We used to implicitly support tensor indices() in get_frames_at(indices), even though the type annotation clearly mention list[int]. In https://github.com/pytorch/torchcodec/pull/746/files#diff-4f95b28922233b6ec6f41a7264a5518af7ea7dfb11329b133d378f9701e3cad0L224-L227 we "broke" that implicit tensor support. This broke transformers (#877), so we re-enabled implicit tensor support in #880.
The current implicit tensor support isn't great: we're converting the tensor to a list which will then be converted to an array of int. We should just natively accept tensors, avoid copies, and make tensor support official and public.