NVIDIA's H100s for Long Contexts

I hope this is a step forward for everyone. We really pushed hard on long context windows. But by and large, it's just a bigger sibling of the MPT model. It means we only need to start dreaming a little bit bigger about what kind of open source models we're going to have. The H100 has more compute units on it in parallel, which means you need wider layers to take full advantage of it.