pub fn check_scale_bindings(scheme: &QuantScheme, bindings: usize)Expand description
Panic when the scale bindings and the scheme’s levels disagree.
Every level binds a scale buffer of its own, so nothing ties the bindings to the scheme: a missing level is dropped from the reconstruction and every value comes back short by that factor, an extra one is a caller quantizing differently than the scheme it passed.
The global level is further constrained by what this reader serves: it binds as f32 rather than being read as f32 bytes. It has one scale for the whole tensor, so a narrower type saves nothing and only reintroduces rounding error.