Exllama v2 Quantizations of FuseLLM-7B

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/Wanfq/FuseLLM-7B

No GQA - VRAM requirements will be higher

Branch	Bits	lm_head bits	Size (4k)	Size (16k)	Description
8_0	8.0	8.0	9.4 GB	15.6 GB	Maximum quality that ExLlamaV2 can produce, near unquantized performance.
6_5	6.5	8.0	8.6 GB	14.8 GB	Near unquantized performance at vastly reduced size, recommended.
5_0	5.0	6.0	7.2 GB	13.4 GB	Slightly lower quality vs 6.5, but usable on 8GB cards with 4k context.
4_25	4.25	6.0	6.5 GB	12.7 GB	GPTQ equivalent bits per weight.
3_5	3.5	6.0	5.9 GB	12.1 GB	Lower quality, not recommended.

VRAM requirements listed for both 4k context and 16k context since without GQA the differences are massive (6.2 GB)

Download instructions

With git:

git clone --single-branch --branch 6_5 https://huggingface.co/bartowski/FuseLLM-7B-exl2 FuseLLM-7B-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 FuseLLM-7B-exl2:

mkdir FuseLLM-7B-exl2
huggingface-cli download bartowski/FuseLLM-7B-exl2 --local-dir FuseLLM-7B-exl2 --local-dir-use-symlinks False

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

mkdir FuseLLM-7B-exl2-6_5
huggingface-cli download bartowski/FuseLLM-7B-exl2 --revision 6_5 --local-dir FuseLLM-7B-exl2-6_5 --local-dir-use-symlinks False