Implement access to device and dtype

[SCiarella]

Closes issue #23

This PR exposes device and dtype such that they can be controlled and set externally by the user. This is particularly relevant for running diffwofost on GPU by setting device="cuda". There are example of this in the notebooks and in the documentation.

The tests have been update by adding a new @pytest.fixture that checks if a GPU is available, and if so, it re-runs the most important tests on the GPU a second time. Notice that GitHub actions do not provide GPU access (for free), so I have been using my local hardware to test the code.

---

Notice that without finalizing the Engine, the GPU device could be slower than the CPU, due to temporary fixes like the use of deepcopy in the TestEngine.

[fnattino]

Nice work @SCiarella! I have left a few comments, note that not all comments are suggestions for changes - some are just notes for future work/discussion. Anyway, feel free to address only what you think makes sense!

[fnattino]

I suspect the default device and dtype will be the same everywhere? So maybe we could define default values of DTYPE and DEVICE at a global level, e.g. in a defaults.py module, instead of having it defined for each module/model? What do yout think?

[SCiarella]

Good point! Yes, they have to be the same everywhere, so it makes sense to follow a more centralized approach that is accessible for the user.

In the next iteration of this PR, I have added to config.py a new class called ComputeConfig that globally controls them. In its docstring, I have reported a small example on how to use that, but we should probably link that somewhere else if we want to use this class to control the global behavior.

[SCiarella]

Thanks @fnattino for the careful review!

I have implemented most of the suggestions that you proposed. In particular, I introduced a new class in config.py called ComputeConfig that is used to set globally dtype and device. Using this global control, I was able to remove most of the redundant type checks that you also pointed out. @SarahAlidoost what do you think of this solution?

There is a point that remains open regarding the kiosk. Currently, I have implemented some type/device conversion for the variables that come from the kiosk. Later, we should ensure that the engine enforces global types and devices, such that the kiosk does not need to be converted every time.

[sonarqubecloud]

Quality Gate passed

Issues
6 New issues
0 Accepted issues

Measures
0 Security Hotspots
87.3% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[fnattino]

Thanks a lot for having a look at all the comments @SCiarella - looks good to me!

[SarahAlidoost]

Thanks @fnattino for the careful review!

I have implemented most of the suggestions that you proposed. In particular, I introduced a new class in config.py called ComputeConfig that is used to set globally dtype and device. Using this global control, I was able to remove most of the redundant type checks that you also pointed out. @SarahAlidoost what do you think of this solution?

I like ComputeConfig implementation. Clean and clear. 👍

There is a point that remains open regarding the kiosk. Currently, I have implemented some type/device conversion for the variables that come from the kiosk. Later, we should ensure that the engine enforces global types and devices, such that the kiosk does not need to be converted every time.

yes lets make an issue for this to not forget it.

[SarahAlidoost]

@SCiarella thanks for the implementation 🥇 All good to me. Nice that you could test the gpu part on your system. Later in #58, we will find alternatives.