Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[ISSUE] --gpu-layers in useless. #1555

Closed
skidd-level-100 opened this issue May 21, 2023 · 4 comments · Fixed by #1625
Closed

[ISSUE] --gpu-layers in useless. #1555

skidd-level-100 opened this issue May 21, 2023 · 4 comments · Fixed by #1625

Comments

@skidd-level-100
Copy link

I am on linux with an nvidia card using CLBLAST (CUBLAST hates me) and --gpu-layers (any number) is no faster than without it,
I checked the nvidia-smi ouput and it is showing me only 80mib vram used (same as without -ngl)

pls help

@Green-Sky
Copy link
Collaborator

Please use the cuda/cublas variant over the OpenCL one. It is always better on nvidia hardware.

@JohannesGaessler
Copy link
Collaborator

The PR for OpenCL GPU acceleration #1459 hasn't been merged yet so setting --n-gpu-layers with LLAMA_CLBLAST does nothing.

@KerfuffleV2
Copy link
Collaborator

KerfuffleV2 commented May 21, 2023

Would it be a good idea to have --n-gpu-layers fail if stuff isn't compiled in a way that enables actually putting layers on the GPU? Could probably just add some #ifdefs around the commandline option unless there's actually a reason to allow the user to use the argument even when there's no effect. (I guess an alternative is just to display a warning/notice in that case rather than raising an error.)

@JohannesGaessler
Copy link
Collaborator

I opened a PR that adds documentation and an error upon trying to use --n-gpu-layers without an implementation. I'll revisit it once my other PR for optimization options gets merged.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging a pull request may close this issue.

4 participants