DistributeTranspilerConfig¶

api_attr

declarative programming (static graph)

class paddle.fluid.transpiler.DistributeTranspilerConfig[source]

A configuration class that provide support for transpiler distributed jobs. Some important parameters are explained as follows:

slice_var_up(bool)¶

Whether to do Tensor slice for parameter servers, default is True.

split_method(PSDispatcher)¶

Methods of dispatching parameters for server, api_fluid_transpiler_RoundRobin or HashName can be used and default is RoundRobin. Try to choose the best method to balance loads for parameter servers.

min_block_size(int)¶

Minimum number of split elements in block, default is 8192.

According to : https://github.com/PaddlePaddle/Paddle/issues/8638#issuecomment-369912156 We can use bandwidth efficiently when data size is larger than 2MB.If you want to change it, please be sure you have read the slice_variable function. You can find the definition of slice_variable in https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/transpiler/distribute_transpiler.py .

Examples

from paddle.fluid.transpiler.ps_dispatcher import RoundRobin
import paddle.fluid as fluid

config = fluid.DistributeTranspilerConfig()
config.slice_var_up = True
config.split_method = RoundRobin
config.min_block_size = 81920