-
Notifications
You must be signed in to change notification settings - Fork 27.5k
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
Fix low cpu mem usage tests #30808
Fix low cpu mem usage tests #30808
Conversation
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. |
9c7725a
to
2fac6e3
Compare
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks for fixing ! 🤗
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks!
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thanks!
Would be nice if an explanation about what happens. Like, if it's a change in accelerate
caused this test starting to fail, and if this is the case, a link to a commit in accelerate
could be provided.
In this PR, I fixed the common tests. The issue was that the tests were not designed correctly ( device placement issue ). As for the failing test on UdopModel, this is an issue on the modeling code. With my fix, we can load UdopModel with low_cpu_mem_usage=True. So it's more an issue on transformers. |
* Fix tests * fix udop failing test * remove skip * style
* Fix tests * fix udop failing test * remove skip * style
* Fix tests * fix udop failing test * remove skip * style
What does this PR do ?
This PR fixes the common tests about
low_cpu_mem_usage
(covered in accelerate_tests). The issue that we had was that we moved one of the model to cuda while the other was on cpu. Hence, we get an error while trying to compare those two models. After fixing the issue on the tests, the only failing model is UdopModel that I fixed (we need to tie the weights in_tie_weights
. Otherwise, it won't work withlow_cpu_mem_usage
)See the run here with most of the tests passing
Furthermore, I'm running accelerate tests on every model to see if we still have failing tests + check if there are any failing tests for model parallelism, cpu_offload, disk_offload.
The other failing tests are fixed here. @ydshieh for viz