Each sample will be loaded and processed in the Dataset.__getitem__.
The common approach is to load and transform a single sample so the transformations would be performed sequentially in the main processes given you are using num_workers=0.
torchvision.transforms also accept tensors and you could use a BatchSampler if you want to load and process multiple tensors at once.
If you use num_workers>0 each process will process and create its own batch.
Samples are passed as a batch to the model and all layers accept batched inputs thus processing all samples at once.