NCA-GENM - Software Development (15% of the exam) - Section 4.2
Work with modern deep learning frameworks.
Identify the key capabilities of modern deep learning frameworks - including dynamic computation graphs, automatic differentiation, and GPU acceleration - that accelerate model development. Use framework APIs to define model architectures, configure training loops, and export models for deployment.
Practice question for this objective
lock_openFree sampleSoftware Developmentmedium
A team is training a large image-text fusion model on NVIDIA A100 GPUs. They enable automatic mixed precision using PyTorch's built-in mixed-precision training utility. Which statement best describes why this approach accelerates training while maintaining numerical stability?
AIt runs forward and backward passes in 16-bit floats for speed while keeping a master copy of weights in 32-bit for gradient accumulationcheck_circle Correct
BIt replaces all tensor operations with integer arithmetic, removing floating-point overhead entirely
CIt dispatches compute-intensive layers to the CPU in 64-bit precision while the GPU handles attention layers in 16-bit
DIt quantises weights to 8-bit integers during the forward pass and dequantises them before the backward pass
Explain how mixed-precision training uses 16-bit and 32-bit floating-point formats to accelerate GPU training without sacrificing numerical stability. Automatic mixed precision exploits NVIDIA Tensor Cores, which deliver significantly higher throughput on FP16 matrix multiplications than on FP32. The GradScaler utility scales the loss upward before the backward pass so that small gradient values do not underflow in FP16, then unscales them before applying the optimiser step. A 32-bit master copy of parameters is updated by the optimiser, preventing the weight drift that would occur if accumulated FP16 updates introduced rounding errors over many steps. This combination gives near-peak Tensor Core utilisation with convergence quality matching full-precision training.
Why A is correct: Tensor Core hardware on Ampere GPUs executes 16-bit matrix multiplications far faster than 32-bit, and the 32-bit master weights prevent the precision loss that would otherwise corrupt small gradient updates.
Why B is wrong: Integer arithmetic is not used; mixed precision keeps floating-point formats throughout, choosing between 16-bit and 32-bit precision per operation rather than switching to integers.
Why C is wrong: Mixed precision keeps all computation on the GPU; offloading layers to CPU would negate the speed benefit and is a separate CPU-offload technique used in memory-constrained scenarios.
Why D is wrong: 8-bit quantisation is a distinct inference-time technique; automatic mixed precision in training uses 16-bit floating-point formats, not 8-bit integers, and does not dequantise during the backward pass.
Answers that look right on this material and are not. Each one is a distractor from a different question in the NCA-GENM bank for this domain.
cuBLAS, because all deep learning bottlenecks reduce to dense matrix multiplications which this library accelerates
errorWhy it is wrong: cuBLAS accelerates general dense matrix-matrix operations but does not include the specialised convolution algorithms, workspace management, and fusion passes that cuDNN provides for neural network layers.
The model's state dictionary and the random-seed value used at training start
errorWhy it is wrong: Tempting because reproducibility is important, but the random seed alone cannot restore the optimiser's accumulated second-moment estimates, so learning rates would be reset on resume.
Parameter-efficient fine-tuning, where small adapter matrices are inserted so only a fraction of the original weights are updated on new data.
errorWhy it is wrong: Adapter-based fine-tuning reduces the number of trained parameters but keeps the same large model; it is tempting as another efficiency technique, yet it produces no smaller deployable model and does not learn from soft outputs.