Worked on stabilizing LoRA training for mlx-vlm qwen3_5-family models in the unslothai/unsloth-zoo repository, focusing on preventing training crashes and improving model reliability. Addressed issues with attention cache handling, gated delta updates, and non-differentiable Metal kernels to ensure smooth training workflows. The solution involved adapting to new mlx-vlm attention keyword arguments and refining the differentiable training path by adjusting fused kernel usage. Using Python and deep learning techniques, the work resolved a critical crash documented in unslothai/unsloth#6002 and contributed to ongoing pull requests, demonstrating a methodical approach to debugging and model training stability in machine learning systems.