Does anyone know of any evidence/forum/paper analyzing benchmark result differences between cloud inference platforms (togetherai) and running models locally with vLLM under greedy decoding?
