mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
Why it matters
Model authors can serve any supported transformers implementation at native vLLM speed with --model-impl transformers, removing the port-to-vLLM step from the path between a new architecture and fast production serving.
