CUDA: use registers instead of smem in topk-moe

[am17an]

Uses the technique used in the vulkan PR #16641. Neat trick!

[am17an]

I am not able to restart tests for CI, The nvidia-cuda test is failing with

The self-hosted runner lost communication with the server. Verify the machine is running and has a healthy network connection. Anything in your workflow that terminates the runner process, starves it for CPU/Memory, or blocks its network access can cause this error.

I'd rather not push another commit to trigger the CI again. @ggerganov is this a permission issue and can it be changed?

[JohannesGaessler]

If one first clicks on a job and then one of the buttons it is possible to manually re-run it. Since permissions were recently tightened I don't know whether you could do that yourself though (I just did).

[am17an]

If one first clicks on a job and then one of the buttons it is possible to manually re-run it. Since permissions were recently tightened I don't know whether you could do that yourself though (I just did).

Yes I was able to do that earlier as well

[ggerganov]

@am17an It seems that approving workflows requires write access, so I just loosened up the approval requirement to "first-time contributors" from "all external contributors":

Your workflows should always run now without additional approval.

If this becomes too heavy for the CI or we think it can become a security concern, we might switch back to manual workflow approval for all PRs.

[am17an]

Thanks. Seems like "collaborator" is not a well thought design by Github, my intuition would be a collaborator would be able to run workflows on their own PRs (at the least)

[am17an]

Actually I'm still not able to re-run jobs. This is the screen I see