Worked on the Liger-Kernel repository to extend multimodal model support by integrating Qwen3-VL, focusing on core operations such as rope, cross entropy, rmsnorm, and flce. Leveraged Python and deep learning techniques to update model configurations and enhance the testing pipeline, ensuring all Qwen3VL tests passed successfully. Investigated convergence issues in the Qwen3VLMoe branch, identifying areas for targeted improvements in future iterations. Expanded test and configuration management for multimodal models, which improved validation speed and traceability of changes. The work emphasized robust model development practices and contributed to the stability and extensibility of the machine learning codebase.