Comment by imtringued

>It’s a peculiar feature of transformer-based LLMs that computing a batch of completions at the same time is almost as fast as computing a single completion. Why is that?

Incorrect. Transformers usually contain a classical MLP layer. Only the MLP layer can be batched. Hence all classical neural networks including convolutional networks (via im2col) can be batched.

If there's anything that the transformer architecture changes, it is that the attention layer cannot be batched.