einsum_op

→TensorImplPtr

int einsum_op(const int & pattern, const int & operands)

source edit

Python wrappers

pylucid.einops.einsumfunction

General tensor contraction via an Einstein summation equation.

The equation uses the same syntax as NumPy's einsum: single-character labels separated by commas for each input, optionally followed by -> and the output labels. If -> is omitted, the output labels are the sorted set of labels that appear exactly once across all inputs (implicit form). Repeated labels in a single operand take the diagonal; labels shared across operands trigger contraction; labels dropped from the output are summed out.

Math

y_{\ell_\text{out}} = \sum_{\ell_\text{contract}} \prod_{k} x^{(k)}_{\ell^{(k)}}

See Also

einops_rearrange_op, einops_reduce_op, einops_repeat_op

Parameters

patternstd::string

Einsum equation, e.g. "ij,jk->ik" for matrix multiplication or "bij,bjk->bik" for batched matmul.

operandsstd::vector<TensorImplPtr>

The input tensors, one per comma-separated label group. All operands must share the same device and dtype.

Returns

TensorImplPtr

Contracted tensor of the shape implied by the output labels.

Raises

EinopsPatternError

If the equation is malformed, label counts do not match operand dimensions, or contracted dimensions disagree in size.

DeviceMismatch

If operands live on different devices.

DtypeMismatch

If operands have different dtypes.

Notes

Ellipsis (...) is not yet supported. Use explicit labels for batch dimensions in the meantime.

Examples

"ij,jk->ik"    — matrix multiplication.
"bhij,bhjk->bhik" — multi-head batched attention contraction.