vLLM adds Day-0 support for Step 5 Preview, including sparse attention

vLLM says it added Day-0 support for the Step 5 Preview model, covering sparse attention, MoE, multi-token prediction and vision.

The serving work runs on Blackwell and Hopper GPUs and includes a sparse indexer, FP8 MegaMoE and cache transfer.

In a follow-up post, vLLM describes how the sparse attention works: it builds summaries of regions, an indexer scores those summaries and selects which regions to attend to, and unfinished regions stay in a local window. vLLM says the backbone combines sliding-window and sparse attention with MoE and multi-token prediction. The first post opens a six-part thread that vLLM says will include kernel results and validation details.