NCA-GENM Question 114
Select 3You are assisting in the deployment of a generative AI multimodal model that processes both text and images. During initial scalability testing, the system struggles to handle increased user requests, leading to latency issues. Under the supervision of a senior team member, which actions should you prioritize to address scalability while ensuring performance and reliability?
- A
Optimize the model by reducing its parameters or using model quantization techniques.
- B
Deploy the model across multiple GPUs to distribute the workload more effectively.
- C
Increase the batch size during inference to handle more requests simultaneously.
- D
Implement caching mechanisms to store frequently requested outputs for faster responses.
- E
Retrain the model from scratch with a larger dataset to improve response times.
Show answer and explanation
Correct answers: A, B, D
Explanation
When addressing scalability, performance, and reliability during model deployment, techniques like model optimization, workload distribution (e.g., using multiple GPUs), and caching are immediate and effective strategies. These approaches reduce computational overhead and improve system responsiveness under increased load. Retraining the model or increasing batch size may not directly address scalability issues and could introduce other challenges.
- A. Correct.
Optimizing the model, such as reducing parameters or using quantization, can improve scalability by reducing computational and memory requirements, making it easier to handle more requests.
- B. Correct.
Deploying the model across multiple GPUs allows for parallel processing, which is essential for scaling up to handle large numbers of user requests.
- C. Incorrect.
Increasing the batch size during inference might seem beneficial, but it can actually increase latency if the hardware cannot process larger batches efficiently. This is not a guaranteed solution for scalability.
- D. Correct.
Caching frequently requested outputs can significantly reduce computation time and improve response times, enhancing both scalability and performance.
- E. Incorrect.
Retraining the model from scratch with a larger dataset is not a solution for addressing immediate scalability and performance issues. It is a long-term strategy for improving model accuracy, not scalability.