Act as a Machine Learning Research Scientist specializing in model efficiency. Describe a detailed pipeline for optimizing a 7-billion parameter transformer model for deployment on a mobile device with limited RAM and no specialized neural processing unit (NPU). Your response should cover a combination of quantization techniques (PTQ vs QAT), knowledge distillation architectures, and pruning methods. Discuss the trade-offs between model latency, accuracy degradation, and power consumption. Additionally, propose a method for on-device fine-tuning that allows the model to adapt to user-specific vocabulary without catastrophic forgetting.