Lossless LLM compression for efficient GPU inference via dynamic-length float