Perplexity Describes Ivy Tulip and ROSE Inference Components

Perplexity's official account posted a reply describing its inference setup. The post states that inference is exposed via standardized APIs. Ivy, a Rust HTTP service, performs CPU work such as parsing, tokenization, and templating, then translates requests to gRPC. Tulip, a Rust gRPC service, schedules batches for the ROSE engine. ROSE, a Python component, implements forward passes and CUDA graph management. The post references an attached flowchart diagram titled Model S.