Skip to content

Support NexusQuant KV cache compression for memory reduction #4506

@jagmarques

Description

@jagmarques

KV cache memory is a bottleneck at long context lengths. NexusQuant offers training-free 7–10x KV compression via E8 lattice quantization + attention-aware token eviction (up to 17x with token merging).

Integration points:

  • After prefill, compress the KV cache in-place before storing in the paged block pool
  • Use attention mask to exclude evicted tokens during generation
  • API: with nexusquant_evict(model): model.generate(...)

Why this matters for lmdeploy:
lmdeploy's TurboMind engine already supports INT4/INT8 KV quant. NexusQuant's E8 lattice VQ is a natural extension — it achieves higher compression than INT4 while maintaining quality, and is drop-in since it doesn't change tensor shapes (only precision).

Validated results:

  • Mistral-7B: 7x compression, -2.26% PPL vs baseline
  • Llama-3-8B: 5.3x compression, -0.002% PPL
  • Training-free, no calibration data required

Library details:

Would you be interested in exploring this as a compression backend for TurboMind? Happy to help with the integration and provide benchmarks on your target models.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions