Hey there,

Thanks for releasing this!

Going through the list of kernels:

1. CrossEntropyLoss
2. RMS NORM
3. RopeEmbedding
4. Swiglu
5. FastLoRA

I'm trying to understand how the various optimizations correlate to performance improvements, is there a chart that shows the gains from #5 alone?

Secondly, could you please explain what's being done/included in the both the PRO/MAX tiers. The wording from the blog post is very imprecise.

Thanks!