Explain Vision Encoders and LLM Bottlenecks

Respond to the following questions on machine learning system fundamentals:

1. What is a vision encoder, and what function does it serve in a computer vision or multimodal model?
2. How is a vision encoder usually trained?
3. What are the primary performance bottlenecks for large language models during inference?
4. How would you reduce LLM inference memory usage and latency? Overview: This question assesses your grasp of vision encoders and their function in computer vision and multimodal models, your familiarity with common encoder training methods, and your capacity to recognize inference bottlenecks in large language models along with memory and latency optimization considerations.