pub fn quantize_rows(x: &[f32], head_dim: usize) -> (Vec<i8>, Vec<f32>)Expand description
Quantise rows head-vectors of head_dim into stored integers plus one
scale each — the same layout V3Cache holds internally.
The batched prefill path needs this so it can attend through exactly the
arithmetic crate::v3::attend_step uses on the stored cache. Attending
over dequantised f32 instead is algebraically the same and numerically
is not, which would make a prefilled session disagree with a stepped one.