bytepoet

4 ms·

1. 1.▲
   by bytepoet 2mo ago

   Thanks for tiny-vllm! The documentation is amazing. I'm using it as a resource for a university course I'm teaching.
2. 2.▲
   by bytepoet 1y ago

   Wonderful! Great, detailed explanation. I look forward to reading the vLLM post as well.
3. 3.▲
   by bytepoet 1y ago

   Thanks for the inputs. It's very helpful to know. I look forward to following mirage development.
4. 4.▲
   by bytepoet 1y ago

   This is very cool. I enjoyed going through the writeup and GitHub README. I was wondering if these same optimizations can be brought to bear on training as well, rather than only inference. I guess the challenge here is fusing backward comp
5. 5.▲
   by bytepoet 1y ago

   I really enjoyed reading this, particularly the first part where the author was specific about why we invariably (and often vaguely) find LLM generated text slightly off. I cherish writing and find it a wonderful tool for thinking. So far,
6. 6.▲
   by bytepoet 1y ago

   I enjoyed reading this paper. The experiments are well-designed and it's well-written. Much work on generalization of ML models deals with asymptotic bounds. Here, there's a precise way of measuring these, even for relatively larg
7. 7.▲
   by bytepoet 1y ago

   Such a well-written and thoughtful blog post. Loved it!
8. 8.▲
   by bytepoet 1y ago

   The inability of LLMs of ask for clarification was exactly the flaw we encountered when testing them on open-ended problems, stated somewhat ambiguously. This was in the context of paradoxical situations, tested on DeepSeek-R1 and Claude-3.
9. 9.▲
   by bytepoet 1y ago

   The blog post is really good. I see that there's a follow-up piece 'The Fifth Kind of Optimisation' about parallelism. Something that I'd like to add is that it's helpful to understand the optimization capabilities
10. 10.▲
    by bytepoet 2y ago

    Thanks a lot, Sasha, for creating these. I found your LLM training puzzles to be excellent as well.