QuATernaryons: Dual-Sided Sub-2-Bit Weight and Dynamic KV-Cache Quantization via 4D Isoclinic Clifford Rotors
We introduce QuATernaryons, a dual-sided sub-2-bit weight and dynamic KV-cache quantization framework utilizing calibrated quaternion rotor butterfly transforms and E8-lattice codebooks for high-throughput, low-precision LLM inference on modern GPU architectures. This is a massive WIP, currently there is work being done on taking these theoretical easures to tolerable, non novice levels that are practical for the real world.
Authors
- Shreyan Dushyant Gondaliya
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-09-29
- DOI
- https://doi.org/10.5281/zenodo.23043501
- Primary Topic
- Parallel Computing and Optimization Techniques
- Type
- preprint