Hey! We are building open-source AI infra (on top of XLA) to make it easier to run AI workloads on non-NVIDIA GPUs (AMD, Google TPU, AWS Trainium, Intel Gaudi). We believe that with some effort, it is possible to achieve similar or better performance per watt compared to NVIDIA GPUs, while also enabling hardware diversity and reducing costs.<p>For the first step in this milestone, we released Felafax RoadRunnerX on Github (pun intended ), which includes a set of notebooks and deployment scripts for fine-tuning LlaMa3 on Google TPUs!<p>What do you all think? Would love to get your feedback!<p>Felafax RoadRunnerX repo - <a href="https://github.com/felafax/RoadrunnerX">https://github.com/felafax/RoadrunnerX</a>
We also built a cloud UI to make it easy to spin-up large clusters of TPUs and run multi-GPU fine-tuning on them! Check that out here - <a href="https://felafax.ai/">https://felafax.ai/</a>