Curious to hear what constraints are there that aren't tackled by the current offering of local runtimes/SDKs for inference.