TransMLA: Migrating GQA Models to MLA with Full DeepSeek Compatibility and Speedup

Neural Information Processing Systems 

Modern large-language models often face communication bottlenecks on current hardware rather than computational limitations.