[ascend300t product] [distributed training function] model_ The script on zoo cannot be trained with multiple cards, and an alarm occurs during single card training

1. Introduction to the environment

The server ：Taishan2280

OS：Euler 2.8

Mindspore edition ：1.1.0

Model selection ：SSD

Package installed ：

A300t-9000-mcu_2.0.8.hpm

A300t-9000-npu-driver_20.2.0.b070_euleros2.8-aarch64.run

A300t-9000-npu-firmware_1.76.22.1.220.run

Ascend-cann-toolkit_20.2.rc1_linux-aarch64.run

SSD Model training cannot be done with two cards at the same time

problem 1：run_distribute_train.sh Execution has no effect

problem 2：NPU Card assignment IP Time display ip Conflict

problem 3：/etc/hccl.conf Caused by empty file rank*.json File cannot be generated

SSD Model single card training , There will be many alarms , Can train successfully

answer ：

Single card training warning It's normal , Look at the tips from the log MindSpore and Ascend The version of does not match , but MindSpore Forward compatible, so it can be executed normally .

Multi card problem 2 Appoint IP What specific instructions are used ,/etc/hccn.conf Null may be due to the specified IP Failure