bartowski/code-millenials-13b-exl2

Exllama v2 Quantizations of code-millenials-13b

Using turboderp's ExLlamaV2 v0.0.11 for quantization.

The "main" branch only contains the measurement.json, download one of the other branches for the model (see below)

Each branch contains an individual bits per weight, with the main one containing only the meaurement.json for further conversions.

Original model: https://huggingface.co/budecosystem/code-millenials-13b

No GQA - VRAM requirements will be higher

Branch	Bits	lm_head bits	Size (4k)	Size (16k)	Description
6_5	6.5	8.0	14.4 GB	24.0 GB	Near unquantized performance at vastly reduced size, recommended.
5_0	5.0	6.0	12.1 GB	21.7 GB	Slightly lower perplexity vs 6.5, can fit in 12 GB card with even lower context.
4_25	4.25	6.0	10.9 GB	20.5 GB	GPTQ equivalent bits per weight.
3_75	3.75	6.0	10.1 GB	19.7 GB	Lower quality but still generally usable.
3_0	3.0	6.0	9.1 GB	18.7 GB	Very low quality, not recommended unless you have to.

VRAM requirements listed for both 4k context and 16k context since without GQA the differences are massive (9.6 GB)

Download instructions

With git:

git clone --single-branch --branch 6_5 https://huggingface.co/bartowski/code-millenials-13b-exl2 code-millenials-13b-exl2-6_5

With huggingface hub (credit to TheBloke for instructions):

pip3 install huggingface-hub

To download the main (only useful if you only care about measurement.json) branch to a folder called code-millenials-13b-exl2:

mkdir code-millenials-13b-exl2
huggingface-cli download bartowski/code-millenials-13b-exl2 --local-dir code-millenials-13b-exl2 --local-dir-use-symlinks False

To download from a different branch, add the --revision parameter:

mkdir code-millenials-13b-exl2-6_5
huggingface-cli download bartowski/code-millenials-13b-exl2 --revision 6_5 --local-dir code-millenials-13b-exl2-6_5 --local-dir-use-symlinks False