The q4_0 comparison at the end of e21u_analyze.py compares stored q4_0 values with unrotated f16 values.
This llama.cpp build stores quantized K rotated by a 256-dim Hadamard transform and quantized V by a 64-dim Hadamard transform,
so that first comparison is invalid. The published q4_0 error numbers apply the same rotation to the f16 values before comparing.