Worked on stabilizing Qwen models within the Blaizzy/mlx-vlm repository, focusing on continuous batching scenarios in Python. Addressed a critical issue where mismatches in batch dimensions could occur due to varying sequence lengths and cache states during streaming or incremental inference. Implemented robust error handling and introduced guards to the attention mechanism, ensuring that discrepancies in batch dimensions and cache states do not lead to failures. This technical approach improved the reliability and stability of model optimization workflows, particularly in environments requiring continuous batching. The work demonstrated depth in attention mechanisms and error handling, enhancing the robustness of streaming inference pipelines.