File size: 1,245 Bytes
3fd0224
 
d578be4
 
 
 
 
 
3fd0224
08c544d
e086b09
d4b0dfe
fc46d14
 
 
a465cbf
 
fc46d14
 
0d92d86
1274c18
 
 
 
 
66fc72a
0d92d86
 
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
---
license: cc-by-nc-4.0
inference: false
pipeline_tag: text-generation
tags:
- gguf
- quantized
- text-generation-inference
---
Simple python script (`gguf-imat.py`) to generate various GGUF-IQ-Imatrix quantizations from a Hugging Face `author/model` input, for Windows and NVIDIA hardware.

This is setup for a Windows machine with 8GB of VRAM, assuming use with an NVIDIA GPU. If you want to change the the `-ngl` (number of GPU layers) amount, you can do so at **line 120**. This is only relevant during the `--imatrix` data generation. If you don't have enough VRAM you can decrease the `-ngl` amount or set it to 0 to only use your System RAM instead for all layers.

Your `imatrix.txt` is expected to be located inside the `imatrix` folder. Included file is considered a good option, [this discussion](https://github.com/ggerganov/llama.cpp/discussions/5263#discussioncomment-8395384) is where it came from.

Adjust `quantization_options` in **line 133**.

**Requirements:**
- Python 3.11
  - `pip install huggingface_hub`
 
**Usage:**
```
python .\gguf-imat.py 
```
Quantizations will be output into the created `models\{model-name}-GGUF` folder.
<br><br>
**If this proves useful for you, feel free to credit and share the repository.**