Add framework fallback ability to execute in sparseml

[markurtz]

No description provided.

[bfineran]

good catch on this necessary functionality. I'm a little worried about how coupled the attempt_num logic is in detect_framework with the execute function. Is there a specific reason for doing it this way? Any thoughts on having a detect function that returns all possible frameworks for an item, and then having execute return the first valid in-framework run of the function between the possible frameworks?

[markurtz]

good catch on this necessary functionality. I'm a little worried about how coupled the attempt_num logic is in detect_framework with the execute function. Is there a specific reason for doing it this way? Any thoughts on having a detect function that returns all possible frameworks for an item, and then having execute return the first valid in-framework run of the function between the possible frameworks?

The original reason why I went this route is that I'm worried about performance for evaluating of a model. Evaluation of model could involve full loading of the model to check if it is tensorflow or keras for example. It is a very fair point that this pathway reduces readability and adds complexity. This is additionally a preoptimization for a performance use case we haven't hit yet, so I'll migrate the code over to return a list of frameworks instead of this more complicated pathway for now

[bfineran]

LGTM. If we see performance issues when trying to load large models in the wrong framework, let's definitely update the detect_frameworks function to use what you had before, or maybe a generator if that would work