devices and devices_list type issues

[zhang090210]

https://github.com/manujosephv/pytorch_tabular/blob/023db2776f96a0f2854e837eef62840be1a12a5e/src/pytorch_tabular/config/config.py#L565C9-L566C45

When specifying a devices_list, this line of code causes

omegaconf.errors.ValidationError: Value '[1, 2, 3, 4]' of type 'list' could not be converted to Integer
    full_key: devices
    object_type=TrainerConfig

[dosubot]

This error happens because in the code, devices is typed as an integer (or None), but when you specify devices_list, it assigns a list to devices, which OmegaConf then rejects as a type mismatch. So, even though the docs and code intend for you to be able to use either devices (as an int, e.g. -1 for all) or devices_list (as a list, e.g. [0, 1]), the current implementation doesn't support this due to strict type validation—hence the ValidationError you see when passing a list to devices via devices_list [code] [docs] [issue].

Right now, the best practice is to use the devices parameter as an integer (like -1 for all available GPUs, or 1 for a single device). If you want to use multiple GPUs, you usually don't need to specify anything—PyTorch Tabular will use all available GPUs by default if you have more than one. Specifying devices_list will trigger this error until the config typing is updated to support both int and list inputs [issue].

If you need to control which GPUs are used, you may need to set the CUDA_VISIBLE_DEVICES environment variable before running your script as a workaround.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}