Fireworks AI

This post introduces the `GLOBAL` option for multi-region deployments, enabling workloads to span across geographies. It details the architectural change that allows a single deployment to draw capacity from multiple regions, addressing the limitation of legacy schedulers that required all nodes to be in a single region. The system ensures that global scheduling does not become a bottleneck for real-time inference pipelines by keeping global scheduling off the request path. Observational data from a production study shows a significant increase in request success rate for multi-region deployments compared to single-region ones.