In my training workflow, I have a cheap register params task that registers and shows the params in flyte deck. We all know provisioning a GPU can take a while, so it was nice to use the UI to see the params. In flyte v2 the parent task is an always on pod, which costs money to keep alive. This lead me to two options for the new workflow: • Pay extra money to have a parent task and see the params • Remove the parent task and turn

register_params

into a trace, but wait until the GPU is provisioned to see the flyte deck. I'm leaning towards the 2nd option, but I'm curious if this also what the community would recommend. Thank you!

@workflow
def wf(ff: FlyteFile) -> str:
    param_uuid = register_params(ff=ff)
    return train_model(param_uuid=param_uuid)

# option 1
@cpu_env
def wf(ff: File) -> str:
    param_uuid = register_params(ff=ff)
    return train_model(param_uuid=param_uuid)

# option 2
def train_model()
    # trace
    param_uuid = register_params(ff=ff)
    # just python from the train_model task
    start_training_job()

Just run parent on a cpu pod for cheap - use low resources and keep it running you can pack a lot