19 lines
1.0 KiB
Markdown
19 lines
1.0 KiB
Markdown
---
|
|
license: gemma
|
|
base_model:
|
|
- ifable/gemma-2-Ifable-9B
|
|
pipeline_tag: text-generation
|
|
---
|
|
## 🛑 Note: not every quant is displayed on the table on the right, you can find everything [here](https://huggingface.co/Hampetiudo/gemma-2-Ifable-9B-i1-GGUF/tree/main).
|
|
|
|
Using <a href="https://github.com/ggerganov/llama.cpp/">llama.cpp</a> release <a href="https://github.com/ggerganov/llama.cpp/releases/tag/b3804">b3804</a> for quantization.
|
|
|
|
Original model: https://huggingface.co/ifable/gemma-2-Ifable-9B
|
|
|
|
All quants were made using the imatrix option (except BF16, that's the original precision). The imatrix was generated with the dataset from [here](https://gist.github.com/tristandruyen/9e207a95c7d75ddf37525d353e00659c), using the BF16 GGUF with a context size of 8192 tokens (default is 512 but higher/same as model context size should improve quality) and 13 chunks.
|
|
|
|
How to make your own quants:
|
|
|
|
https://github.com/ggerganov/llama.cpp/tree/master/examples/imatrix
|
|
|
|
https://github.com/ggerganov/llama.cpp/tree/master/examples/quantize |