Task 2.2: Implement Model Deployment Strategies. AIP-C01 Practice Questions (Page 4)

Questions 16–20

1. 16expert · hard

   A company is deploying a 70B-parameter LLM on AWS for a real-time chat application. The model requires approximately 140GB of GPU memory. They have a cluster of 8x A100 GPUs (80GB each) and need to minimize inter-GPU communication overhead. Which container-based deployment pattern should they use?

   • AUse tensor parallelism to split the model's tensors across multiple GPUs, reducing communication overhead compared to model parallelism.
   • BUse model parallelism to shard the model across multiple GPUs, with each GPU holding a portion of the layers.
   • CDeploy the model on a single GPU using quantization to 8-bit precision.
   • DDeploy the model on a single GPU with lazy loading to load layers on demand.
   Reveal Solution Discussion Select an answer first
2. 17foundation · easy

   When balancing performance and resource usage for a GenAI workload, what is a key trade-off to consider?

   • ASmaller models always provide the same quality as larger models
   • BQuantization always improves model quality
   • CLarger models often provide better quality but require more memory and compute
   • DUsing more GPUs always reduces latency without increasing cost
   Reveal Solution Discussion Select an answer first
4. 18foundation · easy

   Which approach can help balance performance and resource usage for a GenAI workload?

   • ADeploying models only on CPU instances to reduce cost
   • BAlways using the largest available model regardless of cost
   • CChoosing a model size that meets the quality requirements while fitting within the available infrastructure
   • DUsing a single model for all tasks without considering task complexity
   Reveal Solution Discussion Select an answer first
5. 19foundation · easy

   Which scenario is a typical use case for Amazon SageMaker AI endpoints in a hybrid solution?

   • ACreating a serverless function to respond to HTTP requests
   • BRunning a scheduled batch job to process data overnight
   • CStoring training data in a highly durable object storage
   • DHosting a custom-trained model that needs to be called by an application through an API
   Reveal Solution Discussion Select an answer first
6. 20foundation · easy

   An application requires real-time responses with low latency for a variable number of requests. Which deployment strategy is most appropriate for this workload?

   • AA managed endpoint that can scale automatically based on demand
   • BA one-time on-demand invocation using Lambda
   • CA batch processing job that runs nightly
   • DA static website hosted on Amazon S3
   Reveal Solution Discussion Select an answer first