search

tenstorrent / tt-umd

User-Mode Driver for Tenstorrent hardware
Apache License 2.0
9 stars 5 forks source link

Work on adding CI to umd repo #45

Closed broskoTT closed 1 month ago

broskoTT commented 2 months ago

Nothing to add for know, it should be pretty clear. Guard the UMD with tests

vtangTT commented 1 month ago

Just FYI, as a stopgap solution for metal, we are running most/all UMD tests on the metal CI. They run for every push to main and every post-commit

https://github.com/tenstorrent/tt-metal/actions/workflows/umd-unit-tests-wrapper.yaml

broskoTT commented 1 month ago

Every push to tt-metal main, or umd main? In any case we should have a CI here if we plan to ship this as a separate component/endpoint to our products.

vtangTT commented 1 month ago

Every push to tt-metal main will run post-commit which includes umd tests. Example run from most recent main:

https://github.com/tenstorrent/tt-metal/actions/runs/10794191678/job/29937907086

But I agree, I think the CI should ultimately be in this repo.

broskoTT commented 1 month ago

That seems unnecessary, since not every commit on tt-metal will update the UMD commit used. Sounds like you're running same tests on the same version of umd? Also, this doesn't offer any guarantees for UMDs progress, i.e. this doesn't guarantee the current tip of the branch is functional.

vtangTT commented 1 month ago

Yes, I agree it's not the best solution and like you said it only guarantees umd is not broken for the version metal points to.

It was mainly implemented to have some sort of verification for UMD (since at that time we aren't able to get CI machines onto the UMD repo). Also on the off chance that metal exposes an edge-case bug in umd.

It also allowed us to dispatch umd tests on CI through this pipeline: https://github.com/tenstorrent/tt-metal/actions/workflows/umd-unit-tests-wrapper.yaml

The tests themselves run fast enough such that they don't affect the total runtime of post-commit. If development is ramping up on UMD, CI should def be present in this repo, but temporarily it's possible to validate through metal CI if needed.