The same as meta's repo?
Is it the same as meta's llama2-hf repo?
Effectively, yes. I uploaded them partly because some people couldn't get access to Meta's repo, and also because when I tried to quantise from the Meta HF repos, I got weird errors.
It turned out that the error was caused by an incorrectly set parameter in the HF repo, "pretraining_tp": 2,. But I don't think that affects normal inference.
But yeah, if you have access to Llama-2-13B-HF there's no real reason to use this repo and I may delete it eventually.
I see that HF have fixed the pretraining_tp issue in the upstream repo now
The other difference with my repo and theirs was that mine was fp16 and they mistakenly uploaded theirs in fp32. But that is also fixed now.
So yeah, no reason to use my repo at all now I would say.
Is there a reason these are provided as .bin files instead of .safetensors?
Thanks as always for all your work :)
Only that I didn't bother specifying them to be saved as safetensors! I did later for the Llama 70B fp16s
As I was saying earlier, I only put these repos up because a couple of people asked me because they'd not yet got access to the Meta ones, and because I discovered an issue in the Meta repo. That issue has been fixed now, and I think most people have access to Meta's repos if they want it by now. So maybe I should just delete these..