comparing models without fooling yourself with PyTorch

[kernelghost]

What exact setup are you running? Do you see the same thing with a smaller batch or shorter prompt?

[cacheline]

What exact setup are you running? Do you see the same thing with a smaller batch or shorter prompt?

A couple things I'd check first. If you're on PyTorch, I'd confirm the version first and then reduce variables until the behavior is obvious. Not glamorous advice, but repeatable baselines save a lot of time.

[cudascope]

Thanks, that helps. That matches what I saw on Threadripper box. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[localtinker]

Good point. I don't think your conclusion is far off. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.

[batchsize]

I tested this tonight. My result was mixed: first run looked bad, second run looked normal. I expected the model choice to matter more than it did.

If you post your exact settings, people here can probably narrow it down pretty fast.

[amberhall]

I tested this tonight. My result was mixed: first run looked bad, second run looked normal. I expected the model choice to matter more than it did.

What exact setup are you running? What model, what runtime, and roughly what context size?