Rounding Experiments

I'm trying to find a general method to round vectors.

The ideal vector rounding which minimizes the angle with the full-precision vector is NOT a grid. It looks more like Voronoi cells on an n-sphere (for scaled integer vectors without an offset, at least).

I'm trying to find a general way to tractably find the nearest representable scaled integer vector in a reasonable amount of time¹. I'm planning to eventually make a blog post about it to explain it in more detail. The geometric rationale behind it will probably be cool.

This is what the best rounding looks like on a face of a cube (i.e. a vector with 3 components with the max being scaled to 1), for ternary {-1, 0, 1}:

And for pentary {-2, -1, 0, 1, 2}:

The weird stuff starts to happen at heptary {-3, -2, -1, 0, 1, 2, 3}:

Starting from enneary {-4, -3, -2, -1, 0, 1, 2, 3, 4}, floating point errors on the rounding scaling factor become noticeable and have to be considered. Thankfully, it seems like changing the scale by (2**23 + 1) / (2**23) (the next representable number after 1) is enough to make it work as far as I've tested it ([-31, 31]).

From [-63, 63] (7-bit) there are some small artifacts again caused by floating point precision, but this time when sorting fractions when ordering the consecutive rounding scales.

TODO: add higher-precision visualizations.

(These images are best viewed from inside a cube with all faces set to use the desired image as a texture)

---

My main unsolved challenges right now:

• Generalize to more than {-3, -2, -1, 0, 1, 2, 3}
  • Solved, the problem was off-by-one floating point errors introduced by division and multiplication of the rounding scale.
• Prove that a particular algorithm produces the best possible rounding
  • I guess it can be proved informally by noticing that there are no sharp transitions in the errors?
  • Still would eventually need a formal proof, although with floating point numbers it won't ever be perfect.
• Check if nested superblock rounding can be improved
• Remove the need for sorting the components to find the best rounding scale
• Find a fast enough general method to find both the best rounding offset and scale combination
  • I think the anyrize_offset_min_mean function in rounding.py might be it.
• Asymmetric zero-point quantization

Goals

One of the goals of this is to improve the rounding algorithms used in k-quants in llama.cpp.

If this somehow turns out to be equivalent to what's already used in k-quants, then at least this can serve as the basis for a geometric interpretation of k-quants.

Another eventual goal is to try the effect of the "best" rounding schemes on quantization-aware training and to test if it matters or not.

Footnotes

1. This is similar, but quite different from trellis coding. The goal is different. Here, the focus is on rounding. ↩