torch.foreach.norm

torch.foreach.norm(inputs: TensorList, /, *, ord: Scalar = 2, dtype: dtype | None = None) → tuple[Tensor, ...]

Returns the vector norm of each tensor in inputs.

This is semantically equivalent to applying torch.linalg.vector_norm() independently at every list position. Does not mutate its arguments and returns a tuple of result tensors.

Tensor-list arguments must be non-empty. An accelerated multi-tensor implementation is used only when supported by the inputs; otherwise the operation falls back to per-tensor execution.

Every input tensor is reduced over all dimensions. The dim and keepdim options of torch.linalg.vector_norm() are not supported.

Parameters:

• inputs (list or tuple of Tensor) – tensors to reduce.

• ord (Number, optional) – norm order. Default: 2.

• dtype (torch.dtype, optional) – dtype used for the computation.

Returns:

a tuple containing one norm tensor per input tensor.