README.md

---
tags:
- merge
- gguf
- not-for-all-audiences
- storywriting
- text adventure
---

# maid-yuzu-v8-alter-iMat-GGUF

<b>Update:</b> Legacy quants calculated with imatrix showed lower average divergence than expected when compared to their non-imat variants. Uploading those now as well.

<b>Highly requested model.</b> Quantized from fp16 with love.
* <s>1st batch (IQ3_S, IQ3_XS) use a imatrix.dat file calculated from Q8 quant.</s> These have been removed in favor of a newer method. Please see tables below.
* Later files made using .imatrix file from [this](https://huggingface.co/datasets/ikawrakow/imatrix-from-wiki-train) repo (special thanks to [ikawrakow](https://huggingface.co/ikawrakow) again)

<h3>Original IQ3_XS KL-Divergence (calculated vs Q8)</h3>
<img src="https://huggingface.co/InferenceIllusionist/maid-yuzu-v8-alter-iMat-GGUF/resolve/main/IQ3_XS_OLD_KL.JPG?download=true" width="300"/>

<h3>Updated IQ3_XS KL-Divergence (calculated vs Q8)</h3>
<img src="https://huggingface.co/InferenceIllusionist/maid-yuzu-v8-alter-iMat-GGUF/resolve/main/IQ3_XS_NEW_KL.JPG?download=true" width="300"/>
Lower numbers are better

For a brief rundown of iMatrix quant performance please see this [PR](https://github.com/ggerganov/llama.cpp/pull/5747)

<i>All quants are verified working prior to uploading to repo for your safety and convenience. </i>

Please note importance matrix quantizations are a work in progress, IQ3 and above is recommended for best results. 

<b>Tip:</b> Pick a size that can fit in your GPU while still allowing some room for context for best speed. You may need to pad this further depending on if you are running image gen or TTS as well.

Original model card can be found [here](https://huggingface.co/rhplus0831/maid-yuzu-v8-alter)