Alex Redden
Alex Redden
Ah, I guess I had not added it- but I just did.
I merged the changes into the main branch so hopefully that will help (also updated readme)
Ah! Essentially it's just the checkpoint which gets created after loading the model and doing at least 12 steps of inference. You could do something like this in the root...
Ah- You can speed that up by using nightly torch- for me compilation only takes a few (maybe 3-4) seconds at most.
That seems correct, it's possible that it's just related to the cpu- I have a 7950x so everything runs very fast.
Ah- I'll look into this, thanks :)
This commit should fix your issue 0a914297d17f58f64214e1f99d8d3cfb9e791d3e
Ah- yeah that is possible, since it would change the weights and scaling values, not sure how to ensure that they are always identical. I think it might be possible...
Ah cool! @0xtempest If it's a very large PR, could it be done in pieces? I would like to be able to test each change individually since for a very...
Ah yeah I'll look into this, shouldn't be happening. Is caused by the https://github.com/aredden/torch-cublas-hgemm library not being installed. Should be a quick fix. I'll do it today.