Using non-standard (forked) LLaMA C++ branch for quantization.
Using a CLI tool to build KLD evaluation and imatrix calibration datasets for GGUF models, sourced from eaddario/imatrix-calibration.
Using dataset sources: tools, math, code, text_en, text_ru.
Using dataset chunks: 750.
Tensors quantinization F16 instead of BF16, Nvidia Pascal architecture friendly like P100.
Small set of patches added.

Many thanks to Ed Addario for an impressive job.

Quantization comparison

BPW/TGS	PPL correlation	PPL mean ratio	ΔPPL	Mean KLD	Median KLD	Maximum KLD	99.9% KLD	Mean Δp	RMS Δp
3.50	93.18%	1.205301 ± 0.003693	3.066378 ± 0.059410	0.346790 ± 0.003130	0.113532	36.079124	17.036463	-1.548 ± 0.027 %	12.313 ± 0.063 %
4.00	92.50%	1.081636 ± 0.003399	1.219317 ± 0.050468	0.350924 ± 0.003492	0.094924	35.051384	19.111687	-0.644 ± 0.027 %	11.874 ± 0.066 %
4.50	93.92%	1.159525 ± 0.003402	2.382673 ± 0.054462	0.212138 ± 0.003037	0.035486	37.016945	19.404484	-0.722 ± 0.019 %	8.384 ± 0.070 %
5.00	94.32%	1.030399 ± 0.002807	0.454046 ± 0.041667	0.223456 ± 0.003202	0.029363	33.802094	20.041710	-0.307 ± 0.018 %	7.986 ± 0.072 %
5.50	93.59%	0.970038 ± 0.002771	-0.447509 ± 0.042194	0.234948 ± 0.003451	0.024535	32.256840	19.587420	0.123 ± 0.020 %	8.691 ± 0.081 %
6.00	96.85%	1.028155 ± 0.002101	0.420519 ± 0.031506	0.107335 ± 0.002290	0.008626	36.048412	17.149174	-0.060 ± 0.012 %	5.211 ± 0.072 %
6.50	97.55%	1.037597 ± 0.001880	0.561555 ± 0.028552	0.080116 ± 0.001919	0.007975	32.534607	14.545952	-0.128 ± 0.011 %	4.691 ± 0.069 %
7.00	96.92%	1.015746 ± 0.002049	0.235178 ± 0.030606	0.099637 ± 0.002328	0.003383	35.733624	17.560083	0.007 ± 0.010 %	4.330 ± 0.078 %
7.50	97.57%	1.030245 ± 0.001857	0.451735 ± 0.028043	0.067106 ± 0.001900	0.002916	37.250160	14.828805	-0.069 ± 0.009 %	3.827 ± 0.077 %
8.00	97.42%	1.022089 ± 0.001892	0.329923 ± 0.028377	0.077545 ± 0.002062	0.002760	33.393574	16.344349	-0.035 ± 0.009 %	3.835 ± 0.077 %
8.50	97.26%	1.026630 ± 0.001957	0.397746 ± 0.029396	0.082303 ± 0.002123	0.002307	31.664230	17.153591	-0.049 ± 0.009 %	3.815 ± 0.079 %
9.00	98.34%	1.019983 ± 0.001520	0.298461 ± 0.022937	0.044058 ± 0.001500	0.001003	35.295940	10.766310	-0.013 ± 0.007 %	3.132 ± 0.075 %
9.50	98.27%	1.010330 ± 0.001530	0.154286 ± 0.022915	0.051995 ± 0.001669	0.000858	32.079849	13.546452	0.032 ± 0.007 %	3.195 ± 0.078 %
10.00	98.40%	1.013286 ± 0.001478	0.198433 ± 0.022200	0.044456 ± 0.001528	0.000833	31.551548	12.255160	0.002 ± 0.007 %	3.022 ± 0.076 %
10.50	98.30%	1.012990 ± 0.001525	0.194020 ± 0.022882	0.047429 ± 0.001597	0.000826	33.701038	13.457508	0.019 ± 0.007 %	3.073 ± 0.078 %
11.00	98.35%	1.019113 ± 0.001514	0.285470 ± 0.022865	0.042238 ± 0.001490	0.000819	31.194330	11.399836	0.001 ± 0.006 %	2.878 ± 0.075 %

ENOSYS/GLM-4.7-Flash-Uncensored-750-v1-GGUF

Experimental global target bits‑per‑weight quantization of HauhauCS/GLM-4.7-Flash-Uncensored-HauhauCS-Aggressive

Quantization comparison

ENOSYS/
GLM-4.7-Flash-Uncensored-750-v1-GGUF