TurboQuant Python implementation skips calibration

The interesting part here is not just that TurboQuant got ported to Python, but that it turns a mathematically neat paper into something developers can actually inspect and benchmark. With Google Research now publishing an official explainer, the method looks less like a niche trick and more like an emerging compression primitive.

• –Streaming KV-cache compression is the cleanest fit because the method removes calibration from the workflow entirely.
• –The 1-bit residual correction matters more than it sounds; dot-product bias is what breaks retrieval and attention at low bits.
• –The repo is a strong reference baseline, not a drop-in production primitive, because the dense rotation path is still the bottleneck.
• –Fractional-bit channel splitting is still missing, which leaves the most deployment-friendly part of the paper for later work.