[MNT] integration plan with GC.OS

[fkiraly]

We are currently planning a maintenance integration of pytorch-tabular with GC.OS.
This will include homogenization with the torch stack in pytorch-forecasting, and sktime. All packages will remain separate, but with an aim to framework homogenization (and possibly integration where reasonable)

This issue is to plan urgent maintenance items and maintenance integration.

Community input on todos and wishlist is also appreciated, e.g., what are "burning" items, suggested priorities.

Maintenance items

• pyproject migration - [Maintenance] Updating the dependency versions and migration to pyproject.toml #590
• pyproject review and dependabot PRs
• soft dependency review and potential isolation
• stale PR review and triage - highest priority
• failing CI fixes
• python 3.10-3.14 compatibility, EOL for python 3.8, 3.9
• macos compatibility & testing
• numpy 2 support
• tutorial notebooks review

handover items

• issue triage
• stale PR review and triage - all
• operational handover
• pointers to dev channels
• digital assets where required

roadmap items for consideration or wishlist

• architectural review
• test coverage review (not just line coverage but case coverage)
• refactor model level tests to systematic test suite?
• API homogenization with pytorch-forecasting
• scikit-learn integration or scikit-learn-like API

[fkiraly]

FYI @phoeenniixx, @manujosephv

[DM-Berger]

Community input on todos and wishlist is also appreciated, e.g., what are "burning" items, suggested priorities

Updating dependency versions and migrating to pyproject.toml (#590) would be quite helpful to me / us, as we currently run into issues on Windows (ARM64) VMs when trying to uv pip install pytorch_tabular (Python 3.13: mainly, setup.py tries to build Numpy 1.26.4, which fails due to missing compiler and linker toolchains). Works fine on MacOS and Linux though, Python 3.13.

Also, I have to use uv pip install pytorch_tabular because I can't just do uv add pytorch_tabular, and I notice this rolls back a lot of dependencies to pretty old versions, which of course is unideal.

I also suspect migrating the dependencies and shifting to pyproject.toml shouldn't be too bad compared to some of the other tasks you have listed above.

Regardless, thanks for all the work on this. We've found GANDALF useful and to perform surprisingly well in a lot of cases, even compared to tuned LightGBM.

[fkiraly]

Currently, a slate of failing tests is preventing an updated release. Eyes on this would be appreciated! See here:
#600

[fkiraly]

FYI @DM-Berger, new version just released.