0.20.2: faster generation with several requests at once, a sooner first token on long conversations, faster NPU decisions