[Quantization] Add TurboQuant online weight compression (Linear-only)#39970

Open

varjoranta wants to merge 9 commits intovllm-project:mainfrom

varjoranta:feat/turboquant-online-weight-quant

Commits on Apr 17, 2026

[Online Quant] Add TurboQuant weight compression (Linear-only)
varjoranta
committed
Add unit tests for TurboQuant online weight compression
varjoranta
committed
[Docs] Add TurboQuant quantization page and hardware support entry
varjoranta
committed
Reuse existing Lloyd-Max centroid module, drop scipy dependency
varjoranta
committed
Reuse generate_wht_signs from shared turboquant module
varjoranta
committed
Re-attribute algorithm to HIGGS (NAACL 2025) and fix TurboQuant arXiv ID
varjoranta
committed
docs(turboquant): note MXFP4/NVFP4 out of scope
varjoranta
committed
turboquant: plumb bits and group_size through OnlineQuantizationConfig
varjoranta
committed
turboquant: coalesced bulk load for 3-bit dequant
varjoranta
committed