vLLM tests prefill-decode split on 12 GB300 GPUs in pre-release prototype

vLLM reports that its prefill-decode (P/D) prototype ran on 12 GB300 GPUs, connecting DEP4 prefill to DEP8 decode with FP8, NIXL and a 384 GiB cache store.

The test recorded 514,784 prefix-hit tokens (91.24%), and two 1,000-request OCR runs scored 90.4% and 89.4% with zero request errors, according to vLLM's post on X.

The test ran before release and rebase, and the P/D pull request remains open, so the final build may differ.