Description

Apparently the usage of rich when using trl's CLI is currently enforced to True (set to "1"), meaning that rich will always be used.

While this may be a nice enhancement to the CLI display to some, in some cases as logging to an external tracking tool as e.g. GCP's Log Explorer, makes the experience a bit hard since rich is not properly handled so no logs are propagated, as the tqdm bar is being updated.

So on, how the TRL_USE_RICH flag is defined and handled could be improved, so that it can be both enabled and disabled (at the moment it can only be enabled as per

, meaning that when set, no matter what value is is, the condition if TRL_USE_RICH will always be true).

Proposal

• Handle the boolean value for the environment value similarly to https://github.com/python/cpython/blob/48f9d3e3faec5faaa4f7c9849fecd27eae4da213/Lib/distutils/util.py#L308-L321 (already deprecated since Python 3.10, meaning we would need to copy that function into utils.py)
• If TRL_USE_RICH is not set when running trl's CLI, then set it to True by default; otherwise, simply evaluate what's provided using the function mentioned above