Skip to content

[Feature] HiCache for Hybrid and Sparse LLMs #12826

Description

@xiezhq-hermann

Motivation

With the introduction of HiCache, SGLang now leverages multi-level memory hierarchies to maximize the benefits of KV cache reuse. As HiCache continues to mature, our focus is expanding toward broader model support, including hybrid LLMs such as Qwen3-Next (integrating linear and full attention), as well as sparse LLMs like DeepSeek-V3.2-Exp and GPT-OSS (whose sliding-window attention can be viewed as a simple form of sparsity).

Ongoing work

  • Sparse LLM support: Enable selective KV cache loading to GPU memory, reducing memory footprint and allowing larger batch sizes.[Sparse & HICache]: Enables hierarchical sparse KV cache management and scheduling for DeepSeek V32. #14619

    • A unified sparse scheduling framework that easily integrates various sparse algorithms.
    • Integrated with HICache to enable KVCache scheduling between CPU/Remote and GPU, reducing GPU memory consumption per request and Increase inference batch size.
    • Integrated with DeepSeek NSA And SWA
    • Supports various attention backends, such as FA, Triton, and FlashInfer.
    • ...
  • Refactored memory subsystem: Refactoring the mem-cache component to establish a more robust Radix Tree and memory pool foundation. [Feature] Memory Cache System Refactoring Road Map (Mem Cache V2) #12587

  • Linear attention caching: Develop new query and storage mechanisms for KV caches in linear attention layers, building upon the refactored mem-cache.

You’re welcome to join the discussion in #kv-cache-store 😊

Metadata

Metadata

Labels

hicacheHierarchical Caching for SGLanginactive

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions