oobabooga / text-generation-webui

A Gradio web UI for Large Language Models.
GNU Affero General Public License v3.0
40.87k stars 5.34k forks source link

exl2 autosplit #6495

Open noobhunterd opened 1 month ago

noobhunterd commented 1 month ago

Describe the bug

Im using this model: https://huggingface.co/LoneStriker/airoboros-70b-3.3-2.4bpw-h6-exl2

If i load it under 1 GPU, it works perfectly with 2k context + 8bit But if I use autosplit under 2 GPUs as 8k context, it responses in nonsense

Is there some option that I should tick to make it work? Just updated to the latest version of text gen and L2 version of airoboros works fine with autosplit on

Is there an existing issue for this?

Reproduction

latest version of text gen load this model: https://huggingface.co/LoneStriker/airoboros-70b-3.3-2.4bpw-h6-exl2 w/ autosplit and 8k context

Screenshot

No response

Logs

Instruct:

1 GPU log:
Input: hello
Output: Hello! I'm here to help you with any questions or tasks you may have. What can I assist you with today?

Autosplit log:
Input: hello
Output: The string to the text-to-text model 1001

System Info

4090 FE + 3090 MSI via eGPU