Communication Compression for Tensor Parallel LLM Inference [2411.09510]