jax.make_array_from_single_device_arrays(shape, sharding, arrays)[source]#

Returns a jax.Array from a sequence of jax.Arrays on a single device.

jax.Array on a single device is analogous to a DeviceArray. You can use this function if you have already jax.device_put the value on a single device and want to create a global Array. The smaller jax.Arrays should be addressable and belong to the current process.

  • shape (Tuple[int, ...]) – Shape of the jax.Array.

  • sharding (Sharding) – A Sharding instance which describes how the jax.Array is laid out across devices.

  • arrays (Sequence[Array]) – Sequence of jax.Arrays that are on a single device.

Return type



A jax.Array from a sequence of jax.Arrays on a single device.


>>> from jax.experimental.maps import Mesh
>>> from jax.experimental import PartitionSpec as P
>>> import numpy as np
>>> shape = (8, 8)
>>> global_mesh = Mesh(np.array(jax.devices()).reshape(2, 4), ('x', 'y'))
>>> sharding = jax.sharding.NamedSharding(global_mesh, P('x', 'y'))
>>> inp_data = np.arange(prod(shape)).reshape(shape)
>>> arrays = [
...     jax.device_put(inp_data[index], d)
...     for d, index in sharding.addressable_devices_indices_map(shape).items()]
>>> arr = jax.make_array_from_single_device_arrays(shape, sharding, arrays)
>>> arr.addressable_data(0).shape
(4, 2)