sesquipedalianist

I just tried it (I had previously tried DataLoader2 but perhaps with torch 1.13.1) and the spikes still occur. This makes sense to me because it seems the datapipe graph...

For now I have solved my issue by monkeypatching `torch.utils.data.graph_settings.apply_random_seed` as follows: ``` def apply_random_seed_overwrite(datapipe: DataPipe, rng: torch.Generator) -> DataPipe: return datapipe torch.utils.data.graph_settings.apply_random_seed = apply_random_seed_overwrite ``` This effectively disables `apply_random_seed`,...

@andrew-bydlon are you saving anything in memory (like audio samples?). That would likely cause the same issue as I was having.