synthesizer VS pytorch-lightning

The run is a few days in on a 8x 80gb A100 cluster, and I quietly released the first epoch checkpoint here. I am building the model in association with our synthetic data efforts here, at SciPhi.

I am updating because we have another interesting result - by going deeper instead of broader, and by combining new techniques like RAG, we can make incredibly descriptive textbooks. This one here was generated by a ~fully AI pipeline. The pipeline goes MIT OCW -> Syllabus -> Table of Contents -> Textbook. The last step is grounded through vector-lookups over the whole of Wikipedia.

The timing is great, because yesterday I introduced RAG into the synthetic generation pipeline [here](https://github.com/emrgnt-cmplxty/sciphi/tree/main). I'm in the process of indexing the entirety of this pypi dataset using ChromaDB in the cloud. It will be relatively easy to plug this into SciPhi when done.

It's very easy to get started, right in your Terminal, no fees! No credit card at all.

And there are cloud providers like https://replicate.com/ and https://lightning.ai/ that will let you use your LLM via an API key just like you did with OpenAI if you need that.

You don't need OpenAI - nobody does.

https://see.stanford.edu/Course/CS229 https://lightning.ai/ https://www.youtube.com/watch?v=00s9ireCnCw&t=57s https://towardsdatascience.com/

I've been wondering on what is the recommended method of saving logits/activations using PyTorch Lightning. I've looked at Callbacks, Loggers and ModelHooks but none of the use-cases seem to be for this kind of activity (even if I were to create my own custom variants of each utility). The ModelCheckpoint Callback in its utility makes me feel like custom Callbacks would be the way to go but I'm not quite sure. This closed GitHub issue does address my issue to some extent.

Our codebase for the diffusion models builds heavily on OpenAI's ADM codebase , lucidrains, Stable Diffusion, Lightning and Hugging Face. Thanks for open-sourcing!

For lightweight use, pytorch-lightning is too heavy, and its source code will be very difficult for beginners to read, at least for me.

For the A100 GPU, theoretical performance is the same for FP16/BF16 and both rely on the same number of bits, meaning memory should be the same. However since it's quite newly added to PyTorch, performance seems to still be dependent on underlying operators used (pytorch lightning debugging in progress here).