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

Introduce a torch and torchax roller #103

Open
tengyifei opened this issue Feb 10, 2025 · 0 comments
Open

Introduce a torch and torchax roller #103

tengyifei opened this issue Feb 10, 2025 · 0 comments

Comments

@tengyifei
Copy link
Collaborator

tengyifei commented Feb 10, 2025

Currently CI and E2E tests uses the nightly build of PyTorch and PyTorch/XLA: https://github.com/AI-Hypercomputer/torchprime/blob/main/.github/workflows/cpu_test.yml, https://github.com/AI-Hypercomputer/torchprime/blob/main/torchprime/launcher/Dockerfile#L4. This means checks may just stop working at any point if there is an upstream introduced regression.

To prevent these kind of breakages, the standard practice is to use a roller:

  • We pin these docker images at some date
  • A bot (roller) makes a pull request to update the pin every day
  • The pull request is automatically merged if all the checks pass

Because this roller is only run nightly, we may afford to run more extensive regression tests, such as testing a full pod Llama 3.1 405B performance or even on multiple pods. The idea is that we'll run these tests overnight, and by the next morning, we can see if there were any regressions on the PR. Again, if a new nightly build introduces regressions, then that PR can't be merged. An engineer will inspect the profiles to track down the source of the regression in torch_xla.

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

No branches or pull requests

1 participant