jax-ml/jax · error · TypeError

can only specify one unknown axis size with a `-1` value…

Error message

can only specify one unknown axis size with a `-1` value, got {orig_newshape}

What it means

When reshaping with a -1 placeholder, only one axis may be unspecified because its size is inferred from the others. `_compute_newshape` counts int entries equal to -1 and raises if there is more than one. This mirrors NumPy's rule and is checked before any shape arithmetic.

Solutions

  1. Specify all but one dimension explicitly
  2. Compute the inferred dimension yourself: `arr.reshape(arr.shape[0], -1)` for batch-major data

Example fix

# before
arr.reshape(-1, -1)

# after
arr.reshape(arr.shape[0], -1)
Defensive patterns

Strategy: validation

Validate before calling

def count_neg1s(shape):
    return sum(1 for d in shape if isinstance(d, int) and d == -1)

def assert_valid_reshape(shape):
    assert count_neg1s(shape) <= 1, 'at most one -1 allowed'

Prevention

When it happens

Trigger: Calling reshape with two or more -1s, e.g. `arr.reshape(-1, -1)` or `jnp.reshape(arr, (-1, 3, -1))`.

Common situations: Dynamically built reshape shapes where a batch axis and a feature axis are both left as -1; typos when copying shapes.

Related errors


AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27). Data as JSON: /api/errors/15c477a3aca0b81c. Report an issue: GitHub.

Appendix: source

Thrown at jax/_src/numpy/array_methods.py:525

  Refer to :func:`jax.numpy.var` for full documentation.
  """
  return reductions.var(self, axis=axis, dtype=dtype, out=out, ddof=ddof,
                        keepdims=keepdims, where=where, correction=correction)


def _compute_newshape(arr: Array, newshape: DimSize | Shape) -> Shape:
  """Fixes a -1 value in newshape, if present."""
  orig_newshape = newshape  # for error messages
  try:
    iter(newshape)  # pyrefly: ignore[no-matching-overload]
  except TypeError:
    newshape = [newshape]
  else:
    newshape: Sequence[DimSize]  # pyrefly: ignore[redefinition]
  newshape = core.canonicalize_shape(newshape)
  neg1s = [i for i, d in enumerate(newshape) if type(d) is int and d == -1]
  if len(neg1s) > 1:
    raise TypeError("can only specify one unknown axis size with a `-1` value, "
                    f"got {orig_newshape}")
  if neg1s:
    i, = neg1s
    other_sizes = (*newshape[:i], *newshape[i+1:])
    if (all(isinstance(d, int) for d in (*arr.shape, *other_sizes)) and
        arr.size % math.prod(other_sizes) != 0):
      raise TypeError(f"cannot reshape array of shape {arr.shape} (size {arr.size}) "
                      f"into shape {orig_newshape} because the product of "
                      f"specified axis sizes ({math.prod(other_sizes)}) does "
                      f"not evenly divide {arr.size}")
    sz = core.cancel_divide_tracers(arr.shape, other_sizes)
    if sz is not None:
      return (*newshape[:i], sz, *newshape[i+1:])
  else:
    if (all(isinstance(d, int) for d in (*arr.shape, *newshape)) and
        arr.size != math.prod(newshape)):
      raise TypeError(f"cannot reshape array of shape {arr.shape} (size {arr.size}) "
                      f"into shape {orig_newshape} (size {math.prod(newshape)})")

View on GitHub (pinned to 1e1c6a8fc0)