Questions

Select the best answer for each question.

1. What is the primary goal of quantization in large language model deployment?

A) Increase model accuracy by adding more parameters

B) Reduce memory usage and improve inference efficiency

C) Expand the context window of the model

D) Improve reinforcement learning alignment

2. Which statement best describes model distillation?

A) A process that converts floating-point weights into integers

B) A technique that transfers knowledge from a larger model to a smaller model

C) A method for increasing GPU memory allocation

D) A technique for training models without labeled data

3. What is one major advantage of using vLLM for inference?

A) It allows models to train faster on GPUs

B) It improves GPU utilization using optimized memory management

C) It automatically fine-tunes models during inference

D) It replaces the need for tokenization

4. What is the primary purpose of TensorRT-LLM?

A) Providing reinforcement learning training pipelines

B) Optimizing inference performance on NVIDIA GPUs

C) Performing dataset preprocessing for training

D) Creating new transformer architectures

5. Hugging Face Inference Endpoints are best described as:

A) A managed service for deploying machine learning models at scale

B) A dataset hosting platform

C) A framework for training reinforcement learning agents

D) A tokenization library

6. Which metric is commonly used to measure inference responsiveness?

A) BLEU score

B) Perplexity

C) Latency

D) F1 score

7. What does throughput measure in a deployed model system?

A) The number of parameters in a model

B) The number of tokens generated per request

C) The number of requests processed per unit of time

D) The accuracy of the model's responses

8. Why is monitoring token usage important in production systems?

A) It determines whether the model is properly aligned

B) It directly affects the operational cost of inference

C) It improves training speed

D) It prevents GPU overheating

9. Which of the following is an example of a resource monitoring metric?

A) GPU memory usage

B) BLEU score

C) ROUGE score

D) Accuracy

10. What is a common cause of latency spikes in deployed LLM systems?

A) Using tokenizers

B) Excessive batching or heavy GPU workload

C) Having too many training datasets

D) Using LoRA adapters

11. Why is logging prompts and responses useful in production?

A) It increases GPU utilization

B) It allows developers to monitor system behavior and diagnose issues

C) It improves training accuracy

D) It reduces inference latency

12. What is one risk of deploying models without proper monitoring?

A) The model will stop generating tokens

B) Performance or cost issues may go unnoticed

C) The tokenizer may stop working

D) The model will retrain automatically

13. When comparing quantized and non-quantized models, what trade-off often occurs?

A) Lower cost but slightly reduced model precision

B) Higher cost but faster training

C) Increased memory usage with higher accuracy

D) Reduced GPU usage but slower tokenization

14. What is the role of an API layer (such as FastAPI) in an LLM deployment?

A) It trains the model using reinforcement learning

B) It exposes the model as a service that applications can call

C) It compresses model parameters

D) It stores training datasets

15. Why is continuous evaluation important after deployment?

A) Models automatically retrain during inference

B) Real-world usage may expose new errors or alignment issues

C) Tokenizers degrade over time

D) GPUs lose performance after long use