addcdiv

→Tensor

addcdiv(input: Tensor, t1: Tensor, t2: Tensor, value: float = ...)

source

Element-wise fused divide-add.

Computes $\text{input} + \text{value} \cdot (t_1 / t_2)$ in one expression. Common in optimiser updates (e.g. Adam's parameter step uses the form param -= lr * m / sqrt(v)).

Parameters

inputTensor

Accumulator tensor.

t1Tensor

Numerator. Must broadcast with input and t2.

t2Tensor

Denominator. Must broadcast with input and t1.

valuefloat

Scalar multiplier on the quotient. Defaults to 1.0.

Returns

Tensor

Tensor with the broadcast shape of the inputs.

Notes

Element-wise definition:

\text{out}_i = \text{input}_i + \text{value} \cdot \frac{(t_1)_i}{(t_2)_i}.

Division by zero follows standard IEEE 754 rules — no clamping is applied. Gradients flow through all three tensor operands.

Examples

>>> import lucid
>>> a = lucid.tensor([1.0, 2.0])
>>> t1 = lucid.tensor([4.0, 9.0])
>>> t2 = lucid.tensor([2.0, 3.0])
>>> lucid.addcdiv(a, t1, t2, value=0.5)
Tensor([2. , 3.5])