Comment by milansuk

I don't see any explanation for why they trained 8B instead of 7B. I thought that If you have a 16GB GPU, you can put 14GB(7B*16bits) model into it, but how does it fit If the model is exactly 16GB?