Jędrzej Maczan

I built a tiny-vllm in C++ and CUDA - paged attention - continuous batching - educational - 100% human-written™ And now I writing a course where you will build your own vLLM yourself. Still work in progress, I'll finish by the end of April. All for free ofc, just a GitHub repo

@antgr81 appreciated 🫡

@jedmaczan followed

I built a tiny-vllm in C++ and CUDA - paged attention - continuous batching - educational - 100% human-written™ And now I writing a course where you will build your own vLLM yourself. Still work in progress, I'll finish by the end of April. All for free ofc, just a GitHub repo

@EAnsari88820 thx lfg!

@jedmaczan woww keep going !!

@ICE257_ will do

@jedmaczan Followed, keep me updated

@nearlypi thanks man, hope you find it useful

@jedmaczan Great stuff 👍🏾

@YuvrajS9886 Basics of C++ and CUDA and learn on the go. You got this bro!

@jedmaczan Exact prerequisites like c or cuda cus I know nothing ☠️?

Andrej I got inspired by your llm.c and how you explain things from scratch. I’d love if you take a look at my project and the course (I won’t lie, I try to complement LLM101n a bit) @karpathy

@yacinelearning hi brother, recently finished raw C++ and CUDA small vLLM, now writing a tiny-vllm course from scratch x.com/jedmaczan/stat… hbu?

yo guys what you working on this week???

@tatavishnurao thank you king

@jedmaczan Added to elite git repos on my list for 2k26 @jedmaczan

@stephennfern full course soon, breaking down bf16 now

@jedmaczan This is dope ✨️✨️✨️ awaiting ...

@sarthak2143 ❤️

@jedmaczan HELL YEAHHH

@jino_rohit cause it's ml and math 🥰

@jedmaczan this is so cool!

@botir33751732 🫡

@jedmaczan nice good stuff

@Triceratops_b0 code is finished, course on April 30 will be finished I promise bro

@jedmaczan Do it bro 🥳

@gauravkaul tysm! will be much better when full course is finished

@jedmaczan Great work 👏 very well documented

LLMs on WebGPU in PyTorch!!

@anirudhbv_ce @nvidia I recognize the paper, is it courses.cs.washington.edu/courses/cse599…? 😍

Finally got my Softmax kernels running on a @nvidia Blackwell B300 today! A single-pass tiled Softmax and a two-pass streaming Online Softmax. Writing ct.load() feels like cheating compared to manual Triton pointer math when mapping directly to TMA hardware.

The FA4 paper is finally out after a year of work. On Blackwell GPUs, attention now goes about as fast as matmul even though the bottlenecks are so different! Tensor cores are now crazy fast that attn fwd is bottlenecked by exponential, and attn bwd is bottlenecked by shared memory bandwidth.  Some fun stuff in the redesigned algorithm to overcome these bottlenecks: exponential emulation with polynomials, new online softmax to avoid 90% of softmax rescaling, 2CTA MMA instructions that allow two thread blocks to share operands to reduce smem traffic.

AI is writing a growing share of the world's software. No one is formally verifying any of it. New essay: "When AI Writes the World's Software, Who Verifies It?" leodemoura.github.io/blog/2026/02/2…