Free NVIDIA-Certified Associate: Generative AI with LLMs (NCA-GENL) Practice Questions

Sample NCA-GENL Practice Questions

Browse all 15 free NVIDIA-Certified Associate: Generative AI with LLMs practice questions below.

1. In a transformer model, what is the primary purpose of the self-attention mechanism?

   • To apply regularization during training by randomly masking tokens and forcing the network to reconstruct missing inputs from context
   • To compress the input embedding dimension by projecting high-dimensional token representations into a lower-dimensional subspace for memory efficiency
   • To reduce the computational complexity of the model by replacing sequential recurrence with parallel token processing across the entire sequence
   • To allow each token in a sequence to attend to all other tokens and capture contextual relationships

2. Which component of the transformer architecture is responsible for preserving the sequential order information of input tokens?

   • The feed-forward neural network layers that apply independent non-linear transformations to each position's representation after attention computation
   • The multi-head attention mechanism that computes weighted combinations of value vectors using query-key similarity scores across multiple subspaces
   • The layer normalization component that stabilizes training by normalizing activations across the feature dimension for each sample independently
   • Positional encoding

3. In multi-head attention, why does the transformer use multiple attention heads instead of a single attention mechanism?

   • To reduce the total number of parameters in the model by sharing weight matrices across positions and decomposing large attention operations
   • To speed up inference time by distributing the attention computation across multiple GPUs, enabling hardware-level parallelism for each head
   • To allow the model to attend to information from different representation subspaces at different positions
   • To prevent gradient vanishing during backpropagation by creating multiple independent gradient pathways through the attention layers

4. What is the key difference between the encoder and decoder components in an encoder-decoder transformer architecture?

   • The encoder typically has more layers than the decoder to better process and encode input representations before passing them to generation
   • The encoder exclusively uses self-attention while the decoder relies solely on cross-attention to encoder outputs for all token interactions
   • The decoder uses masked self-attention to prevent attending to future tokens during training, while the encoder uses bidirectional attention
   • The decoder uses larger embedding dimensions and wider feed-forward layers than the encoder to improve generation quality and fluency

5. In the attention mechanism, what mathematical operation is performed on the Query (Q) and Key (K) matrices to compute attention scores?

   • Element-wise multiplication of query and key vectors followed by summation across the feature dimension to produce position-wise similarity scores
   • Matrix multiplication (dot product) followed by scaling and softmax
   • A convolution operation applied across query and key matrices to capture local positional patterns and short-range dependencies between tokens
   • Concatenation of the query and key vectors for each position pair followed by a learned linear transformation to compute compatibility scores

6. What is the primary advantage of FlashAttention in modern transformer implementations?

   • It reduces memory usage by computing attention incrementally using tiling and recomputation techniques
   • It replaces multi-head attention with a single consolidated attention head, reducing the total number of learnable parameters across the model's layers significantly
   • It applies post-training quantization to compress attention weight matrices to 8-bit integer precision, reducing memory bandwidth requirements during computation
   • It eliminates the need for softmax normalization in the attention mechanism by using a linear approximation that scales more efficiently with sequence length

7. In decoder-only transformer architectures like GPT, what is the purpose of the KV (Key-Value) cache during inference?

   • To enable efficient batch processing of multiple input sequences simultaneously by sharing intermediate representations across parallel decoding streams
   • To compress the model's weight parameters during deployment by storing them in a lower-precision format that reduces GPU memory consumption
   • To store previously computed key and value vectors from past tokens, avoiding redundant computation during autoregressive generation
   • To implement causal attention masking by dynamically constructing triangular mask matrices that prevent each token from attending to future positions in the sequence

8. What is Multi-Query Attention (MQA) and how does it differ from standard multi-head attention? (Select TWO)

   • MQA uses a single set of key and value projection matrices shared across all attention heads, while maintaining separate query projections per head
   • MQA processes multiple user queries simultaneously in a single forward pass by parallelizing the attention computation across different input sequences in the batch
   • MQA eliminates the query projection matrices entirely and computes attention scores using only the key and value representations, reducing the total parameter count per layer
   • MQA increases the number of attention heads beyond the standard configuration to improve the model's representational capacity and pattern recognition ability
   • MQA significantly reduces the KV cache memory footprint during inference, enabling higher throughput and longer context lengths on the same hardware

9. Why do decoder-only architectures like GPT use causal (masked) self-attention?

   • To compress the KV cache memory footprint during inference by quantizing stored key-value pairs to lower precision formats like INT8 or FP8
   • To reduce the computational complexity of the self-attention mechanism by approximately half, since only the lower triangular portion of the attention matrix needs to be calculated
   • To enable bidirectional attention patterns where each token can attend to both preceding and following positions, improving contextual understanding of the full sequence
   • To prevent the model from attending to future tokens during training, ensuring each position only depends on previous positions

10. What happens during the token generation process when a decoder-only LLM produces an output?

    • The model generates all tokens in the output sequence simultaneously in a single forward pass using parallel decoding, producing the complete response at once rather than sequentially
    • The model outputs dense embedding vectors from its final transformer layer that are directly used as the text representation without any additional projection or sampling step
    • The model outputs logits (raw scores) for each vocabulary token, which are converted to probabilities via softmax and then sampled or selected
    • The model directly outputs text characters one at a time using a character-level mapping function, bypassing any intermediate probability computation or vocabulary-based token selection

11. In a sequence-to-sequence (seq2seq) model for machine translation, what role does the encoder play in the overall architecture?

    • It processes the input sequence and creates a fixed-length or variable-length context representation that captures the semantic meaning
    • It applies cross-attention alignment between source and target sequences during decoding, computing weighted relevance scores to map input to output positions
    • It performs subword tokenization using byte-pair encoding and maps raw input characters to embedding indices before neural network processing begins
    • It generates target language output tokens autoregressively, using previously generated tokens and input context to predict each successive output token

12. When computing cross-entropy loss for a language model with a vocabulary of 50,000 tokens, what does the model output before the loss calculation?

    • A probability distribution over all 50,000 vocabulary tokens, typically produced by a softmax layer
    • An embedding vector representing the semantic meaning of the predicted token, capturing contextual information from surrounding tokens
    • A single integer index representing the predicted token ID, directly mapping the final hidden layer activations to a vocabulary position
    • A binary classification result from a sigmoid activation that determines whether the predicted token matches the ground-truth target token

13. In the context of training recurrent neural networks (RNNs) for sequence modeling, what is 'backpropagation through time' (BPTT)?

    • A technique that reverses the processing order of input sequences during the forward pass to capture backward temporal dependencies, enabling bidirectional context modeling within a single recurrent pass
    • An algorithm that unfolds the RNN across time steps and applies standard backpropagation to compute gradients with respect to parameters shared across all timesteps
    • A weight initialization strategy specifically designed for recurrent architectures that sets initial hidden states and weight matrices to prevent gradient vanishing during early training iterations
    • A cyclic learning rate scheduling method that systematically decreases the step size according to a predetermined decay schedule across training epochs to improve recurrent network convergence

14. Why is Xavier (Glorot) initialization commonly used for initializing weights in deep neural networks for NLP tasks?

    • It initializes all weights to zero across every layer to ensure perfectly symmetric learning signals across neurons, guaranteeing that each unit receives identical gradient updates during early training phases
    • It sets all weights to small constant values like 0.01 across the entire network to prevent gradient explosion, ensuring that activations remain bounded throughout the forward propagation process
    • It sets weights proportional to the number of input and output connections, helping maintain consistent variance of activations and gradients across layers
    • It initializes weights randomly from a standard uniform distribution between -1 and 1 regardless of network dimensions, providing sufficient randomness to break symmetry between neurons in each layer

15. What is the purpose of weight decay (L2 regularization) when training large language models? (Select TWO)

    • To decrease the effective batch size as training progresses, allowing finer-grained gradient updates and improved convergence in later stages
    • To reduce the learning rate exponentially over time according to a decay schedule, ensuring convergence stability in later training phases
    • To gradually remove redundant neurons from the network during training, creating a sparser model that focuses on the most informative pathways
    • To penalize large weight values and encourage the model to learn simpler patterns, reducing overfitting
    • To act as implicit regularization that constrains model complexity, improving generalization to unseen data by encouraging smaller, more distributed weight representations