NCP-ADS Practice Questions: Machine Learning Domain

Correct Answer: C

Explanation: Blue-green deployment is an MLOps practice that involves running two identical production environments. One (blue) serves production traffic while the other (green) is used for testing and updates, allowing for seamless updates without downtime.

Correct Answer: A

Explanation: Data Parallelism involves splitting the data across multiple GPUs, allowing each GPU to process a portion of the data simultaneously, which reduces training time. Model Parallelism is less common and more complex for most ML models. Single GPU Training doesn't leverage multiple GPUs, and Batch Normalization is a technique used within the model, not a multi-GPU strategy.

Correct Answer: C

Explanation: Improper data partitioning can lead to memory overload on specific workers, causing memory issues. It's important to ensure data is evenly distributed across the cluster. Underutilizing CPU resources (A) and overloading network bandwidth (B) are not direct causes of memory issues. Insufficient number of workers (D) may affect performance but not directly cause memory issues.

Correct Answer: C

Explanation: Gradient checkpointing saves memory by not storing all intermediate activations during the forward pass, instead recomputing them during the backward pass. This is effective for training large models on limited GPU resources.

Correct Answer: C

Explanation: Dask with RAPIDS enables distributed computing and scaling of machine learning tasks across multiple GPUs and nodes, facilitating the handling of large datasets and complex computations. It does not inherently provide automatic hyperparameter tuning, real-time deployment, or improved accuracy.

Correct Answer: B

Explanation: Option B is correct because Dask with cuML allows for distributed computing across multiple GPUs, which is essential for handling large datasets that cannot fit into a single GPU's memory. Option A is incorrect because using a single GPU is insufficient for large datasets. Option C is incorrect because cuGraph is used for graph analytics, not general ML training. Option D is incorrect as it relies on CPU-based processing, which is not optimal for large-scale datasets.

Correct Answer: B

Explanation: Enabling mixed precision training reduces GPU memory usage by using fewer bits to represent numbers, allowing for more data to be processed simultaneously. Decreasing the learning rate, increasing hyperparameters, or using a larger validation set would likely increase memory usage or not address the memory issue.

Correct Answer: A

Explanation: Dynamic batching in NVIDIA Triton Inference Server allows for combining multiple inference requests into a single batch, optimizing GPU utilization and improving throughput. Single-thread execution, manual memory management, and static model allocation are not features that directly optimize multi-GPU inference performance.

Correct Answer: B

Explanation: Mixed precision training uses both 16-bit and 32-bit floating-point types to reduce memory usage and improve computational efficiency, leading to reduced training time. It does not inherently increase model accuracy, improve preprocessing speed, or simplify deployment.

Correct Answer: B

Explanation: Mixed precision training reduces training time and memory usage by using lower precision for calculations, which speeds up computation and allows more data to fit into GPU memory. It does not inherently increase model accuracy, simplify architecture, or eliminate the need for data normalization.