Sabitlenmiş Tweet
𝗣𝗿𝗮𝗰𝘁𝗶𝗰𝗮𝗹 𝗱𝗲𝗲𝗽 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗰𝗼𝘂𝗿𝘀𝗲🌻𝘀𝘂𝗺𝗺𝗮𝗿𝘆
🧵Thread with links to all slides, videos, notebooks, and @PyTorch code.
Solutions are now available on the repo of the course: github.com/dataflowr/note…
1/15
English
Marc Lelarge 🌻
1.2K posts
Marc Lelarge 🌻
@marc_lelarge
Researcher in mathematics and machine learning @Inria and @ENS_ULM. Tweeting about maths and AI.
Paris Katılım Nisan 2012
630 Takip Edilen5.9K Takipçiler
Link to the repo: github.com/LLM4Rocq/Putna…
Note: Lean solutions were released Dec 2025—after the model’s May 2025 training cutoff.
English
10/12 Putnam 2025 problems fully formalized in Rocq by Opus 4.6.
Putnam problems were autoformalized from Numina’s Lean versions combined with natural language statements. We then ran Claude Code for 3 days on an isolated VM (no web access) with our new Rocq-MCP server.
English
New hands-on course: implement KV Cache (faster inference) and LoRA (cheap fine-tuning) from scratch on top of @karpathy minGPT. You write the code, benchmark the speedups:
github.com/dataflowr/llm_…
English
@Sri26762339 I did not try on RTX3060 but the first part in Triton is working on T4 on Colab, so it will probably be OK. For the last part full Flash Attention in Triton, it is probably easier on H100...
English
@marc_lelarge Can this be done on 3060 lap in wsl for testing or it can be done in h100 only so that I can do it in my cluster. 🙌
English
@SethHWeidman It depends on your background. For me and my students, Triton is easier (close to Python). I think knowing a bit of Triton allows them to understand how GPUs are working with a minimal effort. Closer to CUDA, I did some Numba too: x.com/marc_lelarge/s…
Marc Lelarge 🌻@marc_lelarge
Learn 𝗚𝗣𝗨 𝗽𝗿𝗼𝗴𝗿𝗮𝗺𝗺𝗶𝗻𝗴 from the ground up: begin with Numba for low-level control, then progress to Triton to write high-performance kernels in a Python-like language. A hands-on Jupyter notebook to get you started quickly.
English
@marc_lelarge It is really not that hard to understand FlashAttention in CUDA directly, I’m not sure why Triton is needed.
English
Here is the code to learn Flash Attention: github.com/dataflowr/gpu_…
English
Marc Lelarge 🌻 retweetledi
Martin Hairer and colleagues released a set of hard maths problems, designed to be test cases for LLMs.
We have *one week* to solve them, using LLMs. They encrypted the solutions at 1stproof.org and will reveal them just after.
arxiv.org/pdf/2602.05192
(1/3)
English
Want to learn formal mathematics with @CoqLang? Start with Volume 1 of Software Foundations.
I’ve created a small Rocq tactics cheatsheet for the first chapters, where you can interactively build your own proofs directly in the browser:
mlelarge.github.io/rocq-proofs/
Thanks @ejgallego
English
Que signifie vraiment « découvrir » une particule invisible ? Du boson de Higgs au seuil des cinq sigma, mon article sur le test d’hypothèse et la prudence statistique au cœur de la physique des particules.
larecherche.fr/physique/lhomm…
Français
@andysingal Notebook is available here: github.com/dataflowr/note…
English
@Jishan_2019 Sorry no recording.
x.com/marc_lelarge/s…
Marc Lelarge 🌻@marc_lelarge
Learn 𝗚𝗣𝗨 𝗽𝗿𝗼𝗴𝗿𝗮𝗺𝗺𝗶𝗻𝗴 from the ground up: begin with Numba for low-level control, then progress to Triton to write high-performance kernels in a Python-like language. A hands-on Jupyter notebook to get you started quickly.
English
@MishraAmogh Yes, I am preparing more advanced materials for my LLM course (next semester).
English
Learn 𝗚𝗣𝗨 𝗽𝗿𝗼𝗴𝗿𝗮𝗺𝗺𝗶𝗻𝗴 from the ground up: begin with Numba for low-level control, then progress to Triton to write high-performance kernels in a Python-like language. A hands-on Jupyter notebook to get you started quickly.
English
@ehsanmok Thanks! It looks super nice. What is the use case for Mojo? compare to Triton?
English
@marc_lelarge Really looking forward to this and diving into this during my weekend. I appreciate your opensource policy!
English
Code (ready to run on @GoogleColab ): github.com/dataflowr/note…
English
@marc_lelarge Woah! Topics really look cool. Are any of these resources online?
English