Worked on the alibaba/rtp-llm repository to deliver a configurable word-embedding tying mechanism for the Qwen3 language model. Introduced a tie_word_embeddings option that enables weight reuse between the embedding layer and the dense lm_head, reducing parameter redundancy and improving memory efficiency for language modeling tasks. Addressed cross-version weight sharing to ensure consistent behavior between Qwen3 and Qwen35 models, preventing misconfigurations when toggling the new feature. The implementation involved Python and leveraged deep learning and natural language processing techniques, with careful attention to configuration wiring and documentation to support maintainability and future experimentation with shared embeddings in model training.