I’ve been messing around with GPTQ models with ExLlama in ooba, and have gotten 33b models @ 3k running smoothly, but was looking to try something bigger than my VRAM can hold.
However, I’m clearly doing something wrong, and the koboldcpp.exe documentation isn’t clear to me. Does anyone have a good setup guide? My understanding is koboldcpp.exe is preferable for GGML, as ooba’s llama.cpp doesn’t support GGML at >4k context yet.
KoboldCpp has documentation on the github page. Maybe just google for other guides if the documentation doesn’t do it for you.
My advice is: Do one step at a time. Get it running first, without fancy stuff. Start with a small model and without gpu acceleration. Then get the acceleration/CUDA working. Then try with a bigger model. And then you can do the elaborate stuff like having some layers in VRAM and others in RAM and blowing up the context size past 2048/default. Don’t do it all at once. That way you might figure out your problem and at which of the steps it happens.
(Edit: And make sure to always use the latest version. You’re playing with pretty recent stuff that still might have bugs.)
I can’t say much about the windows stuff or the state of the integration layers in oobabooga’s.
What’s the problem you’re having with kobold? It doesn’t really require any setup. Download the exe, click on it, select model in the window, click launch. The webui should open in your default browser.
Note this is koboldcpp.exe and not KoboldAI.
The Github describes arguments to use GPU acceleration, but it is fuzzy on what the arguments do and completely neglects to mention what the values for those arguments do. I understand the --gpulayers arg, but the two ints after --useclblast are lost on me. I defaulted to “[path]\koboldcpp.exe --useclblast 0 0 --gpulayers 40”, but it seems to be completely ignoring GPU acceleration, and I’m clueless where the problem lies. I figured it would be easier to ask for a guide and just start my GGML setup from scratch.
Those are OpenCL platform and device identifiers, you can use clinfo to find out which numbers are what on your system.
Also note that if you’re building kobold.cpp yourself, you need to build with LLAMA_CLBLAST=1 for OpenCL support to exist in the first place. Or LLAMA_CUBLAS for CUDA.