pooled is a project that bundles multiple browsers into a single WebRTC mesh to divide inference of large open models. The GitHub commit of November 3, 2026 (45b5ee2, PR #427) addressed the handling for when, among the members other than the room's creator, the device running the model departs or is refreshed. If that device goes away, that model stops, and the surviving devices release the layers they were holding and return to the model selection screen.
Inside a room, one browser acts as the host and owns the conversation, tokenizer, embedding table, LM head, and sampler, while the remaining devices, the workers, divide up contiguous layer ranges. Therefore, when the device that loaded the model disappears, the workers holding the layers could be left in a not-ready state. This commit makes a worker, upon detecting the model host's departure, stop loading at the next tensor point and also cut off the awaited loading pacing. After that, the worker releases the layers it held along with the connection information, and once the room has newly designated the current model host and Start is pressed, the layers are divided and reassigned. This is the part that makes it possible to start a new model without reloading the browser tab.
The documentation (docs/protocol.md) records this situation as a rule. When the member model host departs or is refreshed, that model stops, the remaining devices release the layers and return to the model selection screen, and thereafter anyone can press Start to restart with the model host chosen by the room. The condition is attached that, if only one side of the link breaks, the model does not stop, and as long as the room's creator still names that host, the existing reconnection path is maintained. On the verification side, browser regression scenarios (hostgone, hostreload, hostgone-loading) were added, and tests/unit/host_departure_test.js was newly created. However, this change covers only the handling of connection state within the room, and presents no measurements on loading time or throughput.