Pytorch Mox Logs Are Repeatedly Generated

Symptom

The Pytorch engine of a frequently-used framework is used as an algorithm source of a ModelArts training job. During the running of the training job, Mox versions for each epoch will be printed in the Pytorch Mox log. The log details are as follows:

INFO:root:Using MoXing-v1.13.0-de803ac9
INFO:root:Using OBS-Python-SDK-3.1.2
INFO:root:Using MoXing-v1.13.0-de803ac9
INFO:root:Using OBS-Python-SDK-3.1.2

Possible Cause

Pytorch creates multiple processes in spawn mode. Each process invokes the Mox to download data in multi-process mode. In this case, subprocesses are destroyed and recreated repeatedly, and Mox is imported repeatedly. As a result, a large amount of Mox version information is printed.

Solution

To avoid repeated output of the Pytorch Mox logs of the training job, you need to add the following code to the boot file. When MOX_SILENT_MODE = "1", Mox version information can be blocked in the log.

import os
os.environ["MOX_SILENT_MODE"] = "1"