File size: 1,247 Bytes
7449d16
 
 
 
 
 
 
 
 
 
8943cdc
ae2e095
 
fb2d556
ae2e095
1efcf9c
 
 
 
 
 
 
 
 
 
 
 
 
fb2d556
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
---
license: gemma
language:
- en
pipeline_tag: text-generation
tags:
- google
- gemma
- gguf
- imatrix
base_model: google/gemma-2-9b-it
---

# Quant Infos

- quants done with an importance matrix for improved quantization loss
- Currently requantizing ggufs & imatrix from bf16 for "optimal" accuracy loss
  - initial version was based on f32 gguf provided by google
  - WIP new version should have better metadata as its quantized from scratch with llama.cpp
- Wide coverage of different gguf quant types from Q\_8\_0 down to IQ1\_S
  - experimental custom quant types
    - `_L` with `--output-tensor-type f16 --token-embedding-type f16` (same as bartowski)
    - `_XL` with `--output-tensor-type bf16 --token-embedding-type bf16` (same size as _L, in theory even higher numerical accuracy)
- Quantized with [llama.cpp](https://github.com/ggerganov/llama.cpp) release [b3259](https://github.com/ggerganov/llama.cpp/releases/tag/b3259)
- Imatrix generated with [this](https://gist.github.com/bartowski1182/eb213dccb3571f863da82e99418f81e8) multi-purpose dataset by [bartowski](https://huggingface.co/bartowski).
  ```
  ./imatrix -c 512 -m $model_name-bf16.gguf -f calibration_datav3.txt -o $model_name.imatrix
  ```

# Original Model Card

TODO