[fleet executor] Comm init for dist model inf

[FeixLiu]

PR types

Others

PR changes

Others

Describe

comm init for the dist model inf system.

Test with the following codes:

import paddle.distributed.fleet as fleet
import paddle
from paddle.fluid import core

paddle.enable_static()
fleet.init(is_collective=True)

config = core.DistModelConfig()
config.model_dir = "./inference_model/rank_" + str(fleet.worker_index()) + "/step_0"
config.place = 'GPU'
config.device_id = fleet.worker_index()
config.current_endpoint = "127.0.0.1:700" + str(fleet.worker_index())
config.trainer_endpoints = ["127.0.0.1:7000", "127.0.0.1:7001", "127.0.0.1:7002", "127.0.0.1:7003", "127.0.0.1:7004", "127.0.0.1:7005", "127.0.0.1:7006", "127.0.0.1:7007"]
config.pp_degree = 2
config.mp_degree = 4
config.mp_ring_id = 0
if fleet.worker_index() <= 3:
    config.pp_downstream_ring_id = 20
    config.pp_upstream_ring_id = -1
if fleet.worker_index() >= 4:
    config.pp_downstream_ring_id = -1
    config.pp_upstream_ring_id = 20
config.local_rank = fleet.worker_index()
config.nranks = 8

dist = core.DistModel(config)
dist.init()

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[wangxicoding]

记得CoordSys吗，后面最好抽象一下

[FeixLiu]

感觉没啥必要，inf组网的维度最多只有pp和mp，为了这两个再搞一个coord sys感觉有点多余。其实主要是之前把c++端的coord sys移到python 端了。。。不想再移回来😂

[wangxicoding]

其实可以不用executor执行op来跑的，直接掉api就行，不过这样也没啥问题

[FeixLiu]

这样比较简洁吧，以后如果需要其它op可以直接加在这里