Self hosted LLM
Self hosted LLM
Hello internet users. I have tried gpt4all and like it, but it is very slow on my laptop. I was wondering if anyone here knows of any solutions I could run on my server (debian 12, amd cpu, intel a380 gpu) through a web interface. Has anyone found any good way to do this?
You're viewing a single thread.
I tried Huggingface TGI yesterday, but all of the reasonable models need at least 16 gigs of vram. The only model i got working (on a desktop machine with a amd 6700xt gpu) was microsoft phi-2.
Have you been able to use it with your AMD GPU? I have a 6800 and would like to test something
Yes, since we have similar gpus you could try the following to run it in a docker container on linux, taken from here and slightly modified:
#!/bin/bash model=microsoft/phi-2 # share a volume with the Docker container to avoid downloading weights every run volume=<path-to-your-data-directory>/data docker run -e HSA_OVERRIDE_GFX_VERSION=10.3.0 -e PYTORCH_ROCM_ARCH="gfx1031" --device /dev/kfd --device /dev/dri --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:1.4-rocm --model-id $modelNote how the rocm version has a different tag and that you need to mount your gpu device into the container. The two environment variables are specific to my (any maybe yours also) gpu architecture. It will need a while to download though.
Koboldcpp should allow you to run much larger models with a little bit of ram offloading. There's a fork that supports rocm for AMD cards: https://github.com/YellowRoseCx/koboldcpp-rocm
Make sure to use quantized models for the best performace, q4k_M being the standard.
I know the gpt4all models run fine on my desktop with 8gig vram. It does use a decent chunk of my normal ram though. Could the gpt4all models work on huggingface or do they use different formats? Sorry if I am completely misunderstanding huggingface, I haven't heard of it until now.
Huggingface TGI is just a piece of software handling the models, like gpt4all. Here is a list of models officially supported by TGI, although they state that you can try different ones as well. You follow the link and look for the files section. The size of the model files (safetensors or pickele binaries) gives a good estimate of how much vram you will need. Sadly this is more than most consumer graphics cards have except for santacoder and microsoft phi.
I don't really want to try to get that to work. I wonder how hard it would be to create my own webui using gpt4all's Python package.