simpleT5 icon indicating copy to clipboard operation
simpleT5 copied to clipboard

Published 20 hours ago • verified publisher icon Shivanandroy

Would this work to train flan-t5-xxl on multiple GPUs?

Open experimarketing opened this issue 2 years ago • 5 comments

experimarketing avatar Jan 22 '23 18:01 experimarketing

Bump

emendoza2 avatar Jan 28 '23 18:01 emendoza2

@experimarketing : Right now, it only works with T5 and mT5 on single GPU. I was away from the development for a couple of months. So, I didn't upgrade it to support FlanT5 and multi GPU.

But, I will integrate it ASAP.

Shivanandroy avatar Jan 28 '23 20:01 Shivanandroy

Would it work with the -xxl version? I believe model parrellism would be required to run it. As it is too large to run on a single GPU.

experimarketing avatar Jan 29 '23 15:01 experimarketing

@experimarketing : I'm afraid, It won't!

Shivanandroy avatar Jan 29 '23 19:01 Shivanandroy

@experimarketing : Right now, it only works with T5 and mT5 on single GPU. I was away from the development for a couple of months. So, I didn't upgrade it to support FlanT5 and multi GPU.

But, I will integrate it ASAP.

Thanks for looking into this. Kindly let us know after completion. one more thing really many thanks for developing this library. It simplified the usage of the T5 model.

SomasekharDS avatar Feb 12 '23 04:02 SomasekharDS