[CI] fix and make less flaky#46543
Conversation
vasqu
left a comment
vasqu
left a comment
There was a problem hiding this comment.
thx, included it locally but we should also fix on main 🫡
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
Hmm, CI is flaky today 😢
|
| "confidence_threshold": 0.005, | ||
| "eos_token_id": [1, 106, 50], | ||
| "pad_token_id": 0, | ||
| } |
There was a problem hiding this comment.
released ckpt have it and we shouldn't hardcode it, only actual generation params need a default. If model is created from scratch, it will still be able to generate with above values
There was a problem hiding this comment.
Hmm seems very specific to diffusion gemma either way. I would agree in general if it were affecting multiple models but here it seems negligible
But we can keep if you feel strong about it
There was a problem hiding this comment.
not really, we do the same in all LLM tester by forcing eos=None
There was a problem hiding this comment.
the stopping criteria comes from AR model anyway which causes us to stop when eos is generated. Thus all our tests become flaky with random "generated sizes"
There was a problem hiding this comment.
oki, let's do it then. I guess it doesnt affect the official checkpoints tho, i.e. we need a patch or similar?
There was a problem hiding this comment.
nope, doens't affect them
|
[For maintainers] Suggested jobs to run (before merge) run-slow: diffusion_gemma |
|
CI Dashboard: View test results in Grafana |
|
View the CircleCI Test Summary for this PR: https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=46543&sha=e5268c |
|
Merging as it's flaky and the failing tests are unrelated |
3 participants
What does this PR do?
as per title, deletes duplicated docstring and fixes flaky tests