addbmm

→Tensor

addbmm(input: Tensor, batch1: Tensor, batch2: Tensor, beta: _float = ..., alpha: _float = ...)

source edit

Batched matmul with reduction over the batch axis.

Computes $\beta \cdot \text{input} + \alpha \cdot \sum_{k} (\text{batch1}[k] \cdot \text{batch2}[k])$ . Useful for accumulating multiple parallel matrix products into a single output matrix (e.g. summing per-head attention contributions).

Parameters

inputTensor

Accumulator of shape (M, N).

batch1Tensor

Batched left matrices of shape (B, M, K).

batch2Tensor

Batched right matrices of shape (B, K, N).

betafloat

Scalar multiplier on input. Defaults to 1.0.

alphafloat

Scalar multiplier on the batched matmul sum. Defaults to 1.0.

Returns

Tensor

Tensor of shape (M, N).

Notes

Mathematical definition:

\text{out} = \beta \cdot \text{input} + \alpha \sum_{k = 0}^{B - 1} (\text{batch1}[k] \cdot \text{batch2}[k]).

Differs from baddbmm in that the batch dimension is reduced away; baddbmm keeps the per-batch results.

Examples

>>> import lucid
>>> b1 = lucid.ones((3, 2, 4))
>>> b2 = lucid.ones((3, 4, 2))
>>> M = lucid.zeros((2, 2))
>>> lucid.addbmm(M, b1, b2)
Tensor([[12., 12.],
        [12., 12.]])