Open KenMacD opened 1 month ago
I agree that oneapi support appears to be a good idea, especially given the (admittedly unlikely) possibility of wider industry adoption by amd and nvidia. However, given my lack of access to a supported device for testing, combined with oneapi support being quite new in ollama, possibly being incomplete or buggy for now, as well as my lack of experience with oneapi at the moment, I wouldn't expect oneapi support particularly soon in nix; it'll probably be at least a couple weeks.
If you can help with testing, or want to open a pull request, I'd be happy to work together with you to add support sooner than I'd be able to do myself.
I have it packaged for oneapi here https://github.com/MordragT/nixos/blob/master/pkgs/by-name/ollama-sycl/default.nix Iam pretty sure you will (if you want to compile from source) need the properitary dpcpp compiler or atleast intels open source llvm sycl compiler. My package uses the former. However I wanted to tryout the open source llvm sycl compiler first before I try to upstream it.
I see
ollama
has acuda
androcm
version. Ollama appears to support aoneapi
version now. From some tests it appears this would require access to alibze_intel_gpu.so
library, which I see in theintel-compute-runtime.drivers
.I tested setting
OLLAMA_INTEL_GPU=1
(withOLLAMA_DEBUG=1
) in the module settings and see the following logs:It'd be nice to see an intel/openapi/sycl support added.