How to get gpu global rank in the dataloader model?

[FakeEnd]

Description & Motivation

My custom dataset using a numpy sampler to get data. If I cannot use different seed for different GPU/nodes, it will use the same seed forked from main process, which will sample the same data. So, I want to get global rank in LightningDataModule.train_dataloader to create different dataloader for different GPU/nodes. I see there is a local rank property in here, but when I try to use self.local_rank to get it, it fails.

Pitch

No response

Alternatives

No response

Additional context

No response

cc @Borda

[awaelchli]

Hey
In LightningDataModule, you have to do self.trainer.local_rank.

class MyDataModule(LightningDataModule):
    def train_dataloader(self):
        print(self.trainer.local_rank)
        return DataLoader(MyDataset(...))

[ssharpe42]

@awaelchli when is self.trainer set in the data module?