megatron grpo data_shuffle

[hjh0119]

fix #9199

[gemini-code-assist]

Code Review

This pull request modifies the argument handling for Megatron GRPO by removing the restriction that required dataset_shuffle to be true. It introduces logic to synchronize train_dataloader_shuffle with the dataset_shuffle value. Feedback indicates that the current assignment could unintentionally overwrite user-specified values for train_dataloader_shuffle because of default values, and suggests using a logical AND to respect both configuration flags.

[gemini-code-assist]

The current implementation unconditionally overwrites train_dataloader_shuffle with dataset_shuffle. Since dataset_shuffle defaults to True in RLHFMegatronArgumentsMixin, this will override and ignore a user-specified --train_dataloader_shuffle false when rlhf_type is grpo.

To respect both configuration flags, it is safer to use a logical AND operation, ensuring that shuffling is disabled if either flag is set to False.

Suggested change

	if self.dataset_shuffle is not None:
	self.train_dataloader_shuffle = self.dataset_shuffle
	if self.dataset_shuffle is not None:
	self.train_dataloader_shuffle = self.train_dataloader_shuffle and self.dataset_shuffle