You are responsible for optimizing the inference pipeline of a large language model (LLM) running on edge devices with limited RAM and compute power. Describe techniques you would employ to reduce the model size and inference latency, such as quantization, pruning, knowledge distillation, and operator fusion. Further, explain how you would implement a batching strategy to maximize throughput without violating real-time constraints. Provide a comparative analysis of ONNX Runtime versus TensorRT for this specific use case.