In March 2026, K06aaa contributed to comfyanonymous/ComfyUI by enabling Apple Silicon GPU-accelerated text encoding, focusing on performance optimization for prompt generation workflows. Using Python and leveraging GPU programming techniques, K06aaa modified device-selection logic to support MPS with VRAMState.SHARED, allowing non-quantized text encoders to utilize Apple Silicon GPUs efficiently. The implementation included explicit handling for quantized models, which remain CPU-bound due to CLIP.supports_cast limitations, and provided clear documentation of these trade-offs. This work demonstrated a deep understanding of machine learning model deployment on heterogeneous hardware and laid the foundation for further GPU-accelerated improvements within the repository.