intfloat/multilingual-e5-base fales

[NadaComa]

Dear all,

using the latest release of privateGPT, I get the following error when using intfloat/multilingual-e5-base for the embeddings.
CUDA error 710 at /tmp/pip-install-281klxnb/llama-cpp-python_cf43f0d85af444668d34ca5d9953a909/vendor/llama.cpp/ggml-cuda.cu:7656: device-side assert triggered
current device: 0

This does not happen when using the same model in the previous release. Does anyone have an idea what I would need to change to get rid of this? My guess would be that the chunking works differently.

Btw: This only seems to happen for large Text Files. ~2000 tokens and more. When I use intfloat/multilingual-e5-small, it all runs smoothly.

Hardware: RTX 3060 with 12 GB RAM. In the previous release, even using the large model, only uses about 4GB.

Cheers Nada

[chenxin23233e]

I dont know why the github account banned so much recently. Do you know why?