Using on K80 Tesla Windows - Cuda No Kernal Image

[webclinic017]

Nvidia & its drivers are correctly installed 11.4 I am getting an output also for nvidia-smi, also while running in settings I can see k8o is recognized in koboldcpp

but, I'm getting this error as I run the app .exe

ggml_cuda_compute_forward: ADD failed
CUDA error: no kernel image is available for execution on the device
current device: 0, in function ggml_cuda_compute_forward at d:\a\koboldcpp\koboldcpp\ggml\src\ggml-cuda\ggml-cuda.cu:2325
err
d:\a\koboldcpp\koboldcpp\ggml\src\ggml-cuda\ggml-cuda.cu:75: CUDA error

[LostRuins]

Ah, so K80 was still working before this version? i.e. 1.84.2?

I had just recently dropped K80 support as I could no longer obtain K80 GPUs anywhere on the cloud to test it, and nobody was willing/able to provide K80 testing.

If you are willing to help me verify it works, I can add it back in again. Alternatively, how well does Vulkan work for you?

[LostRuins]

I can prepare a test binary for you in the meantime.

Please try this: https://github.com/LostRuins/koboldcpp/actions/runs/13717728758/artifacts/2710143596

[webclinic017]

@LostRuins Wonderful I tried the file - with cuda and its working without any issues.. Thanks for this.. if u want me to test anything else please let me know

[LostRuins]

Alright, I will retain k80 as a build target then

[LostRuins]

If needed, i'll let you know again and you can help test future versions.

A K80 is nearly impossible to rent outside nowadays, and almost none of our users have it. So If you'd like to help, it will be very useful in future.

[henk717]

@LostRuins in general we only have the 11.4 version these days for K80 support. Without it we could drop the entire version instrad of cutting K80 support. But since K80's work its nice to keep. Vulkan on them will probably be bad, those drivers predate the compute work that went into vulkan so its going to fall back to much slower implementations than the modern cards do.

[LostRuins]

The cu11.4 version is my recommendation to newbies as it has much wider out-of-box support.

For example, a 2023 brand new top end laptop ships with win11 and cuda 12.0, and thus will not work out of box with the cu12 build, whereas the cu11 build will work fine.

The speed difference is at most 10%.

[LostRuins]

K80 support has been reinstated in the latest version (koboldcpp.exe cu11). Please check if it works well for you.