..., Any]](fun: F, *args: Any, **kwargs: Any) -> F: ...
25
+ @overload
26
+ def jit[**P, R](
27
+ fun: Callable[P, R],
28
+ /,
29
+ *,
30
+ in_shardings: Any = ...,
31
+ out_shardings: Any = ...,
32
+ static_argnums: int | Sequence[int] | None = None,
33
+ static_argnames: str | Iterable[str] | None = None,
34
+ donate_argnums: int | Sequence[int] | None = None,
35
+ donate_argnames: str | Iterable[str] | None = None,
36
+ keep_unused: bool = False,
37
+ device: Any = None,
38
+ backend: str | None = None,
39
+ inline: bool | Any = False,
40
+ compiler_options: dict[str, Any] | None = None,
41
+ **kwargs: Any,
42
+ ) -> _Wrapped[P, R]: ...
43
+ @overload
44
+ def jit(
45
+ *,
46
+ in_shardings: Any = ...,
47
+ out_shardings: Any = ...,
48
+ static_argnums: int | Sequence[int] | None = None,
49
+ static_argnames: str | Iterable[str] | None = None,
50
+ donate_argnums: int | Sequence[int] | None = None,
51
+ donate_argnames: str | Iterable[str] | None = None,
52
+ keep_unused: bool = False,
53
+ device: Any = None,
54
+ backend: str | None = None,
55
+ inline: bool | Any = False,
56
+ compiler_options: dict[str, Any] | None = None,
57
+ **kwargs: Any,
58
+ ) -> _JitTransform: ...
59
+ @overload
60
+ def jit[**P, R](
61
+ *,
62
+ fun: Callable[P, R],
63
+ in_shardings: Any = ...,
64
+ out_shardings: Any = ...,
65
+ static_argnums: int | Sequence[int] | None = None,
66
+ static_argnames: str | Iterable[str] | None = None,
67
+ donate_argnums: int | Sequence[int] | None = None,
68
+ donate_argnames: str | Iterable[str] | None = None,
69
+ keep_unused: bool = False,
70
+ device: Any = None,
71
+ backend: str | None = None,
72
+ inline: bool | Any = False,
73
+ compiler_options: dict[str, Any] | None = None,
74
+ **kwargs: Any,
75
+ ) -> _Wrapped[P, R]: ...
14
76
  def vmap(fun: Callable[..., Any], *args: Any, **kwargs: Any) -> Callable[..., Any]: ...
15
77
 
16
78
  # Remaining APIs mirror imports in jax/__init__.py,
@@ -44,7 +106,6 @@ from jax import (
44
106
  from jax._src.ad_checkpoint import (
45
107
  checkpoint as checkpoint,
46
108
  checkpoint_policies as checkpoint_policies,
47
- custom_remat as custom_remat,
48
109
  remat as remat,
49
110
  )
50
111
  from jax._src.api import (
@@ -136,7 +197,6 @@ from jax._src.mesh import (
136
197
  use_abstract_mesh as use_abstract_mesh,
137
198
  )
138
199
  from jax._src.partition_spec import P as P
139
- from jax._src.pjit import reshard as reshard
140
200
  from jax._src.pmap import pmap as pmap
141
201
  from jax._src.shard_map import (
142
202
  shard_map as shard_map,
@@ -168,6 +228,7 @@ from jax.ref import (
168
228
  new_ref as new_ref,
169
229
  Ref as Ref,
170
230
  )
231
+ from jax.sharding import reshard as reshard
171
232
  from jax.version import (
172
233
  __version__ as __version__,
173
234
  __version_info__ as __version_info__,
@@ -9,7 +9,7 @@
9
9
 
10
10
  from collections.abc import Iterator
11
11
  from types import EllipsisType
12
- from typing import Any, overload, Protocol, Sequence, SupportsIndex
12
+ from typing import Any, Callable, overload, Protocol, Sequence, SupportsIndex
13
13
 
14
14
  import numpy as np
15
15
  from jax._shapes import (
@@ -50,7 +50,7 @@ from shape_extensions import (
50
50
  type _Shape = IntTuple
51
51
  type _Axis = int | tuple[int, ...] | None
52
52
  type _SingleAxis = int | None
53
- type _Scalar = bool | int | float | complex | np.number
53
+ type _Scalar = bool | int | float | complex | np.number | np.bool_
54
54
  # Note: when using _ArrayLike in an annotation, the Shape passed to it must
55
55
  # have a default value of [] for scalars to be handled properly.
56
56
  type ArrayLike[Shape: _Shape] = Array[Shape] | np.ndarray[Shape] | _Scalar
@@ -85,6 +85,8 @@ class Array[Shape: _Shape = _Shape]:
85
85
  self,
86
86
  index: _BasicIndex | _ArrayIndex | tuple[_BasicIndex | _ArrayIndex, ...],
87
87
  ) -> Array[IntTuple]: ...
88
+ @property
89
+ def at(self) -> _IndexUpdateHelper[Shape]: ...
88
90
  # JAX reverses every axis, at any rank, so this is not 2-D only.
89
91
  @property
90
92
  def T(self) -> Array[reverse_shape(Shape)]: ...
@@ -118,12 +120,44 @@ class Array[Shape: _Shape = _Shape]:
118
120
  def real(self) -> Array[Shape]: ...
119
121
  @property
120
122
  def imag(self) -> Array[Shape]: ...
123
+ @property
124
+ def aval(self) -> Any: ...
125
+ @property
126
+ def weak_type(self) -> bool: ...
127
+ @property
128
+ def traceback(self) -> Any: ...
129
+ @property
130
+ def format(self) -> Any: ...
121
131
  def __len__[N: IntVar, Rest: _Shape = []](
122
132
  self: Array[[N, *Rest]],
123
133
  ) -> Int[N]: ...
124
- def __iter__[N: IntVar, Rest: _Shape = []](
134
+ def __iter__[N: IntVar, Rest: _Shape](
125
135
  self: Array[[N, *Rest]], /
126
136
  ) -> Iterator[Array[Rest]]: ...
137
+ def __reversed__[N: IntVar, Rest: _Shape](
138
+ self: Array[[N, *Rest]], /
139
+ ) -> Iterator[Array[Rest]]: ...
140
+ def __contains__(self, key: object) -> bool: ...
141
+ def __bool__(self) -> bool: ...
142
+ def __int__(self: Array[[]], /) -> int: ...
143
+ def __float__(self: Array[[]], /) -> float: ...
144
+ def __complex__(self: Array[[]], /) -> complex: ...
145
+ def __index__(self: Array[[]], /) -> int: ...
146
+ def __round__(self, ndigits: int | None = None) -> Array[Shape]: ...
147
+ def __copy__(self) -> Array[Shape]: ...
148
+ def __deepcopy__(self, memo: Any) -> Array[Shape]: ...
149
+ def __format__(self, format_spec: str) -> str: ...
150
+ def __array__(
151
+ self,
152
+ dtype: np.dtype | None = None,
153
+ context: None = None,
154
+ copy: bool | None = None,
155
+ ) -> np.ndarray[Shape]: ...
156
+ def __array_namespace__(self, *, api_version: str | None = None) -> Any: ...
157
+ def __dlpack__(self, *, stream: Any = None) -> Any: ...
158
+ def __dlpack_device__(self) -> tuple[int, int]: ...
159
+ def __buffer__(self, flags: int) -> memoryview: ...
160
+ def __release_buffer__(self, view: memoryview) -> None: ...
127
161
  def copy(self) -> Array[Shape]: ...
128
162
  def conj(self) -> Array[Shape]: ...
129
163
  def conjugate(self) -> Array[Shape]: ...
@@ -145,6 +179,18 @@ class Array[Shape: _Shape = _Shape]:
145
179
  ) -> Array[Shape]: ...
146
180
  def addressable_data(self, index: int) -> Array[IntTuple]: ...
147
181
  def copy_to_host_async(self) -> None: ...
182
+ def block_until_ready(self) -> Array[Shape]: ...
183
+ def clone(self) -> Array[Shape]: ...
184
+ def delete(self) -> None: ...
185
+ def is_deleted(self) -> bool: ...
186
+ def is_ready(self) -> bool: ...
187
+ def devices(self) -> set[_Device]: ...
188
+ def platform(self) -> str: ...
189
+ def on_device_size_in_bytes(self) -> int: ...
190
+ def unsafe_buffer_pointer(self) -> int: ...
191
+ def unsafe_raw_buffer(self) -> Any: ...
192
+ def tobytes(self, order: str = "C") -> bytes: ...
193
+ def tolist(self) -> list[Any]: ...
148
194
  def __add__[OtherShape: _Shape = []](
149
195
  self, other: ArrayLike[OtherShape]
150
196
  ) -> Array[broadcast(Shape, OtherShape)]: ...
@@ -181,6 +227,22 @@ class Array[Shape: _Shape = _Shape]:
181
227
  def __rmod__[OtherShape: _Shape = []](
182
228
  self, other: ArrayLike[OtherShape]
183
229
  ) -> Array[broadcast(Shape, OtherShape)]: ...
230
+ @overload
231
+ def __divmod__(self, other: _Scalar, /) -> tuple[Array[Shape], Array[Shape]]: ...
232
+ @overload
233
+ def __divmod__[OtherShape: _Shape = []](
234
+ self, other: ArrayLike[OtherShape], /
235
+ ) -> tuple[
236
+ Array[broadcast(Shape, OtherShape)], Array[broadcast(Shape, OtherShape)]
237
+ ]: ...
238
+ @overload
239
+ def __rdivmod__(self, other: _Scalar, /) -> tuple[Array[Shape], Array[Shape]]: ...
240
+ @overload
241
+ def __rdivmod__[OtherShape: _Shape = []](
242
+ self, other: ArrayLike[OtherShape], /
243
+ ) -> tuple[
244
+ Array[broadcast(OtherShape, Shape)], Array[broadcast(OtherShape, Shape)]
245
+ ]: ...
184
246
  def __pow__[OtherShape: _Shape = []](
185
247
  self, other: ArrayLike[OtherShape]
186
248
  ) -> Array[broadcast(Shape, OtherShape)]: ...
@@ -248,6 +310,9 @@ class Array[Shape: _Shape = _Shape]:
248
310
  def __matmul__[OtherShape: _Shape = []](
249
311
  self, other: ArrayLike[OtherShape]
250
312
  ) -> Array[matmul_shape(Shape, OtherShape)]: ...
313
+ def __rmatmul__[OtherShape: _Shape = []](
314
+ self, other: ArrayLike[OtherShape]
315
+ ) -> Array[matmul_shape(OtherShape, Shape)]: ...
251
316
  @overload
252
317
  def transpose(self) -> Array[reverse_shape(Shape)]: ...
253
318
  @overload
@@ -297,7 +362,12 @@ class Array[Shape: _Shape = _Shape]:
297
362
  order: str = ...,
298
363
  out_sharding: _NamedSharding | _PartitionSpec | None = ...,
299
364
  ) -> Array[IntTuple]: ...
300
- def ravel(self, order: str = "C") -> Array[ravel_shape(Shape)]: ...
365
+ def ravel(
366
+ self,
367
+ order: str = "C",
368
+ *,
369
+ out_sharding: _NamedSharding | _PartitionSpec | None = None,
370
+ ) -> Array[ravel_shape(Shape)]: ...
301
371
  def flatten(
302
372
  self,
303
373
  order: str = "C",
@@ -323,6 +393,7 @@ class Array[Shape: _Shape = _Shape]:
323
393
  axis: Axis = None,
324
394
  *,
325
395
  total_repeat_length: None = None,
396
+ out_sharding: _NamedSharding | _PartitionSpec | None = None,
326
397
  ) -> Array[repeat_shape(Shape, Repeats, Axis)]: ...
327
398
  # TODO(stroxler): Separate dynamic repeat counts from invalid concrete counts and axes.
328
399
  @overload
@@ -332,6 +403,7 @@ class Array[Shape: _Shape = _Shape]:
332
403
  axis: int | None = None,
333
404
  *,
334
405
  total_repeat_length: int | None = None,
406
+ out_sharding: _NamedSharding | _PartitionSpec | None = None,
335
407
  ) -> Array[IntTuple]: ...
336
408
  @overload
337
409
  # Any non-tuple sequence axis is gradual; see `jax/numpy/__init__.pyi`.
@@ -796,3 +868,117 @@ class Array[Shape: _Shape = _Shape]:
796
868
  size: int | None = None,
797
869
  fill_value: Any = 0,
798
870
  ) -> Array[IntTuple]: ...
871
+
872
+ class _IndexUpdateRef[Shape: _Shape, SliceShape: _Shape]:
873
+ def set(
874
+ self,
875
+ values: ArrayLike[Any] | Sequence[Any],
876
+ *,
877
+ mode: str | None = None,
878
+ indices_are_sorted: bool = False,
879
+ unique_indices: bool = False,
880
+ wrap_negative_indices: bool = True,
881
+ **kwargs: Any,
882
+ ) -> Array[Shape]: ...
883
+ def add(
884
+ self,
885
+ values: ArrayLike[Any] | Sequence[Any],
886
+ *,
887
+ mode: str | None = None,
888
+ indices_are_sorted: bool = False,
889
+ unique_indices: bool = False,
890
+ wrap_negative_indices: bool = True,
891
+ **kwargs: Any,
892
+ ) -> Array[Shape]: ...
893
+ def subtract(
894
+ self,
895
+ values: ArrayLike[Any] | Sequence[Any],
896
+ *,
897
+ mode: str | None = None,
898
+ indices_are_sorted: bool = False,
899
+ unique_indices: bool = False,
900
+ wrap_negative_indices: bool = True,
901
+ **kwargs: Any,
902
+ ) -> Array[Shape]: ...
903
+ def multiply(
904
+ self,
905
+ values: ArrayLike[Any] | Sequence[Any],
906
+ *,
907
+ mode: str | None = None,
908
+ indices_are_sorted: bool = False,
909
+ unique_indices: bool = False,
910
+ wrap_negative_indices: bool = True,
911
+ **kwargs: Any,
912
+ ) -> Array[Shape]: ...
913
+ def divide(
914
+ self,
915
+ values: ArrayLike[Any] | Sequence[Any],
916
+ *,
917
+ mode: str | None = None,
918
+ indices_are_sorted: bool = False,
919
+ unique_indices: bool = False,
920
+ wrap_negative_indices: bool = True,
921
+ **kwargs: Any,
922
+ ) -> Array[Shape]: ...
923
+ def power(
924
+ self,
925
+ values: ArrayLike[Any] | Sequence[Any],
926
+ *,
927
+ mode: str | None = None,
928
+ indices_are_sorted: bool = False,
929
+ unique_indices: bool = False,
930
+ wrap_negative_indices: bool = True,
931
+ **kwargs: Any,
932
+ ) -> Array[Shape]: ...
933
+ def min(
934
+ self,
935
+ values: ArrayLike[Any] | Sequence[Any],
936
+ *,
937
+ mode: str | None = None,
938
+ indices_are_sorted: bool = False,
939
+ unique_indices: bool = False,
940
+ wrap_negative_indices: bool = True,
941
+ **kwargs: Any,
942
+ ) -> Array[Shape]: ...
943
+ def max(
944
+ self,
945
+ values: ArrayLike[Any] | Sequence[Any],
946
+ *,
947
+ mode: str | None = None,
948
+ indices_are_sorted: bool = False,
949
+ unique_indices: bool = False,
950
+ wrap_negative_indices: bool = True,
951
+ **kwargs: Any,
952
+ ) -> Array[Shape]: ...
953
+ def apply(
954
+ self,
955
+ func: Callable[..., Any],
956
+ *,
957
+ mode: str | None = None,
958
+ indices_are_sorted: bool = False,
959
+ unique_indices: bool = False,
960
+ wrap_negative_indices: bool = True,
961
+ **kwargs: Any,
962
+ ) -> Array[Shape]: ...
963
+ def get(
964
+ self,
965
+ *,
966
+ mode: str | None = None,
967
+ fill_value: Any = None,
968
+ indices_are_sorted: bool = False,
969
+ unique_indices: bool = False,
970
+ wrap_negative_indices: bool = True,
971
+ strategy: str | None = None,
972
+ **kwargs: Any,
973
+ ) -> Array[SliceShape]: ...
974
+
975
+ class _IndexUpdateHelper[Shape: _Shape]:
976
+ @overload
977
+ def __getitem__[I: Index](
978
+ self, index: I
979
+ ) -> _IndexUpdateRef[Shape, index_shape(Shape, I)]: ...
980
+ @overload
981
+ def __getitem__(
982
+ self,
983
+ index: _BasicIndex | _ArrayIndex | tuple[_BasicIndex | _ArrayIndex, ...],
984
+ ) -> _IndexUpdateRef[Shape, IntTuple]: ...
@@ -124,26 +124,30 @@ def parameter_broadcast_shape(
124
124
  return result_shape
125
125
 
126
126
  @type_shape_dsl_function
127
- def multinomial_shape(
128
- n_shape: IntTuple, p_shape: IntTuple, result_shape: IntTuple
127
+ def multinomial_shape(n_shape: IntTuple, p_shape: IntTuple) -> IntTuple:
128
+ batch_shape = p_shape[:-1]
129
+ return multinomial_sample_shape(n_shape, p_shape, batch_shape)
130
+
131
+ @type_shape_dsl_function
132
+ def multinomial_sample_shape(
133
+ n_shape: IntTuple, p_shape: IntTuple, sample_shape: IntTuple
129
134
  ) -> IntTuple:
130
- if len(p_shape) == 0 or len(result_shape) == 0:
135
+ if len(p_shape) == 0:
131
136
  return dsl.Invalid("multinomial probabilities must have an event dimension")
132
- probability_shapes = dsl.IntTuples((p_shape, result_shape))
137
+ probability_shapes = dsl.IntTuples((p_shape[:-1], sample_shape))
133
138
  probability_shape = gufunc_broadcast("(),()->()", probability_shapes)
134
- if len(probability_shape) != len(result_shape) or any(
135
- probability_shape[index] != result_shape[index]
136
- for index in range(len(result_shape))
139
+ if len(probability_shape) != len(sample_shape) or any(
140
+ probability_shape[index] != sample_shape[index]
141
+ for index in range(len(sample_shape))
137
142
  ):
138
143
  return dsl.Invalid("probabilities cannot broadcast to the requested shape")
139
- count_shapes = dsl.IntTuples((n_shape, result_shape[:-1]))
144
+ count_shapes = dsl.IntTuples((n_shape, sample_shape))
140
145
  count_shape = gufunc_broadcast("(),()->()", count_shapes)
141
- if len(count_shape) != len(result_shape) - 1 or any(
142
- count_shape[index] != result_shape[index]
143
- for index in range(len(result_shape) - 1)
146
+ if len(count_shape) != len(sample_shape) or any(
147
+ count_shape[index] != sample_shape[index] for index in range(len(sample_shape))
144
148
  ):
145
149
  return dsl.Invalid("counts cannot broadcast to the requested batch shape")
146
- return result_shape
150
+ return dsl.concat(sample_shape, p_shape[-1:])
147
151
 
148
152
  @type_shape_dsl_function
149
153
  def multivariate_normal_shape(
@@ -840,6 +844,50 @@ def vecmat_shape(left: IntTuple, right: IntTuple) -> IntTuple:
840
844
  spec = "(n),(n,m)->(m)"
841
845
  return gufunc_broadcast(spec, operands)
842
846
 
847
+ @type_shape_dsl_function
848
+ def tensordot_axes_shape(
849
+ left: IntTuple,
850
+ right: IntTuple,
851
+ left_axes: IntTuple,
852
+ right_axes: IntTuple,
853
+ ) -> IntTuple:
854
+ if len(left_axes) != len(right_axes):
855
+ return dsl.Invalid("tensordot axis sequences must have equal length")
856
+ if any(axis < 0 - len(left) or axis >= len(left) for axis in left_axes):
857
+ return dsl.Invalid("axis out of bounds")
858
+ if any(axis < 0 - len(right) or axis >= len(right) for axis in right_axes):
859
+ return dsl.Invalid("axis out of bounds")
860
+ left_norm = dsl.IntTuple(
861
+ axis + len(left) if axis < 0 else axis for axis in left_axes
862
+ )
863
+ right_norm = dsl.IntTuple(
864
+ axis + len(right) if axis < 0 else axis for axis in right_axes
865
+ )
866
+ left_indices = dsl.IntTuple(i for i in range(len(left)))
867
+ right_indices = dsl.IntTuple(i for i in range(len(right)))
868
+ left_rem = dsl.IntTuple(
869
+ dimension
870
+ for dimension, index in zip(left, left_indices)
871
+ if not any(index == axis for axis in left_norm)
872
+ )
873
+ right_rem = dsl.IntTuple(
874
+ dimension
875
+ for dimension, index in zip(right, right_indices)
876
+ if not any(index == axis for axis in right_norm)
877
+ )
878
+ if len(left_rem) + len(left_norm) != len(left):
879
+ return dsl.Invalid("duplicate axis")
880
+ if len(right_rem) + len(right_norm) != len(right):
881
+ return dsl.Invalid("duplicate axis")
882
+ mismatches = dsl.IntTuple(
883
+ 1
884
+ for left_axis, right_axis in zip(left_norm, right_norm)
885
+ if left[left_axis] != right[right_axis]
886
+ )
887
+ if len(mismatches) != 0:
888
+ return dsl.Invalid("tensordot contracted dimensions must match")
889
+ return dsl.concat(left_rem, right_rem)
890
+
843
891
  @type_shape_dsl_function
844
892
  def tensordot_shape(
845
893
  left: IntTuple,
@@ -1620,6 +1668,29 @@ def lax_scan_shape(shape: IntTuple, axis: int) -> IntTuple:
1620
1668
  return dsl.Invalid("axis out of bounds")
1621
1669
  return shape
1622
1670
 
1671
+ @type_shape_dsl_function
1672
+ def lax_scan_slice_shape(shape: IntTuple) -> IntTuple:
1673
+ # `lax_scan_stacked_shape` reports a missing scan axis for the whole call.
1674
+ if len(shape) == 0:
1675
+ return dsl.IntTuple.gradual()
1676
+ return shape[1:]
1677
+
1678
+ @type_shape_dsl_function
1679
+ def lax_scan_stacked_shape(xs_shapes: IntTuples, out_shape: IntTuple) -> IntTuple:
1680
+ if len(xs_shapes) == 0:
1681
+ return dsl.concat(dsl.IntTuple((dsl.Int.gradual(),)), out_shape)
1682
+ ranks = dsl.IntTuple((1 if len(shape) == 0 else 0 for shape in xs_shapes))
1683
+ if any(r == 1 for r in ranks):
1684
+ return dsl.Invalid("scan requires every xs array to have a leading axis")
1685
+ first = xs_shapes[0]
1686
+ length = first[0]
1687
+ mismatches = dsl.IntTuple((1 if shape[0] != length else 0 for shape in xs_shapes))
1688
+ if any(m == 1 for m in mismatches):
1689
+ return dsl.Invalid(
1690
+ "scan requires every xs array to have the same leading axis size"
1691
+ )
1692
+ return dsl.concat(dsl.IntTuple((length,)), out_shape)
1693
+
1623
1694
  @type_shape_dsl_function
1624
1695
  def lax_associative_scan_shape(shape: IntTuple, axis: int) -> IntTuple:
1625
1696
  rank = len(shape)
@@ -0,0 +1,10 @@
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ #
3
+ # This source code is licensed under the MIT license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ import numpy as np
7
+ from jax._array import Array as Array
8
+
9
+ type StaticScalar = np.bool_ | np.number | bool | int | float | complex
10
+ type ArrayLike = Array | np.ndarray | StaticScalar
@@ -23,6 +23,8 @@ from jax._shapes import (
23
23
  lax_fft_shape,
24
24
  lax_reduce_shape,
25
25
  lax_scan_shape,
26
+ lax_scan_slice_shape,
27
+ lax_scan_stacked_shape,
26
28
  lax_select_n_shape,
27
29
  lax_select_shape,
28
30
  lax_sort_key_val_shape,
@@ -43,6 +45,7 @@ from jax._src.lax.lax import (
43
45
  DotAlgorithm as DotAlgorithm,
44
46
  DotAlgorithmPreset as DotAlgorithmPreset,
45
47
  DotDimensionNumbers as DotDimensionNumbers,
48
+ polynomial as polynomial,
46
49
  Precision as Precision,
47
50
  PrecisionLike as PrecisionLike,
48
51
  RaggedDotDimensionNumbers as RaggedDotDimensionNumbers,
@@ -1343,6 +1346,19 @@ def scan[Carry, N: IntVar, InShape: _Shape, OutShape: _Shape](
1343
1346
  _split_transpose: bool = False,
1344
1347
  ) -> tuple[Carry, _Array[[N, *OutShape]]]: ...
1345
1348
  @overload
1349
+ def scan[Carry, InShapes: IntTuples, OutShape: _Shape](
1350
+ f: Callable[
1351
+ [Carry, MapIntTuples[lambda S: _Array[lax_scan_slice_shape(S)], InShapes]],
1352
+ tuple[Carry, _Array[OutShape]],
1353
+ ],
1354
+ init: Carry,
1355
+ xs: MapIntTuples[lambda S: _Array[S], InShapes],
1356
+ length: int | None = None,
1357
+ reverse: bool = False,
1358
+ unroll: int | bool = 1,
1359
+ _split_transpose: bool = False,
1360
+ ) -> tuple[Carry, _Array[lax_scan_stacked_shape(InShapes, OutShape)]]: ...
1361
+ @overload
1346
1362
  def scan[Carry, X, Y](
1347
1363
  f: Callable[[Carry, X], tuple[Carry, Y]],
1348
1364
  init: Carry,
@@ -10,6 +10,7 @@ from typing import (
10
10
  Literal,
11
11
  NamedTuple,
12
12
  overload,
13
+ Protocol,
13
14
  Sequence,
14
15
  Unpack,
15
16
  )
@@ -80,6 +81,7 @@ from jax._shapes import (
80
81
  swapaxes_shape,
81
82
  take_along_axis_shape,
82
83
  take_shape,
84
+ tensordot_axes_shape,
83
85
  tensordot_shape,
84
86
  tile_shape,
85
87
  top_k_shape,
@@ -209,6 +211,7 @@ def zeros[N: IntVar](
209
211
  dtype: DTypeLike | None = ...,
210
212
  *,
211
213
  device: _Device | _Sharding | None = ...,
214
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
212
215
  ) -> _Array[[N]]: ...
213
216
  @overload
214
217
  def zeros[Shape: _Shape](
@@ -216,6 +219,7 @@ def zeros[Shape: _Shape](
216
219
  dtype: DTypeLike | None = ...,
217
220
  *,
218
221
  device: _Device | _Sharding | None = ...,
222
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
219
223
  ) -> _Array[Shape]: ...
220
224
  @overload
221
225
  def zeros(
@@ -223,6 +227,7 @@ def zeros(
223
227
  dtype: DTypeLike | None = ...,
224
228
  *,
225
229
  device: _Device | _Sharding | None = ...,
230
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
226
231
  ) -> _Array[IntTuple]: ...
227
232
  @overload
228
233
  def ones[N: IntVar](
@@ -230,6 +235,7 @@ def ones[N: IntVar](
230
235
  dtype: DTypeLike | None = ...,
231
236
  *,
232
237
  device: _Device | _Sharding | None = ...,
238
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
233
239
  ) -> _Array[[N]]: ...
234
240
  @overload
235
241
  def ones[Shape: _Shape](
@@ -237,6 +243,7 @@ def ones[Shape: _Shape](
237
243
  dtype: DTypeLike | None = ...,
238
244
  *,
239
245
  device: _Device | _Sharding | None = ...,
246
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
240
247
  ) -> _Array[Shape]: ...
241
248
  @overload
242
249
  def ones(
@@ -244,6 +251,7 @@ def ones(
244
251
  dtype: DTypeLike | None = ...,
245
252
  *,
246
253
  device: _Device | _Sharding | None = ...,
254
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
247
255
  ) -> _Array[IntTuple]: ...
248
256
  @overload
249
257
  def empty[N: IntVar](
@@ -276,6 +284,7 @@ def full[N: IntVar](
276
284
  dtype: DTypeLike | None = ...,
277
285
  *,
278
286
  device: _Device | _Sharding | None = ...,
287
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
279
288
  ) -> _Array[[N]]: ...
280
289
  @overload
281
290
  def full[Shape: _Shape](
@@ -284,6 +293,7 @@ def full[Shape: _Shape](
284
293
  dtype: DTypeLike | None = ...,
285
294
  *,
286
295
  device: _Device | _Sharding | None = ...,
296
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
287
297
  ) -> _Array[Shape]: ...
288
298
  @overload
289
299
  def full(
@@ -292,6 +302,7 @@ def full(
292
302
  dtype: DTypeLike | None = ...,
293
303
  *,
294
304
  device: _Device | _Sharding | None = ...,
305
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
295
306
  ) -> _Array[IntTuple]: ...
296
307
 
297
308
  # `_like` constructors
@@ -447,6 +458,7 @@ def arange[N: IntVar](
447
458
  *,
448
459
  dtype: DTypeLike | None = ...,
449
460
  device: _Device | _Sharding | None = ...,
461
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
450
462
  ) -> _Array[[arange_stop(Int[N])]]: ...
451
463
  @overload
452
464
  def arange(
@@ -454,6 +466,7 @@ def arange(
454
466
  *,
455
467
  dtype: DTypeLike | None = ...,
456
468
  device: _Device | _Sharding | None = ...,
469
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
457
470
  ) -> _Array[[int]]: ...
458
471
  @overload
459
472
  def arange[Start: Flag[int], Stop: Flag[int]](
@@ -461,6 +474,9 @@ def arange[Start: Flag[int], Stop: Flag[int]](
461
474
  stop: Stop,
462
475
  step: Literal[1] = ...,
463
476
  dtype: DTypeLike | None = ...,
477
+ *,
478
+ device: _Device | _Sharding | None = ...,
479
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
464
480
  ) -> _Array[[arange_size(Start, Stop, 1)]]: ...
465
481
  @overload
466
482
  def arange[Start: Flag[int], Stop: Flag[int], Step: Flag[int]](
@@ -468,6 +484,9 @@ def arange[Start: Flag[int], Stop: Flag[int], Step: Flag[int]](
468
484
  stop: Stop,
469
485
  step: Step,
470
486
  dtype: DTypeLike | None = ...,
487
+ *,
488
+ device: _Device | _Sharding | None = ...,
489
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
471
490
  ) -> _Array[[arange_size(Start, Stop, Step)]]: ...
472
491
  @overload
473
492
  def arange(
@@ -475,6 +494,9 @@ def arange(
475
494
  stop: int | float,
476
495
  step: int | float = ...,
477
496
  dtype: DTypeLike | None = ...,
497
+ *,
498
+ device: _Device | _Sharding | None = ...,
499
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
478
500
  ) -> _Array[[int]]: ...
479
501
  @overload
480
502
  def linspace[
@@ -1027,7 +1049,12 @@ def kron[AShape: _Shape = [], BShape: _Shape = []](
1027
1049
  b: _ArrayLike[BShape],
1028
1050
  ) -> _Array[kron_shape(AShape, BShape)]: ...
1029
1051
  def matmul[LeftShape: _Shape, RightShape: _Shape](
1030
- a: _ShapedArrayLike[LeftShape], b: _ShapedArrayLike[RightShape]
1052
+ a: _ShapedArrayLike[LeftShape],
1053
+ b: _ShapedArrayLike[RightShape],
1054
+ *,
1055
+ precision: Any = None,
1056
+ preferred_element_type: Any = None,
1057
+ out_sharding: _NamedSharding | _PartitionSpec | None = None,
1031
1058
  ) -> _Array[matmul_shape(LeftShape, RightShape)]: ...
1032
1059
  def matvec[LeftShape: _Shape, RightShape: _Shape](
1033
1060
  x1: _ShapedArrayLike[LeftShape],
@@ -1046,6 +1073,22 @@ def outer(
1046
1073
  b: _ArrayLike[Any] | Sequence[Any],
1047
1074
  out: None = None,
1048
1075
  ) -> _Array[IntTuple]: ...
1076
+ @overload
1077
+ def tensordot[
1078
+ Left: _Shape,
1079
+ Right: _Shape,
1080
+ LeftAxes: IntTuple,
1081
+ RightAxes: IntTuple,
1082
+ ](
1083
+ a: _ArrayLike[Left],
1084
+ b: _ArrayLike[Right],
1085
+ axes: tuple[LeftAxes, RightAxes],
1086
+ *,
1087
+ precision: Any = None,
1088
+ preferred_element_type: Any = None,
1089
+ out_sharding: _NamedSharding | _PartitionSpec | None = None,
1090
+ ) -> _Array[tensordot_axes_shape(Left, Right, LeftAxes, RightAxes)]: ...
1091
+ @overload
1049
1092
  def tensordot[
1050
1093
  Left: _Shape = [],
1051
1094
  Right: _Shape = [],
@@ -1195,6 +1238,8 @@ def reshape(
1195
1238
  def ravel[Shape: _Shape = []](
1196
1239
  a: _ArrayLike[Shape],
1197
1240
  order: str = "C",
1241
+ *,
1242
+ out_sharding: _NamedSharding | _PartitionSpec | None = None,
1198
1243
  ) -> _Array[ravel_shape(Shape)]: ...
1199
1244
  @overload
1200
1245
  def squeeze[Shape: _Shape = [], Axis: Flag[_Axis] = None](
@@ -1220,11 +1265,15 @@ def expand_dims(
1220
1265
  def broadcast_to[TargetShape: Flag[_NewShape], Shape: _Shape = []](
1221
1266
  array: _ArrayLike[Shape],
1222
1267
  shape: TargetShape,
1268
+ *,
1269
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
1223
1270
  ) -> _Array[broadcast_to_shape(Shape, TargetShape)]: ...
1224
1271
  @overload
1225
1272
  def broadcast_to(
1226
1273
  array: _ArrayLike[Any],
1227
1274
  shape: Sequence[int] | int,
1275
+ *,
1276
+ out_sharding: _NamedSharding | _PartitionSpec | None = ...,
1228
1277
  ) -> _Array[IntTuple]: ...
1229
1278
  @overload
1230
1279
  def broadcast_arrays[Shape: _Shape = []](
@@ -1468,6 +1517,7 @@ def repeat[
1468
1517
  axis: Axis = None,
1469
1518
  *,
1470
1519
  total_repeat_length: None = None,
1520
+ out_sharding: _NamedSharding | _PartitionSpec | None = None,
1471
1521
  ) -> _Array[repeat_shape(Shape, Repeats, Axis)]: ...
1472
1522
 
1473
1523
  # TODO(stroxler): Separate dynamic repeat counts from invalid concrete counts and axes.
@@ -1478,6 +1528,7 @@ def repeat(
1478
1528
  axis: int | None = None,
1479
1529
  *,
1480
1530
  total_repeat_length: int | None = None,
1531
+ out_sharding: _NamedSharding | _PartitionSpec | None = None,
1481
1532
  ) -> _Array[IntTuple]: ...
1482
1533
  @overload
1483
1534
  def resize(a: _ArrayLike[Any], new_shape: tuple[()]) -> _Array[[]]: ...
@@ -1575,19 +1626,37 @@ def atleast_1d[Shape: _Shape = []](
1575
1626
  ary: _ArrayLike[Shape], /
1576
1627
  ) -> _Array[atleast_1d_shape(Shape)]: ...
1577
1628
  @overload
1578
- def atleast_1d(*arys: _ArrayLike[Any]) -> list[_Array[IntTuple]]: ...
1629
+ def atleast_1d[Shapes: IntTuples](
1630
+ *arys: Unpack[MapIntTuples[lambda S: _Array[S], Shapes]],
1631
+ ) -> MapIntTuples[lambda S: _Array[atleast_1d_shape(S)], Shapes]: ...
1632
+ @overload
1633
+ def atleast_1d[Shapes: IntTuples](
1634
+ *arys: Unpack[MapIntTuples[lambda S: _ArrayLike[S], Shapes]],
1635
+ ) -> MapIntTuples[lambda S: _Array[S], Shapes]: ...
1579
1636
  @overload
1580
1637
  def atleast_2d[Shape: _Shape = []](
1581
1638
  ary: _ArrayLike[Shape], /
1582
1639
  ) -> _Array[atleast_2d_shape(Shape)]: ...
1583
1640
  @overload
1584
- def atleast_2d(*arys: _ArrayLike[Any]) -> list[_Array[IntTuple]]: ...
1641
+ def atleast_2d[Shapes: IntTuples](
1642
+ *arys: Unpack[MapIntTuples[lambda S: _Array[S], Shapes]],
1643
+ ) -> MapIntTuples[lambda S: _Array[atleast_2d_shape(S)], Shapes]: ...
1644
+ @overload
1645
+ def atleast_2d[Shapes: IntTuples](
1646
+ *arys: Unpack[MapIntTuples[lambda S: _ArrayLike[S], Shapes]],
1647
+ ) -> MapIntTuples[lambda S: _Array[S], Shapes]: ...
1585
1648
  @overload
1586
1649
  def atleast_3d[Shape: _Shape = []](
1587
1650
  ary: _ArrayLike[Shape], /
1588
1651
  ) -> _Array[atleast_3d_shape(Shape)]: ...
1589
1652
  @overload
1590
- def atleast_3d(*arys: _ArrayLike[Any]) -> list[_Array[IntTuple]]: ...
1653
+ def atleast_3d[Shapes: IntTuples](
1654
+ *arys: Unpack[MapIntTuples[lambda S: _Array[S], Shapes]],
1655
+ ) -> MapIntTuples[lambda S: _Array[atleast_3d_shape(S)], Shapes]: ...
1656
+ @overload
1657
+ def atleast_3d[Shapes: IntTuples](
1658
+ *arys: Unpack[MapIntTuples[lambda S: _ArrayLike[S], Shapes]],
1659
+ ) -> MapIntTuples[lambda S: _Array[S], Shapes]: ...
1591
1660
 
1592
1661
  # JAX accepts any integer sequence for an axis, but only a tuple is a Flag
1593
1662
  # domain, so any other sequence yields a gradual shape. Rejecting it would flag
@@ -3438,6 +3507,7 @@ def histogram[Bins: Flag[int] = 10](
3438
3507
  range: Sequence[Any] | None = None,
3439
3508
  weights: _ArrayLike[Any] | None = None,
3440
3509
  density: bool | None = None,
3510
+ out_sharding: _NamedSharding | _PartitionSpec | None = None,
3441
3511
  ) -> tuple[
3442
3512
  _Array[histogram_counts_shape(Bins)], _Array[histogram_edges_shape(Bins)]
3443
3513
  ]: ...
@@ -3448,6 +3518,7 @@ def histogram(
3448
3518
  range: Sequence[Any] | None = None,
3449
3519
  weights: _ArrayLike[Any] | None = None,
3450
3520
  density: bool | None = None,
3521
+ out_sharding: _NamedSharding | _PartitionSpec | None = None,
3451
3522
  ) -> tuple[_Array[IntTuple], _Array[IntTuple]]: ...
3452
3523
  @overload
3453
3524
  def histogram2d[Bins: Flag[int] = 10](
@@ -3612,44 +3683,53 @@ def roots(p: _ArrayLike[Any], *, strip_zeros: bool = True) -> _Array[IntTuple]:
3612
3683
 
3613
3684
  # Scalar constructors
3614
3685
 
3615
- bool: Any
3616
- bool_: Any
3617
- int_: Any
3618
- int8: Any
3619
- int16: Any
3620
- int32: Any
3621
- int64: Any
3622
- uint: Any
3623
- uint8: Any
3624
- uint16: Any
3625
- uint32: Any
3626
- uint64: Any
3627
- int1: Any
3628
- int2: Any
3629
- int4: Any
3630
- uint1: Any
3631
- uint2: Any
3632
- uint4: Any
3633
- float_: Any
3634
- float16: Any
3635
- float32: Any
3636
- float64: Any
3637
- bfloat16: Any
3638
- single: Any
3639
- double: Any
3640
- csingle: Any
3641
- cdouble: Any
3642
- complex_: Any
3643
- complex64: Any
3644
- complex128: Any
3645
- float4_e2m1fn: Any
3646
- float6_e2m3fn: Any
3647
- float6_e3m2fn: Any
3648
- float8_e3m4: Any
3649
- float8_e4m3: Any
3650
- float8_e4m3b11fnuz: Any
3651
- float8_e4m3fn: Any
3652
- float8_e4m3fnuz: Any
3653
- float8_e5m2: Any
3654
- float8_e5m2fnuz: Any
3655
- float8_e8m0fnu: Any
3686
+ class _ScalarType(Protocol):
3687
+ @property
3688
+ def dtype(self) -> np.dtype: ...
3689
+ @property
3690
+ def __numpy_dtype__(self) -> np.dtype: ...
3691
+ def __call__(self, x: Any, /) -> _Array: ...
3692
+ def __hash__(self, /) -> int: ...
3693
+ def __eq__(self, other: Any, /) -> bool: ...
3694
+
3695
+ bool: _ScalarType
3696
+ bool_: _ScalarType
3697
+ int_: _ScalarType
3698
+ int8: _ScalarType
3699
+ int16: _ScalarType
3700
+ int32: _ScalarType
3701
+ int64: _ScalarType
3702
+ uint: _ScalarType
3703
+ uint8: _ScalarType
3704
+ uint16: _ScalarType
3705
+ uint32: _ScalarType
3706
+ uint64: _ScalarType
3707
+ int1: _ScalarType
3708
+ int2: _ScalarType
3709
+ int4: _ScalarType
3710
+ uint1: _ScalarType
3711
+ uint2: _ScalarType
3712
+ uint4: _ScalarType
3713
+ float_: _ScalarType
3714
+ float16: _ScalarType
3715
+ float32: _ScalarType
3716
+ float64: _ScalarType
3717
+ bfloat16: _ScalarType
3718
+ single: _ScalarType
3719
+ double: _ScalarType
3720
+ csingle: _ScalarType
3721
+ cdouble: _ScalarType
3722
+ complex_: _ScalarType
3723
+ complex64: _ScalarType
3724
+ complex128: _ScalarType
3725
+ float4_e2m1fn: _ScalarType
3726
+ float6_e2m3fn: _ScalarType
3727
+ float6_e3m2fn: _ScalarType
3728
+ float8_e3m4: _ScalarType
3729
+ float8_e4m3: _ScalarType
3730
+ float8_e4m3b11fnuz: _ScalarType
3731
+ float8_e4m3fn: _ScalarType
3732
+ float8_e4m3fnuz: _ScalarType
3733
+ float8_e5m2: _ScalarType
3734
+ float8_e5m2fnuz: _ScalarType
3735
+ float8_e8m0fnu: _ScalarType
@@ -13,6 +13,7 @@ from jax._shapes import (
13
13
  double_sided_maxwell_shape,
14
14
  event_sample_shape,
15
15
  event_shape,
16
+ multinomial_sample_shape,
16
17
  multinomial_shape,
17
18
  multivariate_normal_shape,
18
19
  orthogonal_shape,
@@ -913,7 +914,7 @@ def multinomial[NShape: _Shape = [], PShape: _Shape = []](
913
914
  shape: None = None,
914
915
  dtype: DTypeLike | None = None,
915
916
  unroll: int | bool = 1,
916
- ) -> _Array[multinomial_shape(NShape, PShape, PShape)]: ...
917
+ ) -> _Array[multinomial_shape(NShape, PShape)]: ...
917
918
  @overload
918
919
  def multinomial[NShape: _Shape = [], PShape: _Shape = [], Shape: _Shape = []](
919
920
  key: _Array[IntTuple],
@@ -923,7 +924,7 @@ def multinomial[NShape: _Shape = [], PShape: _Shape = [], Shape: _Shape = []](
923
924
  shape: Shape,
924
925
  dtype: DTypeLike | None = None,
925
926
  unroll: int | bool = 1,
926
- ) -> _Array[multinomial_shape(NShape, PShape, Shape)]: ...
927
+ ) -> _Array[multinomial_sample_shape(NShape, PShape, Shape)]: ...
927
928
  @overload
928
929
  def multinomial(
929
930
  key: _Array[IntTuple],
@@ -0,0 +1,36 @@
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ #
3
+ # This source code is licensed under the MIT license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ from jax._array import Array
7
+ from jax._src.core import ManualAxisType as ManualAxisType
8
+ from jax._src.mesh import (
9
+ AbstractDevice as AbstractDevice,
10
+ AbstractMesh as AbstractMesh,
11
+ AxisType as AxisType,
12
+ get_abstract_mesh as get_abstract_mesh,
13
+ Mesh as Mesh,
14
+ use_abstract_mesh as use_abstract_mesh,
15
+ )
16
+ from jax._src.partition_spec import (
17
+ PartitionSpec as PartitionSpec,
18
+ UnreducedKind as UnreducedKind,
19
+ )
20
+ from jax._src.pjit import (
21
+ auto_axes as auto_axes,
22
+ explicit_axes as explicit_axes,
23
+ )
24
+ from jax._src.sharding import Sharding as Sharding
25
+ from jax._src.sharding_impls import (
26
+ get_mesh as get_mesh,
27
+ make_single_device_sharding as make_single_device_sharding,
28
+ NamedSharding as NamedSharding,
29
+ set_mesh as set_mesh,
30
+ SingleDeviceSharding as SingleDeviceSharding,
31
+ )
32
+ from shape_extensions import IntTuple
33
+
34
+ def reshard[Shape: IntTuple](
35
+ xs: Array[Shape], out_shardings: NamedSharding | PartitionSpec
36
+ ) -> Array[Shape]: ...
@@ -0,0 +1,14 @@
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ #
3
+ # This source code is licensed under the MIT license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ from typing import Any
7
+
8
+ import numpy as np
9
+ from jax._array import Array
10
+ from jax._src.typing import DTypeLike as DTypeLike
11
+
12
+ type ArrayLike = (
13
+ Array[Any] | np.ndarray[Any] | np.bool_ | np.number | bool | int | float | complex
14
+ )
@@ -8,7 +8,7 @@ description = "JAX type stubs with array-shape tracking, maintained by Pyrefly"
8
8
  readme = "README.md"
9
9
  requires-python = ">=3.12"
10
10
  license = { file = "LICENSE" }
11
- version = "1.4.0.dev3"
11
+ version = "1.5.0.dev1"
12
12
  keywords = ["typechecker", "typechecking", "stubs", "jax"]
13
13
  classifiers = [
14
14
  "Development Status :: 4 - Beta",
@@ -18,7 +18,9 @@ classifiers = [
18
18
  "Typing :: Stubs Only",
19
19
  ]
20
20
  dependencies = [
21
- "pyrefly-shape-extensions==1.4.0.dev3",
21
+ "pyrefly-shape-extensions==1.5.0.dev1",
22
+ "pyrefly-numpy-stubs==1.5.0.dev1",
23
+ "jax",
22
24
  ]
23
25
 
24
26
  [project.urls]