cummin_op

→TensorImplPtr

int cummin_op(const int & a, int axis)

Inclusive cumulative minimum along a single axis, $y_k = \min_{j \le k} x_j$ .

Symmetric to cummax_op: gradient flows only to the position that first achieved the running minimum at each $k$ .

Math

y_k = \min_{j \le k} x_j, \qquad \frac{\partial \mathcal{L}}{\partial x_i} = \sum_{k \ge i} dy_k \cdot \mathbb{1}[\,i = \arg\min_{j \le k} x_j\,].

See Also

cummax_op — symmetric maximum scan.

aTensorImplPtr

Input tensor of any non-scalar shape; backward supports F32 / F64.

axisint

Axis along which to accumulate (negative wraps).

TensorImplPtr

Output tensor of the same shape and dtype as a.