EDIT: So, to clarify, the below was written after I followed the recommended setting to use my graphics card memory since I had over 12gb available. However when I chose server, the game works. I'm not sure which option I should choose. But it works for the "server" option.It's not working for download. Purchased it, followed the steps appropriately for koboldcpp.exe. I saw you had a troubleshooting thread already for this issue, and I set the Preset to CuBLAS and put the text model to model.gguf. Unfortunately nothing works, the AI fails to load, it says "failed to load scrip koboldccp" or something. and I'm not sure what else to do.
Does the command prompt open and then do nothing for forever like for kuu24? Everything should set itself up automatically.
- You're prompted to choose server or GPU
- You choose GPU
- The game tells you that both files are missing
- The game downloads KoboldCPP
- The game downloads the AI model
- A command prompt opens and KoboldCPP loads the model, which should take 2 minutes at most depending on your drive speed
You shouldn't have to manually do anything with the .exe at all.
If the command prompt opens and then does nothing after "High priority for Windows Set...", perhaps there's a problem with this older version of KoboldCPP and the newest Nvidia drivers? I have no way of verifying this myself because I use a 1070.
Does the command prompt open and then close by itself?
Can you further describe your issue specifically?
The command prompt does open and close by itself, but I notice at the end it provides an error "failed to load script koboldccp" or something in the MsDOS window. Anyway, I've been able to play via server, so I'll try playing GPU. It worked because I ran it as administrator so I'll try that.
Last time I tried this it failed to load the AI model entirely. Even doing the steps suggested didn't work. I don't use an NVIDIA graphics card, I have an AMD build with an RX6800 graphics card.
So an update:
Repeated the steps, ran the program as administrator, and everything downloaded as you specified. However, like the last few times, it runs its program and at the end it displays "unhandled exception" or something and the window closes. Now I'm staring at the same screen for "failed to load AI model". I'm not even sure where to start or what to do, haha. No experience with this AI stuff.
EDIT:
Do you have a discord server or something for faster troubleshooting experience? If not, this is fine.
I've added a build that uses a more recent version of KoboldCPP. It might fix the issue.
If it doesn't fix it, open a command prompt inside the KoboldCPP folder like on the screenshot
Then run:
--model "full path to the model.gguf file here with one set of quotation marks" --usevulkan --gpulayers 43 --quiet --multiuser 100 --skiplauncher --highpriority
Which will show you the full error instead of closing the window when the exception happens.
Since windows doesn't seem to provide a way to capture the output of processes without redirecting it and hiding the command prompt completely, the game can't display the actual error by itself.
Oops, I forgot the koboldcpp.exe before the arguments, but yes, you already posted the error. There's a single search result which turned out to be blasbatchsize related, but blasbatchsize defaults to 512 for Vulkan which isn't a problem.
You can try this which uses OpenCL instead of Vulkan:
koboldcpp.exe --model "full path to the model.gguf file here with one set of quotation marks" --useclblast 0 0 --gpulayers 43 --quiet --multiuser 100 --skiplauncher --highpriority
Otherwise I'm out of ideas other than a driver update since it's an unspecified exception.
You could try moving the entire game folder to a users folder like downloads so it doesn't need any special permissions, though I don't think that's the problem since it looks like it's creating an access violation trying to read from a 64 bit memory adress.
I can't say anything about this beyond that since searching for this problem doesn't really return anything of substance, and KoboldCPP is third-party software which I didn't develop.
"full path to the model.gguf file here with one set of quotation marks" should be replaced with "G:\Silverpine Alpha 0.9.1 Experimental\Silverpine_Data\StreamingAssets\KoboldCPP\model.gguf" here, assuming you didn't move the folder.