accelerate

f1ca8ac7 - Allow DataLoaderAdapter subclasses to be pickled by implementing `__reduce__` (#3074)

Allow DataLoaderAdapter subclasses to be pickled by implementing `__reduce__` (#3074) * initial fix for breaking accelerator pickling * cleanup * skip_first_batches should be used on raw dls * multigpu sanity test * bugs * does this work with iterable dsets? * fix typo * ignore these commits, i'm just syncing the origin so i can test on my cloud workstation * comment out failing tests, unsure if those are existing bugs or a recent regression * torch 2.4.0? * pickling generator issues * test_pickle_accelerator * test_pickle_accelerator should work now) * base.__len__() -> len(base) * undo reduce * undo super().__reduce__() again * pass args through superclass * remove prints * doc changes + make style && make quality