pytorch/pytorch · error · TypeError

expected a dimension specifyer but found {repr(s)}

Error message

expected a dimension specifyer but found {repr(s)}

What it means

Inside Tensor.index(), each entry of the dims argument is normalized with _wrap_dim; if the result 'is none' (the value could not be interpreted as a positional int or a Dim), TypeError('expected a dimension specifyer but found {s}') is raised. Valid specifyers are ints and Dim objects.

Source

Thrown at functorch/dim/__init__.py:738

        | list[int | slice | torch.Tensor],
    ) -> _Tensor:
        """
        Index tensor using first-class dimensions.
        """
        from ._dim_entry import _match_levels
        from ._getsetitem import getsetitem_flat, invoke_getitem
        from ._wrap import _wrap_dim

        # Helper to check if obj is a dimpack (tuple/list) and extract items
        def maybe_dimpack(obj: Any, check_first: bool = False) -> tuple[Any, bool]:
            if isinstance(obj, (tuple, list)):
                return list(obj), True
            return None, False

        def parse_dim_entry(s: Any) -> Any:
            d = _wrap_dim(s, self.ndim, False)
            if d.is_none():
                raise TypeError(f"expected a dimension specifyer but found {repr(s)}")
            return d

        # Helper for dimension not present errors
        def dim_not_present(d: Any) -> None:
            if d.is_positional():
                raise TypeError(
                    f"dimension {d.position() + self.ndim} not in tensor of {self.ndim} dimensions"
                )
            else:
                raise TypeError(f"dimension {repr(d.dim())} not in tensor")

        dims_list: list[int | Dim] = []
        indices_list: list[int | slice | torch.Tensor] = []

        lhs_list = isinstance(dims, (tuple, list))
        rhs_list = isinstance(indices, (tuple, list))

        if lhs_list and rhs_list:

View on GitHub (pinned to dcd2ecae77)

Solutions

1. Pass only int positions or Dim objects created by dims()/dimlists()
2. Convert numpy scalars: t.index(int(np_axis), ...)
3. For named axes use actual dims: batch = dims(1); t.index(batch, 0)

Example fix

# before
out = t.index('batch', 0)  # TypeError: string not a dimension specifyer

# after
batch = dims(1)
out = t.index(batch, 0)

from functorch.dim import Dim
if not isinstance(s, (int, Dim)):
    raise TypeError(f'dimension specifyer must be int or Dim, got {type(s).__name__}')

from functorch.dim import Dim
import numpy as np
def is_dim_specifyer(s) -> bool:
    if isinstance(s, np.integer):
        s = int(s)
    return isinstance(s, (int, Dim)) and not isinstance(s, bool)

When it happens

Trigger: Passing None, a string, a float, a numpy scalar, or other objects as a dimension to tensor.index(...) / the dims part of first-class-dimension indexing, e.g. t.index('batch', 0) or t.index(None, slice(None)).

Common situations: Mixing string axis names (pandas/xarray habits) with functorch dims, passing unwrapped numpy ints, or None leaking in from optional config for a dim name.

Related errors

• dimension {d.position() + self.ndim} not in tensor of {self.
• dimension {repr(d.dim())} not in tensor
• dims ({len(dims_seq)}) and indices ({len(indices_seq)}) must
• expected a sequence
• expected an int or a slice

AI-assisted analysis of pytorch/pytorch@dcd2ecae77 (2026-08-14). Data as JSON: /api/errors/9cd87803f06c0c00. Report an issue: GitHub.