QJL residual correction needed or harmful
Parent: Mac local LLMs: KV cache sizing and quantization · Published reference · snapshot 2026-10-05
↓ Facts as markdownall context files
Hannecke's survey reports the community independently found that at 3 or more bits putting all bits into Lloyd-Max centroids beats the paper's two-stage QJL scheme, so TurboQuant in practice is random rotation plus Lloyd-Max scalar quantization, while at 2.5 bits QJL remains necessary.
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Facts
- Hannecke's survey reports the community independently found that at 3 or more bits putting all bits into Lloyd-Max centroids beats the paper's two-stage QJL scheme, so TurboQuant in practice is random rotation plus Lloyd-Max scalar quantization, while at 2.5 bits QJL remains necessary. [source]
Children
- No children recorded.