BlockNonCyclic#

class nvmath.distributed.distribution.BlockNonCyclic( process_grid: ProcessGrid, *, first_process: Sequence[int] | None = None, )[source]#

Block distribution without cycles

Methods

__init__( process_grid: ProcessGrid, *, first_process: Sequence[int] | None = None, )[source]#

copy() → Distribution[source]#: This is a common implementation for those distributions that only require a shallow copy.

shape( rank: int, global_shape: Sequence[int] | None = None, ) → tuple[int, ...][source]#

Get the local shape of data on the given rank according to this distribution.

Parameters:

rank – the process rank for which to calculate the local shape.
global_shape – Global shape of data. Required if the distribution is not bound to a global shape, otherwise not required.

to(cls, /, *, ndim=None, copy=False)[source]#

Convert this distribution object to an equivalent distribution of the given type.

Parameters:

cls – the target distribution type.
ndim – dimensionality of the target distribution. Must be compatible with the dimensionality of the source distribution. This may be required if the source distribution doesn’t have associated dimensionality.
copy – Returns a copy if the source and target type are the same.

Raises:

ConvertDistributionError – if the conversion is not possible.

Attributes

first_process#: Index in the process grid of the process who owns the first block of the distributed tensor.