Microsoft's OasisKV boosts LLM inference throughput 1.69x with lookahead sparse prefetching

microsoft · hf · 2026-08-11

Microsoft introduces OasisKV, a memory-centric LLM inference system that exploits sparse attention to keep only relevant KV entries in HBM, prefetching future important tokens via speculative decoding. Built on vLLM, it maintains accuracy within 0.7 points under a 2048-token KV budget, achieves 1.69x throughput over dense vLLM on reasoning workloads (0.1 point accuracy loss), up to 2.1x on multi-GPU long-context serving, and about 2x throughput with 6.5-9.7x less KV admission under prefill-decode disaggregation.

Original post →

More from Infra

Infra channel →