dairus01/crusades

Templar Crusades is a competition where participants earn cryptocurrency by writing the most efficient machine learning training code. Specifically, it runs on the Bittensor network, which is a blockchain that rewards people for contributing useful computing work. The goal is to maximize "MFU" (Model FLOPs Utilization), which is essentially a score for how efficiently you can train a large AI model on high-end GPUs. Here is how the process works. A participant (called a miner) writes a Python training script and hosts it at a public URL, like a GitHub Gist. They submit that URL to the blockchain, where it is locked using a timed encryption mechanism. After a waiting period, a validator downloads the code and runs it inside a controlled, isolated environment on a standardized hardware setup. The validator measures how fast and efficiently the code trains the AI model, calculates a median efficiency score from multiple runs, and reports that score back to the blockchain, which then determines the miner's earnings. This project is designed for machine learning engineers and AI researchers who know how to optimize deep learning code. To compete, you need access to serious hardware, specifically, at least two NVIDIA A100 GPUs with 80 GB of memory each. The benchmark model being trained is Qwen2.5-7B, which is too large to fit on a single GPU. Participants must use advanced techniques to split the model's data across multiple GPUs efficiently, and any standard optimization tricks like mixed precision or custom CUDA kernels are fair game. A key feature is the project's emphasis on security and fairness. Since the validators are running code written by strangers, the system uses a static security scanner to block dangerous patterns like forbidden imports or attempts to tamper with the system timer. The project also provides a local testing tool that perfectly mirrors the validator's environment, so participants can check their efficiency scores and ensure their code passes security checks before officially submitting it to the live competition. There is also a terminal dashboard that lets anyone monitor the live leaderboard and track recent submissions.