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.

Sign up for GitHub

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

fix ollama gpu acceleration #373924

Closed
wants to merge 1 commit into from
Closed

fix ollama gpu acceleration #373924

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 7 additions & 0 deletions pkgs/by-name/ol/ollama/package.nix
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -199,6 +199,13 @@ goBuild {
make ${dist_cmd} -j $NIX_BUILD_CORES
'';

postInstall = lib.optionalString stdenv.hostPlatform.isLinux ''
scd31 marked this conversation as resolved.
Show resolved Hide resolved
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It may be a good idea to replace the condition stdenv.hostPlatform.isLinux with (enableRocm || enableCuda).
The replacement:

  1. Means what it says: if GPU acceleration is enabled, copy GPU related libraries. Currently the copying is always attempted when building on Linux, even on a default, CPU only build (which shouldn't need any GPU libraries to be copied).
  2. Won't activate for CPU builds, which is (probably) good (I'm not entirely certain, though). This script was removed because build errors were encountered when the cp was executed without any arguments to copy (there were no libraries to copy). I think that this probably happened because the copy was executed on CPU builds, and there aren't any files to copy on CPU builds (though, strangely, that would imply that there used to be GPU libraries to copy on CPU builds, since this script used to work).

I'm not actually sure whether or not the CPU build does have any GPU libraries to copy, or whether my theory is at all accurate in the first place, so further testing is probably needed. See also my other comment on the topic.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think switching this would be good as it is more aligned to the intent of the build and isLinux feels like a hack. It may also cut down the build time.

We should test to confirm that there aren't any strange surprises.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Feedback has been addressed (: I will test this on Nvidia and CPU (cannot test on AMD)

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Works on both machines! Nvidia and CPU are both good

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not actually sure whether or not the CPU build does have any GPU libraries to copy, or whether my theory is at all accurate in the first place, so further testing is probably needed. See also my #365718 (comment) on the topic.

There are multiple types of CPU runners that use the same structure. A rocm + CPU build ends up with:

lib/ollama/runners/:
cpu_avx  cpu_avx2  rocm_avx

Official ollama releases end up with cpu_avx cpu_avx2 cuda_v11_avx cuda_v12_avx rocm_avx folders under runners.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Okay, so maybe we want to copy it indiscriminately then?

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, accel flag checking approach means CPU ollama build on this PR is missing cpu_avx/cpu_avx2 😅

$ nix build github:nixos/nixpkgs/5e728b154d0e0f25cbc8f05b7006d001f20df52e#ollama
$ ls result/lib/ollama/runners
ls: cannot access 'result/lib/ollama/runners': No such file or directory

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

lol I will revert

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done!

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

CPU build looks correct now:

$ nix build github:nixos/nixpkgs/6ad8f0640cdf4077335177a93b319b94a7226f15#ollama
$ ls result/lib/ollama/runners
cpu_avx  cpu_avx2

# copy libggml_*.so and runners into lib
# https://github.com/ollama/ollama/blob/v0.4.4/llama/make/gpu.make#L90
mkdir -p $out/lib
cp -r dist/*/lib/* $out/lib/
'';

postFixup =
# the app doesn't appear functional at the moment, so hide it
''
Expand Down