pytorch-sentiment-analysis | getting started with PyTorch and TorchText for sentiment | Machine Learning library

This is an older discussion, which has changed slightly over the years (mainly due to the purpose of with torch.no_grad() as a pattern. An excellent answer that kind of answers your question as well can be found on Stackoverflow already.
However, since the original question is vastly different, I'll refrain from marking as duplicate, especially due to the second part about the memory.

An initial explanation of no_grad is given here:

with torch.no_grad() is a context manager and is used to prevent calculating gradients [...].

requires_grad on the other hand is used

to freeze part of your model and train the rest [...].

Source again the SO post.

Essentially, with requires_grad you are just disabling parts of a network, whereas no_grad will not store any gradients at all, since you're likely using it for inference and not training.
To analyze the behavior of your combinations of parameters, let us investigate what is happening:

• a) and b) do not store any gradients at all, which means that you have vastly more memory available to you, no matter the number of parameters, since you're not retaining them for a potential backward pass.
• c) has to store the forward pass for later backpropagation, however, only a limited number of parameter (3 million) are stored, which makes this still manageable.
• d), however, needs to store the forward pass for all 112 million parameters, which causes you to run out of memory.

When torchtext builds the vocabulary, it aligns the the token indices with the embedding. If your vocabulary doesn't have the same size and ordering as the pre-trained embeddings, the indices wouldn't be guaranteed to match, therefore you might look up incorrect embeddings. build_vocab() creates the vocabulary for your dataset with the corresponding embeddings and discards the rest of the embeddings, because those are unused.

The GloVe-6B embeddings includes a vocabulary of size 400K. For example the IMDB dataset only uses about 120K of these, the other 280K are unused.

I think it is hardly possible to give a 100% perfect answer, but you can certainly get inspiration from the way other scripts are doing it. The best place to start is the examples/ directory of the huggingface repository itself, where you can for example find this excerpt:

The BucketIterator sorts the data to make batches with examples of similar length to avoid having too much padding. For that it needs to know what the sorting criterion is, which should be the text length. Since it is not fixed to a specific data layout, you can freely choose which field it should use, but that also means you must provide that information to sort_key.

In your case, there are two possible fields, text and wage_label, and you want to sort it based on the length of the text.