ray__TurboQuant: Redefining AI efficiency with extreme compression
https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/
I did not understand what polarQuant is.
Is is something like pattern based compression where the algorithm finds repeating patterns and creates an index of those common symbols or numbers?
1. Efficient recursive transform of kv embeddings into polar coordinates
2. Quantize resulting angles without the need for explicit normalization. This saves memory via key insight: angles follow a distribution and have analytical form.
2. Quantize resulting angles without the need for explicit normalization. This saves memory via key insight: angles follow a distribution and have analytical form.
https://mesuvash.github.io/blog/2026/turboquant-interactive/ has a little visualisation
I like the visualization, but I don’t understand the grid quantization. If every point is on the unit circle aren’t all the center grid cords unused?
The way I understand it, it's a way of compressing vectors by switching from their per-component representation to polar coordinates representation, where the nearby vectors are clumped together to a single line, allowing to describe them by different lengths