For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo function

to_numpy

def to_numpy[dtype: DType, LayoutType: TensorLayout, origin: Origin[mut=origin.mut], Engine: TensorEngine](tensor: TileTensor[dtype, LayoutType, origin, Engine=Engine]) -> PythonObject

Copies a TileTensor into a new NumPy array of the same shape.

Tensors whose layout is row-major and gap-free are copied with a single memcpy. Other layouts are gathered element by element.

This calls into CPython, so it is a host-side entry point.

Example:

from layout import TileTensor, row_major, to_numpy
from layout.coord import Coord, Idx

var storage = Array[Float32, 6](uninitialized=True)
var tensor = TileTensor(Span(storage), row_major(Coord(Idx[2], Idx[3])))
var array = to_numpy(tensor)  # a 2x3 NumPy float32 array

Constraints:

LayoutType must be flat: nested layouts have no single extent per axis to hand to NumPy.

Parameters:

• ​dtype (DType): The element dtype of the tensor.
• ​LayoutType (TensorLayout): The layout of the tensor.
• ​origin (Origin[mut=origin.mut]): The origin of the tensor.
• ​Engine (TensorEngine): The engine of the tensor.

Args:

• ​tensor (TileTensor[dtype, LayoutType, origin, Engine=Engine]): The tensor to copy.

Returns:

PythonObject: A NumPy ndarray of dtype dtype with the tensor's shape.

Raises:

If NumPy is unavailable, or if the underlying NumPy calls fail.