LLMs-from-scratch

This is pretty great. PyTorch uses triton as the backend for torch.compile (the big feature of PyTorch 2.0, and the necessary part for making Flex Attention in the about to be released 2.5 usably fast). Triton's team doesn't support Windows, and, worse yet, does not accept community PRs to enable any sort of support. Here's the github issue: https://github.com/triton-lang/triton/issues/1640 And here's the performance comparison of Flex Attention with and without torch.compile (tldr it's 3x slower than a standard MHA when not compiled): https://github.com/rasbt/LLMs-from-scratch/blob/76e9a9ec02a1... EDIT: after taking a look at the repo, the only thing changed in the "46 commits ahead of [official triton]" is the README. Somewhat sketchy.

> Brains do not reason ... They can, at times, and do so best when emotion is not involved. > I'm tired of reading comments from people who keep repeating that LLMs don't think, don't reason, isn't intelligence because it is not human. LLM's represent a category of algorithm. Quite elegant and useful in some circumstances, but an algorithm none the less. A quick search produced this[0] example discussing same. Another reference, which may not be authoritative based on whatever most recent edit the link produces, is[1]: A large language model (LLM) is a computational model capable of language generation or other natural language processing tasks. As language models, LLMs acquire these abilities by learning statistical relationships from vast amounts of text during a self-supervised and semi-supervised training process. > Def

You can try this.https://github.com/rasbt/LLMs-from-scratch

someone here is writing such a book in public, here's the thread with the discussion "Implementing a ChatGPT-like LLM from scratch, step by step" https://news.ycombinator.com/item?id=39156778 you can view the github for the book as they work on it https://github.com/rasbt/LLMs-from-scratch

The attention mechanism we implement in this book* is specific to LLMs in terms of the text inputs, but it's fundamentally the same attention mechanism that is used in vision transformers. The only difference is that in LLMs, you turn text into tokens, and convert these tokens into vector embeddings that go into an LLM. In vision transformers, instead of regarding images as tokens, you use an image patch as a token and turn those into vector embeddings (a bit hard to explain without visuals here). In both text or vision context, it's the same attention mechanism, and it both cases it receives vector embeddings. (*Chapter 3, already submitted last week and should be online in the MEAP soon, in the meantime the code along with the notes is also available here: https://github.com/rasbt/LLMs-from-scratch/blob/main/ch03/01...)