HiSparse
Named by 3 kept items in this workspace.
2 items name it under frameworks and tools and 1 under techniques.
Search every workspace for HiSparse. It matches the name as written, so it can find other things that share it.
The items that name it 3
Stores history in host memory and fetches only the selected entries, using one cache budget at every load; does not handle draft-model speculative decoding.
Hierarchical memory system from the SGLang team that offloads KV cache entries from HBM to host DRAM with LRU eviction and loads them back on top-k cache misses.
All technologies in this workspace.