Jaap Buurman
Jaap Buurman
> Networkmanager need to be built with resolvconf support. Does that mean this is a bug that needs fixing in the archlinux NetworkManager package? Because it's definitely a regression given...
Okay, so clearing cookies between the login calls seem to work, but this is pretty cumbersome having to do this every time. Is there any way around this? Working example...
Did you move the file that was compiled by any chance? Or did you run the binary from the folder it was created in?
Are there any plans to rewrite that code to be optimized for ROCM instead of a CUDA port?
Why does the performance also regress when FA is enabled when not using batching? That's also a bit weird given the fact it's optimized for small batch sizes. Large batch...
Understandable, thanks for the detailed explanation! Hoping to see other devs join the project to optimize AMD performance then :) By the way, do you think (second hand) 7900 xtx...
Wow! I am seeing very good results here. Some observations: 1. With FA turned off, performance between your branch and master is identical. So that's good, it did not introduce...
Thanks for your explanation. Do you have any intention of trying to upstream these FA changes?
Didn't that comment in that PR mention they would be open to merging improvements, as long as someone would commit to maintaining that code? Not sure if that would be...
That's unfortunate to hear. I really think the 7900xtx could be competitive both on cost and performance. It just needs the support to get there. The situation has already improved...