Alright, here’s an attempt to make as trivial and concise an example as possible:

Untitled.py

Is there a way I could change that predict flow so that the @pypi_base decorator doesn’t declare a pinned version of sklearn—but instead uses the same image that was used in the training step? This is a similar idea to MLflow’s models from code packaging format. MLflow records the requirements of a model—whichever are needed for inference. I’d keep the signature/contract/inputs-outputs of model.predict() the same across model versions—but I might swap out the underlying library versions in the Train flow—even the whole framework (e.g. switch from PyTorch to sklearn). And I’d want the predict flow not to require redeployment to accommodate rolling back or forward to the model version trained during any of the run_ids of TrainFlow

cc: @rich-scientist-42991