Benchmarks icon indicating copy to clipboard operation
Benchmarks copied to clipboard

Published 20 hours ago • verified publisher icon ECP-CANDLE

Tdey/intel xpu

Open newtdms opened this issue 1 year ago • 6 comments

Added callbacks to measure performance throughput more precisely. By default, it shows the throughput for each epoch. Updated the optimizer API for keras==2.12 and keras==3.13 version

newtdms avatar Aug 18 '23 19:08 newtdms

LGTM

venkat-1 avatar Dec 13 '23 12:12 venkat-1

Is there anything in the performance callback that is Intel-specific? I may lift it into candle_lib unless there are objections (and after testing)

jmohdyusof avatar Dec 13 '23 15:12 jmohdyusof

Is there anything in the performance callback that is Intel-specific? I may lift it into candle_lib unless there are objections (and after testing)

tested a few and it seems to work on my local machine.

rajeeja avatar Dec 13 '23 15:12 rajeeja

OK. This branch is still using /common instead of candle_lib so I will not try to merge it. We should try to fix the offload syntax so that xpu, gpu etc are supported in the same code

jmohdyusof avatar Dec 13 '23 15:12 jmohdyusof

LGTM

Thanks Venkat. I added you as a reviewer. I need approval to merge the patch.

newtdms avatar Dec 13 '23 16:12 newtdms

Is there anything in the performance callback that is Intel-specific? I may lift it into candle_lib unless there are objections (and after testing)

No, nothing Intel-specific. This is more precise way for measuring performance. The same callbacks work on A100 too.

newtdms avatar Dec 13 '23 16:12 newtdms