From the link: "The same exact GPU model and program versions must be used for compression and decompression". So it seems to me that you need a standardized and deterministic LLM execution model before this becomes useful.