Window size in LSTM: training and testing

I am reading papers on how LSTM is used for regression in time-series. During training, the dataset is split in fixed-size chunks that depend on the window size. The window slides over the training data, and these small sequences are fed into the network. For testing, I would expect the test set to be split likewise in fixed-size sequences, but I read that this is not necessary and often the whole sequence is fed into the network for inference.
Questions
How does this work? What do authors mean with "the whole sequence"? Is the input for the LSTM during the test phase completely independent on the window-size used during training?