Inference has become a networking problem. With disaggregated serving, where prefill runs on one set of GPUs, decode on another, and KV cache streams between them, performance depends on how fast you can move tensors between GPUs, across nodes, and in and out of storage. Performance is only half of it.
These deployments are elastic, so the data path needs dynamic APIs that let workers join and leave, connections form and tear down, and transfers reroute at runtime without stalling the pipeline. NIXL, […]'s open source Inference Xfer Library, is the layer that solves this, and it sits underneath […] Dynamo and a growing list of inference engines.
China is where some of the world's most demanding inference deployments are being built. We are hiring a senior architect to work directly with the platform teams building them, understanding what breaks at scale and turning that into upstream code and roadmap. You will work daily with core NIXL and Dynamo engineers in Israel and the US.