Exllama v2 Quantizations of Qwen2.5-14B-Instruct
Using turboderp's ExLlamaV2 v0.2.2 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/Qwen2.5-14B-Instruct
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) | Description |
---|---|---|---|---|---|
8_0 | 8.0 | 8.0 | 17.4 GB | 20.9 GB | Max quality that ExLlamaV2 can produce, recommended. |
6_5 | 6.5 | 8.0 | 14.6 GB | 17.5 GB | Near unquantized performance at vastly reduced size, recommended. |
5_0 | 5.0 | 6.0 | 11.6 GB | 14.4 GB | Slightly lower quality vs 6.5. |
4_25 | 4.25 | 6.0 | 10.1 GB | 13.0 GB | GPTQ equivalent bits per weight. |
3_5 | 3.5 | 6.0 | 8.7 GB | 11.5 GB | Lower quality, not recommended. |
3_0 | 3.0 | 6.0 | 7.8 GB | 10.5 GB | Low quality, not recommended. |
Download instructions
With git:
git clone --single-branch --branch 6_5 https://huggingface.co/bartowski/Qwen2.5-14B-Instruct-exl2 Qwen2.5-14B-Instruct-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/Qwen2.5-14B-Instruct-exl2 --revision 6_5 --local-dir Qwen2.5-14B-Instruct-exl2-6_5
Windows (which apparently doesn't like _ in folders sometimes?):
huggingface-cli download bartowski/Qwen2.5-14B-Instruct-exl2 --revision 6_5 --local-dir Qwen2.5-14B-Instruct-exl2-6.5
Want to support my work? Visit my ko-fi page here: https://ko-fi.com/bartowski