fix(config): passing gradient_checkpoint_kwargs

[NanoCode012]

According to huggingface/transformers#28339 , setting it to False increases VRAM. My quick testing shows ~1GB increase at lowest settings.

Furthermore, the default in transformers and torch is going to be True huggingface/transformers#29638 (comment)

Finally, removing the default to False in trainer_building to clean old configs. I see that this kwarg is now set in https://github.com/OpenAccess-AI-Collective/axolotl/blob/a914cb37dc455a3fd0368e3a0898867f25b3a6c9/src/axolotl/utils/config/__init__.py#L170-L176

Description

Motivation and Context

How has this been tested?

Screenshots (if appropriate)

Types of changes

Social Handles (Optional)

[winglian]

In my experience, use_reentrant: False used less VRAM prevented OOM in some situations where use_reentrant: True would OOM, hence the current default.

[winglian]

see my comment in #1167

[NanoCode012]

In that code block above, it seems that the validate_config which runs first, sets it to True by default already.

Then, the trainer_builder which runs second, redundantly checks for it again and tries to set it to False.

Either way, I think one of the defaults should be removed to prevent future confusion.

Edit: your linked PR sets it to true, despite the comment saying it's false.

[winglian]

In that code block above, it seems that the validate_config which runs first, sets it to True by default already.

Then, the trainer_builder which runs second, redundantly checks for it again and tries to set it to False.

Either way, I think one of the defaults should be removed to prevent future confusion.

Edit: your linked PR sets it to true, despite the comment saying it's false.

Thanks for digging into this. Good to go!