Skip to content

Commit cf7d81d

Browse files
Add GPU MODE Cholesky sprint (#1829)
Adds a EuroPython sprint for the new GPU MODE Cholesky challenge, led by Bryce Adelstein Lelbach. The proposal highlights: - the single-file Python submission format and PyTorch, Triton, or embedded-CUDA paths - the live B200 challenge and hosted compute, with no local GPU required - a newcomer-friendly path from setup to a first valid leaderboard score - an optional agentic edit-test-benchmark workflow with human direction and correctness checks The mandatory local pre-commit check exposed existing Header.astro nullability errors. This PR includes a narrow, behavior-preserving typing fix so the hook can complete; the sprint content itself is isolated in src/content/sprints/gpu-mode-cholesky.md. Validation: - pnpm astro check - pnpm exec prettier --check src/content/sprints/gpu-mode-cholesky.md - pnpm exec astro build --mode production - pnpm pagefind <!-- readthedocs-preview ep-website start --> 🖼️ Preview available 🖼️ : https://ep-website--1829.org.readthedocs.build/ <!-- readthedocs-preview ep-website end --> --------- Co-authored-by: Bryce Adelstein Lelbach <brycelelbach@gmail.com>
1 parent 32da916 commit cf7d81d

2 files changed

Lines changed: 63 additions & 7 deletions

File tree

‎src/components/Header.astro‎

Lines changed: 11 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -614,14 +614,18 @@ const activeMenus = NAV_MENUS.map((menu) => ({
614614

615615
<script>
616616
document.addEventListener("DOMContentLoaded", function () {
617-
const navbar = document.getElementById("navbar");
618-
if (!navbar) return;
617+
const navbarElement = document.getElementById("navbar");
618+
if (!navbarElement) return;
619+
const navbar = navbarElement;
619620

620621
// ── Mobile hamburger toggle ──
621-
const toggle = navbar.querySelector(".nav-toggle");
622-
const navLinks = navbar.querySelector(".nav-links");
622+
const toggleElement =
623+
navbar.querySelector<HTMLButtonElement>(".nav-toggle");
624+
const navLinksElement = navbar.querySelector<HTMLElement>(".nav-links");
623625

624-
if (toggle && navLinks) {
626+
if (toggleElement && navLinksElement) {
627+
const toggle = toggleElement;
628+
const navLinks = navLinksElement;
625629
function closeMenuMob() {
626630
navLinks.classList.remove("open");
627631
toggle.classList.remove("active");
@@ -697,7 +701,7 @@ const activeMenus = NAV_MENUS.map((menu) => ({
697701
return parseInt(getComputedStyle(navbar).top) || 0;
698702
}
699703

700-
function setBannerState(atTop) {
704+
function setBannerState(atTop: boolean) {
701705
if (!banner) return;
702706
if (atTop) {
703707
banner.style.transform = "translateY(0)";
@@ -760,7 +764,7 @@ const activeMenus = NAV_MENUS.map((menu) => ({
760764
}
761765
window.addEventListener("scroll", requestTick, { passive: true });
762766

763-
let bounceCheck = null;
767+
let bounceCheck: ReturnType<typeof setTimeout> | null = null;
764768
window.addEventListener(
765769
"scroll",
766770
function () {
Lines changed: 52 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,52 @@
1+
---
2+
title: "GPU MODE: Cholesky Challenge"
3+
numberOfPeople: "20"
4+
pythonLevel: "Any"
5+
contactPerson:
6+
name: "Bryce Adelstein Lelbach"
7+
email: "brycelelbach@gmail.com"
8+
github: "brycelelbach"
9+
twitter: "blelbach"
10+
links:
11+
- title: "Cholesky leaderboard"
12+
url: "https://www.gpumode.com/leaderboard/776"
13+
- title: "Cholesky problem and starter submission"
14+
url: "https://github.com/gpu-mode/reference-kernels/tree/main/problems/linalg/cholesky_py"
15+
- title: "Linear algebra competition announcement"
16+
url: "https://www.gpumode.com/news/linear-algebra-kernels-age-of-research"
17+
- title: "Popcorn CLI"
18+
url: "https://github.com/gpu-mode/popcorn-cli"
19+
- title: "GPU MODE Discord"
20+
url: "https://discord.gg/gpumode"
21+
---
22+
23+
Come learn how to compete in a GPU MODE kernel contest by tackling the new
24+
Cholesky problem together. The challenge is to factor batches of symmetric
25+
positive-definite FP32 matrices on an NVIDIA B200. It runs through 30 July, so
26+
we can put what we learn straight onto the live leaderboard.
27+
28+
Every submission is a single Python file. You can begin with the working PyTorch
29+
baseline, then optimize in Python with tools such as Triton or embedded CUDA.
30+
The benchmark ranges from thousands of small matrices to one 32768-by-32768
31+
matrix, leaving room for many different ideas. Tests and benchmarks run on
32+
hosted B200s, so you do not need a local GPU.
33+
34+
Bryce from NVIDIA will help participants install and register the Popcorn CLI,
35+
understand the problem, make a first correct submission, read benchmark and
36+
profiling results, and choose an optimization to try. The goal is for every
37+
newcomer to leave knowing the complete contest workflow; experienced GPU
38+
programmers are welcome to team up and chase the leaderboard.
39+
40+
Participants who want to use a coding agent can have Popcorn scaffold the
41+
agent's contest workflow. Bryce will show how to give an agent a concrete
42+
optimization goal, let it run the iterative edit-test-benchmark loop, inspect
43+
its performance evidence, and keep numerically incorrect changes off the
44+
leaderboard. Working directly by hand, pairing with another person, or bringing
45+
your preferred coding agent are all equally welcome.
46+
47+
NVIDIA will provide participants with a free sandbox VM and inference keys for
48+
the purposes of the contest.
49+
50+
No previous GPU-kernel or agent experience is required. Bring a laptop and a
51+
GitHub or Discord account for authentication; basic Python familiarity will
52+
help.

0 commit comments

Comments
 (0)