Follow up on this is: I currently have hacked around this and instead opted for using

pyflyte register

which packages and registers my workflows for me as I have done locally instead. Doing this requires me to install all the python deps and a runner that has a GPU (because imports and global code are run on packaging). Would using the

pyflyte package

->

flytectl register

route of registration allow me to bypass this?

After testing it, doesn't work either. I do have an alternative idea though that might help me overcome both needing to install all the packages into my CI runner and require a GPU. One is: • Put imports inside of a task so that the imports are not run. • Add some flag that is by false during CI that makes the imports conditional for everything other than the flyte objects

Conditional imports seem to be a solution to my problem 🙂

Just decided to go with a 1-1 of what I do locally, before making things nice and tidy and appropriate 🙂

The hack to get around unwanted/GPU only imports is to specifiy a

if os.environ.get("import", False):

just before the import. And then to make it succeed (haven't tested this yet), I have added an env var in the

task

spec

@average-finland-92144 I have a question still relating to CI as is recommended in the documentation. https://docs.flyte.org/en/latest/user_guide/flyte_fundamentals/registering_workflows.html#productionizing-your-workflows I understand that

pyflyte package

and

flytectl register

are the way to go. However, I am confused about when

pyflyte package

is supposed to be run during the CI process? Is it in a docker container? Am I supposed to run it locally, store the package as an artifact in some blob storage before any CI deployment and then download it during deployment to run

flytectl register

with it?

btw and not to distract your efforts, but the cool people from Adarga recently open sourced an operator that does this from inside K8s. I know it's not what you're looking to achieve but can shed some lights on how others are approaching this challenge

Nice, thanks for sharing!