Partial¶

class paddle.distributed. Partial

The Partial describes Tensor across multiple devices, this type of tensor has the same shape but only a fraction of the value, which can be further reduce (e.g. sum/min/max) to obtain dist_tensor, often used as an intermediate representation.

Parameters: reduce_type (paddle.distributed.ReduceType) – the reduce type of the Partial state, default paddle.distributed.ReduceType.kRedSum.

Examples

           >>> import paddle
>>> import paddle.distributed as dist
>>> mesh = dist.ProcessMesh([0, 1], dim_names=["x"])
>>> a = paddle.ones([10, 20])
>>> 
>>> # distributed tensor
>>> d_tensor = dist.shard_tensor(a, mesh, [dist.Partial()])

          

is_partial ( self: paddle.base.libpaddle.Placement ) → bool is_partial¶

is_replicated ( self: paddle.base.libpaddle.Placement ) → bool is_replicated¶

is_shard ( self: paddle.base.libpaddle.Placement, dim: Optional[int] = None ) → bool is_shard¶

Partial¶

is_partial¶

is_replicated¶

is_shard¶