addmm

→Tensor

addmm(input: Tensor, mat1: Tensor, mat2: Tensor, beta: _float = ..., alpha: _float = ...)

source edit

General matrix multiply with a scaled accumulator (GEMM).

Computes $\beta \cdot \text{input} + \alpha \cdot (\text{mat1} \times \text{mat2})$ , the canonical BLAS-3 gemm operation. Useful for fused linear layers: an output bias plus a weight multiplication in one call.

Parameters

inputTensor

Accumulator of shape (M, N) (or broadcastable to it).

mat1Tensor

Left matrix of shape (M, K).

mat2Tensor

Right matrix of shape (K, N).

betafloat

Scalar multiplier on input. Defaults to 1.0.

alphafloat

Scalar multiplier on mat1 @ mat2. Defaults to 1.0.

Returns

Tensor

Tensor of shape (M, N).

Notes

Mathematical definition:

\text{out} = \beta \cdot \text{input} + \alpha \cdot (\text{mat1} \cdot \text{mat2}).

The inner dimension K must match between mat1 and mat2. Standard dtype promotion applies; the matmul itself uses Accelerate (CPU stream) or MLX (GPU stream).

Examples

>>> import lucid
>>> M = lucid.zeros((2, 2))
>>> a = lucid.tensor([[1., 2.], [3., 4.]])
>>> b = lucid.tensor([[5., 6.], [7., 8.]])
>>> lucid.addmm(M, a, b, beta=0.0, alpha=1.0)
Tensor([[19., 22.],
        [43., 50.]])

>>> import lucid >>> M = lucid.zeros((2, 2)) >>> a = lucid.tensor([[1., 2.], [3., 4.]]) >>> b = lucid.tensor([[5., 6.], [7., 8.]]) >>> lucid.addmm(M, a, b, beta=0.0, alpha=1.0) Tensor([[19., 22.], [43., 50.]])