Take this quiz to assess your understanding of Code of Ethics and Business Conduct.
5 Sample Questions
In the context of training large language models (LLMs) using pipeline parallelism - a technique where a model is split across multiple devices by layers - what is the primary cause of 'pipeline bubbles,' and which scheduling strategy is specifically designed to reduce them? (Difficulty: hard)
A.
Pipeline bubbles are caused by gradient staleness across pipeline stages, and the Asynchronous SGD scheduling strategy is used to eliminate them entirely.
B.
Pipeline bubbles are caused by devices sitting idle while waiting for activations or gradients from adjacent pipeline stages, and the 1F1B (One Forward One Backward) scheduling strategy is designed to reduce them.
C.
Pipeline bubbles are caused by memory overflow on individual GPUs, and the Tensor Parallelism scheduling strategy is used to reduce them by splitting weight matrices across devices.
D.
Pipeline bubbles are caused by imbalanced layer sizes across pipeline stages, and the Dynamic Load Balancing scheduling strategy eliminates them by redistributing layers at runtime.
In the context of the NVIDIA Ampere architecture and its role in generative AI and LLM workloads, the A100 GPU introduced a new memory technology compared to its Volta predecessor. Which of the following correctly describes a key architectural memory advancement in the A100 and explains its significance for large language model training? (difficulty: hard)
A.
The A100 uses the same HBM2 memory as the V100 but with an improved memory controller that increases effective bandwidth by 3x through lossless compression applied transparently to all data types.
B.
The A100 introduced GDDR6X memory with a 384-bit bus, doubling the memory bandwidth of the V100 to better handle the large weight matrices in transformer-based LLMs.
C.
The A100 introduced HBM2e memory, providing up to 2 TB/s of memory bandwidth, which significantly reduces the memory bandwidth bottleneck when processing the large activation tensors and weight matrices in transformer models.
D.
The A100 introduced HBM3 memory stacked directly on the GPU die using 3D packaging, enabling near-zero latency memory access that eliminates the memory wall problem entirely for LLM workloads.
In the context of few-shot and zero-shot learning for large language models (LLMs), a researcher is comparing two prompting strategies for a complex reasoning task: (A) providing 3 input-output examples where each output is just the final answer, and (B) providing 3 input-output examples where each output includes intermediate reasoning steps before the final answer. The model used is a large-scale LLM (e.g., 100B+ parameters). Which strategy is most likely to yield better performance on multi-step reasoning tasks, and why? (difficulty: hard)
A.
Both strategies perform equally well, because few-shot learning performance is determined entirely by the number and diversity of examples, not by the format of the outputs shown in those examples.
B.
Strategy A, because shorter outputs in examples reduce the prompt length, leaving more context window space for the actual query and reducing the risk of the model losing focus on the task.
C.
Strategy A, because providing final answers only forces the model to rely on its pre-trained knowledge for reasoning, which is more reliable than attempting to follow potentially flawed intermediate steps shown in the examples.
D.
Strategy B, because including intermediate reasoning steps in few-shot examples triggers chain-of-thought reasoning in the model, significantly improving performance on multi-step and arithmetic tasks.
In the context of LLM Tool Use and Function Calling, a developer is building an AI assistant that must call external functions. The LLM is provided with the following tool definition: a function 'get_stock_price' that accepts a single parameter 'ticker' (string). The user asks: 'Is Apple's stock price currently higher than Microsoft's?' Which describes the most accurate behavior of a well-implemented LLM function-calling system when handling this request? (difficulty: hard)
A.
The LLM makes two sequential or parallel calls to 'get_stock_price' - one with ticker 'AAPL' and one with ticker 'MSFT' - then uses both results to reason and formulate a comparative answer.
B.
The LLM informs the user that the request cannot be fulfilled because the 'get_stock_price' tool does not support comparative queries between two assets.
C.
The LLM responds directly with an estimated comparison based on its training data, since it already has knowledge of stock prices from its pre-training corpus.
D.
The LLM makes a single call to 'get_stock_price' with both tickers passed as a comma-separated string in the 'ticker' parameter, then compares the returned values.
In the domain of BERT and encoder-based models within generative AI and NLP pipelines, consider BERT's use of bidirectional self-attention combined with Masked Language Modeling (MLM) during pre-training. Why does this combination allow BERT to build richer contextual representations than a left-to-right language model trained on the same text corpus? (difficulty: medium)
A.
Because each token's representation can attend to all other tokens in the sequence, both preceding and following, rather than only to tokens that came before it
B.
Because masked tokens are replaced with a special [MASK] token, which carries more semantic information than real words
C.
Because BERT processes the input sequence twice, once forward and once backward, and concatenates the two resulting hidden states
D.
Because BERT uses a larger vocabulary size than autoregressive language models, allowing finer-grained token representations
