bartowski/QwQ-32B-Preview-exl2

Exllama v2 Quantizations of QwQ-32B-Preview

Using turboderp's ExLlamaV2 v0.2.4 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./Qwen/QwQ-32B-Preview

Prompt format

<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Available sizes

Branch	Bits	lm_head bits	VRAM (4k)	VRAM (16k)	VRAM (32k)	Description
8_0	8.0	8.0	34.9 GB	37.6 GB	41.6 GB	Max quality producable by ExLlamav2, generally unneeded but maximum performance
6_5	6.5	8.0	28.9 GB	31.6 GB	35.6 GB	Near unquantized performance at vastly reduced size, recommended.
5_0	5.0	8.0	22.6 GB	25.3 GB	29.3 GB	Very high quality, usable at 4k context on 24GB.
4_25	4.25	6.0	19.5 GB	22.2 GB	26.2 GB	GPTQ equivalent bits per weight, slightly higher quality.
3_5	3.5	6.0	16.5 GB	19.2 GB	23.2 GB	Lower quality, only use if you have to.
3_0	3.0	6.0	14.3 GB	17.0 GB	21.0 GB	Very low quality, usable with 16gb of VRAM.

Download instructions

With git:

git clone --single-branch --branch 6_5 https://huggingface.co./bartowski/QwQ-32B-Preview-exl2 QwQ-32B-Preview-exl2-6_5

With huggingface hub (credit to TheBloke for instructions):

pip3 install huggingface-hub

To download a specific branch, use the --revision parameter. For example, to download the 6.5 bpw branch:

Linux:

huggingface-cli download bartowski/QwQ-32B-Preview-exl2 --revision 6_5 --local-dir QwQ-32B-Preview-exl2-6_5

Windows (which apparently doesn't like _ in folders sometimes?):

huggingface-cli download bartowski/QwQ-32B-Preview-exl2 --revision 6_5 --local-dir QwQ-32B-Preview-exl2-6.5

Want to support my work? Visit my ko-fi page here: https://ko-fi.com/bartowski