Skip to main content

MatMul

MatMul

Matrix multiplication of two input tensors. Performs batched matrix multiplication following PyTorch broadcasting rules.

Shape Contract:

  • Input a: [*, n, m] first matrix
  • Input b: [*, m, p] second matrix (inner dimension must match)
  • Output: [*, n, p] matrix product

Notes:

  • No learnable parameters (pure operation)
  • Supports batched operations with broadcasting
  • Inner dimensions must match: a.shape[-1] == b.shape[-2]
  • Core operation for attention mechanisms
  • Uses torch.matmul for optimal performance

Signature​

neuron MatMul()

Ports​

Inputs:

  • a: [*, n, m]
  • b: [*, m, p]

Outputs:

  • default: [*, n, p]

Implementation​

Source { source: "core", path: "operations/MatMul" }