jax-ml/jax · error · ValueError
One Hot indexing is only supported for up to 50 leading…
Error message
One Hot indexing is only supported for up to 50 leading dimensions.
What it means
When the array is multi-dimensional, take_along_axis uses a one-hot einsum trick that labels leading axes with ASCII letters; with 52 letters available, only up to 50 leading dimensions before the taken axis are supported.
Solutions
- Reduce array rank: move/reshape axes so the take axis is early or the array is processed per-slice
- Take with a reshaped 2-D view: flatten all non-taken axes into one, take, then reshape back
- Fix upstream logic creating absurdly high-rank arrays (often a vmap/stack bug)
Example fix
// before y = jnp.take_along_axis(a, idx, axis=60) # rank-61 array // after lead = int(np.prod(a.shape[:60])) y = jnp.take_along_axis(a.reshape(lead, -1), idx.reshape(lead, -1), axis=1).reshape(a.shape)
Defensive patterns
Strategy: validation
Validate before calling
axis_int = axis if axis >= 0 else a.ndim + axis assert axis_int <= 50 or a.ndim == 1, 'take_along_axis supports at most 50 leading dims'
Prevention
- Keep take axes early in high-rank arrays
- Flatten leading axes before take_along_axis for huge ranks
- Investigate rank growth from vmap/stack bugs
When it happens
Trigger: Calling jnp.take_along_axis on arrays with rank > ~52 where the take axis is late (axis_int > 50), triggering the einsum-based path instead of the simple 1-D path.
Common situations: Extremely high-rank tensors from excessive stacking/vmap nesting; pathological shapes from broadcasting bugs that inflated rank.
Related errors
- Async copies only support striding up to 5 dimensions
- indices and arr must have the same number of dimensions;
- Run out of characters for batch dimension in einsum.
- take_along_axis indices must be 1D if axis=None, got shape
- take_along_axis indices must be of integer type, got
AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27).
Data as JSON: /api/errors/c724e9b839d5ea95.
Report an issue: GitHub.
Appendix: source
Thrown at jax/_src/numpy/indexing.py:903
if axis_size == 0:
return lax.full(out_shape, 0, a.dtype)
index_dtype = lax_utils.index_dtype_for_axis_size(
dtypes.dtype(indices), axis_size, wrap_negative_indices
)
indices = lax.convert_element_type(indices, index_dtype)
if wrap_negative_indices:
indices = _normalize_index(indices, axis_size)
if mode == "one_hot":
from jax import nn # pyrefly: ignore[missing-import]
hot = nn.one_hot(indices, axis_size, dtype=np.bool_)
if a.ndim == 1:
return einsum.einsum("...b,b->...", hot, a, preferred_element_type=a.dtype)
if axis_int > len(string.ascii_letters) - 2:
raise ValueError(
"One Hot indexing is only supported for up to 50 leading dimensions."
)
labels = "".join([string.ascii_letters[i] for i in range(axis_int)])
eq = labels + "y...z," + labels + "z...->" + labels + "y..."
return einsum.einsum(
eq,
hot,
a,
precision=lax.Precision.HIGHEST,
preferred_element_type=a.dtype,
)
index_dims = [i for i, idx in enumerate(idx_shape) if i == axis_int or not core.definitely_equal(idx, 1)]
gather_index_shape = tuple(np.array(out_shape)[index_dims]) + (1,)
gather_indices = lax.reshape(indices, gather_index_shape)
slice_sizes = []
offset_dims = []View on GitHub (pinned to 1e1c6a8fc0)