Tulip is Perplexity's lightweight Rust gRPC inference server, sitting between the Ivy and ROSE engines. It collects incoming requests, batches them, and sends them to the GPU.
Tulip is Perplexity's lightweight Rust gRPC inference server, sitting between the Ivy and ROSE engines. It collects incoming requests, batches them, and sends them to the GPU.
For small embedding models, runtime depends on the number of tokens rather than the number of queries, so around 512 tokens is enough to fully utilize the GPU.