tonbistudio/turboquant-pytorch
A PyTorch reimplementation of TurboQuant, Google's vector quantization algorithm for compressing LLM key-value caches.
Not currently ranked — collecting fresh signals.
star history
This repository implements Google’s TurboQuant (ICLR 2026) from scratch in PyTorch. The algorithm compresses LLM key-value caches to reduce memory footprint during inference. The project includes original implementation plus an improved V3 version, with tests on NVIDIA GPUs validating compression ratios and attention fidelity at 2-5x compression rates with 3-bit and 4-bit quantization.
Frequently asked
- What is tonbistudio/turboquant-pytorch?
- A PyTorch reimplementation of TurboQuant, Google's vector quantization algorithm for compressing LLM key-value caches.
- Is turboquant-pytorch open source?
- Yes — tonbistudio/turboquant-pytorch is open source, released under the MIT license.
- What language is turboquant-pytorch written in?
- tonbistudio/turboquant-pytorch is primarily written in Python.
- How popular is turboquant-pytorch?
- tonbistudio/turboquant-pytorch has 1k stars on GitHub.
- Where can I find turboquant-pytorch?
- tonbistudio/turboquant-pytorch is on GitHub at https://github.com/tonbistudio/turboquant-pytorch.