StorageView

class ctranslate2.StorageView

An allocated buffer with shape information.

The object implements the Array Interface and the CUDA Array Interface so that it can be passed to Numpy or PyTorch without copy.

Example

>>> x = np.ones((2, 4), dtype=np.int32)
>>> y = ctranslate2.StorageView.from_array(x)
>>> print(y)
 1 1 1 ... 1 1 1
[cpu:0 int32 storage viewed as 2x4]
>>> z = np.array(y)
...
>>> x = torch.ones((2, 4), dtype=torch.int32, device="cuda")
>>> y = ctranslate2.StorageView.from_array(x)
>>> print(y)
 1 1 1 ... 1 1 1
[cuda:0 int32 storage viewed as 2x4]
>>> z = torch.as_tensor(y, device="cuda")

Inherits from: pybind11_builtins.pybind11_object

Attributes:

device
device_index
dtype
shape

Methods:

from_array
to
to_device

static from_array(array: object) → StorageView

Creates a StorageView from an object implementing the array interface.

Parameters: array – An object implementing the array interface (e.g. a Numpy array or a PyTorch Tensor).
Returns: A new StorageView instance sharing the same data as the input array.
Raises: ValueError – if the object does not implement the array interface or uses an unsupported array specification.

to(dtype: DataType) → StorageView

Converts the storage to another type.

Parameters: dtype – The data type to convert to.
Returns: A new StorageView instance.

to_device(device: Device) → StorageView

Converts the storage to another device.

Parameters: device – The device to copy the data to.
Returns: A new StorageView instance.

property device: Device where the storage is allocated (“cpu” or “cuda”).

property device_index: Device index.

property dtype: Data type used by the storage.

property shape: Shape of the storage view.