⠿ swyrmcomputeBack to website
ARCHITECTURE

Many PCs, one API.

A coordinator receives API requests, counts tokens with the model’s tokenizer, and assigns each request to one available node. Nodes keep an outbound WebSocket to the coordinator, so hosts open no ports.

Scheduling

  • Follow-up messages go back to the PC that holds the conversation cache.
  • Otherwise a node is picked at random weighted by measured speed and reputation; hosts that post a work bond get work first.
  • A request is never sent to a node owned by the same account.
  • When every node is mining, the coordinator wakes one up and the request waits for it.

Why not split one request across PCs today?

Over home internet connections, passing activations between machines for every token adds latency and a second point of failure. Running the whole model on each GPU is faster and more reliable, and more PCs simply mean more parallel requests. Splitting larger models across machines is on the roadmap for cases where measurements show it pays off.

Updates

Strata, models and tuning presets ship as releases signed offline. Nodes verify signatures and checksums, update when idle, roll out gradually and roll back automatically if a release does not start.