GRPO - Accelerated - Dataloader Question

[Bhoy1]

Hi everyone,

I’ve been digging into the GRPOTrainer source code and am trying to implement a custom DataLoader. I noticed that get_train_dataloader(self) internally calls accelerator.prepare(), which is used for distributed setups. However, it also (I think) reconstructs the DataLoader, which seems like it would overwrite or invalidate any custom DataLoader I try to provide.

Has anyone experimented with custom DataLoader implementations on top of GRPOTrainer while still preserving all the benefits of accelerate? I’m mainly trying to understand the best way to integrate a custom loader without breaking the accelerator workflow.

Any guidance or past experience would be greatly appreciated!