|
--- |
|
language: |
|
- en |
|
--- |
|
Quantizations for [PygmalionAI/mythalion-13b](https://huggingface.co/PygmalionAI/mythalion-13b) in the [EXL2 format](https://github.com/turboderp/exllamav2) |
|
|
|
Quant|VRAM estimate|Additional |
|
---|---|--- |
|
[4k_hb8_b8](https://huggingface.co/Beinsezii/Mythalion-13b-EXL2/tree/4k_hb8_b8)|18GB|Recommended! |
|
[4k_hb6_b6](https://huggingface.co/Beinsezii/Mythalion-13b-EXL2/tree/4k_hb6_b6)|15GB| |
|
[4k_hb6_b5](https://huggingface.co/Beinsezii/Mythalion-13b-EXL2/tree/4k_hb6_b5)|13GB|Should fit in 12GB cards with 2k context |
|
|
|
|
|
Breaking down the names: |
|
- **4k** is calibrated with 4096 context @ 82 rows (maximum for wikitext) as opposed to the default 2048 context @ 100 rows. |
|
- **hb8** is a header depth of 8 bits |
|
- **b8** is a model weight average of 8.0 bits |
|
|
|
All quantizations were calibrated with [wikitext-2](https://huggingface.co/datasets/wikitext/viewer/wikitext-2-v1/test) |
|
|
|
You can run a model calibrated at 2k with a 4k context or vice versa. The actual difference between 2k and 4k calibrations appears to be very small. |
|
|
|
VRAM estimates are performed with an extremely long chatlog in [oobabooga webui](https://github.com/oobabooga/text-generation-webui) on a 7900 XTX using [nvtop](https://github.com/Syllo/nvtop) to monitor **pytorch usage only**, rounded up. Systems with lots of extra background processes may use more. Additionally, NVIDIA based systems with [flash attention 2](https://github.com/Dao-AILab/flash-attention) **will use less VRAM** than otherwise estimated. |
|
|
|
The measurement files are provided in the main branch so you can [make your own quants](https://github.com/turboderp/exllamav2/blob/master/doc/convert.md) at other bit depths without going through the 2-3 hours of measuring. |
|
|