-
Notifications
You must be signed in to change notification settings - Fork 42
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 conda build #2458
base: main
Are you sure you want to change the base?
Fix conda build #2458
Conversation
@pbchekin the test was failing in a normal run due to psutil missed in pip cache; I have restored the way it worked before (without caching) |
- name: Install transformers | ||
if: (inputs.env_manager != 'base') && (steps.conda-cache.outputs.status == 'miss') | ||
run: | | ||
git clone --recursive https://github.com/huggingface/transformers |
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.
Use actions/checkout?
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.
Pavel @pbchekin , what do you think if I add e2e test (including installation of required parts) to the test-triton.sh
script, so this would work uniformly with other tests? Do developers run this test?
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.
It could work, but what's your idea for installing transformers, torch{vision,audio,text} from cache? Compiling them every time from sources will be time consuming.
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.
I can use gh to download the pre-built artifacts. That way this will work for developers as well. Aren't these artifacts already available in pip? If yes, I can install them using pip
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.
They are available as artifacts in our nightly wheels, but the workflow should also work when we change the pytorch pin (pytorch has pins for all these dependencies).
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.
ok, I see; this definitely requires some rework (note, your existing actions do not work for conda workflow, I cannot just reuse them); I suggest splitting the task into two - first I do the "fixing", than I can implement caching, or rebuilt wheels if they are not found
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.
why do we need both wheels and a cache?
I've tried using pre-built wheels, it does not work (jit compiler cannot find oneapi) |
No-basekit requires more work