The Multi VE benchmark does not utilize the VEs

[keichi]

Hi, I'm trying to run the multi VE benchmark (multi.py) under multi-ve/, but it seems like it's not using the VE at all. The single VE version (single.py) is using the VE. Please find the job scripts and the results in the following gist:
https://gist.github.com/keichi/0f5c00551d91ecc3d5e14f8e4799f0c8

Looking at the PROGINF output for the multi VE benchmark, both the V. Op. Ratio (%) and MFLOPS rate are zero. Also, the runtime per step is much slower than the single VE benchmark, suggesting the training is not running on the VE.

How can I run the training on the VEs? Any advice would be appreciated.

[k-ishizaka]

I found this issue today. This script might not run as intended but I have no idea. sorry.