Perplexity Describes Tulip Inference Server

Perplexity AI posted a reply describing Tulip, its lightweight Rust gRPC inference server positioned between Ivy and the ROSE engine. The post states that Tulip collects incoming requests, batches them, and sends them to the GPU. For small embedding models, runtime depends on tokens rather than query count, with roughly 512 tokens filling the GPU. The reply includes an attached two-panel line chart.