Modelpublic

tooltd/Qwen3.8-27B-GSQ-RCO-BuffedMod-GGUF

sourceHugging Faceapache-2.0updated 7d agoView on Hugging Face

1likes1.5kdownloads

Model Card

Qwen3.8-27B-GSQ-RCO — BuffedMod GGUF

This repository contains modified versions of the Qwen3.8-27B-GSQ-RCO-IQ3_S GGUF model.

Unlike traditional re-quantization workflows which require quantizing the entire model from unquantized FP16/BF16 weights, these files were created by directly modifying specific tensor weights within the binary GGUF file.

By replacing critical layer tensors (specifically output.weight) with higher-precision quantizations, we achieve better Perplexity (PPL), lower KL Divergence (KLD), and increased top-token agreement while maintaining a compact footprint.

⚡ What Changed in BuffedMod?

In lower-bit quantizations like IQ3_S, the `output.weight` tensor (LM Head) often serves as a major bottleneck for final text generation precision. Upgrading this tensor improves top-token probability matching without adding massive overhead to the overall file size.

📦 Available Variants

1.`Qwen3.8-27B-GSQ-RCO-BuffedMod-IQ3_S_XL-mtp-Q6K.gguf`
2.Modifications: Replaced output.weight from Q4_K to `Q6_K`.
3.Target: Maximum quality retention for IQ3_S baseline.

1.`Qwen3.8-27B-GSQ-RCO-BuffedMod-IQ3_S_XL-mtp-IQ4XS.gguf`
2.Modifications: Replaced output.weight from Q4_K to `Q6_K` AND down-quantized MTP (Multi-Token Prediction) tensors from Q6_K to `IQ4_XS`.
3.Target: Optimized size-to-performance balance by reclaiming memory from non-essential auxiliary tensors.

📊 Benchmark & Evaluation Results

All variants were evaluated using llama-perplexity with original BF16 base model (q38f16baseline.kld) to measure Perplexity (PPL) degradation, KL Divergence (KLD), and Token Probability shifts.

bash

llama-perplexity.exe -f /wikitext-2-raw/wiki.test.raw \
    --kl-divergence \
    --kl-divergence-base q38f16baseline.kld \
    -ngl 99 \
    -m <model_variant>.gguf

📈 Comparison Summary

Metric	Original (`IQ3_S`)	BuffedMod (`Q6_K` Output)	Improvement
Mean PPL(Q)	`7.062697`	`7.032142`	🟢 -0.030555 (Lower is better)
Mean PPL Diff vs Base	`0.112205`	`0.081650`	🟢 -27.2% Error Reduction
Mean KLD	`0.055475`	`0.051738`	🟢 -6.7% Divergence
Median KLD	`0.022928`	`0.018899`	🟢 -17.5% Divergence
Same Top Token Probability	`89.657%`	`90.535%`	🟢 +0.878% Match
RMS Δp	`6.537%`	`6.285%`	🟢 -0.252% Variance

📐 Model Variant Comparison

The table below highlights the trade-off between file size efficiency and accuracy gains across the original and BuffedMod variants:

Model Variant	File Size	PPL(Q) ↓	PPL Diff vs Base ↓	Mean KLD ↓	Median KLD ↓	Top-Token Match ↑
Original (`IQ3_S-mtp`)	11.29 GB	7.0627	0.1122	0.0555	0.0229	89.66%
BuffedMod (`IQ4XS MTP`)	11.49 GB	7.0321	0.0817	0.0517	0.0189	90.54%
BuffedMod (`Q6_K MTP`)	11.59 GB	7.0321	0.0817	0.0517	0.0189	90.54%

📌 Key Highlights: `BuffedMod (IQ4XS MTP)` offers the optimal sweet spot: it gains the full accuracy boost (+0.88% top-token match, -27.2% PPL error vs base) with only +200 MB size overhead compared to the original file. `BuffedMod (Q6_K MTP)` retains maximum fidelity on the auxiliary MTP tensors for an extra +100 MB, maintaining identical core perplexity and divergence performance.

📋 Full Metrics Breakdown

🔴 Baseline Original Metrics (`IQ3_S` with `Q4_K` output.weight)

text

====== Perplexity statistics ======
Mean PPL(Q)                   :   7.062697 ±   0.045484
Mean PPL(base)                :   6.950493 ±   0.044934
Cor(ln(PPL(Q)), ln(PPL(base))):  98.65%
Mean ln(PPL(Q)/PPL(base))     :   0.016014 ±   0.001059
Mean PPL(Q)/PPL(base)         :   1.016143 ±   0.001076
Mean PPL(Q)-PPL(base)         :   0.112205 ±   0.007438

====== KL divergence statistics ======
Mean    KLD:   0.055475 ±   0.000453
Maximum KLD:  15.811345
99.9%   KLD:   2.067696
99.0%   KLD:   0.593792
95.0%   KLD:   0.188752
90.0%   KLD:   0.110145
Median  KLD:   0.022928

====== Token probability statistics ======
Mean    Δp: -0.513 ± 0.017 %
RMS Δp    :  6.537 ± 0.051 %
Same top p: 89.657 ± 0.079 %

🟢 BuffedMod Metrics (`IQ3_S` with `Q6_K` output.weight)