pyrefly-torch-stubs 1.3.0.dev3__tar.gz → 1.3.0.dev4__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 (23) hide show
  1. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/PKG-INFO +2 -2
  2. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/pyproject.toml +2 -2
  3. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/__init__.pyi +176 -100
  4. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/_shapes.pyi +562 -367
  5. pyrefly_torch_stubs-1.3.0.dev4/torch-stubs/fft.pyi +94 -0
  6. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/nn/__init__.pyi +191 -114
  7. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/nn/functional.pyi +244 -121
  8. pyrefly_torch_stubs-1.3.0.dev3/torch-stubs/fft.pyi +0 -142
  9. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/LICENSE +0 -0
  10. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/README.md +0 -0
  11. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/distributions/__init__.pyi +0 -0
  12. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/distributions/beta.pyi +0 -0
  13. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/distributions/categorical.pyi +0 -0
  14. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/distributions/constraints.pyi +0 -0
  15. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/distributions/transformed_distribution.pyi +0 -0
  16. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/distributions/transforms.pyi +0 -0
  17. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/linalg.pyi +0 -0
  18. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/nn/attention/__init__.pyi +0 -0
  19. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/nn/attention/flex_attention.pyi +0 -0
  20. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/nn/init.pyi +0 -0
  21. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/optim/__init__.pyi +0 -0
  22. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/py.typed +0 -0
  23. {pyrefly_torch_stubs-1.3.0.dev3 → pyrefly_torch_stubs-1.3.0.dev4}/torch-stubs/quantization/__init__.pyi +0 -0
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.5
2
2
  Name: pyrefly-torch-stubs
3
- Version: 1.3.0.dev3
3
+ Version: 1.3.0.dev4
4
4
  Summary: PyTorch type stubs with tensor-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 +33,7 @@ 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==1.3.0.dev3
36
+ Requires-Dist: pyrefly-shape-extensions==1.3.0.dev4
37
37
  Description-Content-Type: text/markdown
38
38
 
39
39
  # pyrefly-torch-stubs
@@ -8,7 +8,7 @@ description = "PyTorch type stubs with tensor-shape tracking, maintained by Pyre
8
8
  readme = "README.md"
9
9
  requires-python = ">=3.12"
10
10
  license = { file = "LICENSE" }
11
- version = "1.3.0.dev3"
11
+ version = "1.3.0.dev4"
12
12
  keywords = ["typechecker", "typechecking", "stubs", "torch", "pytorch"]
13
13
  classifiers = [
14
14
  "Development Status :: 4 - Beta",
@@ -18,7 +18,7 @@ classifiers = [
18
18
  "Typing :: Stubs Only",
19
19
  ]
20
20
  dependencies = [
21
- "pyrefly-shape-extensions==1.3.0.dev3",
21
+ "pyrefly-shape-extensions==1.3.0.dev4",
22
22
  ]
23
23
 
24
24
  [project.urls]
@@ -6,18 +6,27 @@
6
6
  """
7
7
  Comprehensive type stubs for PyTorch with shape inference.
8
8
 
9
- New shape inference rules use `@type_shape_dsl_function` definitions from
10
- `torch/_shapes.pyi`, called directly from public return annotations. Some existing rules still use
11
- the older `@uses_shape_dsl(...)` decorator while the stub library is migrated; new rules should not
12
- extend that legacy path.
9
+ Shape inference is expressed through type-level functions in annotations. Library-specific
10
+ functions are defined in `torch/_shapes.pyi`.
13
11
  """
14
12
 
15
13
  import builtins
16
14
  from collections.abc import Sequence
17
- from typing import Any, overload, Self, TYPE_CHECKING
15
+ from types import EllipsisType
16
+ from typing import Any, overload, Self, TYPE_CHECKING, Unpack
18
17
 
19
18
  import shape_extensions
20
- from shape_extensions import broadcast, Elements, Flag, IntTuple, IntVar, uses_shape_dsl
19
+ from shape_extensions import (
20
+ broadcast,
21
+ Elements,
22
+ Flag,
23
+ Index,
24
+ index_shape,
25
+ IntTuple,
26
+ IntTuples,
27
+ IntVar,
28
+ MapIntTuples,
29
+ )
21
30
 
22
31
  # `Generator` is not defined anywhere in this package, and resolving it relies
23
32
  # on how a partial stub package is looked up. The `py.typed` file here contains
@@ -34,20 +43,18 @@ from torch._C import Generator
34
43
  from torch._shapes import (
35
44
  arange_extent,
36
45
  arange_step_extent,
37
- cat_ir,
38
- chunk_ir,
46
+ cat_shape,
47
+ chunk_shapes,
39
48
  diag_embed_shape,
40
49
  dim_shape,
41
50
  eig_shape,
42
- einsum_ir,
51
+ einsum_shape,
43
52
  expand_shape,
44
53
  flatten_shape,
45
54
  index_select_shape,
46
- item_ir,
47
55
  matmul_shape,
48
- movedim_input_ir,
49
- movedim_ir,
50
56
  movedim_scalar_shape,
57
+ movedim_tuple_shape,
51
58
  multinomial_shape,
52
59
  numel_shape,
53
60
  permute_shape,
@@ -62,9 +69,10 @@ from torch._shapes import (
62
69
  select_shape,
63
70
  size_dim_shape,
64
71
  slogdet_shape,
65
- split_ir,
72
+ split_sections_shapes,
73
+ split_size_shapes,
66
74
  squeeze_shape,
67
- stack_ir,
75
+ stack_shape,
68
76
  tensordot_shape,
69
77
  tile_shape,
70
78
  topk_shape,
@@ -81,6 +89,7 @@ __all__ = ["Tensor"]
81
89
 
82
90
  type _Shape = IntTuple
83
91
  type _AnyShape = tuple[Any, ...]
92
+ type _BasicIndex = builtins.int | slice | list[builtins.int] | None | EllipsisType
84
93
 
85
94
  # ============================================================================
86
95
  # Device Type
@@ -111,7 +120,7 @@ ops: Any
111
120
  # Tensor Class
112
121
  # ============================================================================
113
122
 
114
- @shape_extensions.shaped_array(shape="Shape")
123
+ @shape_extensions.shaped_array(shape="Shape", builtin_indexing=False)
115
124
  class Tensor[Shape: _Shape = _AnyShape]:
116
125
  """
117
126
  PyTorch Tensor with shape type parameter.
@@ -134,24 +143,38 @@ class Tensor[Shape: _Shape = _AnyShape]:
134
143
  imag: Self # Imaginary part of complex tensor (shape-preserving)
135
144
  # Note: Use .dim() method for rank (ndim removed in favor of dim())
136
145
  # ==== Indexing ====
146
+ @overload
147
+ def __getitem__[I: Index](
148
+ self: Tensor[Shape], index: I
149
+ ) -> Tensor[index_shape(Shape, I)]: ...
150
+ @overload
137
151
  def __getitem__(
138
152
  self: Tensor,
139
- index: int
140
- | slice
141
- | tuple[int | slice | Tensor | list[int] | None, ...]
153
+ index: _BasicIndex
154
+ | Sequence[builtins.int]
155
+ | Sequence[Sequence[builtins.int]]
142
156
  | Tensor
143
- | list[int],
144
- ) -> Tensor:
145
- """Index into tensor. Shape inference via meta-shape: torch.Tensor.__getitem__"""
146
- ...
147
-
157
+ | tuple[
158
+ _BasicIndex
159
+ | Sequence[builtins.int]
160
+ | Sequence[Sequence[builtins.int]]
161
+ | Tensor,
162
+ ...,
163
+ ],
164
+ ) -> Tensor[IntTuple]: ...
148
165
  def __setitem__(
149
166
  self: Tensor,
150
- index: int
151
- | slice
152
- | tuple[int | slice | Tensor | list[int] | None, ...]
167
+ index: _BasicIndex
168
+ | Sequence[builtins.int]
169
+ | Sequence[Sequence[builtins.int]]
153
170
  | Tensor
154
- | list[int],
171
+ | tuple[
172
+ _BasicIndex
173
+ | Sequence[builtins.int]
174
+ | Sequence[Sequence[builtins.int]]
175
+ | Tensor,
176
+ ...,
177
+ ],
155
178
  value: Tensor | int | float,
156
179
  ) -> None:
157
180
  """Set values in tensor via indexing. Mutates tensor in-place."""
@@ -403,36 +426,36 @@ class Tensor[Shape: _Shape = _AnyShape]:
403
426
  @overload
404
427
  def repeat_interleave[
405
428
  Shape: IntTuple,
406
- Repeats: IntVar,
407
- OutputSize: IntVar,
429
+ Repeats: _Int,
430
+ OutputSize: _Int,
408
431
  Dim: Flag[builtins.int | None],
409
432
  ](
410
433
  self: Tensor[Shape],
411
- repeats: _Int[Repeats],
434
+ repeats: Repeats,
412
435
  dim: Dim = None,
413
436
  *,
414
- output_size: _Int[OutputSize],
437
+ output_size: OutputSize,
415
438
  ) -> Tensor[repeat_interleave_checked_shape(Shape, Repeats, OutputSize, Dim)]: ...
416
439
  @overload
417
440
  def repeat_interleave[
418
441
  Shape: IntTuple,
419
- OutputSize: IntVar,
442
+ OutputSize: _Int,
420
443
  Dim: Flag[builtins.int | None],
421
444
  ](
422
445
  self: Tensor[Shape],
423
446
  repeats: Tensor,
424
447
  dim: Dim = None,
425
448
  *,
426
- output_size: _Int[OutputSize],
449
+ output_size: OutputSize,
427
450
  ) -> Tensor[repeat_interleave_output_shape(Shape, OutputSize, Dim)]: ...
428
451
  @overload
429
452
  def repeat_interleave[
430
453
  Shape: IntTuple,
431
- Repeats: IntVar,
454
+ Repeats: _Int,
432
455
  Dim: Flag[builtins.int | None],
433
456
  ](
434
457
  self: Tensor[Shape],
435
- repeats: _Int[Repeats],
458
+ repeats: Repeats,
436
459
  dim: Dim = None,
437
460
  *,
438
461
  output_size: None = None,
@@ -570,9 +593,8 @@ class Tensor[Shape: _Shape = _AnyShape]:
570
593
  """Enable/disable gradient tracking in-place. Shape-preserving."""
571
594
  ...
572
595
 
573
- @uses_shape_dsl(item_ir)
574
- def item(self: Tensor) -> float | int:
575
- """Returns Python scalar from 0-dimensional tensor. Shape inference via meta-shape: torch.Tensor.item"""
596
+ def item(self: Tensor[[]]) -> builtins.float | builtins.int:
597
+ """Return a Python scalar from a rank-zero tensor."""
576
598
  ...
577
599
 
578
600
  def tolist(self: Tensor) -> Any:
@@ -594,31 +616,54 @@ class Tensor[Shape: _Shape = _AnyShape]:
594
616
  def narrow[
595
617
  Shape: IntTuple,
596
618
  Dim: Flag[builtins.int],
597
- Length: IntVar,
619
+ Length: _Int,
598
620
  ](
599
- self: Tensor[Shape], dim: Dim, start: int, length: _Int[Length]
621
+ self: Tensor[Shape], dim: Dim, start: int, length: Length
600
622
  ) -> Tensor[replace_axis_extent(Shape, Dim, Length)]:
601
623
  """Narrow tensor along dimension. Shape inference via meta-shape: torch.Tensor.narrow"""
602
624
  ...
603
625
 
604
- @uses_shape_dsl(split_ir)
605
626
  @overload
606
- def split(
607
- self: Tensor, split_size_or_sections: int, dim: int = 0
608
- ) -> tuple[Tensor, ...]:
609
- """Split tensor into chunks. Shape inference via meta-shape: torch.Tensor.split"""
627
+ def split[
628
+ Shape: IntTuple,
629
+ SplitSize: _Int,
630
+ Dim: Flag[builtins.int],
631
+ ](
632
+ self: Tensor[Shape], split_size_or_sections: SplitSize, dim: Dim = 0
633
+ ) -> MapIntTuples[lambda S: Tensor[S], split_size_shapes(Shape, SplitSize, Dim)]:
634
+ """Split tensor into chunks. Shape inference via the type-level DSL."""
610
635
  ...
611
636
 
637
+ @overload
638
+ def split[
639
+ Shape: IntTuple,
640
+ Sections: IntTuple,
641
+ Dim: Flag[builtins.int],
642
+ ](
643
+ self: Tensor[Shape], split_size_or_sections: Sections, dim: Dim = 0
644
+ ) -> MapIntTuples[lambda S: Tensor[S], split_sections_shapes(Shape, Sections, Dim)]:
645
+ """Split tensor into variable-sized chunks. Shape inference via the type-level DSL."""
646
+ ...
647
+
648
+ # A list is mutable, so V2 cannot preserve its element values; this trailing
649
+ # overload keeps the documented list spelling checking without shape inference.
650
+ # TODO(stroxler): Route lists through `split_sections_shapes` once list values
651
+ # survive to the type level.
612
652
  @overload
613
653
  def split(
614
654
  self: Tensor, split_size_or_sections: list[int], dim: int = 0
615
655
  ) -> tuple[Tensor, ...]:
616
- """Split tensor into variable-sized chunks. Shape inference via meta-shape: torch.Tensor.split"""
656
+ """Split tensor into variable-sized chunks. Shape inference unavailable for lists."""
617
657
  ...
618
658
 
619
- @uses_shape_dsl(chunk_ir)
620
- def chunk(self: Tensor, chunks: int, dim: int = 0) -> tuple[Tensor, ...]:
621
- """Split tensor into chunks. Shape inference via meta-shape: torch.Tensor.chunk"""
659
+ def chunk[
660
+ Shape: IntTuple,
661
+ Chunks: _Int,
662
+ Dim: Flag[builtins.int],
663
+ ](
664
+ self: Tensor[Shape], chunks: Chunks, dim: Dim = 0
665
+ ) -> MapIntTuples[lambda S: Tensor[S], chunk_shapes(Shape, Chunks, Dim)]:
666
+ """Split tensor into chunks. Shape inference via the type-level DSL."""
622
667
  ...
623
668
 
624
669
  def index_select[
@@ -666,11 +711,10 @@ class Tensor[Shape: _Shape = _AnyShape]:
666
711
  """Move a single dimension to a new position."""
667
712
  ...
668
713
 
669
- @uses_shape_dsl(movedim_ir)
670
714
  @overload
671
- def movedim(
672
- self: Tensor, source: tuple[int, ...], destination: tuple[int, ...]
673
- ) -> Tensor:
715
+ def movedim[Shape: IntTuple, Source: IntTuple, Destination: IntTuple](
716
+ self: Tensor[Shape], source: Source, destination: Destination
717
+ ) -> Tensor[movedim_tuple_shape(Shape, Source, Destination)]:
674
718
  """Move multiple dimensions to new positions. Shape inference via meta-shape: torch.Tensor.movedim"""
675
719
  ...
676
720
 
@@ -685,11 +729,10 @@ class Tensor[Shape: _Shape = _AnyShape]:
685
729
  """Alias for movedim with a single source and destination."""
686
730
  ...
687
731
 
688
- @uses_shape_dsl(movedim_ir)
689
732
  @overload
690
- def moveaxis(
691
- self: Tensor, source: tuple[int, ...], destination: tuple[int, ...]
692
- ) -> Tensor:
733
+ def moveaxis[Shape: IntTuple, Source: IntTuple, Destination: IntTuple](
734
+ self: Tensor[Shape], source: Source, destination: Destination
735
+ ) -> Tensor[movedim_tuple_shape(Shape, Source, Destination)]:
693
736
  """Alias for movedim. Shape inference via meta-shape: torch.Tensor.moveaxis"""
694
737
  ...
695
738
 
@@ -911,9 +954,9 @@ class Tensor[Shape: _Shape = _AnyShape]:
911
954
  """Mode along dimension. Returns (values, indices). Shape inference via meta-shape: torch.Tensor.mode"""
912
955
  ...
913
956
 
914
- def topk[Shape: IntTuple, K: IntVar, Dim: Flag[builtins.int]](
957
+ def topk[Shape: IntTuple, K: _Int, Dim: Flag[builtins.int]](
915
958
  self: Tensor[Shape],
916
- k: _Int[K],
959
+ k: K,
917
960
  dim: Dim = -1,
918
961
  largest: bool = True,
919
962
  sorted: bool = True,
@@ -1471,9 +1514,9 @@ class Tensor[Shape: _Shape = _AnyShape]:
1471
1514
  """Sample from Bernoulli distribution in-place. Shape inference via generic fixture signature."""
1472
1515
  ...
1473
1516
 
1474
- def multinomial[Shape: IntTuple, NumSamples: IntVar](
1517
+ def multinomial[Shape: IntTuple, NumSamples: _Int](
1475
1518
  self: Tensor[Shape],
1476
- num_samples: _Int[NumSamples],
1519
+ num_samples: NumSamples,
1477
1520
  replacement: bool = False,
1478
1521
  ) -> Tensor[multinomial_shape(Shape, NumSamples)]:
1479
1522
  """Sample from multinomial distribution. Shape inference via meta-shape: torch.Tensor.multinomial"""
@@ -1513,18 +1556,21 @@ def matmul[Left: IntTuple, Right: IntTuple](
1513
1556
  """Matrix multiplication function. Shape inference via meta-shape: torch.matmul"""
1514
1557
  ...
1515
1558
 
1516
- @uses_shape_dsl(cat_ir)
1517
- def cat(tensors: list[Tensor] | tuple[Tensor, ...], dim: int = 0) -> Tensor:
1559
+ def cat[Shapes: IntTuples, Dim: Flag[builtins.int]](
1560
+ tensors: MapIntTuples[lambda S: Tensor[S], Shapes], dim: Dim = 0
1561
+ ) -> Tensor[cat_shape(Shapes, Dim)]:
1518
1562
  """Concatenate tensors. Shape inference via meta-shape: torch.cat"""
1519
1563
  ...
1520
1564
 
1521
- @uses_shape_dsl(cat_ir)
1522
- def concat(tensors: list[Tensor] | tuple[Tensor, ...], dim: int = 0) -> Tensor:
1565
+ def concat[Shapes: IntTuples, Dim: Flag[builtins.int]](
1566
+ tensors: MapIntTuples[lambda S: Tensor[S], Shapes], dim: Dim = 0
1567
+ ) -> Tensor[cat_shape(Shapes, Dim)]:
1523
1568
  """Alias for concatenate/cat. Shape inference via meta-shape: torch.cat"""
1524
1569
  ...
1525
1570
 
1526
- @uses_shape_dsl(stack_ir)
1527
- def stack(tensors: list[Tensor] | tuple[Tensor, ...], dim: int = 0) -> Tensor:
1571
+ def stack[Shapes: IntTuples, Dim: Flag[builtins.int]](
1572
+ tensors: MapIntTuples[lambda S: Tensor[S], Shapes], dim: Dim = 0
1573
+ ) -> Tensor[stack_shape(Shapes, Dim)]:
1528
1574
  """Stack tensors (adds new dimension)."""
1529
1575
  ...
1530
1576
 
@@ -1568,32 +1614,32 @@ def unsqueeze[Shape: IntTuple, Dim: Flag[builtins.int]](
1568
1614
  @overload
1569
1615
  def repeat_interleave[
1570
1616
  Shape: IntTuple,
1571
- Repeats: IntVar,
1572
- OutputSize: IntVar,
1617
+ Repeats: _Int,
1618
+ OutputSize: _Int,
1573
1619
  Dim: Flag[builtins.int | None],
1574
1620
  ](
1575
1621
  input: Tensor[Shape],
1576
- repeats: _Int[Repeats],
1622
+ repeats: Repeats,
1577
1623
  dim: Dim = None,
1578
1624
  *,
1579
- output_size: _Int[OutputSize],
1625
+ output_size: OutputSize,
1580
1626
  ) -> Tensor[repeat_interleave_checked_shape(Shape, Repeats, OutputSize, Dim)]: ...
1581
1627
  @overload
1582
1628
  def repeat_interleave[
1583
1629
  Shape: IntTuple,
1584
- OutputSize: IntVar,
1630
+ OutputSize: _Int,
1585
1631
  Dim: Flag[builtins.int | None],
1586
1632
  ](
1587
1633
  input: Tensor[Shape],
1588
1634
  repeats: Tensor,
1589
1635
  dim: Dim = None,
1590
1636
  *,
1591
- output_size: _Int[OutputSize],
1637
+ output_size: OutputSize,
1592
1638
  ) -> Tensor[repeat_interleave_output_shape(Shape, OutputSize, Dim)]: ...
1593
1639
  @overload
1594
- def repeat_interleave[Shape: IntTuple, Repeats: IntVar, Dim: Flag[builtins.int | None]](
1640
+ def repeat_interleave[Shape: IntTuple, Repeats: _Int, Dim: Flag[builtins.int | None]](
1595
1641
  input: Tensor[Shape],
1596
- repeats: _Int[Repeats],
1642
+ repeats: Repeats,
1597
1643
  dim: Dim = None,
1598
1644
  *,
1599
1645
  output_size: None = None,
@@ -1831,7 +1877,7 @@ def full[Shape: IntTuple](size: Shape, fill_value: float) -> Tensor[Shape]:
1831
1877
  @overload
1832
1878
  def arange[End: IntVar](
1833
1879
  end: _Int[End], *, dtype: int | None = None, device: Any = None
1834
- ) -> Tensor[[arange_extent(End)]]:
1880
+ ) -> Tensor[[arange_extent(_Int[End])]]:
1835
1881
  """Create 1D tensor with range [0, end). Shape inference via meta-shape: torch.arange"""
1836
1882
  ...
1837
1883
 
@@ -1843,7 +1889,7 @@ def arange[Start: IntVar, End: IntVar, Step: Flag[builtins.int]](
1843
1889
  *,
1844
1890
  dtype: int | None = None,
1845
1891
  device: Any = None,
1846
- ) -> Tensor[[arange_step_extent(Start, End, Step)]]:
1892
+ ) -> Tensor[[arange_step_extent(_Int[Start], _Int[End], Step)]]:
1847
1893
  """Create 1D tensor with range [start, end) with step. Shape inference via meta-shape: torch.arange"""
1848
1894
  ...
1849
1895
 
@@ -1897,22 +1943,53 @@ def select[Shape: IntTuple, Dim: Flag[builtins.int]](
1897
1943
  """Select along dimension. Shape inference via meta-shape: torch.select"""
1898
1944
  ...
1899
1945
 
1900
- def narrow[Shape: IntTuple, Dim: Flag[builtins.int], Length: IntVar](
1901
- self: Tensor[Shape], dim: Dim, start: int, length: _Int[Length]
1946
+ def narrow[Shape: IntTuple, Dim: Flag[builtins.int], Length: _Int](
1947
+ self: Tensor[Shape], dim: Dim, start: int, length: Length
1902
1948
  ) -> Tensor[replace_axis_extent(Shape, Dim, Length)]:
1903
1949
  """Narrow tensor along dimension. Shape inference via meta-shape: torch.narrow"""
1904
1950
  ...
1905
1951
 
1906
- @uses_shape_dsl(split_ir)
1952
+ @overload
1953
+ def split[
1954
+ Shape: IntTuple,
1955
+ SplitSize: _Int,
1956
+ Dim: Flag[builtins.int],
1957
+ ](
1958
+ self: Tensor[Shape], split_size_or_sections: SplitSize, dim: Dim = 0
1959
+ ) -> MapIntTuples[lambda S: Tensor[S], split_size_shapes(Shape, SplitSize, Dim)]:
1960
+ """Split tensor into chunks. Shape inference via the type-level DSL."""
1961
+ ...
1962
+
1963
+ @overload
1964
+ def split[
1965
+ Shape: IntTuple,
1966
+ Sections: IntTuple,
1967
+ Dim: Flag[builtins.int],
1968
+ ](
1969
+ self: Tensor[Shape], split_size_or_sections: Sections, dim: Dim = 0
1970
+ ) -> MapIntTuples[lambda S: Tensor[S], split_sections_shapes(Shape, Sections, Dim)]:
1971
+ """Split tensor into variable-sized chunks. Shape inference via the type-level DSL."""
1972
+ ...
1973
+
1974
+ # A list is mutable, so V2 cannot preserve its element values; this trailing
1975
+ # overload keeps the documented list spelling checking without shape inference.
1976
+ # TODO(stroxler): Route lists through `split_sections_shapes` once list values
1977
+ # survive to the type level.
1978
+ @overload
1907
1979
  def split(
1908
- self: Tensor, split_size_or_sections: int, dim: int = 0
1980
+ self: Tensor, split_size_or_sections: list[int], dim: int = 0
1909
1981
  ) -> tuple[Tensor, ...]:
1910
- """Split tensor into chunks. Shape inference via meta-shape: torch.split"""
1982
+ """Split tensor into variable-sized chunks. Shape inference unavailable for lists."""
1911
1983
  ...
1912
1984
 
1913
- @uses_shape_dsl(chunk_ir)
1914
- def chunk(self: Tensor, chunks: int, dim: int = 0) -> tuple[Tensor, ...]:
1915
- """Split tensor into chunks. Shape inference via meta-shape: torch.chunk"""
1985
+ def chunk[
1986
+ Shape: IntTuple,
1987
+ Chunks: _Int,
1988
+ Dim: Flag[builtins.int],
1989
+ ](
1990
+ self: Tensor[Shape], chunks: Chunks, dim: Dim = 0
1991
+ ) -> MapIntTuples[lambda S: Tensor[S], chunk_shapes(Shape, Chunks, Dim)]:
1992
+ """Split tensor into chunks. Shape inference via the type-level DSL."""
1916
1993
  ...
1917
1994
 
1918
1995
  def index_select[
@@ -1960,11 +2037,10 @@ def movedim[
1960
2037
  """Move a single dimension to a new position."""
1961
2038
  ...
1962
2039
 
1963
- @uses_shape_dsl(movedim_input_ir)
1964
2040
  @overload
1965
- def movedim(
1966
- input: Tensor, source: tuple[int, ...], destination: tuple[int, ...]
1967
- ) -> Tensor:
2041
+ def movedim[Shape: IntTuple, Source: IntTuple, Destination: IntTuple](
2042
+ input: Tensor[Shape], source: Source, destination: Destination
2043
+ ) -> Tensor[movedim_tuple_shape(Shape, Source, Destination)]:
1968
2044
  """Move multiple dimensions to new positions. Shape inference via meta-shape: torch.movedim"""
1969
2045
  ...
1970
2046
 
@@ -1979,11 +2055,10 @@ def moveaxis[
1979
2055
  """Alias for movedim with a single source and destination."""
1980
2056
  ...
1981
2057
 
1982
- @uses_shape_dsl(movedim_input_ir)
1983
2058
  @overload
1984
- def moveaxis(
1985
- input: Tensor, source: tuple[int, ...], destination: tuple[int, ...]
1986
- ) -> Tensor:
2059
+ def moveaxis[Shape: IntTuple, Source: IntTuple, Destination: IntTuple](
2060
+ input: Tensor[Shape], source: Source, destination: Destination
2061
+ ) -> Tensor[movedim_tuple_shape(Shape, Source, Destination)]:
1987
2062
  """Alias for movedim. Shape inference via meta-shape: torch.moveaxis"""
1988
2063
  ...
1989
2064
 
@@ -2114,9 +2189,9 @@ def mode[Shape: IntTuple, Dim: Flag[builtins.int], Keepdim: Flag[builtins.bool]]
2114
2189
  """Mode along dimension. Returns (values, indices). Shape inference via meta-shape: torch.mode"""
2115
2190
  ...
2116
2191
 
2117
- def topk[Shape: IntTuple, K: IntVar, Dim: Flag[builtins.int]](
2192
+ def topk[Shape: IntTuple, K: _Int, Dim: Flag[builtins.int]](
2118
2193
  self: Tensor[Shape],
2119
- k: _Int[K],
2194
+ k: K,
2120
2195
  dim: Dim = -1,
2121
2196
  largest: bool = True,
2122
2197
  sorted: bool = True,
@@ -2590,9 +2665,10 @@ def tensordot(self: Tensor, other: Tensor, dims: tuple[list[int], list[int]]) ->
2590
2665
  """
2591
2666
  ...
2592
2667
 
2593
- @uses_shape_dsl(einsum_ir)
2594
- def einsum(spec: str, *operands: Tensor) -> Tensor:
2595
- """Einstein summation convention. Shape inference via meta-shape: torch.einsum"""
2668
+ def einsum[Spec: Flag[str], Shapes: IntTuples](
2669
+ spec: Spec, *operands: Unpack[MapIntTuples[lambda S: Tensor[S], Shapes]]
2670
+ ) -> Tensor[einsum_shape(Spec, Shapes)]:
2671
+ """Einstein summation convention. Shape inference via type-level DSL."""
2596
2672
  ...
2597
2673
 
2598
2674
  # Eigenvalue decomposition
@@ -2759,9 +2835,9 @@ def bernoulli[Shape: IntTuple](input: Tensor[Shape], p: float = 0.5) -> Tensor[S
2759
2835
  """Sample from Bernoulli distribution. Shape inference via generic fixture signature."""
2760
2836
  ...
2761
2837
 
2762
- def multinomial[Shape: IntTuple, NumSamples: IntVar](
2838
+ def multinomial[Shape: IntTuple, NumSamples: _Int](
2763
2839
  input: Tensor[Shape],
2764
- num_samples: _Int[NumSamples],
2840
+ num_samples: NumSamples,
2765
2841
  replacement: bool = False,
2766
2842
  ) -> Tensor[multinomial_shape(Shape, NumSamples)]:
2767
2843
  """Sample from multinomial distribution. Shape inference via meta-shape: torch.multinomial"""