RLX Tab Cluster

RLX Tab Cluster is a browser-based demonstration of data-parallel training. The project says each tab compiles a backward graph in a Web Worker, computes gradients on its own data shard, and joins other tabs in averaging those gradients after each round. Tabs in one browser connect through BroadcastChannel; devices in a room use WebRTC, with public relays used for signalling. The published demo offers WebGPU, WebGL and CPU backend choices, plus three gradient payload formats. These are the project’s described features, not independently measured training speed or evidence that it can train large models efficiently.

The Hugging Face Space was created on October 13. Its README says cross-device rooms require HTTPS because WebRTC needs a secure context. The page is a self-contained static demo rather than a hosted training service.