SparseEngine unifies 15 sparse attention methods under one lifecycle contract

SparseEngine is an inference engine for LLMs built around a KV cache abstraction, hosting 15 sparse attention methods across four families. Per arXiv:2609.39068v1, it delivers roughly 10x the aggregate decode throughput of vanilla vLLM with SnapKV and 1.5x–2.6x with Quest and OmniKV at matched batch sizes. The code is released under Apache-2.0.

• 15 sparse attention methods across four families in one engine
• About 10x decode throughput over vanilla vLLM with SnapKV
• 1.5x–2.6x with Quest and OmniKV at matched batch sizes
• Chain Cache resumes from compacted history in multi-turn work