The last operations of LLM's are a projection layer that transforms the embeddings of each token into the logits, a softmax calculation that takes these logits and turns them into probabilities and ...