AI Infrastructure

Hackathon

Hands-on challenges in AI infrastructure.

Hackathon #1

Explore six NPU kernel challenges on the Ascend 910C, with operator definitions, interfaces, data ranges, and numerical and implementation requirements. The competition has ended; the challenges and final leaderboard remain available for reference.

Open Hackathon #1

Challenge instructions and final results are available here on the course website.

Hackathon #2

Decentralized Reinforcement Learning over the Internet

Train a Qwen3-1.7B language model to play Wordle using four GPUs: one trainer updates the model while three workers generate complete games. A simulated network delay slows the delivery of model updates. Design a scheduling policy that reduces waiting without producing too many games from outdated model versions, and compare it with a synchronous reference under the same training and resource settings.

Measure waiting time, games accepted for training, completed updates, and Wordle wins. Submit one trained adapter with your scheduling code, configurations, logs, and a short report. Grading uses the model’s win rate on a separate set of Wordle games; see the full instructions for setup and submission details.

Open Hackathon #2